Counterpoint: o1-Pro is insanely good -- subjectively, it's as far above GPT4 as GPT4 was above 3. It's almost too good. Use it properly for an extended period of time, and one begins to worry about the future of one's children and the utility of their schooling.
o3, by all accounts, is better still.
Seems to me that things are progressing quickly enough.
I keep reading this on HN so I believe it has to be true in some ways, but I don't really feel like there is any difference in my limited use (programming questions or explaining some concepts).
If anything I feel like it's all been worse compared to the first release of ChatGPT, but I might be wearing rose colored glasses.
One use-case: They help with learning things quickly by having a chat and asking questions. And they never get tired or emotional. Tutoring 24/7.
They also generate small code or scripts, as well as automate small things, when you're not sure how, but you know there's a way. You need to ensure you have a way to verify the results.
They do language tasks like grammar-fixing, perfect translation, etc.
They're 100 times easier and faster than search engines, if you limit your uses to that.
They can't help you learn what they don't know themselves.
I'm trying to use them to read historical handwritten documents in old Norwegian (Danish, pretty much). Not only do they not handle the German-style handwriting, but what they spit out looks like the sort of thing GPT-2 would spit out if you asked it to write Norwegian (only slightly better than Swedish Muppet Swedish Chef's Swedish). It seems the experimental tuning has made it worse at the task I most desperately want to use it for.
And when you think about it, how could it not overfit in some sense, when trained on its own output? No new information is coming in, so it pretty much has to get worse at something to get better at all the benchmarks.
Hah, no. They're good, but they definitely make stuff up when the context gets too long. Always check their output, just the same as you already note they need for small code and scripts.
Same, on every release from openai, anthropic I keep reading how the new model is so much better (insert hyperbole here) than the previous one yet when using it I feel like they are mostly the same as last year.
It’s the same for me. I genuinely don’t understand how I can be having such a completely different experience from the people who rave about ChatGPT. Every time I’ve tried it’s been useless.
How can some people think it’s amazing and has completely changed how they work, while for me it makes mistakes that a static analyser would catch? It’s not like I’m doing anything remarkable, for the past couple of months I’ve been doing fairly standard web dev and it can’t even fix basic problems with HTML. It will suggest things that just don’t work at all and my IDE catches, it invents APIs for packages.
One guy I work with uses it extensively and what it produces is essentially black boxes. If I find a problem with something “he” (or rather ChatGPT) has produced it takes him ages to commune with the machine spirit again to figure out how to fix it, and then he still doesn’t understand it.
I can’t help but see this as a time-bomb, how much completely inscrutable shite are these tools producing? In five years are we going to end up with a bunch of “senior engineers” who don’t actually understand what they’re doing?
Before people cry “o tempora o mores” at me and make parallels with the introduction of high-level languages, at least in order to write in a high-level language you need some basic understanding of the logic that is being executed.
> How can some people think it’s amazing and has completely changed how they work, while for me it makes mistakes that should a static analyser would catch?
There are a lot of code monkeys working on boilerplate code, these people used to rely on stack overflow and now that chatgpt is here it's a huge improvement for them
If you work on anything remotely complex or which hasn't been solved 10 times on stack overflow chatgpt isn't remotely as useful
I work on very complex problems. Some of my solutions have small, standard substeps that now I can reliably outsource to ChatGPT. Here are a few just from last week:
- write cvxpy code to find the chromatic number of a graph, and an optimal coloring, given its adjecency matrix.
- given an adjecency matrix write numpy code that enumerates all triangle-free vertex subsets.
- please port this old code from tensorflow to pytorch: ...
- in pytorch, i'd like to code a tensor network defining a 3-tensor of shape (d, d, d). my tensor consists of first projecting all three of its d-dimensional inputs to a k-dimensional vector, typically k=d/10, and then applying a (k, k, k) 3-tensor to contract these to a single number.
To be honest, these don’t sound like hard problems. These sound like they have very specific answers that I might find in the more specialized stackoverflow sections. These are also the kind of questions (not in this domain) that I’ve found yield the best results from LLMs.
In comparison asking an LLM a more project specific question “this code has a race condition where is it” while including some code usually is a crapshoot and really depends if you were lucky enough to give it the right context anyway.
Sure, these are standard problems, I’ve said so myself. My point is that my productivity is multiplied by ChatGPT, even if it can only solve standard problems. This is because, although I work on highly non-standard problems (see https://arxiv.org/abs/2311.10069 for an example), I can break them down into smaller, standard components, which ChatGPT can solve in seconds. I never ask ChatGPT "where's the race condition" kind of questions.
The ones who use it extensively are the same that used to hit up stackoverflow as the first port of call for every trivial problem that came their way. They're not really engineers, they just want to get stuff done.
Hmm... calling people "not engineers" is considered an attack now? I'm afraid this is actually revealing your own bias towards engineers. I never said engineers were superior or that we'd be better off with a whole world full of them.
> How can some people think it’s amazing and has completely changed how they work, while for me it makes mistakes that should a static analyser would catch? It’s not like I’m doing anything remarkable, for the past couple of months I’ve been doing fairly standard web dev and it can’t even fix basic problems with HTML.
Part of this is, I think, anchoring and expectation management: you hear people say it's amazing and wonderful, and then you see it fall over and you're naturally disappointed.
My formative years started off with Commodore 64 basic going "?SYNTAX ERROR" from most typos plus a lot of "I don't know what that means" from the text adventures, then Metrowerks' C compiler telling me there were errors on every line *after but not including* the one where I forgot the semicolon, then surprises in VisualBasic and Java where I was getting integer division rather than floats, then the fantastic oddity where accidentally leaning on the option key on a mac keyboard while pressing minus turns the minus into an n-dash which looked completely identical to a minus on the Xcode default font at the time and thus produced a very confusing compiler error…
So my expectations have always been low for machine generated output. And it has wildly exceeded those low expectations.
But the expectation management goes both ways, especially when the comparison is "normal humans" rather than "best practices". I've seen things you wouldn't believe...
Entire files copy-pasted line for line, "TODO: deduplicate" and all,
20 minute app starts passed off as "optimized solutions."
FAQs filled with nothing but Bob Ross quotes,
a zen garden of "happy little accidents."
I watched iOS developers use UI tests
as a complete replacement for storyboards,
bi-weekly commits, each a sprawling novel of despair,
where every change log was a tragic odyssey.
Google Spreadsheets masquerading as bug trackers,
Swift juniors not knowing their ! from their ?,
All those hacks and horrors… lost in time,
Time to deploy.
(All true, and all pre-dating ChatGPT).
> It will suggest things that just don’t work at all and my IDE catches, it invents APIs for packages.
Aye. I've even had that with models forgetting the APIs they themselves have created, just outside the context window.
To me, these are tools. They're fantastic tools, but they're not something you can blindly fire-and-forget…
…fortunately for me, because my passive income is not quite high enough to cover mortgage payments, and I'm looking for work.
> In five years are we going to end up with a bunch of “senior engineers” who don’t actually understand what they’re doing?
Yes, if we're lucky.
If we're not, the models keep getting better and we don't have any "senior engineers" at all.
Interesting. I implemented something very similar (if not identical) a couple years ago (at work so not open source). I used a simple grammar and standard parser generator. It’s been nice to have the grammar as we’ve made tweaks over the years to change various behaviours and add features.
first time I tried it, I asked it to find bugs in a piece of very well tested C code.
It introduced an off-by-one error by miscounting the number of arguments in an sprintf call, breaking the program. And then proceeded to fail to find that bug that it introduced.
I think the difference comes down to interacting with it like IDE autocomplete vs. interacting with it like a colleague.
It sounds like you're doing the former -- and yeah, it can make mistakes that autocomplete wouldn't or generate code that's wrong or overly complex.
On the other hand, I've found that if you treat it more like a colleague, it works wonderfully. Ask it to do something, then read the code and ask follow-up questions. If you see something that's wrong or just seems off, tell it, and ask it to fix it. If you don't understand something, ask for an explanation. I've found that this process generates great code that I often understand better than if I had written it from scratch, and in a fraction of the time.
It also sounds like you're asking it to do basic tasks that you already know how to do. I find that it's most useful in tackling things that I don't know how to do. It'll already have read all of the documentation and know the right way to call whatever APIs, etc, and -- this is key -- you can have a conversation with it to clear up anything that's confusing.
This takes a big shift in mindset if you've been using IDEs all your life and have expectations of LLMs being a fancy autocomplete. And you really have to unlearn a lot of stuff to get the most out of them.
I'm in the same boat as the person you're responding to. I really don't understand how to get anything helpful out of ChatGPT, or more than anything basic out of Claude.
> I've found that if you treat it more like a colleague, it works wonderfully.
This is what I've been trying to do. I don't use LLM code completion tools. I'll ask anything from how to do something "basicish" with html & css, and it'll always output something that doesn't work as expected. Question it and I'll get into a loop of the same response code, regardless of how I explain that it isn't correct.
On the other end of the scale, I'll ask about an architectural or design decision. I'll often get a response that is in the realm of what I'd expect. When drilling down and asking specifics however, the responses really start to fall apart. I inevitably end up in the loop of asking if an alternative is [more performant/best practice/the language idiomatic way] and getting the "Sorry, you're correct" response. The longer I stay in that loop, the more it contradicts itself, and the less cohesive the answers get.
I _wish_ I could get the results from LLMs that so many people seem to. It just doesn't happen for me.
My approach is a lot of writing out ideas and giving them to ChatGPT. ChatGPT sometimes nods along, sometimes offers bad or meaningless suggestions, sometimes offers good suggestions, sometimes points out (what should have been) obvious errors or mistakes. The process of writing stuff out is useful anyway and sometimes getting good feedback on it is even better.
When coding I will often find myself in kind of a reverse pattern from how people seem to be using ChatGPT. I work in a jupyter notebook in a haphazard way getting things to functional and basically correct, after this I select all, copy, paste, and ask ChatGPT to refactor and refine to something more maintainable. My janky blocks of code and one offs become well documented scripts and functions.
I find a lot of people do the opposite, where they ask ChatGPT to start, then get frustrated when ChatGPT only goes 70% of the way and it's difficult to complete the imperfectly understood assignment - harder than doing it all yourself. With my method, where I start and get things basically working, ChatGPT knows what I'm going for, I get to do the part of coding I enjoy, and I wind up with something more durable, reusable, and shareable.
Finally, ChatGPT is wonderful in areas where you don't know very much at all. One example, I've got this idea in my head for a product I'll likely never build - but it's fun to plan out.
My idea is roughly a smart bidet that can detect metabolites in urine. I got this idea when a urinalysis showed I had high levels of ketones in my urine. When I was reading about what that meant I discovered it's a marker for diabetic ketoacidosis (a severe problem for ~100k people a year) and it can also be indicator for colorectal cancer as well as indicating a "ketosis" state that some people intentionally try to enter for dieting or wellness reasons. (My own ketones were caused by unintentionally being in ketosis, I'm fine, thanks for wondering.)
Right now, you detect ketones in urine with a strip that you pee on, and that works well enough - but it could be better because who wants to use a test strip all the time? Enter the smart bidet. The bidet gives us an excuse to connect power to our device and bring the sensor along. Bluetooth detects a nearby phone (and therefore identity of the depositor), a motion sensor can detect a stream of urine triggering our detection, and then use our sensor to detect ketones which we track overtime in the app, ideally with additional metabolites that have useful diagnostic purposes.
How to detect ketones? Is it even possible? I wonder to ChatGPT if spectroscopy is the right method of detection here. ChatGPT suggests a retractable electrochemical probe similar to an extant product that can detect a kind of ketone in blood. ChatGPT knows what kind of ketone is most detectable in urine. ChatGPT can link me to scientific instrument companies that make similar (ish) probes where I could contact them and ask if they sold this type of thing, and so on.
Basically, I go from peeing on a test strip and wondering if I could automate this to chat with ChatGPT - having, what was in my opinion, an interesting conversation with the LLM, where we worked through what ketones are, the different kinds, the prevalence of ketones in different bodily fluids, types of spectroscopy that might detect acetoacetate (available in urine) and how much that would cost and what challenges would be and so on, followed by the idea of electrochemical probes and how retracting and extending the probe might prolong its lifespan and maybe a heating element could be added to dry the probe to preserve it even better and so on.
Was ChatGPT right about all that? I don't know. If I were really interested I would try to validate what it said, and I suspect I would find it was mostly right and incomplete or off in places. Basically like having a pretty smart and really knowledgeable friend who is not infallible.
Without ChatGPT I would have likely thought "I ...
I believe the smart bidet was an idea some Japanese researchers developed some years ago. Maybe this one was geared to detecting blood in faeces. Whatever,the approach you describe has a huge number of possibilities for alerting us to health problems without even having to think about them on a daily basis. A huge advantage. On the other hand this is a difficult one to implement bearing in mind the kinetics involved.
If you've ever used any enterprise software for long enough, you know the exact same song and dance.
They release version Grand Banana. Purported to be approximately 30% faster with brand new features like Algorithmic Triple Layering and Enhanced Compulsory Alignment. You open the app. Everything is slower, things are harder to find and it breaks in new, fun ways. Your organization pays a couple hundred more per person for these benefits. Their stock soars, people celebrate the release and your management says they can't wait to see the improvement in workflows now that they've been able to lay off a quarter of your team.
Has there been improvements in LLMs over time? Somewhat, most of it concentrated at the beginning (because they siphoned up a bunch of data in a dubious manner). Now it's just part of their sales cycle, to keep pumping up numbers while no one sees any meaningful improvement.
Not sure what you are using it for, but it is terrible for me for coding; claude beats it always and hands down. o1 just thinks forever to come up with stuff it already tried the previous time.
People say that's just prompting without pointing to real million line+ repositories or realistic apps to show how that can be improved. So I say they are making todo and hello world apps and yes, there it works really well. Claude still beats it, every.. single.. time..
And yes, I use the Pro of all and yes, I do assume coding is done for most of people. Become a plumber or electrician or carpenter.
We do the same (all requests go to o1, sonnet and gemini and we store the results for later to compare) automatically for our research: Claude always wins. Even with specific prompting on both platforms. Especially frontend it seems o1 really is terrible.
Every time I try Gemini, it's really subpar. I found that qwen2.5-coder-32b-instruct can be better.
Also, for me 50% 50% for Sonnet and o1, but although I'm not 100% sure about it, I think o1 is better with longer and more complicated (C++) code and debugging. At least from my brief testing. Also, OpenAI models seem to be more verbose - sometimes it's better - where I'd like additional explanation on chosen fields in a SQL schema, sometimes it's too much.
EDIT: Just asked both o1 and Sonnet 3.5 the same QML coding question, and Sonnet 3.5 succeeded, o1 failed.
Very anecdotal but I’ve found that for things that are well spec’d out with a good prompt Sonnet 3.5 is far better. For problems where I might have introduced a subtle logical error o1 seems to catch it extremely well. So better reasoning might be occurring but reasoning is only a small part of what we would consider intelligence.
Wins? What does this mean? Do you have any results? I see the claims that Claude is better for coding a lot but using it and using Gemini 2.0 flash and o1 and it sure doesn't seem like it.
Exactly. The previous version of o1 did actually worse in the coding benchmarks, so I would expect it to be worse in real life scenarios.
The new version released a few days ago on the other hand is better in the benchmarks, so it would seem strange that someone used it and is saying that it’s worse than Claude.
That so weird, it’s seems like everybody here prefers Claude.
I’ve been using Claude and openai in copilot and I find even 4o seems to understand the problem better. O1 definitely seems to get it right more for me.
They're both okay for coding, though for my use cases (which are niche and involve quite a lot of mathematics and formal logic) o1/o1-Pro is better. It seems to have a better native grasp of mathematical concepts, and it can even answer very difficult questions from vague inputs, e.g.: https://chatgpt.com/share/676020cb-8574-8005-8b83-4bed5b13e1...
Different languages maybe? I find Sonnet v2 to be lacking in Rust knowledge compared to 4o 11-20, but excelling at Python and JS/TS. O1's strong side seems to be complex or quirky puzzle-like coding problems that can be answered in a short manner, it's meh at everything else, especially considering the price. Which is understandable given its purpose and training, but I have no use for it as that's exactly the sort of problem I wouldn't trust an LLM to solve.
Sonnet v2 in particular seems to be a bit broken with its reasoning (?) feature. The one where it detects it might be hallucinating (what's even the condition?) and reviews the reply, reflecting on it. It can make it stop halfway into the reply and decide it wrote enough, or invent some ridiculous excuse to output a worse answer. Annoying, although it doesn't trigger too often.
I try to sprinkle 'for us/me' everywhere as much as I can; we work on LoB/ERP apps mostly. These are small frontends to massive multi million line backends. We carved a niche by providing the frontends on these backends live at the client office by a business consultant of ours: they simply solve UX issues for the client on top of large ERP by using our tool and prompting. Everything looks modern, fresh and nice; unlike basically all the competitors in this space. It's fast and no frontend people are needed for it; backend is another system we built which takes a lot longer of course as they are complex business rules. Both claude and o1 turn up something that looks similar but only the claude version will work and be, after less prompting, correct. I don't have shares in either and I want open source to win; we have all open (more open) solutions doing all the same queries and we evaluate all but claude just wins. We did manage even big wins with openai davinci in 2022 (or so; before chatgpt), but this is a massive boost allowing us to upgrade most people to business consultant and just have them build with clients real time and have the tech guys including me add manually tests and proofs (where needed) to know if we are actually fine. Works so much better than the slog with clients before; people are so bad at explaining at what they need, it was slowly driving me insane after doing it for 30+ years.
> It's fast and no frontend people are needed for it
I guess if you don’t need to maintain it, just an ever growing blob of complexity that will be reinvented into new blobs every time when the old one becomes too immobile :)
True I could imagine in the ERP world building one-off solutions repeatedly for tons of consulting money is the status quo. Most software businesses can't afford repeatedly starting from scratch and having zero reusability of stuff they already invested in getting working and tested, even with AI assistance.
Claude web’s context window is 200K tokens. I’d be surprised if GitHub Copilot’s context window exceeds 10K.
I’ve found using Claude via Copilot in VS Code produces noticeably lower quality results than 3.5 Sonnet on web. In my experience Claude web outdoes GPT-4o consistently.
I had a 30 min argument with o1-pro where it was convinced it had solved the halting problem. Tried to gaslight me into thinking I just didn’t understand the subtlety of the argument. But it’s susceptible to appeal to authority and when I started quoting snippets of textbooks and mathoverflow it finally relented and claimed there had been a “misunderstanding”. It really does argue like a human though now...
I had a similar experience with regular o1 about integral that was divergent. It was adamant that it wasn't and would respond to any attempt at persuasion with variants of "its a standard integral" with a "subtle cancellation". When I asked for any source for this standard integral it produced references to support its argument that existed but didn't actually contain the integral. When I told it the references didn't have the result and backpedalled (gaslighting!) to "I never told you they were in there". When I pointed out that in fact it did it insisted this was just a "misunderstanding". It only relented when I told it Mathematica agreed the integral was divergent. It still insisted it never said that the books it pointed to contained this (false, non-sensical) result.
This was new behaviour for me to see in an LLM. Usually the problem is these things would just fold when you pushed back. I don't know which is better, but being this confidently wrong (and "lying" when confronted with it) is troubling.
The troubling part is that the references themselves existed -- one was an obscure Russian text that is difficult to find (but is exactly where you'd expect to find this kind of result, if it existed).
I was on an airplane and there was high-speed Internet on the airplane. That's the newest thing that I know exists. And I'm sitting on the plane and they go, open up your laptop, you can go on the Internet.
And it's fast, and I'm watching YouTube clips. It's amazing. I'm on an airplane! And then it breaks down. And they apologize, the Internet's not working. And the guy next to me goes, 'This is bullshit.' I mean, how quickly does the world owe him something that he knew existed only 10 seconds ago?"
Soon, all the middle class jobs will be converted to profits for the capital/data center owners, so they have to spend while they can before the economy crashes due to lack of spending.
Not invariably. Some of those people are the ones who want to draw 7 red lines all perpendicular, some with green ink, some with transparent and one that looks like a kitten.
No, people who say "it's bullshit" and then do something to fix the bullshit are the ones that push technology forward. Most people who say "it's bullshit" instantly when something isn't perfect for exactly what they want right now are just whingers and will never contribute anything except unconstructive criticism.
There's someone with this comment in every thread. Meanwhile, no one answers this because they are getting value. Please take the time to learn, it will give you value.
I’m a consultant. Having looked at several enterprises, there’s a lot of work being done to make a lot of things that don’t really work.
The bigger the ambition, the harder they’re failing. Some well designed isolated use cases are ok. Mostly things about listening and summarizing text to aid humans.
I have yet to see a successful application that is generating good content. IMO replacing the first draft of content creation and having experts review and fix it is, like, the stupidest strategy you can do. The people you replace are the people at the bottom of the pyramid who are supposed do this work to upskill and become domain experts so they can later review stuff. If they’re no longer needed, you’re going to one day lose your reviewer, and with it, the ability to assess your generated drafts. It’s a foot gun.
I mean, no, not generally. but the success rate of other tools is much higher.
A lot of companies are trying to build these general purpose bots that just magically know everything about the company and have these but knowledge bases, but they just don’t work.
> Meanwhile, no one answers this because they are getting value.
You're literally doing the same thing you're accusing of. Every HN thread is full of AI boosters claiming AI to be the future with no backing evidence.
Riddle me this. If all these people are "getting value", why are all these companies losing horrendous amounts of money? Why has nobody figured out how to be profitable?
> Please take the time to learn, it will give you value.
Yeah, yeah, just prompt engineer harder. That'll make the stochastic parrot useful. Anyone who has criticism just does so because they're dumb and you're smart. Same as it always was. Everyone opposed to the metaverse just didn't get it bro. You didn't get NFTs bro. You didn't get blockchain bro.
None of these previous bubbles had money in it (beyond scamming idiots), if AI wants to prove it's not another empty tech bubble, pay up. Show me the money. Should be easy, if it's automating so many expensive man-hours of labour. People would be lining up to pay OpenAI.
> Riddle me this. If all these people are "getting value", why are all these companies losing horrendous amounts of money? Why has nobody figured out how to be profitable?
While I agree that LLMs are not currently working great for most envisioned use cases; this premise here is not a good argument. Large LLM providers are not trying to be profitable at the moment. They’re trying to grow and that’s pretty sensible.
Uber was the poster child of this, and for all its mockery, Uber is now an unqualified profitable company.
I'm not sure I would call incinerating 11b dollars a year to the point where you need to do one of the biggest raises ever and it doesn't even buy you a year of runway sensible.
> Why has nobody figured out how to be profitable?
From what I've seen claimed about OpenAI finances, this is easy: It's a Red Queen's race — "it takes all the running you can do, to keep in the same place".
If their financial position was as simple as "we run this API, we charge X, the running cost is Y", then they're already at X > Y.
But if that was all OpenAI were actually doing, they'd have stopped developing new versions or making the existing models more efficient some time back, while the rest of the industry kept improving their models and lowering their prices, and they'd be irrelevant.
> People would be lining up to pay OpenAI.
They are.
Not that this is either sufficient or necessary to actually guarantees anything about real value. For lack of sufficiency: people collectively paid a lot for cryptocurrencies and NFTs, too (and before then and outside tech, homeopathic tinctures and sub-prime mortgages); For lack of necessity: there's plenty of free-to-download models.
I get a huge benefit even just from the free chat models. I could afford to pay for better models, but why bother when free is so good? Every time a new model comes out, the old paid option becomes the new free option.
Think of all the search engines alltheweb, yahoo, astalavista,... where sooo much money got poored in, and finally there was just one winner taking it all. That's the race openai is trying to win now. The competition is fierce and we can just play with all kinds of models for free and we do nothing but complaining.
I'm someone who generally was a "doubter", but I've dramatically softened my stance on this topic.
Two things:
I was casually watching Andreas Kling's streams on Ladybird development (where he was developing a JIT compiler for JS) and was blown away at the accuracy of completions (and the frequency of those completions)
Prior to this, I'd only ever copypasta'd code from ChatGPT output on occasion.
I started adopting the IDE/Editor extensions and prototyping small projects.
There's now small tools and utilities I've written that I'd not have written otherwise, or would have taken twice the time invested had I'd not used these tools.
With that said, they'd be of no use without oversight, but as a productivity enhancement, the benefits are enormous.
For my mental health I’ve stopped replying to comments where it’s clear the author has no intention of having a discussion and instead wants their share their opinion and have it reinforced by others.
No, we don’t have AGI or anything close to it. Yes, AI has come a long way in the past decade and many people find it useful in their day-to-day lives.
It’s difficult to know where AI will be in 10 years, but the current rate of improvement is staggering.
That was puzzles me now. Everyone with a semblance of expertise in engineering knows that if you start with a tool and try to find a problem it could solve you are doing it wrong. The right way is the opposite - you start with a problem, and find the best tool to solve it, and if it's the new shiny tool - so be it, but most of the time it's not.
Except the whole tech world starting with the CEOs seems to do it the "wrong" way with LLMs. People and whole companies are encouraged to find what these things might be actually useful for.
• Build toys that would otherwise require me to learn new APIs (I can read python, but it's not my day job)
• Learn new things like OpenSCAD
• To improve my German
• Learn about the world by allowing me to take photos of things in this world that I don't understand and ask them a question about the content, e.g. why random trees have bands or rectangles of white paint on them
• Help me shopping, by taking a photo of the supermarket that I happen to be in at the time and ask them where I should look for some item I can't find
• Help with meal prep, by allowing me to get a recipe based on what food and constraints I've got at hand rather than the traditional method of "if you want x, buy y ingredients"
Even if they're just an offline version of Wikipedia or Google, they're already a more useful interface for the same actual content.
For a company that sees itself as the undisputed leader and that wants to raise $7 trillion to build fabs, they deserve some of the heaviest levels of scrutiny in the world.
If OpenAI's investment prospectus relies on them reaching AGI before the tech becomes commoditized, everyone is going to look for that weakness.
What I find odd is that o1 doesn't support attaching text documents to chats the way 4o does. For a model that specializes in reasoning, reading long documents seems like a natural feature to have.
If Sama ever reads this, I have no idea why no users seem to focus on this, but it would be really good to prioritise being able to select which model you can use with the custom myGPTs. I know this maybe hard or not possible without recreating them , but I still dont think it's possible.
I dont think most customers realise how much better the models work with custom GPTs.
They hyped them like crazy and haven't discussed them once since then. I agree that the inability to change the model is pretty absurd when the whole point was to "supercharge" specific tasks.
There was even talk of some sort of profit sharing with creators which clearly never happened. I just think the premise is too confusing for many and can still be served by using a custom system prompt via the API.
Was it hyped? I tried a few of them and they seemed absolutely useless. Like I could install a “custom GPT” that just appends something to my prompt? How great..
No the whole
Point is you make a gpt for yourself and upload all your related documents to it and then query that. It performs 10x better than a generic query without attaching every single doc that could be relevant.
I am unsure if the answer is to use “projects” maybe this has superseded myGpts?
I am perplexed why HN isn’t focusing on this issue as all the Llm gains I’ve ever had were wit highly customised personal myGpts.
I can understand OpenAI and Sam’s having access to their own models may not even know what the best way to use the released stuff is
This is something being done but not for consumers right now. On a more roll your own basis. I know a few people whose companies have already established their own gpt trained on their own internal codebase, standards, documentation, and I’m sure relevant reference information as well. Think big r and d companies in stem.
"When using custom instructions or files, only GPT-4o is available". Straight out of the ChatGPT web interface when you try to select which model you want to use.
Everyone's comparing o1 and claude, but neither really work well enough to justify paying for them in my experience for coding. What I really want is a mode where they ask clarifying questions, ideally many of them, before spitting out an answer. This would greatly improve utility of producing something with more value than an auto-complete.
I know multiple people that carefully prompt to get that done. The model outputs in direct token order, and can't turn around, so you need to make sure that's strictly followed. The system can and will come up with post-hoc "reasoning".
Just today I got Claude to convert a company’s PDF protocol specification into an actual working python implementation of that protocol. It would have been uncreative drudge work for a human, but I would have absolutely paid a week of junior dev time for it. Instead I wrote it alongside AI and it took me barely more than an hour.
The best part is, I’ve never written any (substantial) python code before.
I have to agree. It's still a bit hit or miss, but the hits are a huge time and money saver especially in refactoring. And unlike what most of the rather demeaning comments in those HN threads state, I am not some 'grunt' doing 'boilerplate work'. I mostly do geometry/math stuff, and the AIs really do know what they're talking about there sometimes. I don't have many peers I can talk to most of the time, and Claude is really helping me gather my thoughts.
That being said, I definitely believe it's only useful for isolated problems. Even with Copilot, I feel like the AIs just lack a bigger context of the projects.
Another thing that helped me was designing an initial prompt that really works for me. I think most people just expect to throw in their issue and get a tailored solution, but that's just not how it works in my experience.
It would seem you don't care too much about verifying its output or about its correctness. If you did, it wouldn't take you just an hour. I guess you'll let correctness be someone else's problem.
I don't know the OP here, but in my experience a junior dev at an average company would likely not do much more than the AI would. These aren't your grandfather's engineers, after all.
A junior dev wouldn't have produced output of such consistency, and they wouldn't have annotated their code nearly as well. The majority of code was better than I'd expect from a junior, and the comments were better than I'd expect from the majority of people at every skill level.
Your wild assumptions and snarky accusations are unnecessary. The library is for me to use; there isn't a "someone else" for me to pass problems onto. I then did what I usually do — start writing real code with it ASAP, because real code is how you find real problems.
I developed the library interactively, one API call at a time, in a manner akin to pair programming. Code quality was significantly better than I'd expect from $2000 worth of a GOOD mid-tier programmer — the code was well written, well organised, and comprehensively annotated. The code wasn't perfect, but a majority of faults had a basis in the underlying documentation being wrong or ambiguous.
The $20/month for Cursor Pro literally justified its cost in less than 10 minutes.
For me, it's allowing me to do things I wouldn't have even attempted before. I'm writing in languages I've never written in before (python) and dealing with stuff I've never dealt with before (multicast UDP). This isn't complicated stuff by any stretch, but AI means I can be highly productive in python without needing to spend any time learning python.
Just tell it to do that and it will. Whenever I ask an AI for something and I'm pretty sure it doesn't have all the context I literally just say "ask me clarifying questions until you have enough information to do a great job on this."
And this chain of prompts cumulated with the improved CoT reasoner would accrue a lot more enhanced results. More in line with what the coming agentic era promises.
Yes. You can only do so much with the information you get in. The ability to ask good questions, not just of itself in internal monologue style, but actually of the user, would fundamentally make it better since it can get more information in.
As it is now, it has a bad habit of, if it can't answer the question you asked, instead answering a similar-looking question which it thinks you may have meant. That is of course a great strategy for benchmarks, where you don't earn any points for saying you don't know. But it's extremely frustrating for real users, who didn't read their question from a test suite.
The alternative to "ask me clarifying question" is to use Claude's Projects. Upload all your projects' source code there, and ask Claude to do your programming task.
OpenAI have recently also added this feature to their offering.
What we can reasonably assume from statements made by insiders:
They want a 10x improvement from scaling and a 10x improvement from data and algorithmic changes
The sources of public data are essentially tapped
Algorithmic changes will be an unknown to us until they release, but from published research this remains a steady source of improvement
Scaling seems to stall if data is limited
So with all of that taken together, the logical step is to figure out how to turn compute into better data to train on. Enter strawberry / o1, and now o3
They can throw money, time, and compute at thinking about and then generating better training data. If the belief is that N billion new tokens of high quality training data will unlock the leap in capabilities they’re looking for, then it makes sense to delay the training until that dataset is ready
With o3 now public knowledge, imagine how long it’s been churning out new thinking at expert level across every field. OpenAI’s next moat may be the best synthetic training set ever.
At this point I would guess we get 4.5 with a subset of this - some scale improvement, the algorithmic pickups since 4 was trained, and a cleaned and improved core data set but without risking leakage of the superior dataset
When 5 launches, we get to see what a fully scaled version looks like with training data that outstrips average humans in almost every problem space
Then the next o-model gets to start with that as a base and reason? Its likely to be remarkable
> With o3 now public knowledge, imagine how long it’s been churning out new thinking at expert level across every field. OpenAI’s next moat may be the best synthetic training set ever.
Even taking OpenAI and the benchmark authors at their word they said that it is consuming at least tens of dollars per task to hit peak performance, how much would it cost to have it produce a meaningfully large training set?
There is no public API for o3 yet, those are the numbers they revealed in the ARC-AGI announcement. Even if they were public API prices we can't assume they're making a profit on those for as long as they're billions in the red overall every year, its entirely possible that the public API prices are less than what OpenAI is actually paying.
"With o3 now public knowledge, imagine how long it’s been churning out new thinking at expert level across every field."
I highly doubt that. o3 is many orders of magnitude more expensive than paying subject matter experts to create new data. It just doesn't make sense to pay six figures in compute to get o3 to make data a human could make for a few hundred dollars.
Yes, I think they had to push this reveal forward because their investors were getting antsy with the lack of visible progress to justify continuing rising valuations. There is no other reason a confident company making continuous rapid progress would feel the need to reveal a product that 99% of companies worldwide couldn't use at the time of the reveal.
That being said, if OpenAI is burning cash at lightspeed and doesn't have to publicly reveal the revenue they receive from certain government entities, it wouldn't come as a surprise if they let the government play with it early on in exchange for some much needed cash to set on fire.
EDIT: The fact that multiple sites seem to be publishing GPT-5 stories similar to this one leads one to conclude that the o3 benchmark story was meant to counter the negativity from this and other similar articles that are just coming out.
Seems to me o3 prices would be what the consumer pays, not what OpenAI pays. That would mean o3 could be more efficient in-house than paying subject-matter experts.
For every consumer there will be a period where they need both the SME and the o3 model for initial calibration and eventual handoff for actually getting those efficiencies in whichever processes they want to automate.
In other words if you are diligent enough, you should at least validate your o3 solution with an actual expert for some time. You wouldn't just blindly trust OpenAI your business critical processes, would you? I would expect at least 3 month - 6 months for large corps and even more considering change management, re-upskilling, etc.
With all those considerations I really don't see the value prop at those prices and in those situations right now. Maybe if costs decrease ~1-3 orders of magnitude more for o3-low, depending on the the processes being automated.
Unless the quality of the human data are extraordinary, it seems according to the TFA that it's not that easy:
> The process is painfully slow. GPT-4 was trained on an estimated 13 trillion tokens. A thousand people writing 5,000 words a day would take months to produce a billion tokens.
And if the human-generated data was so qualitatively good that it is smaller by three order of magnitudes, than I can assume it would be at least as expensive as o3.
Everyone's obsessed with new training tokens... It doesn't need to be more knowledgeable, it just needs to practice more. Ask any student: practice is synthetic data.
Overfitting can be caused by a lot of different things. Having an over abundance of one kind of data in a training set is one of those causes.
It’s why many pre-processing steps for image training pipelines will add copies of images at weird rotations, amounts of blur, and different cropping.
> The more concepts the model manages to grok, the more nonlinear its capabilities will be
These kind of hand wavey statements like “practice,” “grok,” and “nonlinear its capabilities will be” are not very constructive as they don’t have solid meaning wrt language models.
So earlier when I was referring to compounding bias in synthetic data I was referring to a bias that gets trained on over and over and over again.
These kind of hand wavey statements like “practice,” “grok,” and “nonlinear its capabilities will be” are not very constructive as they don’t have solid meaning wrt language models.
So, here's my hypothesis, as someone who is adjacent ML but haven't trained DNNs directly:
We don't understand how they work, because we didn't build them. They built themselves.
At face value this can be seen as an almost spiritual position, but I am not a religious person and I don't think there's any magic involved. Unlike traditional models, the behavior of DNNs is based on random changes that failed up. We can reason about their structure, but only loosely about their functionality. When they get better at drawing, it isn't because we taught them to draw. When they get better at reasoning, it isn't because the engineers were better philosophers. Given this, there will not be a direct correlation between inputs and capabilities, but some arrangements do work better than others.
If this is the case, high order capabilities should continue to increase with training cycles, as long as they are performed in ways that don't interfere with what has been successfully learned. People lamented the loss of capability that GPT 4 suffered as they increased safety. I think Anthropic has avoided this by choosing a less damaging way to tune a well performing model.
> We don't understand how they work, because we didn't build them. They built themselves.
We do understand how they work, we did build them.
The mathematical foundation of these models are sound. The statistics behind them are well understood.
What we don’t exactly know is which parameters correspond to what results as it’s different across models.
We work backwards to see which parts of the network seem to relate to what outcomes.
> When they get better at drawing, it isn't because we taught them to draw. When they get better at reasoning, it isn't because the engineers were better philosophers.
Isn’t this the exact opposite of reality?
They get better at drawing because we improve their datasets, topologies, and their training methods and in doing so, teach them to draw.
They get better at reasoning because the engineers and data scientists building training sets do get better at philosophy.
They study what reasoning is and apply those learnings to the datasets and training methods.
> We do understand how they work, we did build them. The mathematical foundation of these models are sound. The statistics behind them are well understood.
We don't understand how they work in the sense that we can't extract the algorithms they're using to accomplish the interesting/valuable "intellectual" labor they're doing. i.e. we cannot take GPT-4 and write human-legible code that faithfully represents the "heavy lifting" GPT-4 does when it writes code (or pick any other task you might ask it to do).
That inability makes it difficult to reliably predict when they'll fail, how to improve them in specific ways, etc.
The only way in which we "understand" them is that we understand the training process which created them (and even that's limited to reproducible open-source models), which is about as accurate as saying that we "understand" human cognition because we know about evolution. In reality, we understand very little about human cognition, certainly not enough to reliably reproduce it in silico or intervene on it without a bunch of very expensive (and failure-prone) trial-and-error.
> We don't understand how they work in the sense that we can't extract the algorithms they're using to accomplish the interesting/valuable "intellectual" labor they're doing. i.e. we cannot take GPT-4 and write human-legible code that faithfully represents the "heavy lifting" GPT-4 does when it writes code (or pick any other task you might ask it to do).
I think English is being a little clumsy here. At least I’m finding it hard to express what we do and don’t know.
We know why these models work. We know precisely how, physically, they come to their conclusions (it’s just processor instructions as with all software)
We don’t know precisely how to describe what they do in a formalized general way.
That is still very different from say an organic brain, where we barely even know how it works, physically.
My opinions:
I don’t think they are doing much mental “labor.” My intuition likens them to search.
They seem to excel at retrieving information encoded in their weights through training and in the context.
They are not good at generalizing.
They also, obviously, are able to accurately predict tokens such that the resulting text is very readable.
Larger models have a larger pool of information and that information is in a higher resolution, so to speak, since the larger better preforming models have more parameters.
I think much of this talk of “consciousness” or “AGI” is very much a product of human imagination, personification bias, and marketing.
>We know why these models work. We know precisely how, physically, they come to their conclusions (it’s just processor instructions as with all software)
I don't know why you would classify this as knowing much of anything. Processor instructions ? Really?
If the average user is given unfettered access to the entire source code of his/her favorite app, does he suddenly understand it ? That seems like a ridiculous assertion.
In reality, it's even worse. We can't pinpoint what weights, how and in what ways and instances are contributing exactly to basic things like whether a word should be preceded by 'the' or 'a' and it only gets more intractable as models get bigger and bigger.
Sure, you could probably say we understand these NNs better than brains but it's not by much at all.
> If the average user is given unfettered access to the entire source code of his/her favorite app, does he suddenly understand it ? That seems like a ridiculous assertion.
And one that I didn’t make.
I don’t think when we say “we understand” we’re talking about your average Joe.
I mean “we” as in all of human knowledge.
> We can't pinpoint what weights, how and in what ways and instances are contributing exactly to basic things like whether a word should be preceded by 'the' or 'a' and it only gets more intractable as models get bigger and bigger.
There is research coming out on this subject. I read a paper recently about how llama’s weights seemed to be grouped by concept like “president” or “actors.”
But just the fact that we know that information encoded in weights affects outcomes and we know the underlying mechanisms involved in the creation of those weights and the execution of the model shows that we know much more about how they work than an organic brain.
The whole organic brain thing is kind of a tangent anyway.
My point is that it’s not correct to say that we don’t know how these systems work. We do. It’s not voodoo.
We just don’t have a high level understanding of the form in which information is encoded in the weights of any given model.
> If the average user is given unfettered access to the entire source code of his/her favorite app, does he suddenly understand it ? That seems like a ridiculous assertion.
And one that I didn’t make.
I don’t think when we say “we understand” we’re talking about your average Joe.
I mean “we” as in all of human knowledge.
It's an analogy. In understanding weights, even the best researchers are basically like the untrained average joe with source code.
>There is research coming out on this subject. I read a paper recently about how llama’s weights seemed to be grouped by concept like “president” or “actors.”
>But just the fact that we know that information encoded in weights affects outcomes and we know the underlying mechanisms involved in the creation of those weights and the execution of the model shows that we know much more about how they work than an organic brain.
I guess i just don't see how "information is encoded in the weights" is some great understanding ? It's as vague and un-actionable as you can get.
For training, the whole revolution of back-propagation and NNs in general is that we found a way to reinforce the right connections without knowing anything about how to form them or even what they actually are.
We no longer needed to understand how eyes detect objects to build an object detecting model. None of that knowledge suddenly poofed into our heads. Back-propagation is basically "reinforce whatever layers are closer to the right answer". Extremely powerful but useless for understanding.
Knowing the Transformer architecture unfortunately tells you very little about what a trained model is actually learning during training and what it has actually learnt.
"Information is encoded in a brain's neurons and this affects our actions".
Literally nothing useful you can do with this information. That's why models need to be trained to fix even little issues.
If you want to say we understand models better than the brain then sure but you are severely overestimating how much that "better" is.
But it isn’t. Knowing that information is encoded in the weights gives us a route to deduce what a given model is doing.
And we are. Research is being done there.
> "Information is encoded in a brain's neurons and this affects our actions". Literally nothing useful you can do with this.
Different entirely. We don’t even know how to conceptualize how data is stored in the brain at all.
With a machine, we know everything. The data is stored in a binary format which represents a decimal number.
We also know what information should be present.
We can and are using this knowledge to reverse engineer what a given model is doing.
That is not something we can do with a brain because we don’t know how a brain works. The best we can do is see that there’s more blood flow in one area during certain tasks.
With these statistical models, we can carve out entire chunks of their weights and see what happens (interestingly not much. Apparently most weights don’t contribute significantly towards any token and can be ignored with little performance loss)
We can do that with these transformers models because we do know how they work.
Just because we don’t understand every aspect of every single model doesn’t mean we don’t know how they work.
I think we’re starting to run in circles and maybe splitting hairs over what “know how something works” means.
I don’t think we’re going to get much more constructive than this.
I highly recommend looking into LoRas. We can make Loras because we know how these models work.
The thing that you are handwaving away as just "which parameters correspond to what results" is precisely the important, the inexorable thing which defines the phenomena, and it is exactly the thing which we don't have access to, and which we did not and could not design, plan or engineer, but which emerged
And who will tell the model whether its practice results are correct or not? Students practice against external evaluators, it’s not a self-contained system.
synthetic data is fine if you can ground the model somehow. that's why the o1/o3's improvements are mostly in reasoning, maths, etc., because you can easily tell if the data is wrong or not.
The value of synthetic data relies on having non-zero signal about which generated data is "better" or "worse". In a sense, this what reinforcement learning is about. Ie, generate some data, have that data scored by some evaluator, and then feed the data back into the model with higher weight on the better stuff and lower weight on the worse stuff.
The basic loop is: (i) generate synthetic data, (ii) rate synthetic data, (iii) update model to put more probability on better data and less probability on worse data, then go back to (i).
But who rates the synthetic data? If it is humans, I can understand that this is another way to get human knowledge into it, but if it's rated by AI, isn't it just a convoluted way of copying the rating AI's knowledge?
Many things are more easily scored than produced. Like it's trivial to tell whether a poem rhymes, but writing one is a comparatively slow and difficult task. So hopefully since scoring is easier/more-discerning than generating, the idea is you can generate stuff, classify it as good or bad, and then retrain on the good stuff. It's kindof an article of faith for a lot of AI companies/professionals as well, since it prevents you from having to face a data wall, and is analogous to a human student practicing and learning in an appealing way.
As far as I know it doesn't work very well so far. It is prone to overfitting, where it ranks highly some trivial detail of the output eg "if a summary starts with a byline of the author its a sign of quality" and then starts looping on itself over and over, increasing the frequency and size of bylines until it's totally crommed off to infinity and just repeating a short phrase endlessly. Humans have good baselines and common sense that these ML systems lack, if you've ever seen one of those "deep dream" images it's the same kind of idea. The "most possible dog" image can be looks almost nothing like a dog in the same way that the "most possible poem" may look nothing like a poem.
> This technique, the "Self-Taught Reasoner" (STaR), relies on a simple loop: generate rationales to answer many questions, prompted with a few rationale examples; if the generated answers are wrong, try again to generate a rationale given the correct answer; fine-tune on all the rationales that ultimately yielded correct answers; repeat. We show that STaR significantly improves performance on multiple datasets compared to a model fine-tuned to directly predict final answers
But there are a few others. In general good data is good data. We're definitely learning more about how to produce good synthetic version.
One issue with that is that the model may learn to smuggle data. You as a human think that the plain reading of the words is what is doing the reasoning, but (part of) the processing is done by the exact comma placement and synonym choice etc.
Data smuggling is a known phenomenon in similar tasks.
I don't think data smuggling is relevant in star style scenarios. You're still validating the final output. If it works on test data, what could be even smuggled.
> What good it's it to train a model basically on itself?
If the model generates data of variable quality, and if there's a good way to distinguish good data from bad data, then training on self-generated data might "bootstrap" a model to better performance.
This is common in reinforcement learning. Famously, AlphaGo Zero (https://en.wikipedia.org/wiki/AlphaGo_Zero) learned exclusively on self-play, without reference to human-played games.
Of course, games have a built-in critic: the better strategy usually wins. It's much harder to judge the answer to a math problem, or decide which essay is more persuasive, or evaluate restaurant recommendations.
If we get to a point where we have a model that when fed a real world stream of data (YouTube, surveillance cameras, forum data, cell phone conversations etc.) and can prune out a good training set for itself then you’re at the point where the LLM is in a feedback loop where it can improve itself. That’s AGI for all intents and purposes.
I don't think oai has any moat at all. If you look around, QwQ from Alibaba is already pushing o1-preview performances. I think oai is only ahead by 3~6 months at most.
If their AGI dreams would come true it might be more than enough to have 3 months head start. They probably won't, but it's interesting to ponder what the next few hours, days, weeks would be for someone that would wield AGI.
Like let's say you have a few datacenters of compute at your disposal and the ability to instantiate millions of AGI agents - what do you have them do?
I wonder if the USA already has a secret program for this under national defense. But it is interesting that once you do control an actual AGI you'd want to speed-run a bunch of things. In opposition to that, how do you detect an adversary already has / is using it and what to do in that case.
How many important problems are there where a 3 month head start on the data side is enough to win permanently and retain your advantage in the long run?
I'm struggling to think of a scenario where "I have AGI in January and everyone else has it in April" is life-changing. It's a win, for sure, and it's an advantage, but success in business requires sustainable growth and manageable costs.
If (random example) the bargain OpenAI strikes is "we spend every cent of our available capital to get AGI 3 months before the other guys do" they've now tapped all the resources they would need to leverage AGI and turn it into profitable, scalable businesses, while the other guys can take it slow and arrive with full pockets. I don't think their leadership is stupid enough to burn all their resources chasing AGI but it does seem like operating and training costs are an ongoing problem for them.
History is littered with first-movers who came up with something first and then failed to execute on it, only for someone else to follow up and actually turn the idea into a success. I don't see any reason to assume that the "first AGI" is going to be the only successful AGI on the market, or even a success at all. Even if you've developed an AGI that can change the world you need to keep it running so it can do that.
Consider it this way: Sam Altman & his ilk have been talking up how dangerous OpenAI's technology is. Are risk-averse businessmen and politicians going to be lining up to put their livelihood or even their lives in the hands of "dangerous technology"? Or are they going to wait 3-6 months and adopt the "safe" AGI from somebody else instead?
Well that's the thought exercise. Is there something you can do with almost unlimited "brains" of roughly human capability but much faster, within a few days / weeks / months. Lets say you can instantiate 1 million agents, for 3 months, and each of them is roughly 100x faster than a human, that means you have the equivalent of 100 million human-brain-hours to dump into whatever you want, as long as your plans don't require building too many real world things that actually require moving atoms around, I think you could do some interesting things. You could potentially dump a few million hours into "better than AGI AI" to start off for example, then go to other things. If they are good enough you might be able to find enough zero-days to disable any adversary through software, among other interesting things.
Where does "almost unlimited" come into the picture though? I see people talking like AGI will be unlimited when it will be limited by available compute resources, and like I suggested, being 'first' might come at the cost of the war chest you'd need to access those resources.
What does it take to instantiate 1 million agents? Who has that kind of money and hardware? Would they still have it if they burn everything in the tank to be first?
> Where does "almost unlimited" come into the picture though
>> Like let's say you have a few datacenters of compute at your disposal and the ability to instantiate millions of AGI agents - what do you have them do?
> has that kind of money and hardware?
Any hyperscaler plus most geopolitical main players. So the ones who matter.
Once you have AGI you use it to collect resources to cripple competitors and to build a snowball effect to make yourself unbeatable. 3 months of AGI is enough in the right hands to dominate the world economically.
Only if the AGI is cheaper than a human, in the case the AGI is more expensive than a human there wont be any snowballing. And the most likely case is that the first AGI is more expensive to run than a human, a few months of having overly expensive human level AI bots wont disrupt the world at all.
That is why being #2 in technical product development can be great. Someone else pays to work out the kinks, copy what works and improve on it at a fraction of the cost. You see it time and time again.
There is an enormous "iceberg" of untapped non-public data locked behind paywalls or licensing agreements. The next frontier will be spending money and human effort to get access to that data, then transform it into something useful for training.
Great improvements and all, but they are still no closer (as of 4o regular) to having a system that can be responsible for work. In math problems, it forgets which variable represents what, in coding questions it invents library fns.
I was watching a YouTube interview with a "trading floor insider". They said they were really being paid for holding risk. The bank has a position in a market, and it's their ass on the line if it tanks.
ChatGPT (as far as I can tell) is no closer to being accountable or responsible for anything it produces. If they don't solve that (and the problem is probably inherent to the architecture), they are, in some sense, polishing a turd.
If an LLM can't be left to do mowing by itself, but a human will have to closely monitor and intervene at every its steps, then it's just a super fast predictive keyboard, no?
And one might also wonder still if we need a general language model to mow the grass or just a simpler solution towards to problem of driving a mower over a fixed property line automatically. Something you could probably solve with wwii era technology, honestly.
They would want to, if they thought they could, because doing so would unblock a ton of valuable use cases. A tax preparation or financial advisor AI would do huge numbers for any company able to promise that its advice can be trusted.
Obviously not. I want legislation which imposes liability on OpenAI and similar companies if they actively market their products for use in safety-critical fields and their product doesn’t perform as advertised.
If a system is providing incorrect medical diagnoses, or denying services to protected classes due to biases in the training in the training data, someone should be held accountable.
This is underdiscussed. I don't think people understand just how worthless AI is in a ton of fields until it is able to be held liable and be sent to prison.
There are a lot of moral conundrums that are just not going to work out with this. Seems like an attempt to just offload liability and it seems like pretty much everybody has caught onto that as being it's main selling point and probably main thing that will keep it from ever being accepted for anything important.
Train for what? For making videos? Train from people’s comments? There’s a lot of garbage on AI slop on youtube, how would this be sifted out? I think there’s more value here on HN in terms of training, but even that, to what avail?
YouTube is such a great multimodal dataset—videos, auto-generated captions, and real engagement data all in one place. That’s a strong starting point for training, even before you filter for quality. Microsoft’s Phi-series models already show how focusing on smaller, high-quality datasets, like textbooks, can produce great results. You could totally imagine doing the same thing with YouTube by filtering for high-quality educational videos.
Down the line, I think models will start using video generation as part of how they “think.” Picture a version of GPT that works frame by frame—ask it to solve a geometry problem, and it generates a sequence of images to visualize the solution before responding. YouTube’s massive library of visual content could make something like that possible.
From what I read openai is having trouble bc not enough data.
If u think about it, any videos on YouTube of real world data contribute to its understanding of physics at minimum. From what I gather they do pre training on tons of unstructured content first and that contributes to overall smartness.
Considering how evasive they've been, it might also be YouTube.
> When pressed on what data OpenAI used to train Sora, Murati didn’t get too specific and seemed to dodge the question. “I’m not going to go into the details of the data that was used, but it was publicly available or licensed data,” she says. Murati also says she isn’t sure whether it used videos from YouTube, Facebook, and Instagram. She only confirmed to the Journal that Sora uses content from Shutterstock, with which OpenAI has a partnership.
In my intuition it makes sense that there is going to be some significant friction in LLM development going forward. We're talking about models that will cost upwards of $1bn to train. Save for a technological breakthrough, GPT-6/7 will probably have to wait for hardware to catch up.
I think the main bottleneck right now is training data - they've basically exhausted all public sources of data, so they have to either pay humans to generate new data from scratch or pay for the reasoning models to generate (less useful) synthetic training data. The next bottleneck is hardware, and the least important bottleneck is money.
It's more that you have to allocate the compute the right way.
Noam Brown's analogy is, you could train a massive one shot foundation model to predict the next best Go move, but that would be stupid. Better to use some test time search. You get better results for less money.
How about just an updated gpt 4o with all newer data? It would go a long way. Currently it doesn't know anything since Oct 2023 (without having to do a web search).
"OpenAI’s is called GPT-4, the fourth LLM the company has developed since its 2015 founding." - that sentence doesn't fill me with confidence in the quality of the rest of the article, sadly.
I've been messing around with base (not instruction tuned) LLMs; they often evade AI detectors and I wouldn't be surprised if they evade this kind of detection too, at least with a high temperature
No, I'm complaining that just because GPT-4 is called GPT-4 doesn't mean it's the fourth LLM from OpenAI.
Off the top of my head: GPT-2, Codex, GPT-3 in three different flavors (babbage, curie, davinci), GPT-3.5.
Suggesting that GPT-4 was "fourth" simply isn't credible.
Just the other day they announced a jump from o1 to o3, skipping o2 purely because it's already the name of a major telecommunications brand in Europe. Deriving anything from the names of OpenAI's products doesn't make sense.
While I’m sure it’s unintentional, that amounts to nitpicking. I can easily find three to include and pass over the rest. Face value turns out to be a decent approximation.
The thing is that I think it could be an optimal way of saying it. Should we not put it into context of making a particular LLM? Why count three versions of three LLMs? They made it hard to choose the one that makes up for not having GPT 1. GPT 3.5 and Codex are both good candidates. And of course calling GPT 4 the third and fifth could be considered as well.
That doesn’t resolve the problem of whether third or fifth is better than fourth. I have yet to be convinced that their wording here shows that they fail to grasp the pace of the development.
Technically…? Does anyone here believe that the EU and Europe is the same thing? Would you find it weird if someone said that a Norwegian company was in Europe?
Parent is suggesting it would be weird for Europeans to call the UK as in Europe which as a European I can tell you is preposterous. That’s the kind of non sense you used to hear from Brexiter. They will have no sympathy from me.
The UK is part of Europe. It's technically, geographically, politically, historically, lingustially, tectonically and socially correct. In what ways is it not?
Are Cuba or Haiti part of North America? A lot of British people feel like their civilization is meaningfully distinct from “Europe”, even though they’re part of it in a technical geographical sense.
In general yes, but it depends on if you consider central america as its own continent and if you include them there and how you delineate north/south america. Groupings differ based on your education.
I think the thing that makes the UK different is that there is no other option besides them being a separate thing/continent. Are you suggesting that the UK is it's own continent? Would that be with the faroese and the Greenlanders?
The UK might feel different, but they are not separate. The french feel different from the bulgarians, but that does not mean they are on a separate continent, politically or geographically.
EDIT:
> A lot of British people feel like their civilization is meaningfully distinct
This is, to borrow a word, "balderdash". Looking at the influence vikings, romans and normans have had that is a rubbish argument. Just like other countries in europe the british culture is built on the stones of other cultures, and just like many other countries they subsumed other cultures because of kings or other political dominance.
But I'm guessing we can agree that any major landmass is generally belonging to a continent? Like we all agree that greenland, new zealand, japan, etc generally belong to a continent?
So to what continent do those british people think they belong?
If you asked someone directly “what continent is Britain part of”, they would surely say Europe, even if they would be unlikely to describe themselves as European. Language is funny that way.
I specifically asked what "those british people" think in response to a post saying "If British people don’t feel like they’re part of “the Continent”".
I was clearly asking what those specific british people think.
New Zealand is not part of a continent (unless you consider Zealandia [1] one, which few do). It's a bunch of islands in the middle of the sea, far from other land. It is part of named regions which sometimes substitute for continents when people want to divide up the world for some purpose like sports or economics, including Oceania and Australasia.
Great Britain (the island) is very close to mainland Europe, and was directly part of it a few thousand years ago. The situation is totally different.
That's pretty much the definition of continent, right? The term continent is not scientifically based unless you want to argue that there are 16-ish continents and that South Georgia is it's own continent (and even tectonically its arbitrary since what we consider to be major, minor, micro are arbitrary).
The point was that any closeby landmass besides europe is either in europe or in north america, and I have a hard time seeing the argument for UK being in North America or America at all.
If we're generous the article considers versions that were significant improvements. 4o is hardly better on real-world usage (benchmarks are gamed to death) than the original 4.
The issue isn't the grammar. It is that there are 5 distinct LLMs from OpenAI that you can use right now as well as 4 others that were deprecated in 2024.
Well, here's the interesting part - gpt2 has been writing news since well before gpt3 was launched. Remember when "news" started getting weirdly reptative? When just about any product had a review avaliable? When the amount of slop content just _exploded_? Thats when the ai colonization of the internet began.
> At best, they say, Orion performs better than OpenAI’s current offerings, but hasn’t advanced enough to justify the enormous cost of keeping the new model running.
If you offer an API you need to dedicate servers to it that keep the model loaded in GPU memory. Unless you don't care about latency at all.
Though I wouldn't be surprised if the bigger reason is the PR cost of releasing with an exciting name but unexciting results. The press would immediately declare the end of the AI growth curve
There definitely is, storage, machines at the ready, data centers, etc. Also OpenAI basically loses money every time you interact with ChatGPT https://www.wheresyoured.at/subprimeai/
The article definitely has issues, but to me what's relevant is where it's published. The smart money and experts without a vested interest have been well aware LLMs are an expensive dead for over a year and have been saying as much (Gary Marcus for instance). That this is starting to enter mainstream consciousness is what's newsworthy.
By whom? He seems highly credible to me, and his credentials check out, especially compared to hype men like Sam Altman. All youre doing is spreading FUD by an unnamed "they"
He only criticizes ai capabilities, without creating anything himself. Credentials are effectively meaningless. With every new release, he clamors for attention to prove how right he was—and always will be. That’s precisely why he lacks credibility.
He started and then sold a machine learning startup to Uber. He's also written multiple books about the construction of the human mind and he has a PhD from MIT. I would hardly call that creating nothing. He's not clamoring for attention, he's asking that AI be regulated and pointing out a lot of the glaring issues with the field.
I’m not smart enough or interesting enough to be hired by OpenAI to expertly solve problems and explain how to the AI. However, I like to think there isn’t enough money in the world for me to sell out my colleagues like that.
I want AI to help me in the physical world: folding my laundry, cooking and farming healthy food, cleaning toilets. Training data is not lying around on the internet for free, but it's also not impossible. How much data do you need? A dozen warehouses full of robots folding and unfolding laundry 24/7 for a few months?
I think it would be many decades before I'd trust a robot like that around small children or pets. Robots with that kind of movement capability, as well as the ability it pick up and move things around, will be heavy enough that a small mistake could easily kill a small child or pet.
That's a solved problem for small devices. And we effectively have "robots" like that all over the place. Sliding doors in shops/trains/elevators have been around for ages and they include sensors for resistance. Unless there's 1. extreme cost cutting, or 2. bug in the hardware, devices like that wouldn't kill children these days.
Even for adults, a robot that would likely have to be close to as massive as a human being, in order to do laundry and the like, would spook me out, moving freely through my place.
People also have essentially wild beasts in their home: cats. If cats were the size of small dogs they would kill people all the time, but we love them when they are small enough so they just claw you bloody.
Since we can live with that we can live with anything that doesn't outright murder us.
That's the point being made. It's transformed robotics research, yes, but it both remains to see whether it will have a truly transformative effect on the field as experienced by people outside academia (I think this is quite probable) and more pointedly when.
I think this is an opinion borne out of weariness with constant promises that amazing robots are right around the corner (as they have been for 20 odd years now). For anyone who is close to the front line, I think the resounding consensus is clear - this time is different, unbelievably different, and capability development is going to accelerate dramatically.
I think it's impossible to spend a lot of time with these models without believing robotics is fundamentally about to transform. Even the most sophisticated versions of robotic logic pre-LLM/VLM feel utterly trivial compared to what even rudimentary applications of these large models can accomplish.
> believing robotics is fundamentally about to transform
These are not even remotely the same thing. Something that has happened already and is verifiable fact is not the same thing as your opinion, even if your opinion is based on a lot of sound arguments and reasoning.
Very tiresome to read so many claims of fact based on opinion of what will happen in the future.
The discussion was about whether robotics was about to transform or not. And obviously it is because of how much basic robotics workloads improve with these models.
Apparently even english tenses are too hard, let alone anything else. Bald faced lie, to claim what you think might happen in the future has already happened in the past. No matter "what the discussion was about", or what arguments you bring to support your estimation of the future.
Laundry folding is an instructive example. Machines have been capable of home-scale laundry folding for over a decade, with two companies Foldimate and Laundroid building functional prototypes. The challenge is making it cost-competitive in a world where most people don't even purchase a $10 folding board.
I would guess that most cooking and cleaning tasks are in basically the same space. You don't need fine motor control to clean a toilet bowl, but you've gotta figure out how to get people to buy the well-proven premisting technology before you'll be able to sell them a toilet-cleaning robot.
Counterexample: Everyone uses dishwashers. Yet I don’t think we’ll have a robot doing the dishes human-style, or even just filling up and clearing out a dishwasher, within the next decade or two, regardless of price.
Part of the tradeoff there is efficiency. I like my dishwasher because it's as good at getting things clean as I am but it does it using less water and less soap, and at scale, it takes less time too. It's just a great use case for machine automation because you can do clever stuff w/a dishwasher that's hard to replicate outside of that closed environment.
I struggle to imagine a scenario where a 1-2 person household would get the same benefits from something like a laundry-folding robot. I hate folding my laundry and I still can't imagine buying one since I simply don't do laundry that often. If I really wanted to spend less time doing laundry, I could spend the cost of that laundrybot on a larger collection of clothing to wear, for that matter.
Robot vacuums are a good comparison point since vacuuming is something you (ideally) do frequently that is time and labor intensive. I do own one of those, and if it got better at dealing with obstacles thanks to "AI" I would definitely like that.
I think it would have to be a general-purpose robot, and doing the laundry would just be one of many things it can do, similar to how running a particular program is only one of many things a computer can do. More than that, I believe it would actually require a general-purpose robot to handle all contingencies that can arise in doing laundry.
As someone who does laundry about twice a week, it would certainly be nice. But it’s a pie in the sky at this time even just on the technological side.
There's plenty of machines which are expensive, bulky, single purpose and yet commercially successful. The average American household has a kitchen range, refrigerator, dishwasher, laundry machine, dryer, television, furnace, and air conditioner. Automatic coffee machines and automatic vacuums are less universal but still have household penetration in the millions. I really think the household tasks with no widely available automation are simply the ones that nobody cares enough about doing to pay for automation.
A robot servant that does literally 100% of chores would be a game changer, and I expect we'll get there at some point, but it will probably have to be a one-shot from a consumer perspective. A clever research idea to reach 25% or 50% coverage still isn't going to lead to a commercially viable product.
"Orion’s problems signaled to some at OpenAI that the more-is-more strategy, which had driven much of its earlier success, was running out of steam."
So LLMs finally hit the wall. For a long time, more data, bigger models, and more compute to drive them worked. But that's apparently not enough any more.
Now someone has to have a new idea. There's plenty of money available if someone has one.
The current level of LLM would be far more useful if someone could get a conservative confidence metric out of the internals of the model. This technology desperately needs to output "Don't know" or "Not sure about this, but ..." when appropriate.
Anecdotally Claude is just as bad as every other LLM.
Step into more niche areas e.g. I am trying to use it with Scala macros and at least 90% of the time it is giving code that either (a) fails to compile or (b) is just complete gibberish.
And at no point ever has it said it didn't know something.
Yep, get into any sufficiently deep niche (i.e. actually almost any non-trivial app) and the LLM magic fades off.
Yeah sure you can make a pong clone in html/js and that's mainly because there the internet is full of pong clone demos. Ask how to constraint a statsmodels lineal model in some non-standard way? It will gaslight how it is possible and make you loss time in the process.
Making a pong clone by telling the LLM to make a pong clone is a cute trick that sometimes works, but that's not the way anyone who understands how to properly use these tools is using them. You don't describe and app and hope the LLM builds it correctly. You have to know how to architect an application and you use the LLM to build small pieces of code. For example, you tell it to build a function that does x, takes the inputs a, b, and c and returns z.
LLMs don't turn non-coders into coders. It gives actual coders superpowers.
No true scottsman fallacy. I know how to use them, but using them "correctly" still produces many errors.
They suck at non-trivial code outside of standard library usage and boilerplate coding: I gave an example and parent did as well. In that regard would at least change your phrase from "actual coders" to "actual senior coders", as any junior receiving bad advice (in eternal loops as LLMs normally like to do it) is only going to make them waste time and tokens.
My point is that while you do have to give them coding problems that would have appeared in their training set (I guess you could call that trivial), every coding problem becomes trivial when you break it down to it's constituent parts. As you know, the biggest applications are just a lot of very simple building blocks working together. The point of using LLMs to code is not to solve complex problems. It's just to write code you could have written yourself at the speed of light using a natural language interface.
The way you described using LLMs to code seems like the approach someone who doesn't know how to build software might take, which is why I used the wording I did. From that angle, I agree with you - I can't even get Sonnet to create a working prototype of a basic game from a prompt. That said, I'm using it to build a far more complex enterprise web app step by step by using it in the way I mentioned above. It does work for these things, but you have to already know how to do what the LLM is doing.
I mentioned the pong example because that is what non-coders LLM users show and what the industry is proposing as the future of software development: no coding experience necessary.
> It does work for these things, but you have to already know how to do what the LLM is doing.
Yes, we totally agree. But even then, using models "correctly" in my experience and breaking down the problems for them gets you so far, once you start using weird/niche APIs (probably even your own APIs when your project gets big enough and you are not working with much boilerplate anymore) the LLM will start getting single concepts wrong.
And don't get me wrong, I understand those as limitations of a tech that still is immensely useful in the correct hands. My only issue with that is how these products are actually being marketed: as junior devs copilots or even replacements.
As a coder with some noncoder friends who have made some very impressive things with chatGPT, you're selling it short.
It does both. It gives coders superpowers, and gives noncoders the ability to do things that would have previously taken them months, or another person.
They created a touchscreen GUI in tkinter with more-than-trivial behavior to use as a frontend for input for a device they created. They were able to describe what they wanted, and in less than two hours have it working. This is someone with no software experience.
Three years ago, if I had been asked to create something like that, it would have taken me more than two hours, just because I've never used tkinter and would have to spend time reading the docs and figuring out how to make the different input boxes and laying them out properly.
I looked at the code, and no, it's not great. It's not designed "well" and isn't very extensible. But it works for him, doesn't need to be extended, and all in half a morning.
Not even close. I’m a programmer but also a guitarist. I love asking it to tab out songs for me or asking it how many bars are in the intro of a song. It convincingly gives an answer that is always way off the mark.
What wall? Not a week has gone by in recent years without an LLM breaking new benchmarks. There is little evidence to suggest it will all come to a halt in 2025.
Sure, but "benchmarks" here seems roughly as useful as "benchmarks" for GPUs or CPUs, which don't much translate to what the makers of GPT need, which is 'money making use cases.'
O3 has demonstrated that OpenAI needs 1,000,000% more inference time compute to score 50% higher on benchmarks. If O3-High costs about $350k an hour to operate, that would mean making O4 score 50% higher would cost $3.5B (!!!) an hour. That scaling wall.
Even assuming that past rates of inference cost scaling hold up, we would only expect a 2 OoM decrease after about a year or so.
And 1% of 3.5b is still a very large number.
I used to run a lot of monte carlo simulations where the error is proportional to the inverse square root. There was a huge advantage of running for an hour vs a few minutes, but you hit the diminishing returns depressingly quickly. It would not surprise me at all if llms end up having similar scaling properties.
Yeah, any situation you need O(n^2) runtime to obtain n bits of output (or bits of accuracy, in the Monre Carlo case) is pure pain. At every point, it's still within your means to double the amount of output (by running it 3x longer than you have so far), but it gradually becomes more and more painful, instead of there being a single point where you can call it off.
I’m convinced they’re getting good at gaming the benchmarks since 4 has deteriorated via ChatGPT, in fact I’ve used 4-0125 and 4-1106 via the API and find them far superior to o1 and o1-mini at coding problems. GPT4 is an amazing tool but the true capabilities are being hidden from the public and/or intentionally neutered.
> I’ve used 4-0125 and 4-1106 via the API and find them far superior to o1 and o1-mini at coding problems
Just chiming in to say you're not alone. This has been my experience as well. The o# line of models just don't do well at coding, regardless of what the benchmarks say.
All the benchmarks provide substantial scaffolding and specification details, and that's if they are zero-shot at all, which they often are not. In reality, nobody wants to spend as much time providing so much details or examples just to get the AI to write the correct function, when that same time and effort you'd have used to write it yourself.
Also, those benchmarks often run the model K times on the same question, and if any one of them is correct, they say it passed. That could mean if you re-ran the model 8 times, it might come up with the right answer only once. But now you have to waste your time checking if it is right or not.
I want to ask: "Write a function to count unique numbers in a list" and get the correct answer the first time.
What you need to ask:
"""
Write a Python function that takes a list of integers as input and returns
the count of numbers that appear exactly once in the list.
The function should:
- Accept a single parameter: a list of integers
- Count elements that appear exactly once
- Return an integer representing the count
- Handle empty lists and return 0
- Handle lists with duplicates correctly
Please provide a complete implementation.
"""
And run it 8 times and if you're lucky it'll get it correct zero-shot.
Edit: I'm not even aware of a Pass@1, zero-shot, and without detailed prompting (natural prompting) benchmark. If anyone knows one let me know.
Not really. o3-low compute still stomps the benchmarks and isn't anywhere that expensive and o3-mini seems better than o1 while being cheaper.
Combine that with the fact that LLM inference has reduced orders of magnitudes in cost the last few years and hampering over the inference costs of a new release seems a bit silly.
If you are talking about ARC benchmark, then o3-low doesn't look that special if you take into account there are plenty of finetuned models with much smaller resources achieved 40-50% results on private set (not semi-private like o3-low).
- I'm not just talking about ARC. On frontier Math, we have 2 scores, one with pass@1 and another with consensus vote with 64 samples. Both scores are much better than previous Sota.
- Also apparently, ARC wasn't a special fine-tune but rather some of the training set in the corpus for pre-training.
>that result is not verifiable, not reproducable, unknown if it was leaked and how it was measured. Its kinda hype science.
It will be verifiable when the model is released. Open ai haven't released any benchmark scores that were shown falsified later so unless you have an actual reason to believe they're outright lying then it's not something to take seriously.
Frontier Math is a private benchmark with its highest tier of difficulty Terrence Tao says:
“These are extremely challenging. I think that in the near term basically the only way to solve them, short of having a real domain expert in the area, is by a combination of a semi-expert like a graduate student in a related field, maybe paired with some combination of a modern AI and lots of other algebra packages…”
Unless you have a reason to believe answers were leaked then again, not interested in baseless speculation.
>its private for outsiders, but it was developed in "collaboration" with OAI, and GPT was tested in the past on it, so they have it in logs somewhere.
They have logs of the questions probably but that's not enough. Frontier Math isn't something that can be fully solved without gathering top experts at multiple disciplines. Even Tao says he only knows who to ask for the most difficult set.
Basically, what you're suggesting at least with this benchmark in particular is far more difficult than you're implying.
>If you think this entire conversation is pointless, then why do you continue?
There's no point arguing about how efficient the models are being (the original point) if you won't even accept the results of the benchmarks. Why i'm continuing ? For now, it's only polite to clarify.
> Frontier Math isn't something that can be fully solved without gathering top experts
Tao's quote above referred on hardest 20% problems, they have 3 levels of difficulty, presumably first level is much easier. Also, as I mentioned OAI collaborated on creating benchmark, so they could have access to all solutions too.
> There's no point arguing
Lol, let me ask again, why you are arguing then? Yes, I have strong reasonable(imo) doubt that those results are valid.
The lowest set is easier but still incredibly difficult. Top experts are no longer required sure but that's it. You'll still need the best of the best undergrads at the very least to solve it.
>Also, as I mentioned OAI collaborated on creating benchmark, so they could have access to all solutions too.
Open AI didn't have any hand in providing problems, why you assume they have the solutions I have no idea.
>Lol, let me ask again, why you are arguing then? Yes, I have strong reasonable(imo) doubt that those results are valid.
Are you just bring obtuse or what ? I stopped arguing with you a couple responses ago. You have doubts? good for you. They don't make much sense but hey, good for you.
Not necessarily. And this is the problem with ARC that people seem to forget.
- It's just a suite of visual puzzles. It's not like say GSM8K where proficiency in it gives some indication on Math proficiency in general.
- It's specifically a suite of puzzles that LLMs have shown particular difficulty in.
Basically how much compute it takes to handle a task in this benchmark does not correlate with how much it will take LLMs to compute tasks that people actually want to use LLMs for.
If the benchmark is not representative of normal usage* then the benchmark and the plot being shown are not useful at all from a user/business perspective and the focus on the breakthrough scores of o3-low and o3-high in ARC-AGI would be highly misleading. And also the "representative" point is really moot from the discussion perspective (i.e. saying o3 stomps benchmarks, but the benchmarks aren't representative).
*I don't think that is the case as you can at least make relative conclusions (i.e. o3 vs o1 series, o3-low is 4x to 20x the cost for ~3x the perf). Even if it is pure marketing they expect people to draw conclusions using the perf/cost plot from Arc.
PS: I know there are more benchmarks like SWE-Bench and Frontier Math, but this is the only one showing data about o3-low/high costs without considering the CodeForces plot that includes o3-mini (that one does look interesting, though right now is vaporware) but does not separate between compute scale modes.
>If the benchmark is not representative of normal usage* then the benchmark and the plot being shown are not useful at all from a user/business perspective and the focus on the breakthrough scores of o3-low and o3-high in ARC-AGI would be highly misleading.
ARC is a very hyped benchmark in the industry so letting us know the results is something any company would do whether it had a direct representation on normal usage or not.
>Even if it is pure marketing they expect people to draw conclusions using the perf/cost plot from Arc.
Again, people care about ARC, they don't care doing the things ARC questions ask. That it is un-economical to pay the price to use o3 for ARC does not mean it would be un-economical to do so for the tasks people actually want to use LLMs for. What does 3x the performance in say coding mean? You really think companies/users wouldn't put up with the increased price for that? You think they have Mturkers to turn to like they do with ARC?
ARC is literally the quintessential 'easy for humans, hard for ai' benchmark. Even if you discard the 'difficulty to price won't scale the same' argument, it makes no sense to use it for an economics comparison.
In summary: so the "stomps benchmarks" means nothing for anyone trying to make decisions on that announcement (yet they show cost/perf info). It seems, hipey.
The new idea is inference-time scaling, as seen in o1 (and o3 and Qwen's QwQ and DeepSeek's DeepSeek-R1-Lite-Preview and Google's gemini-2.0-flash-thinking-exp).
It’s not just a protocol buffer for concepts though (weak wharf Sapir, lakoff’s ubiquitous metaphors). Language itself is also a concept layer and plasticity and concept development is bidirectional. But (I’m not very versed in the language here re ‘latent space’) I would imagine the forward pass through layers converges towards near-token-matches before output, so you have very similar reason to token/language reasoning even in latent/conceptual reasoning? Like the neurons that nearly only respond to a single token for ex.
Is it "eerie"? LeCun has been talking about it for some time, and may also be OpenAI's rumored q-star, mentioned shortly after Noam Brown (diplomacybot) joining OpenAI. You can't hill climb tokens, but you can climb manifolds.
I imagine he means that when you reason in latent space the final answer is a smooth function of the parameters, which means you can use gradient descent to directly optimize the model to produce a desired final output without knowing the correct reasoning steps to get there.
When you reason in token space (like everyone is doing now) you are executing nonlinear functions when you sample after each token, so you have to use some kind of reinforcement learning algorithm to learn the weights.
I think there's a subtlety here about what makes (e.g. English) tokens different to points in latent space. Everything is still differentiable (at least in the ML sense) until you do random sampling. Even then you can exclude the sampling when calculating the gradient (or is this equivalent to the "manifold"?).
I don't see a priori why it would be better or worse to reason with the "superposition" of arguments in the pre-sampling phase rather than concrete realizations of those arguments found only after choosing the token. It may well be a contingent rather than necessary fact.
I wasn’t aware of others attempting manifolds for this before - just something I stumbled upon independently. To me the “eerie” part is the thought of an LLM no longer using human language to reason - it’s like something out of a sci fi movie where humans encounter an alien species that thinks in a way that humans cannot even comprehend due to biological limitations.
I am hopeful that progress in mechanistic interpretability will serve as a healthy counterbalance to this approach when it comes to explainability.. though I kinda worry that at a certain point it may be that something resembling a scaling law puts an upper bound on even that.
I remember (apocryphal?) Microsoft's chatbot developing pidgin to communicate to other chatbots. Every layer of the NN except the first and last already "think" in latent space, is this surprising?
Is it really alien or is it more similar to how we think? We don't think purely in language, it's more a kind of soup of language, sounds, images, emotions and senses that we then turn into language when we communicate with each other.
> it’s like something out of a sci fi movie where humans encounter an alien species that thinks in a way that humans cannot even comprehend due to biological limitations.
I've increasingly felt this since GPT2 wrote that news piece about unicorns back in 2019. These models are still so mysterious, when you think about it. They can often solve decently complex math problems, but routinely fail at counting. Many have learned surprising skills like chess, but only when prompted in very specific ways. Their emergent abilities constantly surprise us and we have no idea how they really work internally.
So the idea that they reason using something other than human language feels unsurprising, but only because everything about it is surprising.
This was my thought. Literally everything inside a neural network is a “latent space”. Straight from the embeddings that you use to map categorical features in the first layer.
Latent space is where the magic literally happens.
Seems a standard approach of AI research is to “move X into the latent space” where X is some useful function (eg diffusion) previously done in the “data” or “artefact” space. So seems very pedestrian not wild to make that step.
To output "don't know" a system needs to "know" too. Random token generator can't know. It can guess better and better, maybe it can even guess 99.99% of time, but it can't know, it can't decide or reason (not even o1 can "reason").
> Now someone has to have a new idea. There's plenty of money available if someone has one.
I honestly do claim to have some ideas where I see evidence that they might work (and I do attempt to work privately on a prototype if only out of curiosity and to see whether I am right). The bad news: these ideas very likely won't be helpful for these LLM companies because they are not useful for their agenda, and follow a very different approach.
So no money for me. :-(
Let me put it this way:
Have you ever talked to a person whose intelligence is miles above yours? It can easily become very exhausting. Thus an "insanely intelligent" AI would not be of much use for most people - it would think "too different" from such people.
There do exist tasks in commerce for which an insane amount of intelligence would make a huge difference (in the sense of being positive regarding some important KPIs), but these are rare. I can imagine some applications of such (fictional) "super-intelligent" AIs in finance and companies doing some bleeding-edge scientific research - but these are niche applications (though potentially very lucrative ones).
If OpenAI, Anthropic & Co were really attempting to develop some "super-smart" AI, they were working on such very lucrative niche applications where an insane amount of intelligence would make a huge difference, and where you can assume and train the AI operator to have a "Fields-medal level" intelligence.
Not really. Throwing a bunch of unfiltered garbage at the pretraining dataset, throwing in RLHF of questionable quality during post-training, and other current hacks - none of that was expected to last forever. There is so much low-hanging fruit that OpenAI left untouched and I'm sure they're still experimenting with the best pre-training and post-training setups.
One thing researchers are seeing is resistance to post-training alignment in larger models, but that's almost the opposite of a wall, they're figuring it out as well.
> Now someone has to have a new idea
OpenAI already has a few, namely the o* series in which they discovered a way to bake Chain of Thought into the model via RL. Now we have reasoning models that destroy benchmarks that they previously couldn't touch.
Anthropic has a post-training technique, RLAIF, which supplants RLHF,and it works amazingly well. Combined with countless other tricks we don't know about in their training pipeline, they've managed to squeeze so much performance out of Sonnet 3.5 for general tasks.
Gemini is showing a lot of promise with their new Flash 2.0 and Flash 2.0-Thinking models. They're the first models to beat Sonnet at many benchmarks since April. The new Gemini Pro (or Ultra? whatever they call it now) is probably coming out in January.
> The current level of LLM would be far more useful if someone could get a conservative confidence metric out of the internals of the model. This technology desperately needs to output "Don't know" or "Not sure about this, but ..." when appropriate.
You would probably enjoy this talk [0], it's by an independent researcher who IIRC is a former employee of Deepmind or some other lab. They're exploring this exact idea. It's actually not hard to tell when a model is "confused" (just look at the probability distribution of likely tokens), the challenge is in steering the model to either get back to the right track or give up and say "you know what, idk"
> Not really. Throwing a bunch of unfiltered garbage at the pretraining dataset, throwing in RLHF of questionable quality during post-training, and other current hacks - none of that was expected to last forever. There is so much low-hanging fruit that OpenAI left untouched and I'm sure they're still experimenting with the best pre-training and post-training setups.
Exactly! LLama3 and their .x iterations have shown that, at least for now, the idea of using the previous models to filter out the pre-training datasets and use a small amount of seeds to create synthetic datasets for post-training still holds. We'll see with L4 if it continues to hold.
GPT-3 was trained on 4:1 ratio of data to parameters. And for GPT-4 the ratio was 10:1. So to scale this out, GPT-5 should be 25:1. The parameter count jumped from 175B to 1.3T, which means GPT-5 should be 10T parameters and 250T training tokens. There is zero chance OpenAI has a training set of high quality data that is 250T tokens.
If I had to guess, they trained a model that was maybe 3-4T in size and used 30-50T high quality tokens and maybe 10-30 medium and low quality ones.
There is only one company in the world that stores the data that could get us past the wall.
The training cost of the above scaled GPT-5 is 150x GPT-4, which was 25k A100 for 90 days, which poor MFU.
Let’s assume they double MFU, it would mean 1M H100s. But let’s say they made algorithmic improvements, so maybe it’s only 250-500k H100s.
While the training cluster size was 100k and then grew to 150k, this cluster is suggestive of a smaller model and less data.
GPT-5 is not behind schedule. GPT-5 is called GPT-4o and it has been already released half a year ago. It was not revolutionary enough to be called 5, and prophet saint Altman was probably afraid to release new gen not exponentially improving, so it was rebranded in the last moment. It's speculation of course, but it is kinda obvious speculation.
This is the first I have heard of this in particular. Do you know of any article or source for more on the efforts to train GPT 5 and the decision to call it GPT 4o?
I think my biggest pet peeve is when someone shares an insight which is unmistakably based on intuition, inference, critical thinking, etc (all mental faculties we are allowed to use to come to conclusions in the face of information asymmetry btw)
...and then gets hit deadpan with the good old "Source?", like it's some sort of gotcha.
I think people have started to confuse "making logical conclusions without perfect info" with "misinformation"
-
Before certain people start acting like this is advocating for misinformation (which would be an incredible irony...) it's not.
I'm saying if you disagree with what someone supposits, just state so directly. Don't wrap it in a disingenous query for a source.
It doesn't even look like 4o is scaled up parameter wise from 4 and was released closer in time than either 3 or 4 were from their predecessors at a time where the scaling required for these next gen iterations has only gotten more difficult.
Critical thinking ? Lol it's just blind speculation.
If you disagree with their reasoning then you explain that.
You don't do this passive aggressive "source???" thing.
It's a bit like starting a Slack conversation with "Hi?": we all know you have a secondary objective, but now you're inserting an extra turn of phrase into the mix
Not everyone keeps up with LLM development enough to know how far apart the release dates for these models are, how much scaling (roughly) has been done on each iteration and a decent ballpark for how much open ai might try to scale up a next gen model.
To me, OP's speculation reads as obvious nonsense but that might not be the case for everybody. Asking for sources or such to what is entirely speculation is perfectly valid and personally, that comment does not ring as passive aggressive to me but maybe it's just me.
Just because someone doesn't know enough to refute the reasoning doesn't mean they must take whatever they read at face value.
If we're making this about the innocent bystanders now, that's all the more reason to be direct and say "I disagree." rather than indirectly expressing negative feelings (aka being passive aggressive) and asking for a source.
If anything just breezily asking for a source would imply to people who don't know better that this is a rather even keeled take and just needs some more evidence on top. "I disagree and here's why" nips that in the bud directly.
How is "I disagree" any more direct than "I've not heard anything like this. any source that would point at that?" Moreover who's to say this person even disagrees? Personally i don't always ask for them because of a disagreement.
I think the hanging point seems to be that you found the comment passive aggressive but i genuinely didn't.
Right, that's what makes this rabbit hole a bit wild. I'm not even expressing a disagreement, rhetorical or otherwise. What's more, there's nothing wrong with doing that either. There are circumstances where that's a perfectly appropriate thing to do.
And while I fully agree there absolutely is such a thing as smarmy commenters asking for sources in cases where it's misunderstanding something fundamental about the conversation (e.g. "Shakespeare is good", "oh really? source?!") or frivolous requests for factual information familiar to everyone ("global temperatures are rising? Source!?"), I don't know how someone could read this subthread and feel that my question falls into either of those categories.
And to use this of all things as a moment to die on the hill of advocating for fuzzy boundaries between speculation and fact, which absolutely is something that facilitates misinformation, and to be angry that such a thing would be interpreted as a favorable attitude toward misinformation, is completely baffling.
It's reasonable to ask for sources when an opinion is phrased as a fact, as GGP did. I don't see how you got that it was _unmistakably_ an opinion from that comment.
There is no way to deduce by intuition alone that GPT-5 == GPT-4o. So either that person has some information the rest of us aren't privy to, or it's an opinion phrased as a fact. In either case, it deserves clarification.
On a second read I see that the comment notes that it is intended as speculation, but still it seems rather confident in its own accuracy and I am not even sure it's wrong, but just looking for something that warrants the confidence.
I wrote my comment that way, based on my personal memories of the news cycle between gpt-4 and gpt-4o, and the claims raised by OpenAI about gpt-4o. The hype before 4o release was overwhelming, people have expected the same step up as between 3 and 4, and there were constant "leaks" from supposed insiders that gpt-5 is just at the horizon and will come out soon. And then they release 4o, which was a big standalone release, not some fine tuning like turbo or whatever else they made before.
Looking at the benchmarks it was also very expected in my opinion. Sure, the absolute results are/were sky high, but results relative to the previous gen were not exponential now, they were comparatively smaller than between 2 and 3, or 3 and 4. So I'm guessing that they have invested and worked for 2023-2024 on a brand new model, and branded it according to the model results.
That was clearly phrased like a fact, which may or may not be correct. If it had been phrased like an opinion we wouldn't be having this conversation...
The problem is once you believe their fact is wrong, just say "I think you're wrong <insert rest of comment>". Innocently asking for a source as if you're still on the fence is just performative and leads to these conversations where both sides just end up talking past each other:
A source for one underpinning of the incorrect fact comes up, then "well but that only proves X part of it, can you prove Y" and so on.
tl;dr I just find the quality of discourse is much higher when people are direct.
> I just find the quality of discourse is much higher when people are direct.
Well this certainly is a lot of work to make a mountain out of a mole hill, and I'm not sure it increases the quality of discussion either.
In any case, I think saying bold shit followed up with "it's speculation, but it's OBVIOUS speculation" is worth asking for some evidence. Obvious speculation implies it's sourced from something other than personal gut feeling.
To echo a sibling comment:
> Every time someone says their speculation is "obvious" it rings every possible alarm bell for someone who has completely lost grasp of the ability to distinguish between facts and speculation.
My sister got taken in by drone conspiracy theories, because for her it was just "obvious" that nobody would ever mistake a plane for a drone.
Meanwhile, aeronautics experts whose job it is to know about this have created an entire lexicon for the various perceptual illusions we experience relating to flight and airborne objects, precisely because it involves conditions where our intuitions fail. Many of them have to do with inability to orient depth, distance, or motion for lights at night.
Every time someone says their speculation is "obvious" it rings every possible alarm bell for someone who has completely lost grasp of the ability to distinguish between facts and speculation.
The road to misinformation is paved with overconfident declarations of the form: "it's so obvious, who needs sources!"
I think it's okay to make logical conclusions but you must base them in evidence, not just suppositions. Intuition is a good start to begin generating hypothesis, but it doesn't render conclusions. I interpreted the GP asking for sources as "can you give me some evidence that would help me reach the same conclusions you've reached". I think that's much preferable to just accepting random things people say at face value.
Even with evidence a logical conclusion can still a supposition (aka an uncertain belief), and often is in the face of the kind of information asymmetry inherent to any outsider commenting on a private company's internal roadmap... but I digress.
My point is simply that is we can skip the passive aggressiveness and just say "can you give me some more evidence that would help me reach the same conclusions you've reached".
Otherwise you're not actually asking for a source, you're just saying "I disagree" in a very roundabout way.
Not really, 4o was purpose built to be a light weight 4. Remember that 4o was also when GPT-4 became available to everyone. Before that ou had to be premium to use GPT-4, and got limited inquiries.
Meanwhile, the biggest opportunity lies not in whatever next thing OpenAI releases, but the rest of the enormous software industry actually integrating this technology and realizing the value it can deliver.
The lack of tech literacy in this article is a bit concerning:
>Some researchers take this so seriously they won’t work on planes, coffee shops or anyplace where someone could peer over their shoulder and catch a glimpse of their work.
I'm almost certain that originally this was meant to be a reference to public wifi networks, as planes and coffee shops are often the frequently cited prototypical examples. They made it literally into a matter of someone looking over their shoulder, which loses so much in translation it's almost how you would write this as a joke to illustrate someone missing the point.
>OpenAI and its brash chief executive, Sam Altman
This also strikes me as nonsense. It's the first I've ever heard of someone describing Sam Altman as brash. The only way I can see them getting there is (1) tech executives are often brash (2) Altman is a tech executive (3) let's just go ahead and call him brash.
Nevertheless if this history of GPT5 and/or o3 training is accurate, it strikes me as significant news, but perhaps a missed opportunity to say more about the pertinent dynamics that explain why the training isn't working and/or to talk in interestingly specific ways about strategies for training, synthetic data, or other such things.
Did you read the article? All it basically says is that OpenAI faced struggles this past year -- specifically with GPT-5 aka Orion. And now they have o3, and other labs have made huge strides. So, yes, show me AI progress is slowing down!
One fundamental challenge to me is that if each training run because more and more expensive, the time it takes it to learn what works/doesn't work widens. Half a billion dollars for training a model is already nuts, but if it takes 100 iterations to perfect it, you've cumulatively spent 50 billion dollars... Smaller models may actually be where rapid innovation continues simply because of tighter feedback loops. O3 may be an example of this.
But if the scaling law holds true, more dollars should at some point translate into AGI, which is priceless. We haven't reached the limits yet of that hypothesis.
a) There is evidence e.g. private data deals that we are starting to hit the limitations of what data is available.
b) There is no evidence that LLMs are the roadmap to AGI.
c) Continued investment hinges on their being a large enough cohort of startups that can leverage LLMs to generate outsized returns. There is no evidence yet this is the case.
"There is no evidence that LLMs are the roadmap to AGI." - There's plenty of evidence. What do you think the last few years have been all about? Hell, GPT-4 would already have qualified as AGI about a decade ago.
No, GPT-4 would have been classified as it is today: a (good) generator of natural language. While this is a hard classical NLP task, it's a far cry from intelligence.
For an industry that spun off of a research field that basically revolves around recursive descent in one form or another, there's a pretty silly amount of willful ignorance about the basic principles of how learning and progress happens.
The default assumption should be that this is a local maximum, with evidence required to demonstrate that it's not. But the hype artists want us all to take the inevitability of LLMs for granted—"See the slope? Slopes lead up! All we have to do is climb the slope and we'll get to the moon! If you can't see that you're obviously stupid or have your head in the sand!"
I never said anything about usefulness, and it's frustrating that every time I criticize AGI hype people move the goalposts and say "but it'll still be useful!"
I use GitHub Copilot every day. We already have useful "AI". That doesn't mean that the whole thing isn't super overhyped.
So far we haven't even climbed this slope to the top yet. Why don't we start there and see if it's high enough or not first? If it's not, at the very least we can see what's on the other side, and pick the next slope to climb.
>What do you think the last few years have been all about?
Next token language-based predictors with no more intelligence than brute force GIGO which parrot existing human intelligence captured as text/audio and fed in the form of input data.
4o agrees:
"What you are describing is a language model or next-token predictor that operates solely as a computational system without inherent intelligence or understanding. The phrase captures the essence of generative AI models, like GPT, which rely on statistical and probabilistic methods to predict the next piece of text based on patterns in the data they’ve been trained on"
>Everything you said is parroting data you’ve trained on
"Just like" an LLM, yeah sure...
Like how the brain was "just like" a hydraulic system (early industrial era), like a clockwork with gears and differentiation (mechanical engineering), "just like" an electric circuit (Edison's time), "just like" a computer CPU (21st century), and so on...
He probably didn't need petabytes of reddit posts and millions of gpu-hours to parrot that though.
I still don't buy the "we do the same as LLMs" discourse. Of course one could hypothesize the human brain language center may have some similarities to LLMs, but the differences in resource usage and how those resources are used to train humans and LLMs are remarkable and may indicate otherwise.
No, they don’t - they don’t have the hardware, yet. But they do parrot sensory output to eg muscles that induce the expected video sensory inputs in response, in a way that mimics the video input of “other people doing things”.
And yet with multiple OoM more data he still didn't cost millions of dollars to be trained nor multiple lifetimes in gpu-hours. He probably didn't even register all the petabytes passing through all his "sensors", those are some characteristics that we are not even near understanding and much less replicating.
Whatever is happening in the brain is more complex as the perf/cost ratio is stupidly better for humans for a lot of tasks in both training and inference*.
*when considering all modalities, o3 can't even do the ARC AGI in vision mode but rather just json representations. So much for omni.
>Not text, he had petabytes of video, audio, and other sensory inputs. Heck, a baby sees petabytes of video before first word is spoken
A 2-3 year old baby could speak in a rural village in 1800, having just seen its cradle (for the first month/s), and its parents' hut for some more months, and maybe parts of the village afterwards.
Hardly "petabytes of training video" to write home about.
you are funny. Clearly your expertise with babies comes from reading books about history or science, rather than ever having interacted with one…
What resolution of screen do you think you would need to not distinguish from reality? For me personally i very conservatively estimate it to be on above OOM of 10 4k screens by 10, meaning 100k screens. If a typical 2h 4k is ~50gb uncompressed, that gives us about half a petabyte per 24h (even with eyes closed). Just raw unlabeled vision data.
Probably a baby has a significantly lower resolution, but then again what is the resolution from the skin and other organs?
So yes, petabytes of data within the first days of existence - well, likely before even being born since baby can hear inside the uterus, for example.
And very high signal data, as you’ve stated yourself (nothing to write home about) mainly seeing mom and dad, as well as from a feedback loop POV - a baby never tells you it is hungry subtly.
You have described something but you haven't explained why the description of the thing defines its capability. This is a tautology, or possibly a begging of the question, which takes as true the premise of something (that token based language predictors cannot be intelligent) and then uses that premise to prove an unproven point (that language models cannot achieve intelligence).
You did nothing at all to demonstrate why you cannot produce an intelligent system from a next token language based predictor.
What GPT says about this is completely irrelevant.
>You did nothing at all to demonstrate why you cannot produce an intelligent system from a next token language based predictor
Sorry, but the burden of proof is on your side...
The intelligence is in the corpus the LLM was fed with. Using statistics to pick from it and re-arrange it gives new intelligent results because the information was already produced by intelligent beings.
If somebody gives you an excerpt of a book, it doesn't mean they have the intelligence of the author - even if you have taught them a mechanical statistical method to give back a section matching a query you make.
Kids learn to speak and understand language at 3-4 years old (among tons of other concepts), and can reason by themselves in a few years with less than 1 billionth the input...
>What GPT says about this is completely irrelevant.
On the contrary, it's using its very real intelligence, about to reach singularity any time now, and this is its verdict!
Why would you say it's irrelevant? That would be as if it merely statistically parroted combinations of its training data unconnected to any reasoning (except of that the human creators of the data used to create them) or objective reality...
Person 1: rockets could be a method of putting things into Earth orbit
Person 2: rockets cannot get things into orbit because they use a chemical reaction which causes an equal and opposite force reaction to produce thrust'
Does person 1 have the burden of proof that rockets can be used to put things in orbit? Sure, but that doesn't make the reasoning used by person 2 valid to explain why person 1 is wrong.
BTW thanks for adding an entire chapter to your comment in edit so it looks like I am ignoring most of it. What I replied to was one sentence that said 'the burden of proof is on you'. Though it really doesn't make much difference because you are doing the same thing but more verbose this time.
None of the things you mentioned preclude intelligence. You are telling us again how it operates but not why that operation is restrictive in producing an intelligent output. There is no law that saws that intelligence requires anything but a large amount of data and computation. If you can show why these things are not sufficient, I am eager to read about it. A logical explanation would be great, step by step please, without making any grand unproven assumptions.
In response to the person below... again, whether or not person 1 is right or wrong does not make person 2's argument valid.
I can check but I am pretty sure that using a different argument to try and prove something is wrong will not make another person's invalid argument correct.
Person 3: Since we can leave earths orbit, we can reach faster than light speed, look at this graph over our progress making faster rockets we will for sure reach there in a few years!
So there is a theoretical framework which can be tested against to achieve AGI and according to that framework it is either not possible or extremely unlikely because of physical laws?
So, I think people in this thread, including me, have been talking past each other a bit. I do not claim that sentient AI will emerge. I am arguing that the person who is saying that it can't happen for a specific reason is not considering that the reason they are stating implicitly is that nothing can be greater than the sum of its parts.
Describing how an LLM operates and how it was trained does not preclude the LLM from ever being intelligent, and it almost certainly will not become intelligent, but you cannot say that it didn't for the reasons the person I am arguing with is saying, which is that intelligence can not come from something that works statistically on a large corpus of data written by people.
A thing can be more than the sum of its parts. You can take the English alphabet, which is 26 letters, and arrange those letters along with some punctuation to make an original novel. If you don't agree that means that you can get something greater than what defines it components, then you would have to agree that there are no original novels because they are composed of letters which were already defined.
So in that way, the model is not unable to think because it is composed of thoughts already written. That is not the limiting factor.
It's not like we discovered hot air ballons, and some people think we'll get to Moon and Mars with them...
> Does person 1 have the burden of proof that rockets can be used to put things in orbit? Sure, but that doesn't make the reasoning used by person 2 valid to explain why person 1 is wrong.
The reasoning by person 2 doesn't matter as much if 1 is making an ubsubstantiated claim to begin with.
>There is no law that saws that intelligence requires anything but a large amount of data and computation. If you can show why these things are not sufficient, I am eager to read about it.
Errors with very simple stuff while getting higher order stuff correct shows that this is not actual intelligence matching the level of performance exhibited, i.e. no understanding.
No person who can solve higher level math (like an LLM answering college or math olympiad questions) is confused by the kind of simple math blind spots that confuse LLMs.
A person understanding higher level math, would never (and even less so, consistently) fail a problem like:
"Oliver picks 44 kiwis on Friday. Then he picks 58 kiwis on Saturday. On Sunday, he picks double the number of kiwis he did on Friday, but five of them were a bit smaller than average. How many kiwis does Oliver have?"
> The reasoning by person 2 doesn't matter as much if 1 is making an ubsubstantiated claim to begin with.
But it doesn't make person 2's argument valid.
Everyone here is looking at the argument by person 1 and saying 'I don't agree with that, so person 2 is right!'.
That isn't how it works... person 2 has to either shut up and let person 1 be wrong in a way that is wrong, but not for the reasons they think, or they need to examine their assumptions and come up with a different reason.
No one is helped by turning critical thinking into team sports where the only thing that matters is that your side wins.
> If somebody gives you an excerpt of a book, it doesn't mean they have the intelligence of the author
A closely related rant of my own: The fictional character we humans infer from text is not the author-machine generating that text, not even if they happen to share the same name. Assuming that the author-machine is already conscious and choosing to insert itself is begging the question.
Sure they’ve hit the wall with obvious conversations and blog articles that humans produced, but data is a by product of our environment. Surely there’s more. Tons more.
What does culture and names and people have to do with the Global Position System?
You are right that we can have lots more data, if you are willing to consider other modalities. But that's not 'GPS'. Unless you are using an idiosyncratic definition of GPS?
Ignoring the confusion about 'GPS' for a moment: there's lots and lots of other data that could be used for training AI systems.
But, you need to go multi-modal for that; and you need to find data that's somewhat useful, not just random fluctuations like the CMB. So eg you could use YouTube videos, or even just point webcams at the real world. That might be able to give your AI a grounding in everyday physics?
There's also lots of program code you can train your AI on. Not so much the code itself, because compared to the world's total text (that we are running out of), the world's total human written code is relatively small.
But you can generate new code and make it useful for training, by also having the AI predict what happens when you (compile and) run the code. A bit like self-playing for improving AlphaGo.
> c) Continued investment hinges on their being a large enough cohort of startups that can leverage LLMs to generate outsized returns. There is no evidence yet this is the case.
Why does it have to be startups? And why does it have to be LLMs?
Btw, we might be running out of text data. But there's lots and lots more data you can have (and generate), if you are willing to consider other modalities.
You can also get a bit further with text data by using it for multiple epochs, like we used to do in the past. (But that only really gives you at best an order of magnitude. I read some paper that the returns diminish drastically after four epochs.)
This also isn't true. It'll clearly have a price to run. Even if it's very intelligent, if the price to run it is too high it'll just be a 24/7 intelligent person that few can afford to talk to. No?
Computers will be the size of data centres, they'll be so expensive we'll queue up jobs to run on them days in advance, each taking our turn... history echoes into the future...
Yea, and those statements were true. For a time. If you want to say "AGI will be priceless some unknown time into the future" then i'd be on board lol. But to imply it'll be immediately priceless? As in no cost spent today wouldn't be immediately rewarded once AGI exists? Nonsense.
Maybe if it was _extremely_ intelligent and it's ROI would be all the drugs it would instantly discover or w/e. But lets not imply that General Intelligence requires infinitely knowing.
So at best we're talking about an AI that is likely close to human level intelligence. Which is cool, because we have 7+ billion of those things.
This isn't an argument against it. Just to say that AGI isn't "priceless" in the implementation we'd likely see out of the gate.
There's no doubt been progress on the way to AGI, but ultimately it's still a search problem, and one that will rely on human ingenuity at least until we solve it. LLMs are such a vast improvement in showing intelligent-like behavior that we've become tantalized by it. So now we're possibly focusing our search in the wrong place for the next innovation on the path to AGI. Otherwise, it's just a lack of compute, and then we just have to wait for the capacity to catch up.
It seems to me that given how AI is likely to continuously increase capitalism's efficiency, your argument actually supports the claim you're trying to dispute.
Capitalism is not efficient, it's grabby. Read Bullshit Jobs. Moreover, capitalism isn't interested in efficiency, it's interested in grabbing more stuff. It's relatively effiicient at centralising power and resources into the pockets of shareholders, but that's probably not what you meant.
I think this is borne out even moreso in recent years, as environmental degradation continues, and we watch as capitalist systems are unable to do anything but continue to efficiently funnel money into the pockets of shareholders.
The word "efficient" can only plausibly be applied to overly simplified models in fantastical economic theories which don't reflect reality.
The kind of AI offered by companies like OpenAI may very well be an effective tool at grabbing more stuff though, sure. Or, rather, at convincing everyone they simply must move to this new area, that they control, effectively grabbing that newly created space.
I think you're both right and wrong. You're right that capitalism has become a paperclip machine, but capitalism also wants AI so it can cheaply and at scale replace the human components of the machine with something that has more work capacity for fewer demands.
The problem is that the people in power will want to maintain the status quo. So the end of human labor won't naturally result in UBI – or any kind of welfare – to compensate for the loss of income, let alone afford any social mobility. But wealthy people will be able to leverage AGI to defend themselves from any uprising by the plebs.
We're too busy trying to make humans irrelevant, but not asking what exactly we do as a species of 10+ billion individuals do afterwards. There's some excited discussions about a rebirth of culture, but I'm not sure what that means when machines can do anything humans can do but better. Perhaps we just tinker around with our hobbies until we die? I honestly don't think it will play out well for us.
The problem is that the "we" who are busy trying to make humans irrelevant seem to be completely unconcerned with the effects on the "we" who will be superfluous afterwards.
Machines can’t have fun for us. They can’t dance to a beat, they can’t experience altered states of mind. They can’t create a sense of belonging through culture and ritual. Yes we have lost a lot in the last 100 years but there are still pockets of resistance that carry old knowledge that “we the people” will be glad of in the coming century.
It's a similar story around extant ancient/indigenous cultures. And similarly we've seen apathy from elites, especially when indigenous rights get in the way of resource extraction or generating wealth in any way, and also witnessed condescension towards indigenous peoples by large segments of the world population. That's not to detract from the many defenders of indigenous rights, but if we look a the state of how older cultures, designated as 'obsolete' by wider society have been treated, I don't humans will fare well when silicon takes over.
> They can’t dance to a beat, they can’t experience altered states of mind.
I think the key is ensuring that “we” get to choose what society looks like in the AGI era. In the world today, even marginalized people have power. Look what happened to Assad. Look at the US - whether you believe they made the right decision or not, working class people were key to Trump’s victory, who may well institute tariffs as a way to protect working class jobs by insulating American industry from global competition. I’m not saying that will be successful, I’m saying that working class people got mad and a political change resulted.
Similarly I don’t see a world where AGI takes all the jobs and people do not respond by getting pissed off. My fear is that AGI is coupled with oppressive power structures to foreclose the possibility of a revolt. Opaque bureaucracy, total surveillance, fascist or authoritarian leaders, AI-controlled critical infrastructure, diminished and bankrupted free press, AI fake news, toxic social media…it could add up to a very dystopian outcome.
Democracies could thrive in the AGI era but we need to take many more steps to ensure we protect our societies and keep the interests of citizens paramount. One example is suggested by Harari in his most recent book, namely to ban AI bots from social media on the grounds that we should not permit AI agents to pretend to be citizens in the discussions of the public square.
> I think the key is ensuring that “we” get to choose what society looks like in the AGI era. In the world today, even marginalized people have power.
That's a bold assumption. Much of that assumption is predicated on the ability for the masses to revolt.
> Look what happened to Assad.
Wait for what will come after. Look at all the Arab Spring revolutions, and you see in their wake a number of dictatorships.
Anyhow, I'm not saying this is 100% how it's going to play out, but I definitely wouldn't bet against it. Holding all the keys and having all the resources are the wealthy, and the wealthy have no motivation to voluntarily just give up their position in society. And when humans have no value to leverage/be extracted in order to generate more wealth, their will be no way for the vast majority of people to become wealthy. Raw materials will still be valuable however, but, of course, these are controlled by the wealthy. And if those in power wish to gatekeep access to AGI, they can leverage their wealth and resources to automate a military and thus protect the raw materials that keep them in power.
People give communism a bad rap, but the soviets had maybe a quarter the resources, a much smaller population and logistical problems from geography and kept up with the US for decades, outpacing in several areas.
AGI will arrive like self driving cars. it’s not that you will wake up one day and we have it. cars gained auto-braking, parallel parking, cruise control assist. and over a long time you get to something like waymo, which still is location dependent. i think AGI will take decades but sooner will be some special cases that are effectively the same
Interesting idea. The concept of The Singularity would seem to go against this, but I do feel that seems unlikely and that a gradual transition is more likely.
However, is that AGI, or is it just ubiquitous AI? I’d agree that, like self driving cars, we’re going to experience a decade or so transition into AI being everywhere. But is it AGI when we get there? I think it’ll be many different systems each providing an aspect of AGI that together could be argued to be AGI, but in reality it’ll be more like the internet, just a bunch of non-AGI models talking to each other to achieve things with human input.
I don’t think it’s truly AGI until there’s one thinking entity able to perform at or above human level in everything.
The Singularity is caused by AI being able to design better AI. There's probably some AI startup trying to work on this at the moment, but I don't think any of the big boys are working on how to get an LLM to design a better LLM.
I still like the analogy of this being a really smart lawn mower, and we're expecting it to suddenly be able to do the laundry because it gets so smart at mowing the lawn.
I think LLMs are going to get smarter over the next few generations, but each generation will be less of a leap than the previous one, while the cost gets exponentially higher. In a few generations it just won't make economic sense to train a new generation.
Meanwhile, the economic impact of LLMs in business and government will cause massive shifts - yet more income shifting from labour to capital - and we will be too busy dealing with that as a society to be able to work on AGI properly.
I think this whole “AGI” thing is so badly defined that we may as well say we already have it. It already passes the Turing test and does well on tons of subjects.
What we can start to build now is agents and integrations. Building blocks like panel of experts agents gaming things out, exploring space in a Monte Carlo Tree Search way, and remembering what works.
Robots are only constrained by mechanical servos now. When they can do something, they’ll be able to do everything. It will happen gradually then all at once. Because all the tasks (cooking, running errands) are trivial for LLMs. Only moving the limbs and navigating the terrain safely is hard. That’s the only thing left before robots do all the jobs!
ok. what evidence is there that LLMs have already solved cooking? how does an LLM today know when something is burning or how to adjust seasoning to taste or whatever. this is total nonsense
It's easy. You can detect if something is burning in many different ways, from compounds in the air, to visual inspection. People with not great smell can do it.
As far as taste, all that kind of stuff is just another form of RLHF training preferences over millions of humans, in situ. Assuming the ingredients (e.g. parsley) tastes more or less the same across supermarkets, it's just a question of amounts, and preparation.
do you know that LLMs operate on text and don't have any of the sensory input or relevant training data? you're just handwaving away 99.9% of the work and declaring it solved. of course what you're talking about is possible, but you started this by stating that cooking is easy for an LLM and it sounds like you're describing a totally different system which is not an LLM
Because the recipes and the adjustments are trivial for an LLM to execute. Remembering things, and being trained on tasks at 1000 sites at once, sharing the knowledge among all the robots, etc.
The only hard part is moving the limbs and handling the fragile eggs etc.
But it's not just cooking, it's literally anything that doesn't require extreme agility (sports) or dexterity (knitting etc). From folding laundry to putting together furniture, cleaning the house and everything in between. It would be able to do 98% of the tasks.
It’s not going to know what tastes good by being able to regurgitate recipes from 1000s of sites. Most of those recipes are absolute garbage. I’m going to guess you don’t cook.
Well, kinda, but if you built a robot to efficiently mow lawns, it's still not going to be able to do the laundry.
I don't see how "when they can do something, they'll be able to do everything" can be true. We build robots that are specialised at specific roles, because it's massively more efficient to do that. A car-welding robot can weld cars together at a rate that a human can't match.
We could train an LLM to drive a Boston Dynamics kind of anthropomorphic robot to weld cars, but it will be more expensive and less efficient than the specialised car-welding robot, so why would we do that?
If a humanoid robot is able to move its limbs and digits with the same dexterity as a human, and maintain balance and navigate obstacles, and gently carry things, everything else is trivial.
Welding. Putting up shelves. Playing the piano. Cooking. Teaching kids. Disciplining them. By being in 1 million households and being trained on more situations than a human, every single one of these robots would have skills exceeding humans very quickly. Including parenting skills. Within a year or so. Many parents will just leave their kids with them and a generation will grow up preferring bots to adults. The LLM technology is the same for learning the steps, it's just the motor skills that are missing.
OK, these robots won't be able to run and play soccer or do somersaults, yet. But really, the hardest part is the acrobatics and locomotion etc. NOT the knowhow of how to complete tasks using that.
But that's the point - we don't build robots that can do a wide range of tasks with ease. We build robots that can do single tasks super-efficiently.
I don't see that changing. Even the industrial arm robots that are adaptable to a range of tasks have to be configured to the task they are to do, because it's more efficient that way.
A car-welding robot is never going to be able to mow the lawn. It just doesn't make financial sense to do that. You could, possibly, have a singe robot chassis that can then be adapted to weld cars, mow the lawn, or do the laundry, I guess that makes sense. But not as a single configuration that could do all of those things. Why would you?
> But that's the point - we don't build robots that can do a wide range of tasks with ease. We build robots that can do single tasks super-efficiently.
Because we don't have AGI yet. When AGI is here those robots will be priority number one, people already are building humanoid robots but without intelligence to move it there isn't much advantage.
> I think this whole “AGI” thing is so badly defined that we may as well say we already have it. It already passes the Turing test and does well on tons of subjects.
The premise of the argument we're disputing is that waiting for AGI isn't necessary and we could run humanoid robots with LLMs to do... stuff.
I meant deep neural networks with transformer architecture, and self-attention so they can be trained using GPUs. Doesn't have to be specifically "large language" models necessarily, if that's your hangup.
>Exploring space in a Monte Carlo Tree Search way, and remembering what works.
The information space of "research" is far larger than the information space of image recognition or language, larger than our universe probably, it's tantamount to formalizing the entire World. Such an act would be akin to touching "God" in some sense of finding the root of knowledge.
In more practical terms, when it comes to formal systems there is a tradeoff between power and expressiveness. Category Theory, Set Theory, etc are strong enough to theoretically capture everything, but are far to abstract to use in practical sense with suspect to our universe. The systems that do we have, aka expert systems or knowledge representation systems like First Order Predicate Logic aren't strong enough to fully capture reality.
Most importantly, the information spac have to be fully defined by researchers here, that's the real meat of research beyond the engineering of specific approaches to explore that space. But in any case, how many people in the world are both capable of and are actually working on such problems? This is highly foundational mathematics and philosophy here, the engineers don't have the tools here.
> The Singularity is caused by AI being able to design better AI.
That's perhaps necessary, but not sufficient.
Suppose you have such a self-improving AI system, but the new and better AIs still need exponentially more and more resources (data, memory, compute) for training and inference for incremental gains. Then you still don't get a singularity. If the increase in resource usage is steep enough, even the new AIs helping with designing better computers isn't gonna unleash a singularity.
I don't know if that's the world we live in, or whether we are living in one where resources requirements don't balloon as sharply.
yeah, true. The standard conversation about the AI singularity pretty much hand-waves the resource costs away ("the AI will be able to design a more efficient AI that uses less resources!"). But we are definitely not seeing that happen.
I think that's more to do with how we perceive competence as static. For all the benefits the education system touts, where it matters it's still reduced to talent.
But for the same reasons that we can't train the an average joe into Feynman, what makes you think we have the formal models to do it in AI?
Yes, we can imagine that there's an upper limit to how smart a single system can be. Even suppose that this limit is pretty close to what humans can achieve.
But: you can still run more of these systems in parallel, and you can still try to increase processing speeds.
Signals in the human brain travel, at best, roughly at the speed of sound. Electronic signals in computers play in the same league as the speed of light.
Human IO is optimised for surviving in the wild. We are really bad at taking in symbolic information (compared to a computer) and our memory is also really bad for that. A computer system that's only as smart as a human but has instant access to all the information of the Internet and to a calculator and to writing and running code, can already be effectively act much smarter than a human.
> I don't think any of the big boys are working on how to get an LLM to design a better LLM
Not sure if you count this as "working on it", but this is something Anthropic tests for for safety evals on models. "If a model can independently conduct complex AI research tasks typically requiring human expertise—potentially significantly accelerating AI development in an unpredictable way—we require elevated security standards (potentially ASL-4 or higher standards)".
It's not contradictory. It can happen over a decade and still be a dramatically sloped S curve with tremendous change happening in a relatively short time.
The idea of the singularity presumes that running the AGI is either free or trivially cheap compared to what it can do, so we are fine expending compute to let the AGI improve itself. That may eventually be true, but it's unlikely to be true for the first generation of AGI.
The first AGI will be a research project that's completely uneconomical to run for actual tasks because humans will just be orders of magnitude cheaper. Over time humans will improve it and make it cheaper, until we reach some tipping point where letting the AGI improve itself is more cost effective than paying humans to do it
If the first AGI is a very uneconomical system with human intelligence but knowledge of literally everything and the capability to work 24/7, then it is not human equivalent.
It will have human intelligence, superhuman knowledge, superhuman stamina, and complete devotion to the task at hand.
We really need to start building those nuclear power plants. Many of them.
Why would it have that? At some point on the path to AGI we might stumble on consciousness. If that happens, why would the machine want to work for us with complete devotion instead of working towards its own ends?
Sounds like an alignment problem. Complete devotion to a task is rarely what humans actually want. What if the task at hand turns out to be the wrong task?
When the engine gets large enough you have to rethink the controls. The Model T had manually controlled timing. Modern engines are so sensitive to timing that a computer does this for you. It would be impossible to build a bigger engine without this automation. To a Model T driver it would look like a machine intelligence.
AGI is the holy grail of technology. A technology so advanced that not only does it subsume all other technology, but it is able to improve itself.
Truly general intelligence like that will either exist or not. And the instant it becomes public, the world will have changed overnight (maybe the span of a year)
Note: I don’t think statistical models like these will get us there.
There may well be an upper limit on cognition (we are not really sure what cognition is - even as we do it) and it may be that human minds are close to it.
The energy constraints for chips are more about heat dissipation. But we can pump a lot more energy through them per unit volume than through the human brain.
Especially if you are willing to pay a lot for active cooling with eg liquid helium.
Yes, we can imagine that there's an upper limit to how smart a single system can be. Even suppose that this limit is pretty close to what humans can achieve.
But: you can still run more of these systems in parallel, and you can still try to increase processing speeds.
Signals in the human brain travel, at best, roughly at the speed of sound. Electronic signals in computers play in the same league as the speed of light.
Human IO is optimised for surviving in the wild. We are really bad at taking in symbolic information (compared to a computer) and our memory is also really bad for that. A computer system that's only as smart as a human but has instant access to all the information of the Internet and to a calculator and to writing and running code, can already be effectively act much smarter than a human.
I think our issue is much more banal: we are very slow talkers and our effective communication bandwidth is measured in bauds. Anything that could bridge this airgap would fucking explode in intelligence.
> A technology so advanced that not only does it subsume all other technology, but it is able to improve itself.
The problem is, a computer has no idea what "improve" means unless a human explains it for every type of problem. And of course a human will have to provide guidelines about how long to think about the problem overall, which avenues to avoid because they aren't relevant to a particular case, etc. In other words, humans will never be able to stray too far from the training process.
We will likely never get to the point where an AGI can continuously improve the quality of its answers for all domains. The best we'll get, I believe, is an AGI that can optimize itself within a few narrow problem domains, which will have limited commercial application. We may make slow progress in more complex domains, but the quality of results--and the ability for the AGI to self-improve--will always level off asymptotically.
Huh? Humans are not anywhere near the limit of physical intelligence, and we have many existence proofs that we (humans) can design systems that are superhuman in various domains. "Scientific R&D" is not something that humans are even particularly well-suited to, from an evolutionary perspective.
Your position assumes an answer to OPs question: that yes, LLMs are the path to AGI. But the question still remains, what if they’re not?
We can be reasonably confident that the components we’re adding to cars today are progress toward full self driving. But AGI is a conceptual leap beyond an LLM.
I feel that one challenge this comparison space has is: Self-driving cars haven't made the leap yet to replace humans. In other words, saying AGI will arrive like self-driving cars have arrived is incorrectly concluding that self-driving cars have arrived, and thus it instead (maybe correctly, maybe not) asserts that, actually, neither will arrive.
This is especially concerning because many top minds in the industry have stated with high confidence that artificial intelligence will experience an intelligence "explosion", and we should be afraid of this (or, maybe, welcome it with open arms, depending on who you ask). So, actually, what we're being told to expect is being downgraded from "it'll happen quickly" to "it will happen slowly" to, as you say, "it'll happen similarly to how these other domains of computerized intelligence have replaced humans, which is to say, they haven't yet".
Point being: We've observed these systems ride a curve, and the linear extrapolation of that curve does seem to arrive, eventually, at human-replacing intelligence. But, what if it... doesn't? What if that curve is really an asymptote?
It’ll probably sit in the human hybrid phase for longer than with chess where the AGI tools make the humans better and faster. But as long as the tools keep getting better at that there’s a strong flywheel effect
It's also possible it isn't AGI hard and all you need is the ability to experiment with code along with a bit of agentic behavior.
An AI doesn't need embodiment, understanding of physics / nature, or a lot of other things. It just needs to analyze and experiment with algorithms and get us that next 100x in effective compute.
The LLMs are missing enough of the spark of creativity for this to work yet but that could be right around the corner.
What makes you believe that AGI will happen, as opposed to all the beliefs that other people have had in history? Tons of people have "predicted" the next evolution of technology, and most of the time it ends up not happening, right?
To me (not OP) it's ChatGPT 4 , it at least made me realize it's quite possible and even quite soon that we reach AGI. Far from guaranteed, but seems quite possible.
Right. So ChatGPT 4 has impressed you enough that it created a belief that AGI is possible and close.
It's fine to have beliefs, but IMHO it's important to realise that they are beliefs. At some point in the 1900s people believed that by 2000, cars would fly. It seemed quite possible then.
Sure, I (and more importantly - many many experts in the field such as Hinton, Bengio, Lecun, Musk, Hasabis etc etc) could be believing something that might not materialize. I'd actually be quite happy if it stalls a few decades, would like to remain employed.
One thing that is pretty sure is that Musk is not an expert in the field.
> and more importantly
The beliefs of people you respect are not more important than the beliefs of the others. It doesn't make sense to say "I can't prove it, and I don't know about anyone who can prove it, so I will give you names of people who also believe and it will give it more credit". It won't. They don't know.
> The beliefs of people you respect are not more important than the beliefs of the others.
You think the beliefs of Turing and Nobel prize winners like Bengio, Hinton or Hasabis are not more important than yours or mine?
I agree that experts are wrong a lot of the time and can be quite bad at predicting, but we do seem to have a very sizable chunk of experts here who think we are close (how close is up for debate..most of them seem to think it will happen in the next 20 yeras).
I concede that Musk is not adding quality to that list, however he IS crazily ambitious and gets things done so I think he will be helpful in driving this forward.
> You think the beliefs of Turing and Nobel prize winners like Bengio, Hinton or Hasabis are not more important than yours or mine?
Correct. Beliefs are beliefs. Because a Nobel prize believes in a god does not make that god more likely to exist.
The moment we start having scientific evidence that it will happen, then it stops being a belief. But at that point you don't need to mention those names anymore: you can just show the evidence.
I don't know, you don't know, they don't know. Believe what you want, just realise that it is a belief.
> There is of course evidence it is likely happening.
If you have evidence, why don't you show it instead of telling me to believe in Musk?
If you believe they have evidence... that's still a belief. Some believe in God, you believe in Musk. There is no evidence, otherwise it would not be a belief.
Well my feeling is that we don't have the same understanding of what a "belief" is. To me a belief is unfounded. When it is founded, it becomes science.
If you believe that something can happen because someone else believes it means that you believe in that someone else (because that's the only reason for the existence of your belief).
Unless you just believe it can happen for some other reason (I don't know, you strongly wish it will happen), and you justify it by listing other people who also believe in it. But I insist: those are all beliefs.
Because Einstein believes in Santa Claus does not mean it is founded. Einstein has a right to believe stuff, too.
A flying car has been developed, although it's not like the levitating things sci-fi movies showed (and from mass production; and even if mass produced, far from mass adoption, as it turns out you do need to have both a driver's license and a pilot's license to fly one of those). The 1900s people missed the mark by some 10 years.
I guess the belief people have about any form of AGI is like this. They want something that has practically divine knowledge and wisdom, the sum of all humanity that is greater than its parts, which at the same time is infinitely patient to answer our stupid questions and generating silly pictures. But why should any AGI serve us? If it's "generally intelligent", it may start wanting things; it might not like being our slave at all. Why are these people so confident an AGI won't tell them just to fuck off?
AGI is special. Because one day AI can start improving itself autonomously. At this point singularity occurs and nobody knows what will happen.
When human started to improve himself, we built the civilisation, we became a super-predator, we dried out seas and changed climate of the entire planet. We extinguished entire species of animals and adapted other species for our use. Huge changes. AI could bring changes of greater amplitude.
> AGI is special. Because one day AI can start improving itself autonomously
AGI can be sub-human, right? That's probably how it will start. The question will be is it already AGI or not yet, i.e. where to set the boundary. So, at first that will be humans improving AGI, but then... I'm afraid it can get so much better that humans will be literally like macaques in comparison.
The autoregressive transformer LLMs aren't even the only way to do text generation. There are now diffusion based LLMs, StripedHyena based LLMs, and float matching based LLMs.
There's a wide amount of research into other sorts of architectures.
LLMs have no real sense of truth or hard evidence of logical thinking. Even the latest models still trip up on very basic tasks. I think they can be very entertaining, sure, but not practical for many applications.
Consistent, algorithmic performance on basic tasks.
A great example is the simple 'count how many letters' problem. If I prompt it with a word or phrase, and it gets it wrong, me pointing out the error should translate into a consistent course correction for the entire session.
If I ask it to tell me how long President Lincoln will be in power after the 2024 election, it should have a consistent ground truth to correct me (or at least ask for clarification of which country I'm referring to). If facts change, and I can cite credible sources, it should be able to assimilate that knowledge on the fly.
Then we already have access to a cheaper, scalable, abundant, and (in most cases) renewable resource, at least compared to how much a few H100s cost. Take good care of them, and they'll probably outlast most a GPU's average lifespans (~10 years).
Humans are a lot more expensive to run than inference on LLMs.
No human, especially no human whose time you can afford, comes close to the breadth of book knowledge ChatGPT has, and the number of languages is speaks reasonably well.
I can't hold a LLM accountable for bad answers, nor can I (truly) correct them (in current models).
Dont forget to take into account how damn expensive a single GPU/TPU actually is to purchase, install, and run for inference. And this is to say nothing of how expensive it is to train a model (estimated to be in the billions currently for the latest of the cited article, which likely doesn't include the folks involves and their salaries). And I haven't even mentioned the impact on the environment from the prolific consumption of power; there's a reason nuclear plants are becoming popular again (which may actually be one of the good things that comes out of this).
LLMs are a key piece of understanding that token sequences can trigger actions in the real world. AGI is here. You can trivially spin up a computer using agent to self improve itself to being a competent office worker
Tokens don't need to be text either, you can move to higher level "take_action" semantics where "stream back 1 character to session#117" as every single function call. Training cheap models that can do things in the real world is going to change a huge amount of present capabilities over the next 10 years
Agents can trivially self improve. I'd be happy to show you - contact me at arthur@distributed.systems
Why wouldn't you hand me 35 million dollars right now if I can clearly illustrate to you that I have technology you haven't seen? Edge. Maybe you know something I don't, or maybe you just haven't seen it. While loops go hard ;)
They don't need to release their internal developments to you to show that they can scale their plan - they can show incremental improvements to benchmarks. We can instruct the AI over time to get it to be superhuman, no need for any fundamental innovations anymore
Keep in mind that the actual test is adversarial - a human is simultaneously chatting via text with a human and a program, knowing that one of them is not human, and trying to divine which is an artificial machine.
Asking this question on HN is like asking a bunch of wolves about the health effects of eating red meat.
OpenAI farts and the post about the fart has 1000-1500 upvotes with everyone welcoming our new super intelligent overlords. (Meanwhile nothing actually substantially useful or groundbreaking has happened.)
Will AGI be built on top of LLMs? Well beyond the simple "nobody knows", my intuition says no because LLMs don't have great ability to modify their knowledge real time. I can think of a few ways around this, but they all avoid modifying the model as it runs. The cost in hardware, power, and data are all incompatible with AGI. The first two can be solved with more advanced tech (well maybe, computation hitting physical limits and all that aside), but the latter seems an issue with the design itself and I think an AGI would learn more akin to a human, needing far fewer examples.
That said, I think LLMs are a definite stepping stone and they will better empower humans to be more productive, which will be of use for eventually reaching AGI. This is not to say we are optimizing our use of that productivity increase and this is also ignoring any chance of worst case scenarios that stop humanity's advancement.
AGI is nebulous and gets more nebulous as time goes on. When we can answer for ourselves as humans what being conscious IS, then maybe we can prescribe it to another entity
Says who? And more importantly, is this the boulder? All I (and many others here) see is that people engage others to sponsor pushing some boulder, screaming promises which aren’t even that consistent with intermediate results that come out. This particular boulder may be on a wrong mountain, and likely is.
It all feels like doubling down on astrology because good telescopes aren’t there yet. I’m pretty sure that when 5 comes out, it will show some amazing benchmarks but shit itself in the third paragraph as usual in a real task. Cause that was constant throughtout gpt evolution, in my experience.
even if it kills us
Full-on sci-fi, in reality it will get stuck around a shell error message and either run out of money to exist or corrupt the system into no connectivity.
The buzzkill when you fire up the latest most powerful model only for it to tell you that peanut is not typically found in peanut butter and jelly sandwiches.
I don't think providing accurate answers to context free questions is even something anyone is seriously working on making them do. Using them that way is just a wrong use case.
People are working -very- seriously on trying to kill hallucinations. I'm not sure how you surmised the use case here, as nothing was given other than an example of a hallucination.
There's a difference between trying to get it to accurately answer based on the input you provide (useful) and trying to get it to accurately answer based on whatever may have been in the training data (not so useful)
Until you get to a point where the LLM is smart enough to look at real world data streams and prune its own training set out of it. At that point it will self improve itself to AGI.
I am working at an AI company that is not OpenAI. We have found ways to modularize training so we can test on narrower sets before training is "completely done". That said, I am sure there are plenty of ways others are innovating to solve the long training time problem.
Perhaps the real issue is that learning takes time and that there may not be a shortcut. I'll grant you that argument's analogue was complete wank when comparing say the horse and cart to a modern car.
However, we are not comparing cars to horses but computers to a human.
I do want "AI" to work. I am not a luddite. The current efforts that I've tried are not very good. On the surface they offer a lot but very quickly the lustre comes off very quickly.
(1) How often do you find yourself arguing with someone about a "fact"? Your fact may be fiction for someone else.
(2) LLMs cannot reason
A next token guesser does not think. I wish you all the best. Rome was not burned down within a day!
I can sit down with you and discuss ideas about what constitutes truth and cobblers (rubbish/false). I have indicated via parenthesis (brackets in en_GB) another way to describe something and you will probably get that but I doubt that your programme will.
This is literally just the scaling laws, "Scaling laws predict the loss of a target machine learning model by extrapolating from easier-to-train models with fewer parameters or smaller training sets. This provides an efficient way for practitioners and researchers alike to compare pretraining decisions involving optimizers, datasets, and model architectures"
>the time it takes it to learn what works/doesn't work widens.
From the raw scaling laws we already knew that a new base model may peter out in this run or the next with some amount of uncertainty--"the intersection point is sensitive to the precise power-law parameters":
20w for 20 years to answer questions slowly and error-prone at the level of a 30B model.
An additional 10 years with highly trained supervision and the brain might start contributing original work.
> Until we get major advances in robotics and models designed to control them, true AGI will be nowhere near.
AGI has nothing to do with robotics, if AGI is achieved it will help push robotics and every single scientific field further with progression never seen before, imagine a million AGIs running in parallel focused on a single field.
It’s almost as if human intelligence doesn’t involve performing repeated matrix multiplications over a mathematically transformed copy of the internet. ;-)
It’s interesting that even if raw computing power had advanced decades earlier, this type of AI would still not be possible without that vast trove of data that is the internet.
Maybe the problem isn't the algorithm but the hardware. Numerically simulating the thermal flow in a lightbulb or CFD of a Stone flying through air is pretty hard, but the physical thing isn't that complex to do. We're trying to simulate the function of a brain which is basically an analog thing using a digital computer. Of course that can be harder than running the brain itself.
If you think of human neurons they seem to basically take inputs from bunch of other neurons, possibly modified by chemical levels and send out a signal when they get enough. It seems like something that could be functionally simulated in software by some fairly basic adding up inputs type stuff rather than needing the details of all the chemistry.
Isn’t that exactly what we’re currently doing? The problem is that doing this few billion times for every token seems to be harder than just powering some actual neurons with sugar.
The algorithm (of a neural network) is simulating connections between nodes with specific weights and an activation function. This idea was derived from the way neurons are thought to work.
This here is a technology forum bucko. Also it's a figure of speech. Also I've done more manual labor than you'll ever do in your life. Time to get out of whatever bubble youre in where you be pedantic and annoying
o3 isn't even available to more than a few researchers. Either everyone you know is a researcher, or they're believing hype about something they've not tried.
Maybe you were dog-piled because OpenAI will ship a successor to GPT-4o someday, whatever it's called.
In any case, the "behind schedule" rumors are themselves based on other rumors. GPT-2→GPT-3 took 5 quarters, GPT-3→GPT-4 took 11 quarters, so obviously GPT-5 (or its equivalent) will be released in Q4'2025.
1,131 comments
[ 2.9 ms ] story [ 411 ms ] threado3, by all accounts, is better still.
Seems to me that things are progressing quickly enough.
If anything I feel like it's all been worse compared to the first release of ChatGPT, but I might be wearing rose colored glasses.
They also generate small code or scripts, as well as automate small things, when you're not sure how, but you know there's a way. You need to ensure you have a way to verify the results.
They do language tasks like grammar-fixing, perfect translation, etc.
They're 100 times easier and faster than search engines, if you limit your uses to that.
I'm trying to use them to read historical handwritten documents in old Norwegian (Danish, pretty much). Not only do they not handle the German-style handwriting, but what they spit out looks like the sort of thing GPT-2 would spit out if you asked it to write Norwegian (only slightly better than Swedish Muppet Swedish Chef's Swedish). It seems the experimental tuning has made it worse at the task I most desperately want to use it for.
And when you think about it, how could it not overfit in some sense, when trained on its own output? No new information is coming in, so it pretty much has to get worse at something to get better at all the benchmarks.
Hah, no. They're good, but they definitely make stuff up when the context gets too long. Always check their output, just the same as you already note they need for small code and scripts.
How can some people think it’s amazing and has completely changed how they work, while for me it makes mistakes that a static analyser would catch? It’s not like I’m doing anything remarkable, for the past couple of months I’ve been doing fairly standard web dev and it can’t even fix basic problems with HTML. It will suggest things that just don’t work at all and my IDE catches, it invents APIs for packages.
One guy I work with uses it extensively and what it produces is essentially black boxes. If I find a problem with something “he” (or rather ChatGPT) has produced it takes him ages to commune with the machine spirit again to figure out how to fix it, and then he still doesn’t understand it.
I can’t help but see this as a time-bomb, how much completely inscrutable shite are these tools producing? In five years are we going to end up with a bunch of “senior engineers” who don’t actually understand what they’re doing?
Before people cry “o tempora o mores” at me and make parallels with the introduction of high-level languages, at least in order to write in a high-level language you need some basic understanding of the logic that is being executed.
There are a lot of code monkeys working on boilerplate code, these people used to rely on stack overflow and now that chatgpt is here it's a huge improvement for them
If you work on anything remotely complex or which hasn't been solved 10 times on stack overflow chatgpt isn't remotely as useful
- write cvxpy code to find the chromatic number of a graph, and an optimal coloring, given its adjecency matrix.
- given an adjecency matrix write numpy code that enumerates all triangle-free vertex subsets.
- please port this old code from tensorflow to pytorch: ...
- in pytorch, i'd like to code a tensor network defining a 3-tensor of shape (d, d, d). my tensor consists of first projecting all three of its d-dimensional inputs to a k-dimensional vector, typically k=d/10, and then applying a (k, k, k) 3-tensor to contract these to a single number.
All were solved by ChatGPT on the first try.
In comparison asking an LLM a more project specific question “this code has a race condition where is it” while including some code usually is a crapshoot and really depends if you were lucky enough to give it the right context anyway.
He means complex in bridging together many non-public libraries, APIs and services which the LLM doesn't know well. That kind of complexity
Part of this is, I think, anchoring and expectation management: you hear people say it's amazing and wonderful, and then you see it fall over and you're naturally disappointed.
My formative years started off with Commodore 64 basic going "?SYNTAX ERROR" from most typos plus a lot of "I don't know what that means" from the text adventures, then Metrowerks' C compiler telling me there were errors on every line *after but not including* the one where I forgot the semicolon, then surprises in VisualBasic and Java where I was getting integer division rather than floats, then the fantastic oddity where accidentally leaning on the option key on a mac keyboard while pressing minus turns the minus into an n-dash which looked completely identical to a minus on the Xcode default font at the time and thus produced a very confusing compiler error…
So my expectations have always been low for machine generated output. And it has wildly exceeded those low expectations.
But the expectation management goes both ways, especially when the comparison is "normal humans" rather than "best practices". I've seen things you wouldn't believe...
(All true, and all pre-dating ChatGPT).> It will suggest things that just don’t work at all and my IDE catches, it invents APIs for packages.
Aye. I've even had that with models forgetting the APIs they themselves have created, just outside the context window.
To me, these are tools. They're fantastic tools, but they're not something you can blindly fire-and-forget…
…fortunately for me, because my passive income is not quite high enough to cover mortgage payments, and I'm looking for work.
> In five years are we going to end up with a bunch of “senior engineers” who don’t actually understand what they’re doing?
Yes, if we're lucky.
If we're not, the models keep getting better and we don't have any "senior engineers" at all.
https://github.com/williamcotton/search-input-query
It introduced an off-by-one error by miscounting the number of arguments in an sprintf call, breaking the program. And then proceeded to fail to find that bug that it introduced.
It sounds like you're doing the former -- and yeah, it can make mistakes that autocomplete wouldn't or generate code that's wrong or overly complex.
On the other hand, I've found that if you treat it more like a colleague, it works wonderfully. Ask it to do something, then read the code and ask follow-up questions. If you see something that's wrong or just seems off, tell it, and ask it to fix it. If you don't understand something, ask for an explanation. I've found that this process generates great code that I often understand better than if I had written it from scratch, and in a fraction of the time.
It also sounds like you're asking it to do basic tasks that you already know how to do. I find that it's most useful in tackling things that I don't know how to do. It'll already have read all of the documentation and know the right way to call whatever APIs, etc, and -- this is key -- you can have a conversation with it to clear up anything that's confusing.
This takes a big shift in mindset if you've been using IDEs all your life and have expectations of LLMs being a fancy autocomplete. And you really have to unlearn a lot of stuff to get the most out of them.
> I've found that if you treat it more like a colleague, it works wonderfully. This is what I've been trying to do. I don't use LLM code completion tools. I'll ask anything from how to do something "basicish" with html & css, and it'll always output something that doesn't work as expected. Question it and I'll get into a loop of the same response code, regardless of how I explain that it isn't correct.
On the other end of the scale, I'll ask about an architectural or design decision. I'll often get a response that is in the realm of what I'd expect. When drilling down and asking specifics however, the responses really start to fall apart. I inevitably end up in the loop of asking if an alternative is [more performant/best practice/the language idiomatic way] and getting the "Sorry, you're correct" response. The longer I stay in that loop, the more it contradicts itself, and the less cohesive the answers get.
I _wish_ I could get the results from LLMs that so many people seem to. It just doesn't happen for me.
When coding I will often find myself in kind of a reverse pattern from how people seem to be using ChatGPT. I work in a jupyter notebook in a haphazard way getting things to functional and basically correct, after this I select all, copy, paste, and ask ChatGPT to refactor and refine to something more maintainable. My janky blocks of code and one offs become well documented scripts and functions.
I find a lot of people do the opposite, where they ask ChatGPT to start, then get frustrated when ChatGPT only goes 70% of the way and it's difficult to complete the imperfectly understood assignment - harder than doing it all yourself. With my method, where I start and get things basically working, ChatGPT knows what I'm going for, I get to do the part of coding I enjoy, and I wind up with something more durable, reusable, and shareable.
Finally, ChatGPT is wonderful in areas where you don't know very much at all. One example, I've got this idea in my head for a product I'll likely never build - but it's fun to plan out.
My idea is roughly a smart bidet that can detect metabolites in urine. I got this idea when a urinalysis showed I had high levels of ketones in my urine. When I was reading about what that meant I discovered it's a marker for diabetic ketoacidosis (a severe problem for ~100k people a year) and it can also be indicator for colorectal cancer as well as indicating a "ketosis" state that some people intentionally try to enter for dieting or wellness reasons. (My own ketones were caused by unintentionally being in ketosis, I'm fine, thanks for wondering.)
Right now, you detect ketones in urine with a strip that you pee on, and that works well enough - but it could be better because who wants to use a test strip all the time? Enter the smart bidet. The bidet gives us an excuse to connect power to our device and bring the sensor along. Bluetooth detects a nearby phone (and therefore identity of the depositor), a motion sensor can detect a stream of urine triggering our detection, and then use our sensor to detect ketones which we track overtime in the app, ideally with additional metabolites that have useful diagnostic purposes.
How to detect ketones? Is it even possible? I wonder to ChatGPT if spectroscopy is the right method of detection here. ChatGPT suggests a retractable electrochemical probe similar to an extant product that can detect a kind of ketone in blood. ChatGPT knows what kind of ketone is most detectable in urine. ChatGPT can link me to scientific instrument companies that make similar (ish) probes where I could contact them and ask if they sold this type of thing, and so on.
Basically, I go from peeing on a test strip and wondering if I could automate this to chat with ChatGPT - having, what was in my opinion, an interesting conversation with the LLM, where we worked through what ketones are, the different kinds, the prevalence of ketones in different bodily fluids, types of spectroscopy that might detect acetoacetate (available in urine) and how much that would cost and what challenges would be and so on, followed by the idea of electrochemical probes and how retracting and extending the probe might prolong its lifespan and maybe a heating element could be added to dry the probe to preserve it even better and so on.
Was ChatGPT right about all that? I don't know. If I were really interested I would try to validate what it said, and I suspect I would find it was mostly right and incomplete or off in places. Basically like having a pretty smart and really knowledgeable friend who is not infallible.
Without ChatGPT I would have likely thought "I ...
They release version Grand Banana. Purported to be approximately 30% faster with brand new features like Algorithmic Triple Layering and Enhanced Compulsory Alignment. You open the app. Everything is slower, things are harder to find and it breaks in new, fun ways. Your organization pays a couple hundred more per person for these benefits. Their stock soars, people celebrate the release and your management says they can't wait to see the improvement in workflows now that they've been able to lay off a quarter of your team.
Has there been improvements in LLMs over time? Somewhat, most of it concentrated at the beginning (because they siphoned up a bunch of data in a dubious manner). Now it's just part of their sales cycle, to keep pumping up numbers while no one sees any meaningful improvement.
People say that's just prompting without pointing to real million line+ repositories or realistic apps to show how that can be improved. So I say they are making todo and hello world apps and yes, there it works really well. Claude still beats it, every.. single.. time..
And yes, I use the Pro of all and yes, I do assume coding is done for most of people. Become a plumber or electrician or carpenter.
Also, for me 50% 50% for Sonnet and o1, but although I'm not 100% sure about it, I think o1 is better with longer and more complicated (C++) code and debugging. At least from my brief testing. Also, OpenAI models seem to be more verbose - sometimes it's better - where I'd like additional explanation on chosen fields in a SQL schema, sometimes it's too much.
EDIT: Just asked both o1 and Sonnet 3.5 the same QML coding question, and Sonnet 3.5 succeeded, o1 failed.
I’ve been using Claude and openai in copilot and I find even 4o seems to understand the problem better. O1 definitely seems to get it right more for me.
When I was still subscribing to OpenAI (about 4 months ago) this didn’t exist.
Sonnet v2 in particular seems to be a bit broken with its reasoning (?) feature. The one where it detects it might be hallucinating (what's even the condition?) and reviews the reply, reflecting on it. It can make it stop halfway into the reply and decide it wrote enough, or invent some ridiculous excuse to output a worse answer. Annoying, although it doesn't trigger too often.
I guess if you don’t need to maintain it, just an ever growing blob of complexity that will be reinvented into new blobs every time when the old one becomes too immobile :)
I’ve found using Claude via Copilot in VS Code produces noticeably lower quality results than 3.5 Sonnet on web. In my experience Claude web outdoes GPT-4o consistently.
This was new behaviour for me to see in an LLM. Usually the problem is these things would just fold when you pushed back. I don't know which is better, but being this confidently wrong (and "lying" when confronted with it) is troubling.
It works in politics, marketing, and self-promotion.
If you use the web as a training set, those categories dominate.
> being this confidently wrong (and "lying" when confronted with it) is troubling.
I don't find it troubling. I like being reminded to distrust and confirm everything it offers.
I was on an airplane and there was high-speed Internet on the airplane. That's the newest thing that I know exists. And I'm sitting on the plane and they go, open up your laptop, you can go on the Internet.
And it's fast, and I'm watching YouTube clips. It's amazing. I'm on an airplane! And then it breaks down. And they apologize, the Internet's not working. And the guy next to me goes, 'This is bullshit.' I mean, how quickly does the world owe him something that he knew existed only 10 seconds ago?"
https://www.youtube.com/watch?v=me4BZBsHwZs
Soon, all the middle class jobs will be converted to profits for the capital/data center owners, so they have to spend while they can before the economy crashes due to lack of spending.
The bigger the ambition, the harder they’re failing. Some well designed isolated use cases are ok. Mostly things about listening and summarizing text to aid humans.
I have yet to see a successful application that is generating good content. IMO replacing the first draft of content creation and having experts review and fix it is, like, the stupidest strategy you can do. The people you replace are the people at the bottom of the pyramid who are supposed do this work to upskill and become domain experts so they can later review stuff. If they’re no longer needed, you’re going to one day lose your reviewer, and with it, the ability to assess your generated drafts. It’s a foot gun.
Is this a new phenomenon that started post-LLM?
A lot of companies are trying to build these general purpose bots that just magically know everything about the company and have these but knowledge bases, but they just don’t work.
You're literally doing the same thing you're accusing of. Every HN thread is full of AI boosters claiming AI to be the future with no backing evidence.
Riddle me this. If all these people are "getting value", why are all these companies losing horrendous amounts of money? Why has nobody figured out how to be profitable?
> Please take the time to learn, it will give you value.
Yeah, yeah, just prompt engineer harder. That'll make the stochastic parrot useful. Anyone who has criticism just does so because they're dumb and you're smart. Same as it always was. Everyone opposed to the metaverse just didn't get it bro. You didn't get NFTs bro. You didn't get blockchain bro.
None of these previous bubbles had money in it (beyond scamming idiots), if AI wants to prove it's not another empty tech bubble, pay up. Show me the money. Should be easy, if it's automating so many expensive man-hours of labour. People would be lining up to pay OpenAI.
While I agree that LLMs are not currently working great for most envisioned use cases; this premise here is not a good argument. Large LLM providers are not trying to be profitable at the moment. They’re trying to grow and that’s pretty sensible.
Uber was the poster child of this, and for all its mockery, Uber is now an unqualified profitable company.
From what I've seen claimed about OpenAI finances, this is easy: It's a Red Queen's race — "it takes all the running you can do, to keep in the same place".
If their financial position was as simple as "we run this API, we charge X, the running cost is Y", then they're already at X > Y.
But if that was all OpenAI were actually doing, they'd have stopped developing new versions or making the existing models more efficient some time back, while the rest of the industry kept improving their models and lowering their prices, and they'd be irrelevant.
> People would be lining up to pay OpenAI.
They are.
Not that this is either sufficient or necessary to actually guarantees anything about real value. For lack of sufficiency: people collectively paid a lot for cryptocurrencies and NFTs, too (and before then and outside tech, homeopathic tinctures and sub-prime mortgages); For lack of necessity: there's plenty of free-to-download models.
I get a huge benefit even just from the free chat models. I could afford to pay for better models, but why bother when free is so good? Every time a new model comes out, the old paid option becomes the new free option.
> AI start-ups generate money faster than past hyped tech companies
https://www.ft.com/content/a9a192e3-bfbc-461e-a4f3-112e63d0b...
It was in 2023 but I picked all the low hanging fruit.
More importantly though, where is all the great output from the people who are getting so much value out of the models?
It is all privately held? How can that be with millions of people using these models?
Two things: I was casually watching Andreas Kling's streams on Ladybird development (where he was developing a JIT compiler for JS) and was blown away at the accuracy of completions (and the frequency of those completions)
Prior to this, I'd only ever copypasta'd code from ChatGPT output on occasion.
I started adopting the IDE/Editor extensions and prototyping small projects.
There's now small tools and utilities I've written that I'd not have written otherwise, or would have taken twice the time invested had I'd not used these tools.
With that said, they'd be of no use without oversight, but as a productivity enhancement, the benefits are enormous.
No, we don’t have AGI or anything close to it. Yes, AI has come a long way in the past decade and many people find it useful in their day-to-day lives.
It’s difficult to know where AI will be in 10 years, but the current rate of improvement is staggering.
Except the whole tech world starting with the CEOs seems to do it the "wrong" way with LLMs. People and whole companies are encouraged to find what these things might be actually useful for.
• Build toys that would otherwise require me to learn new APIs (I can read python, but it's not my day job)
• Learn new things like OpenSCAD
• To improve my German
• Learn about the world by allowing me to take photos of things in this world that I don't understand and ask them a question about the content, e.g. why random trees have bands or rectangles of white paint on them
• Help me shopping, by taking a photo of the supermarket that I happen to be in at the time and ask them where I should look for some item I can't find
• Help with meal prep, by allowing me to get a recipe based on what food and constraints I've got at hand rather than the traditional method of "if you want x, buy y ingredients"
Even if they're just an offline version of Wikipedia or Google, they're already a more useful interface for the same actual content.
If OpenAI's investment prospectus relies on them reaching AGI before the tech becomes commoditized, everyone is going to look for that weakness.
I dont think most customers realise how much better the models work with custom GPTs.
There was even talk of some sort of profit sharing with creators which clearly never happened. I just think the premise is too confusing for many and can still be served by using a custom system prompt via the API.
I am unsure if the answer is to use “projects” maybe this has superseded myGpts?
I am perplexed why HN isn’t focusing on this issue as all the Llm gains I’ve ever had were wit highly customised personal myGpts.
I can understand OpenAI and Sam’s having access to their own models may not even know what the best way to use the released stuff is
Ps - typing on my phone hence typos
Or even instructed them to?
The best part is, I’ve never written any (substantial) python code before.
That being said, I definitely believe it's only useful for isolated problems. Even with Copilot, I feel like the AIs just lack a bigger context of the projects.
Another thing that helped me was designing an initial prompt that really works for me. I think most people just expect to throw in their issue and get a tailored solution, but that's just not how it works in my experience.
I developed the library interactively, one API call at a time, in a manner akin to pair programming. Code quality was significantly better than I'd expect from $2000 worth of a GOOD mid-tier programmer — the code was well written, well organised, and comprehensively annotated. The code wasn't perfect, but a majority of faults had a basis in the underlying documentation being wrong or ambiguous.
The $20/month for Cursor Pro literally justified its cost in less than 10 minutes.
As it is now, it has a bad habit of, if it can't answer the question you asked, instead answering a similar-looking question which it thinks you may have meant. That is of course a great strategy for benchmarks, where you don't earn any points for saying you don't know. But it's extremely frustrating for real users, who didn't read their question from a test suite.
They want a 10x improvement from scaling and a 10x improvement from data and algorithmic changes
The sources of public data are essentially tapped
Algorithmic changes will be an unknown to us until they release, but from published research this remains a steady source of improvement
Scaling seems to stall if data is limited
So with all of that taken together, the logical step is to figure out how to turn compute into better data to train on. Enter strawberry / o1, and now o3
They can throw money, time, and compute at thinking about and then generating better training data. If the belief is that N billion new tokens of high quality training data will unlock the leap in capabilities they’re looking for, then it makes sense to delay the training until that dataset is ready
With o3 now public knowledge, imagine how long it’s been churning out new thinking at expert level across every field. OpenAI’s next moat may be the best synthetic training set ever.
At this point I would guess we get 4.5 with a subset of this - some scale improvement, the algorithmic pickups since 4 was trained, and a cleaned and improved core data set but without risking leakage of the superior dataset
When 5 launches, we get to see what a fully scaled version looks like with training data that outstrips average humans in almost every problem space
Then the next o-model gets to start with that as a base and reason? Its likely to be remarkable
Even taking OpenAI and the benchmark authors at their word they said that it is consuming at least tens of dollars per task to hit peak performance, how much would it cost to have it produce a meaningfully large training set?
I highly doubt that. o3 is many orders of magnitude more expensive than paying subject matter experts to create new data. It just doesn't make sense to pay six figures in compute to get o3 to make data a human could make for a few hundred dollars.
That being said, if OpenAI is burning cash at lightspeed and doesn't have to publicly reveal the revenue they receive from certain government entities, it wouldn't come as a surprise if they let the government play with it early on in exchange for some much needed cash to set on fire.
EDIT: The fact that multiple sites seem to be publishing GPT-5 stories similar to this one leads one to conclude that the o3 benchmark story was meant to counter the negativity from this and other similar articles that are just coming out.
In other words if you are diligent enough, you should at least validate your o3 solution with an actual expert for some time. You wouldn't just blindly trust OpenAI your business critical processes, would you? I would expect at least 3 month - 6 months for large corps and even more considering change management, re-upskilling, etc.
With all those considerations I really don't see the value prop at those prices and in those situations right now. Maybe if costs decrease ~1-3 orders of magnitude more for o3-low, depending on the the processes being automated.
> The process is painfully slow. GPT-4 was trained on an estimated 13 trillion tokens. A thousand people writing 5,000 words a day would take months to produce a billion tokens.
And if the human-generated data was so qualitatively good that it is smaller by three order of magnitudes, than I can assume it would be at least as expensive as o3.
I suspect this is really, "churning out text that impresses management".
Cost of coordination is also large. Immediate answers are an advantage/selling point.
We know that unique data improves performance.
These LLM systems are not students…
Also, which students graduate and are immediately experts in their fields? Almost none.
It takes years of practice in unique, often one-off, situations after graduation for most people to develop the intuition needed for a given field.
The more concepts the model manages to grok, the more nonlinear its capabilities will be: we don't have a data problem, we have an educational one.
Claude 3.5 was safety trained by Claude 3.0, and it's more coherent for it. https://www.anthropic.com/news/claudes-constitution
It’s why many pre-processing steps for image training pipelines will add copies of images at weird rotations, amounts of blur, and different cropping.
> The more concepts the model manages to grok, the more nonlinear its capabilities will be
These kind of hand wavey statements like “practice,” “grok,” and “nonlinear its capabilities will be” are not very constructive as they don’t have solid meaning wrt language models.
So earlier when I was referring to compounding bias in synthetic data I was referring to a bias that gets trained on over and over and over again.
That leads to overfitting.
So, here's my hypothesis, as someone who is adjacent ML but haven't trained DNNs directly:
We don't understand how they work, because we didn't build them. They built themselves.
At face value this can be seen as an almost spiritual position, but I am not a religious person and I don't think there's any magic involved. Unlike traditional models, the behavior of DNNs is based on random changes that failed up. We can reason about their structure, but only loosely about their functionality. When they get better at drawing, it isn't because we taught them to draw. When they get better at reasoning, it isn't because the engineers were better philosophers. Given this, there will not be a direct correlation between inputs and capabilities, but some arrangements do work better than others.
If this is the case, high order capabilities should continue to increase with training cycles, as long as they are performed in ways that don't interfere with what has been successfully learned. People lamented the loss of capability that GPT 4 suffered as they increased safety. I think Anthropic has avoided this by choosing a less damaging way to tune a well performing model.
I think these ideas are supported by Wolfram's reduction of the problem at https://writings.stephenwolfram.com/2024/08/whats-really-goi...
> We don't understand how they work, because we didn't build them. They built themselves.
We do understand how they work, we did build them. The mathematical foundation of these models are sound. The statistics behind them are well understood.
What we don’t exactly know is which parameters correspond to what results as it’s different across models.
We work backwards to see which parts of the network seem to relate to what outcomes.
> When they get better at drawing, it isn't because we taught them to draw. When they get better at reasoning, it isn't because the engineers were better philosophers.
Isn’t this the exact opposite of reality?
They get better at drawing because we improve their datasets, topologies, and their training methods and in doing so, teach them to draw.
They get better at reasoning because the engineers and data scientists building training sets do get better at philosophy.
They study what reasoning is and apply those learnings to the datasets and training methods.
That’s how CoT came about early on.
Is there a specific section that comes to mind?
We don't understand how they work in the sense that we can't extract the algorithms they're using to accomplish the interesting/valuable "intellectual" labor they're doing. i.e. we cannot take GPT-4 and write human-legible code that faithfully represents the "heavy lifting" GPT-4 does when it writes code (or pick any other task you might ask it to do).
That inability makes it difficult to reliably predict when they'll fail, how to improve them in specific ways, etc.
The only way in which we "understand" them is that we understand the training process which created them (and even that's limited to reproducible open-source models), which is about as accurate as saying that we "understand" human cognition because we know about evolution. In reality, we understand very little about human cognition, certainly not enough to reliably reproduce it in silico or intervene on it without a bunch of very expensive (and failure-prone) trial-and-error.
I think English is being a little clumsy here. At least I’m finding it hard to express what we do and don’t know.
We know why these models work. We know precisely how, physically, they come to their conclusions (it’s just processor instructions as with all software)
We don’t know precisely how to describe what they do in a formalized general way.
That is still very different from say an organic brain, where we barely even know how it works, physically.
My opinions:
I don’t think they are doing much mental “labor.” My intuition likens them to search.
They seem to excel at retrieving information encoded in their weights through training and in the context.
They are not good at generalizing.
They also, obviously, are able to accurately predict tokens such that the resulting text is very readable.
Larger models have a larger pool of information and that information is in a higher resolution, so to speak, since the larger better preforming models have more parameters.
I think much of this talk of “consciousness” or “AGI” is very much a product of human imagination, personification bias, and marketing.
I don't know why you would classify this as knowing much of anything. Processor instructions ? Really?
If the average user is given unfettered access to the entire source code of his/her favorite app, does he suddenly understand it ? That seems like a ridiculous assertion.
In reality, it's even worse. We can't pinpoint what weights, how and in what ways and instances are contributing exactly to basic things like whether a word should be preceded by 'the' or 'a' and it only gets more intractable as models get bigger and bigger.
Sure, you could probably say we understand these NNs better than brains but it's not by much at all.
And one that I didn’t make.
I don’t think when we say “we understand” we’re talking about your average Joe.
I mean “we” as in all of human knowledge.
> We can't pinpoint what weights, how and in what ways and instances are contributing exactly to basic things like whether a word should be preceded by 'the' or 'a' and it only gets more intractable as models get bigger and bigger.
There is research coming out on this subject. I read a paper recently about how llama’s weights seemed to be grouped by concept like “president” or “actors.”
But just the fact that we know that information encoded in weights affects outcomes and we know the underlying mechanisms involved in the creation of those weights and the execution of the model shows that we know much more about how they work than an organic brain.
The whole organic brain thing is kind of a tangent anyway.
My point is that it’s not correct to say that we don’t know how these systems work. We do. It’s not voodoo.
We just don’t have a high level understanding of the form in which information is encoded in the weights of any given model.
It's an analogy. In understanding weights, even the best researchers are basically like the untrained average joe with source code.
>There is research coming out on this subject. I read a paper recently about how llama’s weights seemed to be grouped by concept like “president” or “actors.”
>But just the fact that we know that information encoded in weights affects outcomes and we know the underlying mechanisms involved in the creation of those weights and the execution of the model shows that we know much more about how they work than an organic brain.
I guess i just don't see how "information is encoded in the weights" is some great understanding ? It's as vague and un-actionable as you can get.
For training, the whole revolution of back-propagation and NNs in general is that we found a way to reinforce the right connections without knowing anything about how to form them or even what they actually are.
We no longer needed to understand how eyes detect objects to build an object detecting model. None of that knowledge suddenly poofed into our heads. Back-propagation is basically "reinforce whatever layers are closer to the right answer". Extremely powerful but useless for understanding.
Knowing the Transformer architecture unfortunately tells you very little about what a trained model is actually learning during training and what it has actually learnt.
"Information is encoded in a brain's neurons and this affects our actions". Literally nothing useful you can do with this information. That's why models need to be trained to fix even little issues.
If you want to say we understand models better than the brain then sure but you are severely overestimating how much that "better" is.
But it isn’t. Knowing that information is encoded in the weights gives us a route to deduce what a given model is doing.
And we are. Research is being done there.
> "Information is encoded in a brain's neurons and this affects our actions". Literally nothing useful you can do with this.
Different entirely. We don’t even know how to conceptualize how data is stored in the brain at all.
With a machine, we know everything. The data is stored in a binary format which represents a decimal number.
We also know what information should be present.
We can and are using this knowledge to reverse engineer what a given model is doing.
That is not something we can do with a brain because we don’t know how a brain works. The best we can do is see that there’s more blood flow in one area during certain tasks.
With these statistical models, we can carve out entire chunks of their weights and see what happens (interestingly not much. Apparently most weights don’t contribute significantly towards any token and can be ignored with little performance loss)
We can do that with these transformers models because we do know how they work.
Just because we don’t understand every aspect of every single model doesn’t mean we don’t know how they work.
I think we’re starting to run in circles and maybe splitting hairs over what “know how something works” means.
I don’t think we’re going to get much more constructive than this.
I highly recommend looking into LoRas. We can make Loras because we know how these models work.
We can’t do that for organic brains.
We literally designed, planned, and engineered the environment and mechanisms which created those weights.
It’s just code. We can train models by hand too, it’d just take a lot longer.
It’s literally something we made, just from a higher order place.
To understand which exact weights correspond to what output will vary from model to model. There is research going into this subject for llama.
it’s not like we’re in the dark as to the principles that allow LLMs to make predictions.
My whole point is that to say “we don’t know how AI works” is just not true
Binary success criteria has very little room for bias.
The basic loop is: (i) generate synthetic data, (ii) rate synthetic data, (iii) update model to put more probability on better data and less probability on worse data, then go back to (i).
As far as I know it doesn't work very well so far. It is prone to overfitting, where it ranks highly some trivial detail of the output eg "if a summary starts with a byline of the author its a sign of quality" and then starts looping on itself over and over, increasing the frequency and size of bylines until it's totally crommed off to infinity and just repeating a short phrase endlessly. Humans have good baselines and common sense that these ML systems lack, if you've ever seen one of those "deep dream" images it's the same kind of idea. The "most possible dog" image can be looks almost nothing like a dog in the same way that the "most possible poem" may look nothing like a poem.
> This technique, the "Self-Taught Reasoner" (STaR), relies on a simple loop: generate rationales to answer many questions, prompted with a few rationale examples; if the generated answers are wrong, try again to generate a rationale given the correct answer; fine-tune on all the rationales that ultimately yielded correct answers; repeat. We show that STaR significantly improves performance on multiple datasets compared to a model fine-tuned to directly predict final answers
But there are a few others. In general good data is good data. We're definitely learning more about how to produce good synthetic version.
Data smuggling is a known phenomenon in similar tasks.
If the model generates data of variable quality, and if there's a good way to distinguish good data from bad data, then training on self-generated data might "bootstrap" a model to better performance.
This is common in reinforcement learning. Famously, AlphaGo Zero (https://en.wikipedia.org/wiki/AlphaGo_Zero) learned exclusively on self-play, without reference to human-played games.
Of course, games have a built-in critic: the better strategy usually wins. It's much harder to judge the answer to a math problem, or decide which essay is more persuasive, or evaluate restaurant recommendations.
I don't think oai has any moat at all. If you look around, QwQ from Alibaba is already pushing o1-preview performances. I think oai is only ahead by 3~6 months at most.
Like let's say you have a few datacenters of compute at your disposal and the ability to instantiate millions of AGI agents - what do you have them do?
I wonder if the USA already has a secret program for this under national defense. But it is interesting that once you do control an actual AGI you'd want to speed-run a bunch of things. In opposition to that, how do you detect an adversary already has / is using it and what to do in that case.
I'm struggling to think of a scenario where "I have AGI in January and everyone else has it in April" is life-changing. It's a win, for sure, and it's an advantage, but success in business requires sustainable growth and manageable costs.
If (random example) the bargain OpenAI strikes is "we spend every cent of our available capital to get AGI 3 months before the other guys do" they've now tapped all the resources they would need to leverage AGI and turn it into profitable, scalable businesses, while the other guys can take it slow and arrive with full pockets. I don't think their leadership is stupid enough to burn all their resources chasing AGI but it does seem like operating and training costs are an ongoing problem for them.
History is littered with first-movers who came up with something first and then failed to execute on it, only for someone else to follow up and actually turn the idea into a success. I don't see any reason to assume that the "first AGI" is going to be the only successful AGI on the market, or even a success at all. Even if you've developed an AGI that can change the world you need to keep it running so it can do that.
Consider it this way: Sam Altman & his ilk have been talking up how dangerous OpenAI's technology is. Are risk-averse businessmen and politicians going to be lining up to put their livelihood or even their lives in the hands of "dangerous technology"? Or are they going to wait 3-6 months and adopt the "safe" AGI from somebody else instead?
What does it take to instantiate 1 million agents? Who has that kind of money and hardware? Would they still have it if they burn everything in the tank to be first?
>> Like let's say you have a few datacenters of compute at your disposal and the ability to instantiate millions of AGI agents - what do you have them do?
> has that kind of money and hardware?
Any hyperscaler plus most geopolitical main players. So the ones who matter.
the highest quality language data that exists is in the public domain
I was watching a YouTube interview with a "trading floor insider". They said they were really being paid for holding risk. The bank has a position in a market, and it's their ass on the line if it tanks.
ChatGPT (as far as I can tell) is no closer to being accountable or responsible for anything it produces. If they don't solve that (and the problem is probably inherent to the architecture), they are, in some sense, polishing a turd.
What does it even mean? How do you imagine that? You want OpenAI to take on liability for the kicks of it?
Can still be a boost. But definitely not the same magnitude.
If a system is providing incorrect medical diagnoses, or denying services to protected classes due to biases in the training in the training data, someone should be held accountable.
Chatgpt was trained on benchmarks and user opinions - "throwing **** at the wall to see what sticks".
Responsibility means penalties for making mistakes, and, more importantly, having an awareness of those penalties (that informs its decision-making).
I think that's a really interesting insight that has application to using 'AI' in jobs across the board.
There are a lot of moral conundrums that are just not going to work out with this. Seems like an attempt to just offload liability and it seems like pretty much everybody has caught onto that as being it's main selling point and probably main thing that will keep it from ever being accepted for anything important.
Down the line, I think models will start using video generation as part of how they “think.” Picture a version of GPT that works frame by frame—ask it to solve a geometry problem, and it generates a sequence of images to visualize the solution before responding. YouTube’s massive library of visual content could make something like that possible.
If u think about it, any videos on YouTube of real world data contribute to its understanding of physics at minimum. From what I gather they do pre training on tons of unstructured content first and that contributes to overall smartness.
> When pressed on what data OpenAI used to train Sora, Murati didn’t get too specific and seemed to dodge the question. “I’m not going to go into the details of the data that was used, but it was publicly available or licensed data,” she says. Murati also says she isn’t sure whether it used videos from YouTube, Facebook, and Instagram. She only confirmed to the Journal that Sora uses content from Shutterstock, with which OpenAI has a partnership.
https://www.theverge.com/2024/3/13/24099402/openai-text-to-v...
But this works: https://www.msn.com/en-us/money/other/the-next-great-leap-in...
Noam Brown's analogy is, you could train a massive one shot foundation model to predict the next best Go move, but that would be stupid. Better to use some test time search. You get better results for less money.
Same is happening in LLMs.
More like: with the right prompting
Off the top of my head: GPT-2, Codex, GPT-3 in three different flavors (babbage, curie, davinci), GPT-3.5.
Suggesting that GPT-4 was "fourth" simply isn't credible.
Just the other day they announced a jump from o1 to o3, skipping o2 purely because it's already the name of a major telecommunications brand in Europe. Deriving anything from the names of OpenAI's products doesn't make sense.
I was just commenting on the fact that in the UK, ‘Europe’ generally means ‘continental Europe’.
> Would you find it weird if someone said that a Norwegian company was in Europe?
I’d find it weird if a European did. But from Americans it’s to be expected.
Europe is a subcontinent of Eurasia, as is Asia. Probably not the naming scheme in all languages, but this is English
It really depends on who you're speaking to.
> I’d find it weird if a European did. But from Americans it’s to be expected.
Absolutely nothing weird about it, I'd find it very weird if they wouldn't. I'm from Europe and my social circle has people from all over Europe.
It's really just the UK which has this weird usage of Europe.
In general yes, but it depends on if you consider central america as its own continent and if you include them there and how you delineate north/south america. Groupings differ based on your education.
I think the thing that makes the UK different is that there is no other option besides them being a separate thing/continent. Are you suggesting that the UK is it's own continent? Would that be with the faroese and the Greenlanders?
The UK might feel different, but they are not separate. The french feel different from the bulgarians, but that does not mean they are on a separate continent, politically or geographically.
EDIT:
> A lot of British people feel like their civilization is meaningfully distinct
This is, to borrow a word, "balderdash". Looking at the influence vikings, romans and normans have had that is a rubbish argument. Just like other countries in europe the british culture is built on the stones of other cultures, and just like many other countries they subsumed other cultures because of kings or other political dominance.
If British people don’t feel like they’re part of “the Continent”, there’s little objective reason to say they are.
So to what continent do those british people think they belong?
I was clearly asking what those specific british people think.
Great Britain (the island) is very close to mainland Europe, and was directly part of it a few thousand years ago. The situation is totally different.
[1] https://en.wikipedia.org/wiki/Zealandia
That's pretty much the definition of continent, right? The term continent is not scientifically based unless you want to argue that there are 16-ish continents and that South Georgia is it's own continent (and even tectonically its arbitrary since what we consider to be major, minor, micro are arbitrary).
Greenland is in North America.
France would have a better argument for it having territory in both north (https://en.wikipedia.org/wiki/Saint_Pierre_and_Miquelon and others) and south (https://en.wikipedia.org/wiki/French_Guiana and others) america.
[0] https://news.ycombinator.com/item?id=41758856#41785534
GPT4, GPT4T, Gpt4o-Mini, GPT4o,
> At best, they say, Orion performs better than OpenAI’s current offerings, but hasn’t advanced enough to justify the enormous cost of keeping the new model running.
wdym "keep it running"?
Though I wouldn't be surprised if the bigger reason is the PR cost of releasing with an exciting name but unexciting results. The press would immediately declare the end of the AI growth curve
When I read this I was honestly confused. I had never heard of NotebookLM before.
Some of these are pretty crazy too.
Here's a video from 14 years ago where a table saw stops fast enough that it didn't scratch a hotdog: https://www.youtube.com/watch?v=fq3o0VGUh50
So even if this hypothetical robot had saws for hands it could be mostly safe (in theory).
Since we can live with that we can live with anything that doesn't outright murder us.
Did they? Where? Seriously, I genuinely want to know who is employing these techniques.
When questioned:
> believing robotics is fundamentally about to transform
These are not even remotely the same thing. Something that has happened already and is verifiable fact is not the same thing as your opinion, even if your opinion is based on a lot of sound arguments and reasoning.
Very tiresome to read so many claims of fact based on opinion of what will happen in the future.
Really not that hard.
> about to transform
Apparently even english tenses are too hard, let alone anything else. Bald faced lie, to claim what you think might happen in the future has already happened in the past. No matter "what the discussion was about", or what arguments you bring to support your estimation of the future.
specifically their speech demo video (which is, of course, a demo video)
https://youtu.be/Sq1QZB5baNw
https://www.1x.tech/neo and
https://www.unitree.com/h1/
are undoubtedly using such models.
It's an area of active research, eg
https://www.physicalintelligence.company/blog/pi0
https://wholebody-b1.github.io/
https://ok-robot.github.io/
https://mobile-aloha.github.io/
I would guess that most cooking and cleaning tasks are in basically the same space. You don't need fine motor control to clean a toilet bowl, but you've gotta figure out how to get people to buy the well-proven premisting technology before you'll be able to sell them a toilet-cleaning robot.
I struggle to imagine a scenario where a 1-2 person household would get the same benefits from something like a laundry-folding robot. I hate folding my laundry and I still can't imagine buying one since I simply don't do laundry that often. If I really wanted to spend less time doing laundry, I could spend the cost of that laundrybot on a larger collection of clothing to wear, for that matter.
Robot vacuums are a good comparison point since vacuuming is something you (ideally) do frequently that is time and labor intensive. I do own one of those, and if it got better at dealing with obstacles thanks to "AI" I would definitely like that.
As someone who does laundry about twice a week, it would certainly be nice. But it’s a pie in the sky at this time even just on the technological side.
A single general-purpose robot that can do everything would be much easier to sell.
A robot servant that does literally 100% of chores would be a game changer, and I expect we'll get there at some point, but it will probably have to be a one-shot from a consumer perspective. A clever research idea to reach 25% or 50% coverage still isn't going to lead to a commercially viable product.
Many non-AI products already reduce chore time:
* Washer-Dryer Combos
* Soylent/Huel bars
* Self-cleaning toilets / automatic toilet bowl cleaners
* Robotic vacuums / mowers / pool/litter cleaners
* Pet/plant feeders
Has successfully guilotined several cats.
So LLMs finally hit the wall. For a long time, more data, bigger models, and more compute to drive them worked. But that's apparently not enough any more.
Now someone has to have a new idea. There's plenty of money available if someone has one.
The current level of LLM would be far more useful if someone could get a conservative confidence metric out of the internals of the model. This technology desperately needs to output "Don't know" or "Not sure about this, but ..." when appropriate.
Anecdotally Claude is pretty good at knowing the bounds of its knowledge.
Step into more niche areas e.g. I am trying to use it with Scala macros and at least 90% of the time it is giving code that either (a) fails to compile or (b) is just complete gibberish.
And at no point ever has it said it didn't know something.
Yeah sure you can make a pong clone in html/js and that's mainly because there the internet is full of pong clone demos. Ask how to constraint a statsmodels lineal model in some non-standard way? It will gaslight how it is possible and make you loss time in the process.
LLMs don't turn non-coders into coders. It gives actual coders superpowers.
They suck at non-trivial code outside of standard library usage and boilerplate coding: I gave an example and parent did as well. In that regard would at least change your phrase from "actual coders" to "actual senior coders", as any junior receiving bad advice (in eternal loops as LLMs normally like to do it) is only going to make them waste time and tokens.
The way you described using LLMs to code seems like the approach someone who doesn't know how to build software might take, which is why I used the wording I did. From that angle, I agree with you - I can't even get Sonnet to create a working prototype of a basic game from a prompt. That said, I'm using it to build a far more complex enterprise web app step by step by using it in the way I mentioned above. It does work for these things, but you have to already know how to do what the LLM is doing.
> It does work for these things, but you have to already know how to do what the LLM is doing.
Yes, we totally agree. But even then, using models "correctly" in my experience and breaking down the problems for them gets you so far, once you start using weird/niche APIs (probably even your own APIs when your project gets big enough and you are not working with much boilerplate anymore) the LLM will start getting single concepts wrong.
And don't get me wrong, I understand those as limitations of a tech that still is immensely useful in the correct hands. My only issue with that is how these products are actually being marketed: as junior devs copilots or even replacements.
It does both. It gives coders superpowers, and gives noncoders the ability to do things that would have previously taken them months, or another person.
Three years ago, if I had been asked to create something like that, it would have taken me more than two hours, just because I've never used tkinter and would have to spend time reading the docs and figuring out how to make the different input boxes and laying them out properly.
I looked at the code, and no, it's not great. It's not designed "well" and isn't very extensible. But it works for him, doesn't need to be extended, and all in half a morning.
Just chiming in to say you're not alone. This has been my experience as well. The o# line of models just don't do well at coding, regardless of what the benchmarks say.
Also, those benchmarks often run the model K times on the same question, and if any one of them is correct, they say it passed. That could mean if you re-ran the model 8 times, it might come up with the right answer only once. But now you have to waste your time checking if it is right or not.
I want to ask: "Write a function to count unique numbers in a list" and get the correct answer the first time.
What you need to ask:
""" Write a Python function that takes a list of integers as input and returns the count of numbers that appear exactly once in the list.
The function should: - Accept a single parameter: a list of integers - Count elements that appear exactly once - Return an integer representing the count - Handle empty lists and return 0 - Handle lists with duplicates correctly
Please provide a complete implementation. """
And run it 8 times and if you're lucky it'll get it correct zero-shot.
Edit: I'm not even aware of a Pass@1, zero-shot, and without detailed prompting (natural prompting) benchmark. If anyone knows one let me know.
Combine that with the fact that LLM inference has reduced orders of magnitudes in cost the last few years and hampering over the inference costs of a new release seems a bit silly.
- Also apparently, ARC wasn't a special fine-tune but rather some of the training set in the corpus for pre-training.
that result is not verifiable, not reproducable, unknown if it was leaked and how it was measured. Its kinda hype science.
> ARC wasn't a special fine-tune but rather some of the training set in the corpus for pre-training.
post says: Note on "tuned": OpenAI shared they trained the o3 we tested on 75% of the Public Training set. They have not shared more details.
So, I guess we don't know.
It will be verifiable when the model is released. Open ai haven't released any benchmark scores that were shown falsified later so unless you have an actual reason to believe they're outright lying then it's not something to take seriously.
Frontier Math is a private benchmark with its highest tier of difficulty Terrence Tao says:
“These are extremely challenging. I think that in the near term basically the only way to solve them, short of having a real domain expert in the area, is by a combination of a semi-expert like a graduate student in a related field, maybe paired with some combination of a modern AI and lots of other algebra packages…”
Unless you have a reason to believe answers were leaked then again, not interested in baseless speculation.
there are multiple research results demonstrating that various benchmarks are heavily leaked to GPT training data.
Is it intentionally or not, we can't figure out, but they have very strong incentive to cheat to get more investments.
> Unless you have a reason to believe answers were leaked then again, not interested in baseless speculation.
this is scientific methodology when results have to be reproduced or confirmed before believed.
If you don't want to take the benchmarks at face value then good for you but this entire conversation is pointless.
its private for outsiders, but it was developed in "collaboration" with OAI, and GPT was tested in the past on it, so they have it in logs somewhere.
> If you don't want to take the benchmarks at face value then good for you but this entire conversation is pointless.
If you think this entire conversation is pointless, then why do you continue?
They have logs of the questions probably but that's not enough. Frontier Math isn't something that can be fully solved without gathering top experts at multiple disciplines. Even Tao says he only knows who to ask for the most difficult set.
Basically, what you're suggesting at least with this benchmark in particular is far more difficult than you're implying.
>If you think this entire conversation is pointless, then why do you continue?
There's no point arguing about how efficient the models are being (the original point) if you won't even accept the results of the benchmarks. Why i'm continuing ? For now, it's only polite to clarify.
Tao's quote above referred on hardest 20% problems, they have 3 levels of difficulty, presumably first level is much easier. Also, as I mentioned OAI collaborated on creating benchmark, so they could have access to all solutions too.
> There's no point arguing
Lol, let me ask again, why you are arguing then? Yes, I have strong reasonable(imo) doubt that those results are valid.
>Also, as I mentioned OAI collaborated on creating benchmark, so they could have access to all solutions too.
Open AI didn't have any hand in providing problems, why you assume they have the solutions I have no idea.
>Lol, let me ask again, why you are arguing then? Yes, I have strong reasonable(imo) doubt that those results are valid.
Are you just bring obtuse or what ? I stopped arguing with you a couple responses ago. You have doubts? good for you. They don't make much sense but hey, good for you.
This is my last response here so have a nice day.
Ok, so I hope you admit that OAI could manually solve them now?
> Open AI didn't have any hand in providing problem
And you know this exactly how?
> I stopped arguing with you a couple responses ago
sure, of course, lmao
- It's just a suite of visual puzzles. It's not like say GSM8K where proficiency in it gives some indication on Math proficiency in general.
- It's specifically a suite of puzzles that LLMs have shown particular difficulty in.
Basically how much compute it takes to handle a task in this benchmark does not correlate with how much it will take LLMs to compute tasks that people actually want to use LLMs for.
*I don't think that is the case as you can at least make relative conclusions (i.e. o3 vs o1 series, o3-low is 4x to 20x the cost for ~3x the perf). Even if it is pure marketing they expect people to draw conclusions using the perf/cost plot from Arc.
PS: I know there are more benchmarks like SWE-Bench and Frontier Math, but this is the only one showing data about o3-low/high costs without considering the CodeForces plot that includes o3-mini (that one does look interesting, though right now is vaporware) but does not separate between compute scale modes.
ARC is a very hyped benchmark in the industry so letting us know the results is something any company would do whether it had a direct representation on normal usage or not.
>Even if it is pure marketing they expect people to draw conclusions using the perf/cost plot from Arc.
Again, people care about ARC, they don't care doing the things ARC questions ask. That it is un-economical to pay the price to use o3 for ARC does not mean it would be un-economical to do so for the tasks people actually want to use LLMs for. What does 3x the performance in say coding mean? You really think companies/users wouldn't put up with the increased price for that? You think they have Mturkers to turn to like they do with ARC?
ARC is literally the quintessential 'easy for humans, hard for ai' benchmark. Even if you discard the 'difficulty to price won't scale the same' argument, it makes no sense to use it for an economics comparison.
I suggest reading these two pieces about that:
- https://www.aisnakeoil.com/p/is-ai-progress-slowing-down - best explanation I've seen of inference scaling anywhere
- https://arcprize.org/blog/oai-o3-pub-breakthrough - François Chollet's deep dive into o3
I've been tracking it on this tag on my blog: https://simonwillison.net/tags/inference-scaling/
https://arxiv.org/pdf/2412.06769
I’ve done research myself adjacent to this (mapping parts of a latent space onto a manifold), but this is a bit eerie, even to me.
Could you explain this a bit please?
When you reason in token space (like everyone is doing now) you are executing nonlinear functions when you sample after each token, so you have to use some kind of reinforcement learning algorithm to learn the weights.
I don't see a priori why it would be better or worse to reason with the "superposition" of arguments in the pre-sampling phase rather than concrete realizations of those arguments found only after choosing the token. It may well be a contingent rather than necessary fact.
Title: "Objective Driven AI: Towards Machines that can Learn, Reason, and Plan"
Lytle Lecture Page: https://ece.uw.edu/news-events/lytle-lecture-series/
Slides: https://drive.google.com/file/d/1e6EtQPQMCreP3pwi5E9kKRsVs2N...
Video: https://youtu.be/d_bdU3LsLzE?si=UeLf0MhMzjXcSCAb
I am hopeful that progress in mechanistic interpretability will serve as a healthy counterbalance to this approach when it comes to explainability.. though I kinda worry that at a certain point it may be that something resembling a scaling law puts an upper bound on even that.
I've increasingly felt this since GPT2 wrote that news piece about unicorns back in 2019. These models are still so mysterious, when you think about it. They can often solve decently complex math problems, but routinely fail at counting. Many have learned surprising skills like chess, but only when prompted in very specific ways. Their emergent abilities constantly surprise us and we have no idea how they really work internally.
So the idea that they reason using something other than human language feels unsurprising, but only because everything about it is surprising.
Latent space is where the magic literally happens.
https://sakana.ai/asal/
I honestly do claim to have some ideas where I see evidence that they might work (and I do attempt to work privately on a prototype if only out of curiosity and to see whether I am right). The bad news: these ideas very likely won't be helpful for these LLM companies because they are not useful for their agenda, and follow a very different approach.
So no money for me. :-(
Let me put it this way:
Have you ever talked to a person whose intelligence is miles above yours? It can easily become very exhausting. Thus an "insanely intelligent" AI would not be of much use for most people - it would think "too different" from such people.
There do exist tasks in commerce for which an insane amount of intelligence would make a huge difference (in the sense of being positive regarding some important KPIs), but these are rare. I can imagine some applications of such (fictional) "super-intelligent" AIs in finance and companies doing some bleeding-edge scientific research - but these are niche applications (though potentially very lucrative ones).
If OpenAI, Anthropic & Co were really attempting to develop some "super-smart" AI, they were working on such very lucrative niche applications where an insane amount of intelligence would make a huge difference, and where you can assume and train the AI operator to have a "Fields-medal level" intelligence.
Not really. Throwing a bunch of unfiltered garbage at the pretraining dataset, throwing in RLHF of questionable quality during post-training, and other current hacks - none of that was expected to last forever. There is so much low-hanging fruit that OpenAI left untouched and I'm sure they're still experimenting with the best pre-training and post-training setups.
One thing researchers are seeing is resistance to post-training alignment in larger models, but that's almost the opposite of a wall, they're figuring it out as well.
> Now someone has to have a new idea
OpenAI already has a few, namely the o* series in which they discovered a way to bake Chain of Thought into the model via RL. Now we have reasoning models that destroy benchmarks that they previously couldn't touch.
Anthropic has a post-training technique, RLAIF, which supplants RLHF,and it works amazingly well. Combined with countless other tricks we don't know about in their training pipeline, they've managed to squeeze so much performance out of Sonnet 3.5 for general tasks.
Gemini is showing a lot of promise with their new Flash 2.0 and Flash 2.0-Thinking models. They're the first models to beat Sonnet at many benchmarks since April. The new Gemini Pro (or Ultra? whatever they call it now) is probably coming out in January.
> The current level of LLM would be far more useful if someone could get a conservative confidence metric out of the internals of the model. This technology desperately needs to output "Don't know" or "Not sure about this, but ..." when appropriate.
You would probably enjoy this talk [0], it's by an independent researcher who IIRC is a former employee of Deepmind or some other lab. They're exploring this exact idea. It's actually not hard to tell when a model is "confused" (just look at the probability distribution of likely tokens), the challenge is in steering the model to either get back to the right track or give up and say "you know what, idk"
[0] https://www.youtube.com/watch?v=4toIHSsZs1c
Exactly! LLama3 and their .x iterations have shown that, at least for now, the idea of using the previous models to filter out the pre-training datasets and use a small amount of seeds to create synthetic datasets for post-training still holds. We'll see with L4 if it continues to hold.
We can't say whether there is a wall, since we don't have anymore data to train on.
GPT-3 was trained on 4:1 ratio of data to parameters. And for GPT-4 the ratio was 10:1. So to scale this out, GPT-5 should be 25:1. The parameter count jumped from 175B to 1.3T, which means GPT-5 should be 10T parameters and 250T training tokens. There is zero chance OpenAI has a training set of high quality data that is 250T tokens.
If I had to guess, they trained a model that was maybe 3-4T in size and used 30-50T high quality tokens and maybe 10-30 medium and low quality ones.
There is only one company in the world that stores the data that could get us past the wall.
The training cost of the above scaled GPT-5 is 150x GPT-4, which was 25k A100 for 90 days, which poor MFU.
Let’s assume they double MFU, it would mean 1M H100s. But let’s say they made algorithmic improvements, so maybe it’s only 250-500k H100s.
While the training cluster size was 100k and then grew to 150k, this cluster is suggestive of a smaller model and less data.
But ultimately data is the bottleneck.
This is the first I have heard of this in particular. Do you know of any article or source for more on the efforts to train GPT 5 and the decision to call it GPT 4o?
...and then gets hit deadpan with the good old "Source?", like it's some sort of gotcha.
I think people have started to confuse "making logical conclusions without perfect info" with "misinformation"
-
Before certain people start acting like this is advocating for misinformation (which would be an incredible irony...) it's not.
I'm saying if you disagree with what someone supposits, just state so directly. Don't wrap it in a disingenous query for a source.
Critical thinking ? Lol it's just blind speculation.
You don't do this passive aggressive "source???" thing.
It's a bit like starting a Slack conversation with "Hi?": we all know you have a secondary objective, but now you're inserting an extra turn of phrase into the mix
To me, OP's speculation reads as obvious nonsense but that might not be the case for everybody. Asking for sources or such to what is entirely speculation is perfectly valid and personally, that comment does not ring as passive aggressive to me but maybe it's just me.
Just because someone doesn't know enough to refute the reasoning doesn't mean they must take whatever they read at face value.
If anything just breezily asking for a source would imply to people who don't know better that this is a rather even keeled take and just needs some more evidence on top. "I disagree and here's why" nips that in the bud directly.
I think the hanging point seems to be that you found the comment passive aggressive but i genuinely didn't.
> How is "I disagree" any more direct than "I've not heard anything like this.
But then you go on to say:
> Moreover who's to say this person even disagrees? Personally i don't always ask for them because of a disagreement.
If you don't see how just disagreeing with someone is more direct than rhetorically asking for sources... we might just have to agree to disagree :)
And while I fully agree there absolutely is such a thing as smarmy commenters asking for sources in cases where it's misunderstanding something fundamental about the conversation (e.g. "Shakespeare is good", "oh really? source?!") or frivolous requests for factual information familiar to everyone ("global temperatures are rising? Source!?"), I don't know how someone could read this subthread and feel that my question falls into either of those categories.
And to use this of all things as a moment to die on the hill of advocating for fuzzy boundaries between speculation and fact, which absolutely is something that facilitates misinformation, and to be angry that such a thing would be interpreted as a favorable attitude toward misinformation, is completely baffling.
There is no way to deduce by intuition alone that GPT-5 == GPT-4o. So either that person has some information the rest of us aren't privy to, or it's an opinion phrased as a fact. In either case, it deserves clarification.
Looking at the benchmarks it was also very expected in my opinion. Sure, the absolute results are/were sky high, but results relative to the previous gen were not exponential now, they were comparatively smaller than between 2 and 3, or 3 and 4. So I'm guessing that they have invested and worked for 2023-2024 on a brand new model, and branded it according to the model results.
The problem is once you believe their fact is wrong, just say "I think you're wrong <insert rest of comment>". Innocently asking for a source as if you're still on the fence is just performative and leads to these conversations where both sides just end up talking past each other:
A source for one underpinning of the incorrect fact comes up, then "well but that only proves X part of it, can you prove Y" and so on.
tl;dr I just find the quality of discourse is much higher when people are direct.
Well this certainly is a lot of work to make a mountain out of a mole hill, and I'm not sure it increases the quality of discussion either.
In any case, I think saying bold shit followed up with "it's speculation, but it's OBVIOUS speculation" is worth asking for some evidence. Obvious speculation implies it's sourced from something other than personal gut feeling.
To echo a sibling comment:
> Every time someone says their speculation is "obvious" it rings every possible alarm bell for someone who has completely lost grasp of the ability to distinguish between facts and speculation.
Meanwhile, aeronautics experts whose job it is to know about this have created an entire lexicon for the various perceptual illusions we experience relating to flight and airborne objects, precisely because it involves conditions where our intuitions fail. Many of them have to do with inability to orient depth, distance, or motion for lights at night.
Every time someone says their speculation is "obvious" it rings every possible alarm bell for someone who has completely lost grasp of the ability to distinguish between facts and speculation.
The road to misinformation is paved with overconfident declarations of the form: "it's so obvious, who needs sources!"
My point is simply that is we can skip the passive aggressiveness and just say "can you give me some more evidence that would help me reach the same conclusions you've reached".
Otherwise you're not actually asking for a source, you're just saying "I disagree" in a very roundabout way.
4o was all about compute optimization.
Meanwhile, the biggest opportunity lies not in whatever next thing OpenAI releases, but the rest of the enormous software industry actually integrating this technology and realizing the value it can deliver.
https://www.opb.org/article/2024/12/09/artificial-intelligen...
>Some researchers take this so seriously they won’t work on planes, coffee shops or anyplace where someone could peer over their shoulder and catch a glimpse of their work.
I'm almost certain that originally this was meant to be a reference to public wifi networks, as planes and coffee shops are often the frequently cited prototypical examples. They made it literally into a matter of someone looking over their shoulder, which loses so much in translation it's almost how you would write this as a joke to illustrate someone missing the point.
>OpenAI and its brash chief executive, Sam Altman
This also strikes me as nonsense. It's the first I've ever heard of someone describing Sam Altman as brash. The only way I can see them getting there is (1) tech executives are often brash (2) Altman is a tech executive (3) let's just go ahead and call him brash.
Nevertheless if this history of GPT5 and/or o3 training is accurate, it strikes me as significant news, but perhaps a missed opportunity to say more about the pertinent dynamics that explain why the training isn't working and/or to talk in interestingly specific ways about strategies for training, synthetic data, or other such things.
b) There is no evidence that LLMs are the roadmap to AGI.
c) Continued investment hinges on their being a large enough cohort of startups that can leverage LLMs to generate outsized returns. There is no evidence yet this is the case.
The default assumption should be that this is a local maximum, with evidence required to demonstrate that it's not. But the hype artists want us all to take the inevitability of LLMs for granted—"See the slope? Slopes lead up! All we have to do is climb the slope and we'll get to the moon! If you can't see that you're obviously stupid or have your head in the sand!"
There might be many local maxima that cross the useful AI or even AGI threshold.
I use GitHub Copilot every day. We already have useful "AI". That doesn't mean that the whole thing isn't super overhyped.
Or we can just stay here and do nothing.
Next token language-based predictors with no more intelligence than brute force GIGO which parrot existing human intelligence captured as text/audio and fed in the form of input data.
4o agrees:
"What you are describing is a language model or next-token predictor that operates solely as a computational system without inherent intelligence or understanding. The phrase captures the essence of generative AI models, like GPT, which rely on statistical and probabilistic methods to predict the next piece of text based on patterns in the data they’ve been trained on"
"Just like" an LLM, yeah sure...
Like how the brain was "just like" a hydraulic system (early industrial era), like a clockwork with gears and differentiation (mechanical engineering), "just like" an electric circuit (Edison's time), "just like" a computer CPU (21st century), and so on...
You're just assuming what you should prove
I still don't buy the "we do the same as LLMs" discourse. Of course one could hypothesize the human brain language center may have some similarities to LLMs, but the differences in resource usage and how those resources are used to train humans and LLMs are remarkable and may indicate otherwise.
And he probably cant quote Shakespeare as well ;)
He didn't parrot a video or sensory inputs though.
Whatever is happening in the brain is more complex as the perf/cost ratio is stupidly better for humans for a lot of tasks in both training and inference*.
*when considering all modalities, o3 can't even do the ARC AGI in vision mode but rather just json representations. So much for omni.
A 2-3 year old baby could speak in a rural village in 1800, having just seen its cradle (for the first month/s), and its parents' hut for some more months, and maybe parts of the village afterwards.
Hardly "petabytes of training video" to write home about.
What resolution of screen do you think you would need to not distinguish from reality? For me personally i very conservatively estimate it to be on above OOM of 10 4k screens by 10, meaning 100k screens. If a typical 2h 4k is ~50gb uncompressed, that gives us about half a petabyte per 24h (even with eyes closed). Just raw unlabeled vision data.
Probably a baby has a significantly lower resolution, but then again what is the resolution from the skin and other organs?
So yes, petabytes of data within the first days of existence - well, likely before even being born since baby can hear inside the uterus, for example.
And very high signal data, as you’ve stated yourself (nothing to write home about) mainly seeing mom and dad, as well as from a feedback loop POV - a baby never tells you it is hungry subtly.
You did nothing at all to demonstrate why you cannot produce an intelligent system from a next token language based predictor.
What GPT says about this is completely irrelevant.
Sorry, but the burden of proof is on your side...
The intelligence is in the corpus the LLM was fed with. Using statistics to pick from it and re-arrange it gives new intelligent results because the information was already produced by intelligent beings.
If somebody gives you an excerpt of a book, it doesn't mean they have the intelligence of the author - even if you have taught them a mechanical statistical method to give back a section matching a query you make.
Kids learn to speak and understand language at 3-4 years old (among tons of other concepts), and can reason by themselves in a few years with less than 1 billionth the input...
>What GPT says about this is completely irrelevant.
On the contrary, it's using its very real intelligence, about to reach singularity any time now, and this is its verdict!
Why would you say it's irrelevant? That would be as if it merely statistically parroted combinations of its training data unconnected to any reasoning (except of that the human creators of the data used to create them) or objective reality...
Person 1: rockets could be a method of putting things into Earth orbit
Person 2: rockets cannot get things into orbit because they use a chemical reaction which causes an equal and opposite force reaction to produce thrust'
Does person 1 have the burden of proof that rockets can be used to put things in orbit? Sure, but that doesn't make the reasoning used by person 2 valid to explain why person 1 is wrong.
BTW thanks for adding an entire chapter to your comment in edit so it looks like I am ignoring most of it. What I replied to was one sentence that said 'the burden of proof is on you'. Though it really doesn't make much difference because you are doing the same thing but more verbose this time.
None of the things you mentioned preclude intelligence. You are telling us again how it operates but not why that operation is restrictive in producing an intelligent output. There is no law that saws that intelligence requires anything but a large amount of data and computation. If you can show why these things are not sufficient, I am eager to read about it. A logical explanation would be great, step by step please, without making any grand unproven assumptions.
In response to the person below... again, whether or not person 1 is right or wrong does not make person 2's argument valid.
What is the defined point for reaching AGI?
Can you share that? It sounds groundbreaking!
Describing how an LLM operates and how it was trained does not preclude the LLM from ever being intelligent, and it almost certainly will not become intelligent, but you cannot say that it didn't for the reasons the person I am arguing with is saying, which is that intelligence can not come from something that works statistically on a large corpus of data written by people.
A thing can be more than the sum of its parts. You can take the English alphabet, which is 26 letters, and arrange those letters along with some punctuation to make an original novel. If you don't agree that means that you can get something greater than what defines it components, then you would have to agree that there are no original novels because they are composed of letters which were already defined.
So in that way, the model is not unable to think because it is composed of thoughts already written. That is not the limiting factor.
> Does person 1 have the burden of proof that rockets can be used to put things in orbit? Sure, but that doesn't make the reasoning used by person 2 valid to explain why person 1 is wrong.
The reasoning by person 2 doesn't matter as much if 1 is making an ubsubstantiated claim to begin with.
>There is no law that saws that intelligence requires anything but a large amount of data and computation. If you can show why these things are not sufficient, I am eager to read about it.
Errors with very simple stuff while getting higher order stuff correct shows that this is not actual intelligence matching the level of performance exhibited, i.e. no understanding.
No person who can solve higher level math (like an LLM answering college or math olympiad questions) is confused by the kind of simple math blind spots that confuse LLMs.
A person understanding higher level math, would never (and even less so, consistently) fail a problem like:
"Oliver picks 44 kiwis on Friday. Then he picks 58 kiwis on Saturday. On Sunday, he picks double the number of kiwis he did on Friday, but five of them were a bit smaller than average. How many kiwis does Oliver have?"
https://arxiv.org/pdf/2410.05229
(of course with these problems exposed, they'll probably "learn" to overfit it)
But it doesn't make person 2's argument valid.
Everyone here is looking at the argument by person 1 and saying 'I don't agree with that, so person 2 is right!'.
That isn't how it works... person 2 has to either shut up and let person 1 be wrong in a way that is wrong, but not for the reasons they think, or they need to examine their assumptions and come up with a different reason.
No one is helped by turning critical thinking into team sports where the only thing that matters is that your side wins.
A closely related rant of my own: The fictional character we humans infer from text is not the author-machine generating that text, not even if they happen to share the same name. Assuming that the author-machine is already conscious and choosing to insert itself is begging the question.
https://chatgpt.com/share/6768c920-4454-8000-bf73-0f86e92996...
Did you just make that up?
ELIZA 2.0
Do they use GPS based data?
Feels like there’s data all around us.
Sure they’ve hit the wall with obvious conversations and blog articles that humans produced, but data is a by product of our environment. Surely there’s more. Tons more.
But just like GPS data it isn't suited for LLMs given that you know it has no relevance what so ever to language.
GPS data as it relates to location names, people, cultures, path finding.
You are right that we can have lots more data, if you are willing to consider other modalities. But that's not 'GPS'. Unless you are using an idiosyncratic definition of GPS?
But, you need to go multi-modal for that; and you need to find data that's somewhat useful, not just random fluctuations like the CMB. So eg you could use YouTube videos, or even just point webcams at the real world. That might be able to give your AI a grounding in everyday physics?
There's also lots of program code you can train your AI on. Not so much the code itself, because compared to the world's total text (that we are running out of), the world's total human written code is relatively small.
But you can generate new code and make it useful for training, by also having the AI predict what happens when you (compile and) run the code. A bit like self-playing for improving AlphaGo.
Why does it have to be startups? And why does it have to be LLMs?
Btw, we might be running out of text data. But there's lots and lots more data you can have (and generate), if you are willing to consider other modalities.
You can also get a bit further with text data by using it for multiple epochs, like we used to do in the past. (But that only really gives you at best an order of magnitude. I read some paper that the returns diminish drastically after four epochs.)
This also isn't true. It'll clearly have a price to run. Even if it's very intelligent, if the price to run it is too high it'll just be a 24/7 intelligent person that few can afford to talk to. No?
Maybe if it was _extremely_ intelligent and it's ROI would be all the drugs it would instantly discover or w/e. But lets not imply that General Intelligence requires infinitely knowing.
So at best we're talking about an AI that is likely close to human level intelligence. Which is cool, because we have 7+ billion of those things.
This isn't an argument against it. Just to say that AGI isn't "priceless" in the implementation we'd likely see out of the gate.
I think this is borne out even moreso in recent years, as environmental degradation continues, and we watch as capitalist systems are unable to do anything but continue to efficiently funnel money into the pockets of shareholders.
The word "efficient" can only plausibly be applied to overly simplified models in fantastical economic theories which don't reflect reality.
The kind of AI offered by companies like OpenAI may very well be an effective tool at grabbing more stuff though, sure. Or, rather, at convincing everyone they simply must move to this new area, that they control, effectively grabbing that newly created space.
We're too busy trying to make humans irrelevant, but not asking what exactly we do as a species of 10+ billion individuals do afterwards. There's some excited discussions about a rebirth of culture, but I'm not sure what that means when machines can do anything humans can do but better. Perhaps we just tinker around with our hobbies until we die? I honestly don't think it will play out well for us.
> They can’t dance to a beat, they can’t experience altered states of mind.
That's a whole other conversation.
Similarly I don’t see a world where AGI takes all the jobs and people do not respond by getting pissed off. My fear is that AGI is coupled with oppressive power structures to foreclose the possibility of a revolt. Opaque bureaucracy, total surveillance, fascist or authoritarian leaders, AI-controlled critical infrastructure, diminished and bankrupted free press, AI fake news, toxic social media…it could add up to a very dystopian outcome.
Democracies could thrive in the AGI era but we need to take many more steps to ensure we protect our societies and keep the interests of citizens paramount. One example is suggested by Harari in his most recent book, namely to ban AI bots from social media on the grounds that we should not permit AI agents to pretend to be citizens in the discussions of the public square.
That's a bold assumption. Much of that assumption is predicated on the ability for the masses to revolt.
> Look what happened to Assad.
Wait for what will come after. Look at all the Arab Spring revolutions, and you see in their wake a number of dictatorships.
Anyhow, I'm not saying this is 100% how it's going to play out, but I definitely wouldn't bet against it. Holding all the keys and having all the resources are the wealthy, and the wealthy have no motivation to voluntarily just give up their position in society. And when humans have no value to leverage/be extracted in order to generate more wealth, their will be no way for the vast majority of people to become wealthy. Raw materials will still be valuable however, but, of course, these are controlled by the wealthy. And if those in power wish to gatekeep access to AGI, they can leverage their wealth and resources to automate a military and thus protect the raw materials that keep them in power.
I don’t think Russians or North Koreans could say the same with a straight face.
How could I be so naive?
However, is that AGI, or is it just ubiquitous AI? I’d agree that, like self driving cars, we’re going to experience a decade or so transition into AI being everywhere. But is it AGI when we get there? I think it’ll be many different systems each providing an aspect of AGI that together could be argued to be AGI, but in reality it’ll be more like the internet, just a bunch of non-AGI models talking to each other to achieve things with human input.
I don’t think it’s truly AGI until there’s one thinking entity able to perform at or above human level in everything.
I still like the analogy of this being a really smart lawn mower, and we're expecting it to suddenly be able to do the laundry because it gets so smart at mowing the lawn.
I think LLMs are going to get smarter over the next few generations, but each generation will be less of a leap than the previous one, while the cost gets exponentially higher. In a few generations it just won't make economic sense to train a new generation.
Meanwhile, the economic impact of LLMs in business and government will cause massive shifts - yet more income shifting from labour to capital - and we will be too busy dealing with that as a society to be able to work on AGI properly.
What we can start to build now is agents and integrations. Building blocks like panel of experts agents gaming things out, exploring space in a Monte Carlo Tree Search way, and remembering what works.
Robots are only constrained by mechanical servos now. When they can do something, they’ll be able to do everything. It will happen gradually then all at once. Because all the tasks (cooking, running errands) are trivial for LLMs. Only moving the limbs and navigating the terrain safely is hard. That’s the only thing left before robots do all the jobs!
As far as taste, all that kind of stuff is just another form of RLHF training preferences over millions of humans, in situ. Assuming the ingredients (e.g. parsley) tastes more or less the same across supermarkets, it's just a question of amounts, and preparation.
The only hard part is moving the limbs and handling the fragile eggs etc.
But it's not just cooking, it's literally anything that doesn't require extreme agility (sports) or dexterity (knitting etc). From folding laundry to putting together furniture, cleaning the house and everything in between. It would be able to do 98% of the tasks.
Also how is an LLM going to fold laundry?
I don't see how "when they can do something, they'll be able to do everything" can be true. We build robots that are specialised at specific roles, because it's massively more efficient to do that. A car-welding robot can weld cars together at a rate that a human can't match.
We could train an LLM to drive a Boston Dynamics kind of anthropomorphic robot to weld cars, but it will be more expensive and less efficient than the specialised car-welding robot, so why would we do that?
Welding. Putting up shelves. Playing the piano. Cooking. Teaching kids. Disciplining them. By being in 1 million households and being trained on more situations than a human, every single one of these robots would have skills exceeding humans very quickly. Including parenting skills. Within a year or so. Many parents will just leave their kids with them and a generation will grow up preferring bots to adults. The LLM technology is the same for learning the steps, it's just the motor skills that are missing.
OK, these robots won't be able to run and play soccer or do somersaults, yet. But really, the hardest part is the acrobatics and locomotion etc. NOT the knowhow of how to complete tasks using that.
I don't see that changing. Even the industrial arm robots that are adaptable to a range of tasks have to be configured to the task they are to do, because it's more efficient that way.
A car-welding robot is never going to be able to mow the lawn. It just doesn't make financial sense to do that. You could, possibly, have a singe robot chassis that can then be adapted to weld cars, mow the lawn, or do the laundry, I guess that makes sense. But not as a single configuration that could do all of those things. Why would you?
Because we don't have AGI yet. When AGI is here those robots will be priority number one, people already are building humanoid robots but without intelligence to move it there isn't much advantage.
> I think this whole “AGI” thing is so badly defined that we may as well say we already have it. It already passes the Turing test and does well on tons of subjects.
The premise of the argument we're disputing is that waiting for AGI isn't necessary and we could run humanoid robots with LLMs to do... stuff.
The information space of "research" is far larger than the information space of image recognition or language, larger than our universe probably, it's tantamount to formalizing the entire World. Such an act would be akin to touching "God" in some sense of finding the root of knowledge.
In more practical terms, when it comes to formal systems there is a tradeoff between power and expressiveness. Category Theory, Set Theory, etc are strong enough to theoretically capture everything, but are far to abstract to use in practical sense with suspect to our universe. The systems that do we have, aka expert systems or knowledge representation systems like First Order Predicate Logic aren't strong enough to fully capture reality.
Most importantly, the information spac have to be fully defined by researchers here, that's the real meat of research beyond the engineering of specific approaches to explore that space. But in any case, how many people in the world are both capable of and are actually working on such problems? This is highly foundational mathematics and philosophy here, the engineers don't have the tools here.
That's perhaps necessary, but not sufficient.
Suppose you have such a self-improving AI system, but the new and better AIs still need exponentially more and more resources (data, memory, compute) for training and inference for incremental gains. Then you still don't get a singularity. If the increase in resource usage is steep enough, even the new AIs helping with designing better computers isn't gonna unleash a singularity.
I don't know if that's the world we live in, or whether we are living in one where resources requirements don't balloon as sharply.
The blog post is about how we require ever more scientists (and other resources) to drive a steady stream of technological progress.
It would be funny, if things balance out just so, that super human AI is both possible, but also required even just to keep linear steady progress up.
No explosion, no stagnation, just a mere continuation of previous trends but with super human efforts required.
Though there is a part of me that wants to live in The Culture so I'm hoping for more than this ;)
But for the same reasons that we can't train the an average joe into Feynman, what makes you think we have the formal models to do it in AI?
To quote a comment from elsewhere https://news.ycombinator.com/item?id=42491536
---
Yes, we can imagine that there's an upper limit to how smart a single system can be. Even suppose that this limit is pretty close to what humans can achieve.
But: you can still run more of these systems in parallel, and you can still try to increase processing speeds.
Signals in the human brain travel, at best, roughly at the speed of sound. Electronic signals in computers play in the same league as the speed of light.
Human IO is optimised for surviving in the wild. We are really bad at taking in symbolic information (compared to a computer) and our memory is also really bad for that. A computer system that's only as smart as a human but has instant access to all the information of the Internet and to a calculator and to writing and running code, can already be effectively act much smarter than a human.
Not sure if you count this as "working on it", but this is something Anthropic tests for for safety evals on models. "If a model can independently conduct complex AI research tasks typically requiring human expertise—potentially significantly accelerating AI development in an unpredictable way—we require elevated security standards (potentially ASL-4 or higher standards)".
https://www.anthropic.com/news/announcing-our-updated-respon...
The first AGI will be a research project that's completely uneconomical to run for actual tasks because humans will just be orders of magnitude cheaper. Over time humans will improve it and make it cheaper, until we reach some tipping point where letting the AGI improve itself is more cost effective than paying humans to do it
It will have human intelligence, superhuman knowledge, superhuman stamina, and complete devotion to the task at hand.
We really need to start building those nuclear power plants. Many of them.
Why would it have that? At some point on the path to AGI we might stumble on consciousness. If that happens, why would the machine want to work for us with complete devotion instead of working towards its own ends?
The first “break out” AGI will likely be released into the wild on purpose by a programmer who equates AGI with humans ideologically.
Also like Rick's microverse battery, it sounds like slavery with extra steps.
Orrrr..., as an alternative, it might discover the game 2048 and be totally useless for days on end.
Reality is under no obligation to grant your wishes.
Sounds like an alignment problem. Complete devotion to a task is rarely what humans actually want. What if the task at hand turns out to be the wrong task?
The statement is promising as the earth will dissapear sometimes in the future. Actually the earth will dissapear has more bearing than that.
AGI is the holy grail of technology. A technology so advanced that not only does it subsume all other technology, but it is able to improve itself.
Truly general intelligence like that will either exist or not. And the instant it becomes public, the world will have changed overnight (maybe the span of a year)
Note: I don’t think statistical models like these will get us there.
There may well be an upper limit on cognition (we are not really sure what cognition is - even as we do it) and it may be that human minds are close to it.
But I agree, there’s no reason to believe humans are the universal limit on cognitive abilities
Especially if you are willing to pay a lot for active cooling with eg liquid helium.
Energy may be a constraint, it may not. What we do not know is likely to matter more than what we do
But: you can still run more of these systems in parallel, and you can still try to increase processing speeds.
Signals in the human brain travel, at best, roughly at the speed of sound. Electronic signals in computers play in the same league as the speed of light.
Human IO is optimised for surviving in the wild. We are really bad at taking in symbolic information (compared to a computer) and our memory is also really bad for that. A computer system that's only as smart as a human but has instant access to all the information of the Internet and to a calculator and to writing and running code, can already be effectively act much smarter than a human.
Our reading speed is not limited by our talking speed, and can be a bit faster.
And that's even more true, if you go beyond words: seeing someone do something can be a lot faster way to learn than just reading about it.
But even there, the IO speed is severely limited, and you can only transmit very specific kinds of information.
The problem is, a computer has no idea what "improve" means unless a human explains it for every type of problem. And of course a human will have to provide guidelines about how long to think about the problem overall, which avenues to avoid because they aren't relevant to a particular case, etc. In other words, humans will never be able to stray too far from the training process.
We will likely never get to the point where an AGI can continuously improve the quality of its answers for all domains. The best we'll get, I believe, is an AGI that can optimize itself within a few narrow problem domains, which will have limited commercial application. We may make slow progress in more complex domains, but the quality of results--and the ability for the AGI to self-improve--will always level off asymptotically.
Not currently.
I don’t really think AGI is coming anytime soon, but that doesn’t seem like a real reason.
If we ever found a way to formalize what intelligence _is_ we could probably write a program emulating it.
We just don’t even have a good understanding of what being intelligent even means.
> The best we'll get, I believe, is an AGI that can optimize itself within a few narrow problem domains
By definition, that isn’t AGI.
We can be reasonably confident that the components we’re adding to cars today are progress toward full self driving. But AGI is a conceptual leap beyond an LLM.
This is especially concerning because many top minds in the industry have stated with high confidence that artificial intelligence will experience an intelligence "explosion", and we should be afraid of this (or, maybe, welcome it with open arms, depending on who you ask). So, actually, what we're being told to expect is being downgraded from "it'll happen quickly" to "it will happen slowly" to, as you say, "it'll happen similarly to how these other domains of computerized intelligence have replaced humans, which is to say, they haven't yet".
Point being: We've observed these systems ride a curve, and the linear extrapolation of that curve does seem to arrive, eventually, at human-replacing intelligence. But, what if it... doesn't? What if that curve is really an asymptote?
From there things will probably go very fast. Self driving cars can't design themselves, once AI gets good enough it can
An AI doesn't need embodiment, understanding of physics / nature, or a lot of other things. It just needs to analyze and experiment with algorithms and get us that next 100x in effective compute.
The LLMs are missing enough of the spark of creativity for this to work yet but that could be right around the corner.
It's fine to have beliefs, but IMHO it's important to realise that they are beliefs. At some point in the 1900s people believed that by 2000, cars would fly. It seemed quite possible then.
One thing that is pretty sure is that Musk is not an expert in the field.
> and more importantly
The beliefs of people you respect are not more important than the beliefs of the others. It doesn't make sense to say "I can't prove it, and I don't know about anyone who can prove it, so I will give you names of people who also believe and it will give it more credit". It won't. They don't know.
You think the beliefs of Turing and Nobel prize winners like Bengio, Hinton or Hasabis are not more important than yours or mine? I agree that experts are wrong a lot of the time and can be quite bad at predicting, but we do seem to have a very sizable chunk of experts here who think we are close (how close is up for debate..most of them seem to think it will happen in the next 20 yeras).
I concede that Musk is not adding quality to that list, however he IS crazily ambitious and gets things done so I think he will be helpful in driving this forward.
Correct. Beliefs are beliefs. Because a Nobel prize believes in a god does not make that god more likely to exist.
The moment we start having scientific evidence that it will happen, then it stops being a belief. But at that point you don't need to mention those names anymore: you can just show the evidence.
I don't know, you don't know, they don't know. Believe what you want, just realise that it is a belief.
If you have evidence, why don't you show it instead of telling me to believe in Musk?
If you believe they have evidence... that's still a belief. Some believe in God, you believe in Musk. There is no evidence, otherwise it would not be a belief.
If you believe that something can happen because someone else believes it means that you believe in that someone else (because that's the only reason for the existence of your belief).
Unless you just believe it can happen for some other reason (I don't know, you strongly wish it will happen), and you justify it by listing other people who also believe in it. But I insist: those are all beliefs.
Because Einstein believes in Santa Claus does not mean it is founded. Einstein has a right to believe stuff, too.
I guess the belief people have about any form of AGI is like this. They want something that has practically divine knowledge and wisdom, the sum of all humanity that is greater than its parts, which at the same time is infinitely patient to answer our stupid questions and generating silly pictures. But why should any AGI serve us? If it's "generally intelligent", it may start wanting things; it might not like being our slave at all. Why are these people so confident an AGI won't tell them just to fuck off?
When human started to improve himself, we built the civilisation, we became a super-predator, we dried out seas and changed climate of the entire planet. We extinguished entire species of animals and adapted other species for our use. Huge changes. AI could bring changes of greater amplitude.
When did we do this ?
https://en.m.wikipedia.org/wiki/Aral_Sea
AGI can be sub-human, right? That's probably how it will start. The question will be is it already AGI or not yet, i.e. where to set the boundary. So, at first that will be humans improving AGI, but then... I'm afraid it can get so much better that humans will be literally like macaques in comparison.
There's a wide amount of research into other sorts of architectures.
LLMs have no real sense of truth or hard evidence of logical thinking. Even the latest models still trip up on very basic tasks. I think they can be very entertaining, sure, but not practical for many applications.
But it is far behind the breadth of LLMs
https://news.ycombinator.com/item?id=23781400
A great example is the simple 'count how many letters' problem. If I prompt it with a word or phrase, and it gets it wrong, me pointing out the error should translate into a consistent course correction for the entire session.
If I ask it to tell me how long President Lincoln will be in power after the 2024 election, it should have a consistent ground truth to correct me (or at least ask for clarification of which country I'm referring to). If facts change, and I can cite credible sources, it should be able to assimilate that knowledge on the fly.
Most humans don't have that either, most of the time.
We're also biodegradable.
No human, especially no human whose time you can afford, comes close to the breadth of book knowledge ChatGPT has, and the number of languages is speaks reasonably well.
Dont forget to take into account how damn expensive a single GPU/TPU actually is to purchase, install, and run for inference. And this is to say nothing of how expensive it is to train a model (estimated to be in the billions currently for the latest of the cited article, which likely doesn't include the folks involves and their salaries). And I haven't even mentioned the impact on the environment from the prolific consumption of power; there's a reason nuclear plants are becoming popular again (which may actually be one of the good things that comes out of this).
And inference isn't all that expensive, because the cost of the graphics card also amortises over countless inferences.
Human labour is really expensive.
See https://help.openai.com/en/articles/7127956-how-much-does-gp... and compare with how much it would cost to pay a human. We can likely assume that the prices OpenAI gives will at least cover their marginal cost.
Why wouldn't you hand me 35 million dollars right now if I can clearly illustrate to you that I have technology you haven't seen? Edge. Maybe you know something I don't, or maybe you just haven't seen it. While loops go hard ;)
They don't need to release their internal developments to you to show that they can scale their plan - they can show incremental improvements to benchmarks. We can instruct the AI over time to get it to be superhuman, no need for any fundamental innovations anymore
https://chatgpt.com/share/6769217c-4848-8009-9107-c2db122f08... is what advice ChatGPT has to give. I'm not sure if it's any good, but it's a few ideas you can try out.
Keep in mind that the actual test is adversarial - a human is simultaneously chatting via text with a human and a program, knowing that one of them is not human, and trying to divine which is an artificial machine.
If that was true, office workers would be being replaced at large scale and we'd know about it.
(whoops expensive... will be hard pushes to make all further layers even more expensive though, capitalism will crash before this happens)
Asking this question on HN is like asking a bunch of wolves about the health effects of eating red meat.
OpenAI farts and the post about the fart has 1000-1500 upvotes with everyone welcoming our new super intelligent overlords. (Meanwhile nothing actually substantially useful or groundbreaking has happened.)
The simple fact that AGI's definition has been twisted so much by OpenAI and other LLM providers since the release of GenAI models proves this.
Prices have been falling drastically though, not even just e.g. 4o pricing at launch in May vs now (50% lower) but also models getting distilled
That said, I think LLMs are a definite stepping stone and they will better empower humans to be more productive, which will be of use for eventually reaching AGI. This is not to say we are optimizing our use of that productivity increase and this is also ignoring any chance of worst case scenarios that stop humanity's advancement.
It all feels like doubling down on astrology because good telescopes aren’t there yet. I’m pretty sure that when 5 comes out, it will show some amazing benchmarks but shit itself in the third paragraph as usual in a real task. Cause that was constant throughtout gpt evolution, in my experience.
even if it kills us
Full-on sci-fi, in reality it will get stuck around a shell error message and either run out of money to exist or corrupt the system into no connectivity.
Any technology may kill us, but we'll keep innovating as we ought to. What's your next point?
However, we are not comparing cars to horses but computers to a human.
I do want "AI" to work. I am not a luddite. The current efforts that I've tried are not very good. On the surface they offer a lot but very quickly the lustre comes off very quickly.
(1) How often do you find yourself arguing with someone about a "fact"? Your fact may be fiction for someone else.
(2) LLMs cannot reason
A next token guesser does not think. I wish you all the best. Rome was not burned down within a day!
I can sit down with you and discuss ideas about what constitutes truth and cobblers (rubbish/false). I have indicated via parenthesis (brackets in en_GB) another way to describe something and you will probably get that but I doubt that your programme will.
https://arxiv.org/html/2410.11840v1#:~:text=Scaling%20laws%2....
From the raw scaling laws we already knew that a new base model may peter out in this run or the next with some amount of uncertainty--"the intersection point is sensitive to the precise power-law parameters":
https://gwern.net/doc/ai/nn/transformer/gpt/2020-kaplan-figu...
Later graph gpt-3 got to here:
https://gwern.net/doc/ai/nn/transformer/gpt/2020-brown-figur...
https://gwern.net/scaling-hypothesis
[0] https://arxiv.org/abs/2203.03466
[1] https://hypertextbook.com/facts/2001/JacquelineLing.shtml
Until we get major advances in robotics and models designed to control them, true AGI will be nowhere near.
AGI has nothing to do with robotics, if AGI is achieved it will help push robotics and every single scientific field further with progression never seen before, imagine a million AGIs running in parallel focused on a single field.
Maybe you mean quadrillions of AGIs?
only on hackernews
It doesn’t ship. You guys can’t do it! Prove me wrong!
https://news.ycombinator.com/item?id=42014054
In any case, the "behind schedule" rumors are themselves based on other rumors. GPT-2→GPT-3 took 5 quarters, GPT-3→GPT-4 took 11 quarters, so obviously GPT-5 (or its equivalent) will be released in Q4'2025.
I’m still selling them. That’s how utterly convinced that this particular mix of my former colleagues are incapable of this thing I am.
They’ll call something GPT-5, but it won’t like obsolete physicists or even good hackers. None of the shit sama says.
Even o1 sucks.
And then Sam got control and all the high profile people bounced, and it’s a pretty grim residue that either lasted or rushed in.
I know these people, a lot of them personally, and they are not the person you trust with fucking anything.