This is one of the most tilted pieces I’ve ever read. For months Ed has predicted The AI Bubble will burst “any day now” frequently citing ai company’s revenue as a sign the product is not viable and the valuations are too high. The valuations seem to be primarily based on the R&D progress instead of on a theory that widespread adoption of the existing product will experience an uptick. The current landscape imho should be viewed as an R&D race amongst private actors.
> The valuations seem to be primarily based on the R&D progress
There hasn't been much R&D progress, though. Sure, as another commenter pointed out, context lengths have gotten longer and chat models can interpret images now, but the industry figureheads have been pushing agents, and we're not much closer to those than we were two years ago when GPT-4 came out. Current models simply are not consistent enough to do the kind of agentic stuff that AI valuations are predicated upon, nor is there any sign that a significantly smarter GPT-5 is just around the corner. Multi-modal chat is cute, but OpenAI is burning money. They're all burning money, and they don't have a product. They imply and imply that there's something big on the horizon, but it's been years, and there just isn't a killer app yet. Their platform isn't good enough, and it's not improving in the ways it would need to in order for Godot to arrive and for agents to be feasible.
By this I mean it’s a bet on what R&D might yield, current progress being some kind of signal. No one has certainty here. It’s an emergent technology and no one knows for certain how far it can be pushed.
I think you can simultaneously think that there is some real value being made with LLMs and also look at OpenAI losing $5B a year or thereabouts and really wonder how they're not going to run out of money.
That said, I'm learning a new sdk and I've moved 500-1k searches a month from kagi and google to llms.
Recent results are showing exponential improvement in reasoning and dramatic decreases in the time and cost to train models. O3 now ranks 50th on code forces according to openai staff. Are you aware of all of this and still say R&D hasn’t progressed?
You can invest in building bigger and more complicated pipe structures, but until you show the field that is supposed to be irrigated, you can't say you're disrupting farming business.
In the context of Moore's law exponential growth was measured in the number of transistors per integrated circuit. This seems vigorous and straightforward.
With AI the improvements have certainly been impressive but it isn't straightforward how you can define "reasoning" to measure whether or not the reasoning is exponentially "improving".
Again: Those are incremental improvements. The valuations are based on the promise that agents are just around the corner, and we just haven't seen the kind of categorical shift in intelligence that agents would require.
Spotting bubbles and predicting they will burst at some point is not a particularly useful skill. Housing in Amsterdam was in a bubble for 37 years in the 1700s; identifying the bubble early on would have been completely pointless.
> This is one of the most tilted pieces I’ve ever read.
He comes across as just a ludicrously unpleasant, spite-filled person.
> I'm fucking tired of having to write this sentence.
> I am so very bored of having this conversation
> I don't care about this number!
> Shut the fuck up!
> This isn't the early days of shit.
> Didn't we just talk about this? Fine, fine.
> $3.25 billion a quarter is absolutely pathetic.
> This isn’t real business! Sorry!
> He said in one of his stupid and boring blogs that
> This man is full of shit! Hey, tech media people reading this — your readers hate this shit! Stop printing it! Stop it!
> It's here where I'm going to choose to scream.
> Dario Amodei — much like Sam Altman — is a liar, a crook, a carnival barker and a charlatan, and the things he promises are equal parts ridiculous and offensive.
> Why are we humoring these oafs?
> Despite Newton's fawning praise
> Nobody talks like this! This isn’t how human beings sound! I don’t like reading it!
> Ewww.
> I'm sorry, I know I sound like a hater, and perhaps I am, but this shit doesn't impress me even a little.
> I know, I know, I'm a hater, I'm a pessimist, a cynic, but I need you to fucking listen to me: everything I am describing is unfathomably dangerous
> expensive, stupid, irksome, quasi-useless new product
> I know this has been a rant-filled newsletter, but I'm so tired of being told to be excited about this warmed-up dogshit.
> I refuse to sit here and pretend that any of this matters.
> I'm tired of the delusion. I'm tired of being forced to take these men seriously.
When I read this kind of thing, it’s very apparent that this is being driven entirely by spite not insight. He’s just so angry about everything. There are 57 exclamation marks in this article!
There’s substance under the brashness, though. He’s just upset that his reason is contradicting what everyone around him is saying, struggling to cope with the dissonance. Like being gaslighted. It’s a natural reaction, but I agree, annoying to read.
> a theory that widespread adoption of the existing product will experience an uptick
because profit of this can not cover the investment in this industry
adoption of iphone/smartphone/internet brought new products, including those for reproduction and those for consumption
but generative AI is totally different with iPhone, consumers maybe willing to buy a new ai-powered iphone __just like how they bought new iPhones for every 2years before__
> The current landscape imho should be viewed as an R&D race amongst private actors
in fact, it's a CapEx race, you don't need to R&D anything (ofc you must pretent you do)
that's why it's a con
> The AI Bubble will burst “any day now”
"The canary in the coal mine to look at is when Satya Nadella or Sundar or Zuckerberg say, ‘You know that $80bn of capex I said I was going to do? I think I’m going to cut that by two-thirds.’ That’s what you need to look for."
I can’t help but be reminded of Greenspan’s remarks on the housing market in 2006 while reading this comment:
While he was chairman of the central bank through January 2006, Greenspan always denied there was a
bubble in the nationwide U.S. real estate market,
saying only that a certain number of metropolitan real estate markets could
see declines in home values because of a localized run-up in prices. That view of any real estate bubble as
a merely a local phenomenon is a condition he
termed as "froth" in congressional testimony in 2005, as well as in subsequent comments.
The writer is a journalist, runs a media firm and podcasts. His job is to get attention. You get attention by being outrageous. The “AI will kill us all” take is covered by too many people, so he’s taking the “AI is doomed” path. No one is going to engage with a reasonable middle-of-the-road article. He’s got no credibility on this subject, but he knows how to turn attention into dollars. Everyone here keeps falling for it.
As a 40yo software engineer (at Microsoft) with no specific domain expertise other than using genAI for fun and some code completion, this essay/blog post articulates my gut feelings about where we are at, very well.
I am a math PhD student and I already draw some value from recent reasoning models. I strongky believe than in 1-2 years LLMs will become established tool for scientists to help with coding and math. I really don't think you can call this a con...
41 here, working in healthtech… and Devin has committed more code and closed more tickets on my behalf in the past week at my behest than I’ve done on my own in a month.
It’s basically functioning as a team of entry-level junior engineers at this point.
Previously I was having to spend a fair amount of time writing tickets and providing context, but lately I’ve fed all my meeting transcripts and such into an LLM and it interactively creates Jira tickets for me. Each one takes me maybe 30s to read before I confirm them and the assistant creates the actual tickets.
Sure. One task I gave it a couple of days ago was to upgrade the version of Python used in a project. In this case, that was a task suited for a junior engineer - it was simple enough to be described fully, but complex enough to require effort.
Devin was able to recognize that the project used Poetry, was Dockerized, and that the Python version was specific in multiple places (.python-version, pyproject.toml, Dockerfile). It saw that a couple of minor dependencies didn’t support the new version of Python, so it went back and upgraded those to the most recent matching version first.
Devin had never touched the repository in question before getting this task.
I’ve given it more and less complex tasks, and yeah, it struggles with some things. I’d estimate that it consumes about 5-10% of my time but multiples my overall output by ~3x.
It excites me. The only way it would really terrify me is if I were a very junior engineer right now or in college to be one.
I think we’ll see a ton of complaints about how bad the job market is in the next couple of years. That will be true, but only for juniors or for seniors who don’t embrace the tech. For seniors who do embrace it and specialize in implementing these systems, it’ll be a gold mine.
Then, over 5-10 years, our seniors will start to retire or leave the field. No one will be there to replace them. At that point we’ll see a resurgence in the job market.
Things like autocompletion and “chat with your codebase” help juniors more than seniors; agents help seniors much more than juniors. As these systems improve, their failure cases get more and more complex/nuanced - you will always need senior people with the insight necessary to figure out what’s wrong when it breaks. For a while that will help seniors and hurt juniors… right up until businesses realize that they don’t have replacements for their existing senior engineers, at which point they’ll be desperate to hire again.
You have massively misunderstood what I’m terrified by. In fact you’ve described something I find the least terrifying of anything I’ve ever read because it’s all pure fantasy.
I would be very curious about the size and complexity of this codebase. Every review of Devin I’ve seen has been very negative (burns a ton of money, gets stuck, doesn’t implement the changes you want).
For large codebases (greater than 15k or 20k LOC) the context size seems like a real problem right now.
I’ve used it for everything from “change this text on a webpage” to squashing complex migrations in multiple apps in a Django monolith where migrations in one app depends on migrations in other apps.
My apologies if anyone finds this offensive, but I sorta see Devin as a fresh junior SWE hire. It doesn’t do well with tasks that require deep knowledge sometimes, but it has shallow or better knowledge of everything. I would describe it as working with a brand new SWE with an IQ of about 85 who is also on the low end of being high-functioning autistic. By that I mean that it takes most things literally and sometimes has difficulty with nuance.
> burns a ton of money, gets stuck, doesn’t implement the changes you want
The first time you use it, I think that’s pretty fair. Every time it gets stuck or does the wrong thing, when you correct it, it gives you the option to add to its “knowledge base”. That’s a bunch of additional context that it applies in only certain situations. Within a week or so of using it regularly, it’s significantly more valuable. It “learns” much faster than a human.
Example:
About a dozen of our projects all rely on a shared repository (“Enki”) that contains a Composefile, configs, and some light automation. Tests are run in Docker, and you have to navigate to the other repo’s directory to bring up the service. Some of those projects have service names in the Composefile that differ from the project name. I was able to run the steps interactively on “Devin’s machine”, tell Devin what I had done, and then tell it that this is the correct approach for any project that depends on that repository. I didn’t tell it what projects those are, or how to find out.
The next time I used Devin on a project like that, it tried to run the tests directly in a local Python environment. That didn’t work, but it tried the correct approach next. That worked, so it added a line to its knowledge base “Project <foo> uses Enki.” From that point forward it did the right thing the first time.
> For large codebases (greater than 15k or 20k LOC) the context size seems like a real problem right now.
The primary project I’m working on is a Django app. I don’t have it in front of me right now, but it’s about five years old, has been under very active development the entire time, and is comprised of about twenty apps. It’s not the largest codebase I’ve worked on, but it’s far from the smallest. I can do a line count tomorrow if you’d like.
> When you put aside the hype and anecdotes, generative AI has languished in the same place, even in my kindest estimations, for several months, though it's really been years. The one "big thing" that they've been able to do is to use "reasoning" to make the Large Language Models "think" [...]
This is missing the most interesting changes in generative AI space over the last 18 months:
- Multi-modal: LLMs can consume images, audio and (to an extent) video now. This is a huge improvement on the text-only models of 2023 - it opens up so many new applications for this tech. I use both image and audio models (ChatGPT Advanced Voice) on a daily basis.
- Context lengths. GPT-4 could handle 8,000 tokens. Today's leading models are almost all 100,000+ and the largest handle 1 or 2 million tokens. Again, this makes them far more useful.
- Cost. The good models today are 100x cheaper than the GPT-3 era models and massively more capable.
The "iPhone moment" gets used a lot, but maybe it's more analogous to the early internet: we have the basics, but we're still learning what we can do with this new protocol and building the infrastructure around it to be truly useful. And as you've pointed out, our "bandwidth" is increasing exponentially at the same time.
If nothing else, my workflows as a software developer have changed significantly in these past two years with just what's available today, and there is so much work going into making that workflow far more productive.
But if this is like the internet, it’s not refuting the idea that this is a huge bubble. The internet did have a massive investment bubble.
And I’d argue it took decades to actually achieve some of the things we were promised in the early days of the internet. Some have still not come to fruition (the tech behind end to end encrypted emails was developed decades ago, yet email as most people use it is still ridiculously primitive and janky)
Can it be an investment bubble but also a hugely promising technology? The FOMO-frothing herd will over-invest in whatever is new and shiny, regardless of its merits?
While there was certainly a software bubble during the early internet, it still took obscene amounts of investments in brand new technologies in the late 90's. Entire datacenters full of hardware modems. In fact, 'datacenters' had to become a thing.
Then came DSL, then came cable, then came fiber. Countless billions of dollars invested into all these different systems.
This AI stuff is something else. Lots of hardware investment, sure, but also lots of software investment. It is becoming so good and so cheap its showing up on every single search engine result.
Anyway, my point is, while there may have been aspects of the early internet being a bubble, there were real dollars chasing real utility, and I think AI is quite similar in that regard.
Yes. But this article argues two things at once - that the technology is itself not useful and not used, and that this won't change in the future. And it also argues that this is a bad investment, at least in the form of OpenAI.
I have very little idea of the second - it's totally possible OpenAI is a bad investment. I think this article is massively wrong about the first part though - this is an incredible technology, and this should be evident to everyone (I'm a little shocked we're still having an argument of the form "I'm a world-class developer and this increases my productivity" vs. "no, you're wrong!" on the other).
> If nothing else, my workflows as a software developer have changed significantly in these past two years with just what's available today, and there is so much work going into making that workflow far more productive.
this is exactly the problem
The more productivity AI brings to workers, the fewer employees employers need to hire, the less salary employers need to pay, and the less money workers have for consumption.
I think the answer is that open source made developers more valuable because they could build you a whole lot more functionality for the same amount of effort thanks to not having to constantly reinvent every wheel that they needed.
More effective developers results in more demand for custom software, resulting in more jobs for developers.
My hope is that AI-assisted programming will have similar results.
I don't really know myself, but I think there's a decent change that most developer jobs will actually disappear. Your argument isn't wrong, but when we're nearing (though still far from) the state where all productive tech work can be handled by LLMs. Once it can effectively and correctly fix bugs and add new well-defined features to a real codebase, things start to look very different for most developers.
> This is missing the most interesting changes in generative AI space over the last 18 months
I agree, though personally I'm liking the "big thing" as well. R1 is able to one-shot a lot of work for me, churning away in the background while I do other things.
> Multi-modal
IMO this is still early days and less reliable. What are some of your daily use cases?
> Context lengths
This is the biggest thing IMO (Models remaining coherent at > 32k contexts)
And whatever improvements have caused models like Qwen2.5 to be able to write valid code reliably vs the GPT-4 and earlier days.
There are a whole lot of useful smaller niche projects HF like extracting vocals/drums/piano from music, etc
Multi-modal audio is great. I talk to ChatGPT when I'm cooking or walking the dog.
For images I use it for things like helping draft initial alt text for images, extracting tables from screenshots, translating photos of signs in languages I don't speak - and then really fun stuff like "invent a recipe to recreate this plate of food" or "my CSS renders like this, what should I change?" or "How do you think I turn on this oven?" (in an Airbnb).
I've recently started using the share-screen feature provided for Gemini by https://aistudio.google.com/live when I'm reading academic papers and I want help understanding the math. I can say "What does this symbol with the squiggle above it?" out loud and Gemini will explain it for me - works really well.
Just last night I was digging around in my basement, pulling apart my furnace, showing pics of the inside of it, having GPT explain how it works and what I needed to do to fix it.
If there are no reputable sources to point to, then where exactly is GPT deriving its answer from? And how can we be assured GPT is correct about the furnace in question?
I mean.. I fed it all the photos of the unit and every diagram and instruction panels from the thing. I was confident in the information it was giving me about what parts did what and where to look and what to look for. You have to evaluate its output, certainly.
Getting it to fix a mower now. It's surfacing some good YouTube vids.
I use it like that all the time. There's so much information in the world which assumes you have a certain level of understanding already - you can decipher the jargon terms it uses, you can fill in the blanks when it doesn't provide enough detail.
I don't have 100% of the "common sense" knowledge about every field, but good LLMs probably have ~80% of that "common sense" baked in. Which makes them better at interpreting incomplete information than I am.
A couple of examples: a post on some investment forum mentions DCA. A cooking recipe tells me "boil the pasta until done".
I absolutely buy that feeding in a few photos of dusty half-complete manual pages found near my water heater would provide enough context for it to answer questions usefully.
Oh right, yeah I've done things like this (phone calls to ChatGPT) or the openwebui Whisper -> LLM -> TTS setup. I thought there might be something more than this by now
As a society we choose to let the excess wealth pile up into the hands of people that are investing to bring about their own utopia.
If we're stretching, we can talk about opportunity cost. But the people spending and creating the "bubble" don't have better opportunities. They're not nations that see a ROI on things like transportation infrastructure or literacy.
So unless the discussion is taken more broadly and higher taxes are on the table, there really isn't a cost or subsidy imo.
This. IIUC to serve an LLM is to perform an O(n^2) computation on the model weights for every single character of user input. These models are 40+GB so that means I need to provision about 40GB RAM per concurrent user and perform hundreds of TB worth of computations per query.
How much would I have to charge for this? Are there any products where the users would actually get enough value out of it to pay what it costs?
Compare to the cost of a user session in a normal database backed web app. Even if that session fans out thousands of backend RPCs across a hundred services, each of those calls executes in milliseconds and requires only a fraction of the LLM's RAM. So I can support thousands of concurrent users per node instead of one.
> IIUC to serve an LLM is to perform an O(n^2) computation on the model weights for every single character of user input.
The computations are not O(n^2) in terms of model weights (parameters), but linear. If it were quadratic, the number would be ludicrously large. Like, "it'll take thousands of years to process a single token" large.
(The classic transformers are quadratic on the context length, but that's a much smaller number. And it seems pretty obvious from the increases in context lengths that this is no longer the case in frontier models.)
> These models are 40+GB so that means I need to provision about 40GB RAM per concurrent user
The parameters are static, not mutated during the query. That memory can be shared between the concurrent users. The non-shared per-query memory usage is vastly smaller.
> How much would I have to charge for this?
Empirically, as little as 0.00001 cents per token.
For context, the Bing search API costs 2.5 cents per query.
Interesting. There's obviously been a precipitous drop in the sticker price, but has there really been a concomitant efficiency increase? It's hard to believe the sticker price these companies are charging has anything to do with reality given how massively they're subsidized (free Azure compute, billions upon billions in cash, etc). Is this efficiency trend real? Do you know of any data demonstrating it?
People are comparing the current rush to the investments made in the early days of the internet while (purposely?) forgetting how expensive access to it was back then. Not saying that AI companies should make a profit today, but I don't see or hear that AI usage is becoming essentials in any way or form.
Yeah that's the big problem. The Internet (e-commerce, specifically) is an obviously good idea. Technologies which facilitate it are profitable because they participate in an ecosystem which is self sustaining. Brick and mortar businesses have something to gain by investing in an online presence. As far as I can tell, there's nothing similar with AI. The speech to text technology in my phone that I'm using right now to write this post is cool but it's not a killer app in the same way that an online shopping cart is.
I have personal anecdotal evidence that they're getting more efficient: I've had the same 64GB M2 laptop for three years now. Back in March 2023 it could just about run LLaMA 1, a rubbish model. Today I'm running Mistral Small 3 on the same hardware and it's giving me a March-2023-GPT-4-era experience and using just 12GB of RAM.
People who I trust in this space have consistently and credibly talked about these constant efficiency gains. I don't think this is a case of selling compute for less than it costs to run.
I've heard from insiders that AWS Nova and Google Gemini - both incredibly cheap - are still charging more for inference than they spend on the server costs to run a query. Since those are among the cheapest models I expect this is true of OpenAI and Anthropic as well.
The subsidies are going to the training costs. I don't know if any model is running at a profit once training/research costs are included.
Not very well in my experience. Last time I checked ChatGPT/DALL-E couldn't understand the its own output to know that what it had drawn was incorrect. Nor could it correct mistakes that were pointed out to it.
For example, I ask it to draw an image of a bike with rim brakes it could not, nor could it "see" that what was wrong with the brakes that it had drawn. For all intents and purposes it was just remixing the images it had been trained on without much understanding.
Generating images and consuming images are very different challenges, which for most models use entirely different systems (ChatGPT constructs prompts to DALL-E for example: https://simonwillison.net/2023/Oct/26/add-a-walrus/ )
Evaluating vision LLMs on their ability to improve their own generation of images doesn't make sense to me. That's why I enjoy torturing new models with my pelican on a bicycle SVG benchmark!
There's a little grain of salt with respect to context lengths: the number has grown, but performance seems to degrade with larger context windows.
Anecdote:
I often front-load a bunch of package.jsons from a monorepo when making tooling / CI focused changes. Even 10 or 20k tokens in, Claude says things like "we should look at the contents of somepackage/package.json to check the specifics of the `dev` script."
But its already in the context window! Given the reminder (not reloading it, just saying "its in there"), Claude makes the inference it needs for the immediate problem.
This seems to approximate a 'working memory' for the assistant or models themselves. Curious whether the model is imposing this on the assistant as part of its schema for simulating a thoughtful (but fallible) agent, or if the model itself has the limitation.
Technology is largely a function of human imagination, tempered by constraints imposed by time on the one hand and amplified by new discoveries unfolding on the other. Imagination being a fuel, what does the imagination project is possible in 100 years?
Why in the world would you want a car?? They are horrible to maintain, difficult to operate, expensive, smell bad, and are slower than horses! -somebody salty about automobiles, circa 1900.
Tbh they'd have been right. Cars are expensive to buy and to maintain. They are difficult to operate—people die every day. They do smell bad, and the particulate matter they emit is extremely bad for you. And they are slow: Traffic is an unsolvable problem in urban centres.
Redesigning our cities around cars was one of the big mistakes of the 20th century.
Horses weren’t cheap when they were a primary mode of transportation. Lots of people have died riding, driving, and breaking horses. They definitely smell bad, and their “particulate matter” was so bad that houses had to be set back and elevated from the street.
Cities designed around cars are far superior to cities designed around horses.
No, they were not. Trains are good for taking large numbers of people from one place where they don't want to be to another place where they don't want to be. In certain situations that can be a useful thing to do, but you can't design a city around it. In every case you have in mind, rest assured, the city was there first.
??? No, they don't. They only smell bad when they're kept standing in their urine, which is still not worse than a hairdresser. Compared to dogs (or ICE cars) horses smell way less. They do sweat to regulate temperature, which has a distinct smell, but it's way less irritating to a human's nose than the sweat of the rider.
There's a set of distinct smells associated with the horses, but other than the piss, none of them are particularly "bad". In my experience, humans tend to smell way worse overall (from food to body odor to the excrement) than horses.
Cars replaced streetcars, not horses. Few people in big cities were getting on horseback to commute to work in the morning in the early 20th century. People generally walked or took the streetcar.
> They are difficult to operate—people die every day. They do smell bad, and the particulate matter they emit is extremely bad for you. And they are slow
Sarcastically, one might say that horses are even harder to operate (they have minds of their own), they smell worse that automobiles (esp EVs), and the particulate matter they excrete would be unhealthy to consume. They are also very slow.
More seriously, the trajectory that our imagination pushes towards seems to be “overcoming” biological limitations. Perhaps a symbiosis of machine and biology and consciousness will take us to the next level by opening up vast new universe state.
>Redesigning our cities around cars was one of the big mistakes of the 20th century.
Cities have been designed around Carriages for millenia. You can go and walk in Pompeii and observe pavements for pedestrians, roads for wheeled carriages with crossing spots of elevated stones for pedestrians.
It turns out that cities require a lot of goods to be moved through - more than a pedestrian can carry, and over inclines that human muscle power doesn't like.
The reason why cities are designed around cars is that cars were designed to fit in contemporary cities and they co-evolved over the 20th century. It was the slow kind of evolution, with each step being easier and cheaper than the big redesign.
> The reason why cities are designed around cars is that cars were designed to fit in contemporary cities and they co-evolved over the 20th century
No, the switch to car-centric infrastructure was a deliberate policy choice lobbied for by the automotive industry. [1] We ripped up a lot of good transit to lay down roads this wide and fast.
> It turns out that cities require a lot of goods to be moved through - more than a pedestrian can carry, and over inclines that human muscle power doesn't like.
Hence the utility of public transit, which kills substantially fewer people and is much cheaper. Though goods are mostly moved with trucks, and trucks aren't my concern. Urban congestion isn't caused by 18-wheelers.
The more pertinant observation is that cars are a great tool for mobility, but going _all in_ on cars causes a whole bunch of issues at society scale. If you zone cities and design infrastructure with the assumption that everyone drives, it forces everybody into the least space efficient mode of transport. You have to designate huge amounts of valuable real estate to keeping all those cars somewhere. People who _can't_ drive will have much more difficulty navigating life. When cars arrived, some parts of the world made sure their cities were still easy to navigate by foot, bike and transit, and I'd argue they're more pleasant places to be.
The point isn't to say that new tech is bad, but that there can be adverse consequences to jumping in wholesale.
People like the horse joke, but it works with computers in general. The best argument in favor of computers in the 80s was to save paper, otherwise they were expensive and overcomplicated.
They'd be right too soon. The ICE has been a very useful technology but car dependency has been a nightmare for society. Forget a bicycle for the mind, maybe LLMs are like an SUV for the mind.
> How does this industry actually continue? Do OpenAI and Anthropic continue to raise tens of billions of dollars every six months until they work this out?
It's been made fairly clear that the insiders are setting the stage for governments to back them (bail them out).
Nah you don't to know to build the same thing precisely. Just the other day I wanted to write a vanilla JS component that could let you select a picture from something like a carousel and be able to blow up the selected picture when clicked. I know JS / HTML but am not used to working with vanilla JS. Copilot didn't write it all by itself but it did teach me things I didn't know like making a custom tag in vanilla JS by extending an HTMLElement.
The code isn't the most readable because I don't need it to be however if you make me write it from scratch in an interview style setting I'd have trouble doing it. If I read the code I can follow it and it makes sense + it's an easy component to manually test. So.. no, I don't need to know how to precisely build the same thing.
And before you worry that I'm committing code I can't build from scratch.. This is a simple component for a 5 page landing page build with astro where I'm the "main" dev ( wrote like 80% of the code). The web-page won't even need maintainance once it's deployed
It’s not a common use case though, dipping into an unfamiliar tech stack only to dip out after committing the code. Typically, when you learn a new stack (eg. for a job), you’ll be living in it for at least a few months, and at that point, you’d be better served by perusing the docs and getting deeply familiar with the API.
The copilots get you going quick at the expense of your learning, which is great for one-offs, but not lasting work quality.
I believe if you spend months / hundreds of hours using any framework / tool you will eventually read so much of it that it's easy to get "deeply familiar" and even then.. it's often times faster to have an LLM write 80-90% of the code for you and just refine / finish it.
I think autocomplete alone would be enough to make coding a killer app for AI.
I agree the tools are overhyped for allowing non-developers to write code. It’s not (today) a replacement for a dev agency that takes a set of requirements and runs with it, it’s a replacement for a junior developer who you need to micromanage a bit. But that’s still a boon to productivity!
by giving the models durable memory, they become "agentic". in any case, they don't make as much of a mess when their output is getting written into git.
I keep on trying them, but if they are useful they are useful for only a small fraction of engineers at the moment. I'm not sure if this is due to the nature of the work, or the nature of the user.
I have heard "top" engineers at various places say it makes them 2x faster, or whatever, but I would like to see this assessed by timed testing, as is sometimes done for evaluating software engineering.
Copilot may let me type less, but I have not seen the wall clock effects, which is a very hard thing to measure (time perception is very unreliable).
The vast majority of software work is not greenfielding a PoC or reimplementing an existing, small, well-specced project. We’ve had OpenAPI client generators for years after all.
The majority of software work is maintaining large, existing products: adding features, fixing bugs, improving performance, etc., or building new software in problem domains that aren’t so well-defined.
In my opinion, AI only really helps you (a lot) if you are bottlenecked by the actual code-writin. I have not been in a such a position since... I dont even know. Maybe In my 20s, 15 or so years ago? Even if AI wrote my code 100x faster it would not appreciably change my working days.
If it could test and verify things though... ideally physically since Im in embedded and pulling SD-cards etc is a thing.
Yeah I’m in the same boat with it. I do keep trying, but so far it has been far from earth-shattering.
I’d love it if I could get it to write decent unit tests given a basic description of what I’m testing but I at least cannot get it to output anything useful for our codebase. It’s too unaware of the broader context of the code, what objects need to be instantiated from some other internal library and passed in, etc. It can do a decent job if all I have is a totally isolated function that doesn’t touch the rest of the codebase or use any domain objects, but that’s a rare enough case as to be essentially useless to me.
I think it also really accelerates learning of a new language or framework, when that language or framework is really well documented on the web. For novel programming frameworks, obviously it's a bit more challenging to get help from an LLM.
One of more recent attempts at using LLM code assist was to try to fix a bug in a Swift SSH Agent's connection handling that was causing hangs. I know zero Swift, much less the networking frameworks. So I pumped the output of `tree` on the git repo into the LLM, asked for which file likely handled connections, and it found it right away. That's probably 15 minutes saved. Before putting in the file I asked for likely reasons for deadhangs, got that list, then put in the Swift file that handled connections, and it pointed to what the likely problem was. That's probably 1 hour+ of reading documentation to try to figure out what the code was doing wrong with the networking framework, assuming the LLM was not hallucinating. And that "not hallucinating" probability is high enough in my experience that I spend >50% of my time trying to verify I'm not getting bullshitted.
The LLM proposed a fix (~10-20 minute savings), but even as somebody who doesn't use Swift it seemed like >99% chance that it had just introduced a bunch of race conditions in the data structures it used to track connection status. So I asked about it, and it said "Oh yeah of course how could I forget" and then significantly complicated the solution with something that I thought looked like it probably worked. But was the LLM just being obsequious or was it correct the first time? So hard to tell...
So in about 20 minutes I probably accomplished in a language I didn't know, in a code base I didn't know, about 2 hours+ of learning.
But if I knew the language, it would have saved me very little time, and may have cost me some time.
I agree it's impressive and stuff, but I wouldn't consider a JS POC as a serious project. I have never done that in my whole life and would rather see results from a 10 years old application with a million lines of code of C++. That's would be realistic. What you did is refactoring a pet project and I don't know why we're wasting $billions for that.
> I would strongly argue that coding assistants are AI’s first killer app. Copilot, Cursor, Windsurf etc.
These IMO are relatively useful things. But probably (in their current state) will not justify the valuation of the companies involved and the massive investment occurring right now.
I don't know how the future will unfold. I do think it is reasonable to be somewhat bearish on what has been promised vs. what has been released.
The killer app is entertainment. Since LLMs emerged people have consistently loved getting them to say whatever they want or roleplay with them. Once integrated into games, it will be very fun to have natural conversations with the inhabitants of game worlds. Imagine a goomba talking to Super Mario before he stomps its head or delivering your pithy one liner in response to some final boss’s villainous monologue.
and strangely in two years no one has demonstrated anything like this that people found of value. In fact for all the "entertainment" sectors it has been injected into, we have gotten poisoning of self-publishing and soulless generic pornography. Remember the twitch channels that had AI generated content? Where did those go? Surely by your rationale the market would have taken over by now. Surely there would be something.
It's almost like entertainment requires some humanity and thought and true creativity behind it.
The primary reason is because LLMs are expensive computationally and financially.
In an open world game, it’s trivial to assign memories and facts an AI learns about its world from interactions or in response to game events. All an LLM has to do is be fine tuned to take data from that internal knowledge base and express it as natural language text, in order to have intelligent and useful conversations with a player. It’s not difficult.
This only works if your game doesn't use the GPU. Then there's the whole problem with nondeterminism. But I'm sure when those small problems are solved people will use this technology in games /s
Say you reduced their revenue to only that application. Would it be sustainable? Would it be worth the billions upon billions of dollars that have been shoveled at it? Would it add more than the billions upon billions of dollars in the end?
By your logic I could claim a quantum computer with qubits on the scale of the mass of the sun is a killer app for doing RSA encryption breaking. And I would be making an equally useless statement.
This is moving the goalpost on what "killer app" means. Code assistants are a compelling use of the tech that has quickly shown real-world value, which is the point I'm trying to make here.
Whether the companies that are leading the market today will end up being the ones who capture that value is anyone's bet.
It’s only going to get cheaper over time. It’s already cheap enough that if these services disappeared overnight I’d switch to an open source alternative with a local model. The industry needed VC backing to pay the fixed cost of the research, but the cost of running inference is not insane compared to the volume it provides.
I think the argument this article makes though is that generative AI isn't generating enough value to justify the sky high valuations and investment. Sure, if all of these services disappeared overnight, the remaining users could self host or run a model locally. I feel like that speaks to the lack of value any one company provides though?
Maybe this era of AI will be remembered as a wealth transfer from VCs back to everyday consumers lol
Only insofar as $1 Uber rides were a transfer of VC funding to everyday consumers. When the funding dries up and there's a need for revenue, then we'll see what they charge. Hopefully for them, they've become inextricable from our lives, like Uber has, and we'll gladly pay the new price. Not everyone will be hooked though, but the VC bet is that enough people will be that their horse wins the race and they'll make money.
There are a lot of horses in this race and the literally billion dollar question is who's Amazon this time around, and who's Webvan. Who's Uber and who's Flywheel (which doesn't even have a Wikipedia page anymore,
ouch). Not knowing which horse is going to win doesn't negate the fact that a horse is going to win.
Model available LLMs, like Llama and Deepseek and StableDiffusion are totally a wealth transfer to consumers. Better make use of them!
I’m a little shocked at how much negativity there is around LLMs among developers. It’s a new tool that requires some learning, and it’s sometimes not so great, but if you’ve used an IDE with real coding assistance built in (eg. VS Code in Edit with Copilot mode - NOT Chat mode, using Claude 3.5), it’s honestly not much worse than a junior dev and 100x faster. And if the code is bad you throw it away and try again 10 seconds later. The amount of speed up I see as a very experienced dev is astronomical. And just like 6 months ago it was awful. How great is it gonna be in a year or two? It doesn’t even have access to running unit tests or reading console errors or IDE hints, and it still generates mostly correct code. Once it gets more deeply embedded it’s just going to improve more and more.
Not totally. But you might be surprised at the things you can do. Cursor has some template-like files where you can basically teach the AI “when we do X, do it this way.” Or you can change the global prompt to add the things it should keep in mind when working with you.
If you actually take the time to tell it “hey, don’t do it this way,” it can definitely do it differently the next time.
On top of that, is anyone training models on their own codebase, and noting to AI which patterns are best practice and which aren’t?
There are a ton of ways to make it better than the baseline copilot experience
If the old metric is right, that it is ten times harder to debug code than to write it, having something that writes buggy code 100x faster than you can understand it is a problem.
Especially given that you can ask an LLM to optimise code and on multiple runs it can not tell if it's is improving or degenerating.
I think the usefulness is just very domain specific. If you're writing some types of boilerplate or often-tutorialized code it can spit out something very reasonable. Other types of code, like say in game dev, it stumbles around and never produces anything usable.
But like you said, in a few more years we'll see! It does feel like there's some missing pieces yet to be figured out to truly "reason" and generalize.
> If you're writing some types of boilerplate or often-tutorialized code it can spit out something very reasonable. Other types of code, like say in game dev, it stumbles around and never produces anything usable.
This makes me think of a quirk I discovered recently which is that ChatGPT simply won't generate a picture of a 'full glass of wine'. It generates pictures with all sorts of crazy waves/splashes in the glass but the glass is always half full no matter how you prompt it.
I'm not enough of an expert to make any deductions from this, but I think it hints at what the limitations of the currently models are.
Yeah I'm surprised by all the negativity as well. I'm listening to the post right now (using xtts-v2 finetuned on a voice I like lol). Sounds like these companies are overvalued / over hyped. Maybe they are / some of these companies go the way of myspace, but LLMs are incredibly useful for me.
I'm able to do a so much more using LLMs (Mistral-Large, Qwen2.5 and R1 locally, Claude via API) than without them.
Personally, I've found DeepSeek R1 to be a profoundly good model for thinking through problems across fields.
I had a complex finance situation that I was struggling with, both from a mathematical/taxation perspective and a personal psychological finance hangup. I spent a few good hours talking to it through everything and had a mental breakthrough. To get the same kind of insight, I would have to pay a financial advisor AND a psychologist for several hours.
That all of this was free while someone calls it a "con" seems completely wrong
(I got my CFA cousin to look over the numbers and he agreed with R1's advice, fwiw)
Yeah, I've had similar experiences. I still hesitate if it's a field I don't know too well of course (never trust an LLM), but R1 has been able to solve things I've been stuck on. And watching it's <think></think> process has been insightful. Only issue is that it ties up all my GPUs while I run it.
Hopefully Mistral can copy their technique and give us a 123b reasoning model.
LLMs are pretty good at the aspects of coding that I consider to be "the fun part". Using them has made me more productive, but also made my job less fun, because I can't justify spending time using my own brain to do "the fun part" on my employer's dime. And that was something I was particularly good at, which is why I was able to be paid well to do it.
So now my company makes more money, and the work gets done faster, but I can't say I feel appreciative. I'm sure it's great for founders though, for whom doing the work is merely an obstacle to having the finished product. For me, the work is the end goal, because I'm not hired to own the result.
It's kind of analogous to the old taxi drivers who took pride in having a sixth sense knowing which route to take you, vs uber drivers who just blindly follow their navigation
Some of them might have had a really good mental map; but the majority would just take inefficient routes (and charge you some random price that they put into their counter) — plenty of reasons to dislike Uber but having a pre-set price, vetted/rated drivers, and clear routing for a taxi service is a massive plus in my opinion.
Looks like you haven't used a decent IDE: these things have been standard for decades, locally and with minimal requirements. But wait, now it happens in the Cloud (meh, that's not gonna fly anymore, too last decade)...AND requires massive amounts of power AND cooling, PLUS it's FUBAR about 50/50.
For an incremental improvement...not great, not terrible.
I think LLMs are vastly overhyped and mostly useless, but I use Copilot as glorified autocomplete and like it.
It does what the other poster said: it automates the boring parts of "this db model has eight fields that are mostly what you expect" and it autocompletes them mostly accurately.
You're really comparing an IDE's autocomplete with something that can, at minimum, write out entire functions for you?
You're either completely misremembering what IDEs have been able to do up until 3 years ago, or completely misunderstanding what is available now. Even the very basic "autocomplete" functionality of IDEs is meaningfully better now.
Bit of a boomer statement here but maybe this will encourage devs such as yourself to contribute more to open source passion projects that will help dethrone the monopolies. Looking at Valve's investment into Linux via Proton as a great example.
It would be so nice to have a productivity Linux OS that just works on all my devices without tinkering. I want to stop supporting the closed source monopolies, but the alternatives aren't up to par yet. I am extremely hopeful that they will be once mega corps inevitably decay and people tire of the boom-bust cycle.
As technologists, we all want beautifully designed tools, and I'm increasingly seeing that these are only created by passionate and talented people who truly care about tech, unlike megacorps that only care about enriching their board and elite shareholders.
Honestyl. It has its drawbacks but I am usually at 50x with few different agents running side by side. What we need is better GPU competition with tons of ram.
Doesn't that just scream "bad design" at you? Shouldn't we be aiming for agents that require less GPU? And agents that are good enough that we don't have to shop around for "competing prices" on answers?
For easy things, LLM assist has sped things up a lot for me.
For medium complexity things, I can get them done quickly without manual coding if I have a clear understanding in mind of what the implementation should look like. I supply the requirements, design and strategy and it's fairly easy to "keep things on the rails". The "write a PRD first" hack (https://www.aiagentshub.net/blog/how-to-10x-your-development...) works pretty well. Agent with YOLO mode and terminal access rips, particularly if you have good tests.
For tasks where I know the spec of the feature but don't clearly understand how I would design / implement the feature myself, it's hit-and-miss. Mostly miss for me.
I also haven't had much success with niche libraries, have to stick to the most popular library/tool/framework choices or it will quickly get stuck.
Whereas I've been disabling AI assist features because I find them actively disruptive to the development process. When it ghost pops up text suggesting what I should do, it's sometimes right...but it breaks flow. It forces me to read and parse apparently correct code, and decide if it is correct or it's just a mirage which is valid but not actually what I'm doing at all.
This is the crux. A cool thing has been invented, with real usages. Unfortunately, it's cost hundreds of billions of dollars and it has absolutely zero hope of making the trillions needed to justify that.
Now someone will respond about how it's just a stepping stone, and how the billions are justified by _something completely imaginary, and not invented yet, and maybe not ever_ e.g. agents.
>it's cost hundreds of billions of dollars and it has absolutely zero hope of making the trillions needed to justify that.
The BigTech companies have been flush with liquidity and poured those hundreds of billions into the promising tech, and as result we got a wonderful new technology. There is not much need for those trillions in return - just look at liquidity positions of those companies, they are just fine. If those trillions come in eventually - even better.
>There is not much need for those trillions in return
Whilst you are correct that big tech cos do not need the return to survive, that's not how public markets work at all, and thus not how the incentives for those in charge of the companies work, and so making you actually wrong.
If i were wrong, those companies would be distributing that cash to shareholders instead of chasing any promise of any big chance.
If investment in AI don't pan out (i do think that it will pan out, and those trillions will come) then those companies would just pour even more billions into whatever big thing/promise would come next. Rinse and repeat. Because some of those things do generate tremendous returns, and thus not playing that game is what really constitute true loss of money.
US right now is run by someone whose explicit promises, if actually implemented, have an obvious immedidiate 13-14% reduction in GDP — literally, never mind side effects, I'm not counting any businesses losing confidence in the idea that America is a place to invest, this is just direct impact.
DOGE + deportation by themselves do most of that percentage. The tariffs are a rounding error in comparison, but still bad on the kind of scale that gets normal politicians kicked out.
cash on hands GOOG - 100B, AMZN - 80B, FB - 70B, and their core businesses are basically printing money, so they pretty much do have to invest into new things. If somebody sees a multi-billion dollar sink better than AI right now ...
Nobody has ever been punished for choosing IBM. It’s the same story here. Nobody is going to blame them for following the zeitgeist, but you bet they’d be punished if they didn’t and it doesn’t pan out.
The whole thing is like bitcoin. There’s too many people that benefit from maintaining the collective illusion.
Ideally it was thought to have shortened a very expensive war, and may have prevented the USSR from taking over Europe by leveraging its unquestioned postwar conventional forces advantage.
The profit was made by the private sector in supplying goods to the program. Today, private companies do a lot and earn a lot of money from stockpile maintenance.
I don't know why it is so hard to understand. I mean money doesn't really exist without a government[0] and while government plays a role in the market and economy, this role is VERY different than that of a business. A government isn't trying to "make money", is isn't trying to make investors happy, and it certainty can't take existential risks that could make "the company" go bankrupt (or it shouldn't lol).
But I do think (and better understand) there is a failure to understand this at a higher abstraction. One part is simply "money is a proxy." This is an uncontestable fact. But one must ask "proxy for what?" and I think people only accept the naive simple answer. Unfortunately, this "is a proxy" concept is extremely generalization. Everything is an estimation, everything is an approximation, and most things are realistically intractable. We use sibling problems or similar problems to work with that are concrete, but there are always assumptions made and ignoring these can have disastrous consequences. Approximations are good (they're necessary even) but the more advanced a {topic,field,civilization,etc} gets, the more important it is to include higher order terms. Frankly, I don't think humans were built for that (though by some miracle we have the capacity to deal with it).
My partner and her dad are both economists, and one thing I've learned is that what many people think are "economics questions" are actually "business questions". I think a story from her dad makes this extremely clear. A government agency hired him to look at the cost benefit analysis of some stuff (like building a few hospitals and some other unambiguously beneficial institutions), and when he presented everyone was happy but had a final question "should we build them?" The answer? "That's not the role of an economist." The reason for this is because money can't actually be accurately attributed to these things. You can project monetary costs for construction, staffing, and bills, and you can make projections about how many people this will benefit, how it can reduce burdens elsewhere, and as well as make /some/ projections about potential cost savings. But you can't answer "should you." Because the weight of these values is not something that can be codified with any data. It is an importance determined by the public and more realistically their representatives. Very few times can you give a strong answer to a question like "should we build a new hospital" and essentially in only the extreme cases. I'll give another example. In my town there was an ER that was closed due to budget constraints. This ER was across the street to the local university, which students represent ~15% of the population. The next nearest ER? A 15 minute ambulance ride away and in the next town over. Did the city save money? Yes. Did the sister city's ER become even busier? Also yes. Did people lose access to medicine? Yes. Did people die? Also yes. Have economists put a price on human life? Also yes, but they are very clear that this is not a real life and a very naive assumptions[1]. It is helpful in the same way drawing random squiggles on a board can help a conversation. Any squiggles can really be drawn but the existence of _something_ helps create some point to start from.
[0] okay crypto bros, you're not wrong but low volatility is critical as well as some other aspects. Let's not get off topic
Doesn't matter. The Manhattan project was a breakthrough in fundamental science that changed the world. Current generative AI are a solid degree improvement on previous technology that is not remotely as big a leap as the amount of money poured into it assumes it to be.
The Manhattan Project was driven by the U.S. Government, which doesn't need a VC-tier return. The entire business model of VCs is based on the idea that they'll have the occasional 100x return, and if none of the AI companies do that it would destroy the VC model.
> I use Cursor sometimes, and VSCode + Continue with llama.cpp, and it's great. That's not worth billions. It's definitely not worth trillions.
That seems like a suspect claim. If you're saying that you, personally, cannot create billions of dollars in value with Cursor & friends that is certainly true - but you are in no position to make a judgement call about where the cap on value creation is for the LLM market is worth based on your personal use cases. LLMs don't just do code completion. We really can't estimate how much potential value is being created without doing some serious data diving and studying of cases.
A better argument would be that the DeepSeek experience suggests these companies have no moat and therefore no way to earn a return on capital. But LLMs are probably going to generate at least trillions of dollars in value because they're on par or ahead of Wikipedia and Google for answering many queries then they also have hundreds of ancillary uses like answering medical questions at weird hours or creative/professional writing.
It's possible to grow an economy by trillions of real value without any actor being able to extract that as a profit or it even showing up in the books as money.
Consider that Wikipedia is much bigger than Encyclopedia Britanica, but because it is given away to everyone for free, it is not counted as E.B.'s max sale price ($2900 in 1989?) times the world's internet connected population (5.6e9?) — $16 trillion.
AI, regardless of value, are priced at the marginal cost to reproduce weights or run inference depending on which you care about.
But I do mean "reproduce" not "invent" — it doesn't matter if DeepSeek's "a few million" was only possible because they benefited from published research, it just matters that they could.
And if the hardware is the bottleneck for inference, that profit goes to the hardware manufacturer, not to the top ten companies who made models.
Where are you getting this from? Outside of o3, every AI provider's API is super cheap, with most productive queries I do coming in under 2c. We have no reason to believe any of them are selling API requests at a loss. I think <2c per query hardly counts as "quite expensive".
The reasoning people have for them selling API requests at a loss is simply their financial statements. Anthropic burned $3B this year. ChatGPT lost $5B. Microsoft has spent $19B on AI and Google has spent close to $50B. Given that revenue for the market leader ChatGPT is $3.7B, it's safe to say that they're losing massive amounts of money.
These companies are heavily subsidized by investors and their cloud service providers (like Microsoft and Google) in an attempt to gain market share. It might actually work - but this situation, where a product is sold under cost to drum up usage and build market share, with the intent to gain a monopoly and raise prices later on - is sort of the definition of a bubble, and is exactly how the mobile app bubble, the dot-com bubble, and previous AI bubbles have played out.
Not sure if it matters at this point. There will need to be many more rounds of CapEx to realize the promises that have been put forth about these models.
The implication would be that those API requests are being sold at a loss. Amodei wrote in January that Claude 3.5 Sonnet was trained for only a few $10Ms, but Anthropic has been losing billions.
That would be a killer for the current and near future generations of LLM as a business. If they are having to pay many times in compute what they are able to get for the API use (due to open models being near comparable?), then you definitely can't "make up for it in volume".
How? I get that many devs like using them for writing code. Personally I don't, but maybe someday someone will invent a UX for this that I don't despise, and I could be convinced.
So what? That's a tiny market. Where in the landscape of b2b and b2c software do LLMs actually find market fit? Do you have even one example? All the ideas I've heard so far are either science fiction (just wait any day now we'll be able to...) or just garbage (natural language queries instead of SQL). What is this shit for?
Various minor thing so far. For example I heard about ChatGPT being evaluated as a tool for providing answers for patients in therapy. ChatGPT answers were evaluated as more empathetic, more human and more aligned with guidelines of therapy than answers given by human therapists.
Providing companionship to lonely people is another potential market.
It's not as good as people at solving problems yet but it's already better than humans at bullshiting them.
Are people actually satisfied by that? I personally find "chatting" with an LLM grating and dissatisfying because it often makes very obvious and incongruous errors, and it can't reason. It has no logical abilities at all, really. I think you're really underestimating what a therapist actually does, and what human communication actually is. It's more than word patterns.
I could see this being useful in a "dark pattern" sense, but only if it's incredibly cheap, to increase the cost to the user of engaging with customer support. If you have to argue with the LLM for an hour before being connected to an actual person who can help you, then very few calls will make it to the support staff and you can therefore have a much smaller team. But that only works if you hate your users.
Subjective evaluation of "humanity" and "empathy" in responses is much less important than clinical outcome. I don't think an online chat with a nebulous entity will ever be as beneficial as interactions that can, at least occasionally, be in-person. Especially as the trust of online conversations degrade. Erosion of trust online seems like a major negative consequence of all the generative AI slop (LLM or otherwise).
Clinical outcome of humans doing therapy would be better if for some reason doing therapy worse (less according to taught guidelines) was better. But, sure, we can wait for another research or follow up. It might be true. Therapy has dismal outcomes anyways and the outcomes are mostly independent of which theoretical framework the therapy is done according to. It might be the case that the only value in therapy is human connection that AI fails to simulate. But it seem that for some people it simulates connection pretty well.
Anecdotally, almost every day I’ll overhear conversations at my local coffee shop of non-developers gushing about how much ChatGPT has revolutionized their work: church workers for writing bulletins and sermons, small business owners for writing loan applications or questions about taxes, writers using it for proofreading, etc. And this is small town Colorado.
Not since the advent of Google have I heard people rave so much about the usefulness of a new technology.
These are not the sort of uses we need to make this thing valuable. To be worthwhile it needs to add value to existing products. Can it do that meaningfully well? If not it's nothing more than a curiosity.
To make money though it just needs to have a large or important audience and a means of convincing people to think, want, or do things that people with money will pay to make people think, want or do.
Can you get enough revenue from ads to pay the cost of serving LLM queries? Has anyone demonstrated this is a viable business yet?
A related question: has anyone figured out how to monetize LLM input? When a user issues a Google search query they're donating extremely valuable data to Google that can be used to target relevant ads to that user. Is anyone doing this successfully with LLM prompt text?
I bet Google is utilizing the value of the LLM input prompts with close to the same efficiency they are monetizing search. I that case, there are two questions -- 1) will LLM overtake search? and 2) can anyone beat Google at monetizing these inputs? I think the answer to both is no. Google already has a wide experience lead monetizing queries. And personally, I'd rather have a search engine that does a better job of excluding spam without having to worry whether or not it's making stuff up. Kagi has a better search than any of the LLMs (except for local results like restaurants/maps).
LLMs are incredible at editing my writing. Every email I write is improved by LLMs. My executive summaries are improved by LLMs. It wont be long until every single office worker is using LLMs as an integral part of their daily stack, people just have to try it and theyll see how useful it is for writing.
Microsoft turned itself into a trillion dollar company off the back of enterprise SAAS products and LLMs are among the most useful.
My company uses them for a fuckton of things that were previously too intractable for static logic to work (because humans are involved).
This is mostly in the realm of augmented customer support (e.g. customer says something, and the support agent immediately gets the summarized answer on their screen)
It’s nothing that can’t be done without, but when the whole problem can be simplified to “write a good prompt” a lot of use cases are suddenly within reach.
It’s a question if they’ll keep it around when they realize it doesn’t always quite work, but at least right now MS is making good money off of it.
> The article is about how the economics of the LLM market is making all tech look bad.
No, it's not. The first half of the article talks about how useless the actual product is, how the only reason we hear about it is because the media loves to talk about it.
I feel like you missed the first third of this article that was quite clear they are not saying there are no uses cases. They are saying there doesn't seem to be an economic model that makes sense.
I was thinking the same but it's not really what the post is about. They talk about there are use cases for LLMs and devs can be benefiting.
What it goes into is how over hyped and over valued these companies are. They've blown through $5bn of compute each in a year and their revenue is abysmal. Microsoft won't report on ai separately, probably because it's abysmal.
I'm positive on LLMs for coding. But I think I have to agree with their assessment. Coding seems like the best area for these tools and what we see now is great. It's probably even worth $10b to the IT industry maybe eventually. But they're not paying for it yet, clearly. And I also think it's just not going to have huge significance outside our industry. The people I rub shoulders with outside of work have not mentioned or asked about it once, which is not necessarily meaningful but it does reveal the limits of hype too.
> I’m a little shocked at how much negativity there is around LLMs among developers.
While the timeline is unclear; it seems likely that LLMs will obsolete precisely the skills that developers use to earn their income. I imagine a lot of them feel rather threatened by the rapid rate of progress.
Pointing out that it is already operating at junior dev quality and rapidly improving is unlikely to quiet the discontent.
I use LLMs in coding. There are Junior Devs in my team.
If you think LLMs operate at "junior dev" capacity you either don't work with junior devs and is just bullshitting your way around here, or you just pick pretty awful junior devs.
LLMs are alright. An okay productivity tool, although its inconsistencies many times nullify productivity gains - By design they often spit out wrong results that look and sound very plausible. A productivity blackhole. Its mistakes are sometimes hard to spot, but pervasive.
Beyond that, if your think that all a dev does is spit out code, and since LLMs can spit out code it can replace devs in some imaginary timeline, you are sorely mistaken. The least part of my work is actually spitting out code, although it is the part I enjoy the most.
I honestly feel way nore threatened by economic downturns and the looming threat of recession. The only way LLMs threaten me is by being a wasteful technology that may precipitate a downturn in tech companies, causing more layoffs, etc nd so forth.
If you define "junior" based mostly on age, then LLM's aren't yet at the level of a good "junior".
If you base it on ability, then an LLM can be be more useful to a good developer than 1 or more less competent "junior" team members (regardless of their age).
Not because it can do all the things like any "junior" can (like make coffee), but because the things it can do on top of what a "junior" can do, more than makes up for it.
The value of developers is not the code they output. It's the mental models they develop of the problem domain and the systems they build. LLMs can output code without developing the mental models.
Code is liability. The knowledge inside developers' heads is the corresponding asset. If you just produce code without the mental models being developed and refined, you're just increasing liability without the counterpart increase in assets.
>> If you think LLMs operate at "junior dev" capacity you either don't work with junior devs and is just bullshitting your way around here, or you just pick pretty awful junior devs.
I’ve hired lots of junior devs, some of them very capable. I’ve been in this industry for more than 15 years. LLMs operate at junior dev capacity, that’s pretty clear to me at this moment.
Yep. There are people who love programming, it's the best part of the work anyhow! And then there are people who come and tell that whatever you do doesn't matter and they are more content on getting a new app by writing a prompt and deploying possibly buggy code. Two different crowds of people.
I'm in a middle. I enjoy Zed and its predictions, I utilize R1 to help me to reason. I do _not_ ever want to stop programming. And I see so often whenever somebody less experienced than me shows me look how Cursor did this with three prompts, can we merge? And the solution is just wrong and doesn't solve the hard issues.
For me the biggest issues are the people who want to see the craft of programming gone. But I do enjoy the tooling.
> it seems likely that LLMs will obsolete precisely the skills that developers use to earn their income
I’m not particularly worried. I think it’s obvious that software engineering is definitely an “intelligence complete” problem. Any system that can do software engineering can solve any problem that requires intelligence. So, either my job is safe or I get to live through the fall of almost all white collar disciplines. There’s not a huge middle ground.
Although perhaps this is just the programmer stereotype of thinking that if someone can code, they can do anything.
> Any system that can do software engineering can solve any problem that requires intelligence. So, either my job is safe or I get to live through the fall of almost all white collar disciplines. There's not a huge middle ground.
How about the middle ground where a human using AI replaces you?
Conversely I'm shocked the negativity hasn't graduated to naked hostility among developers. A group that tends to pride itself on clarity of thought entranced by bullshit generators? A group that tends to pride itself on correctness of work cheerfully adding tools to their workflow that provably fuck up in unpredictable ways and that have to be monitored constantly for just such behavior? Why not hire a few junior devs instead if that's your jam, at least you can train a human towards competence.
That experience is heavily subsidized and is unprofitable for these companies providing it based on what we know. Even with all of the other developers who are also using the same work flow and espousing how great it is. Even with all of the monthly subscribers at various tiers. It has been unprofitable for several years and continues to be unprofitable and will likely remain unprofitable given current trends.
The author spends a good amount of bytes telling us that they don't want to hear this argument even though they expect it.
Perhaps, and the externalities often unaccounted for or hand-waved away.
Even the US Government is getting involved in subsidizing these companies and all of the infrastructure and resources needed to keep it expanding. We can look forward to even more methane power plants, more drilling, more fracking, more noisy data-centres sucking up fresh water from local reserves and increased damage to the environment that will come out of the pocket books of... ?
Update: And for what? "Deep Research"? Apparently it's not that great or world-changing for the costs involved. It seems that the author is tired of the yearly promise that everything is just a year or two away as long as we keep shovelling more money and resources into the furnace.
Inference isn’t that expensive. A single junior dev costs orders of magnitude more than the amount of inference I use. Companies in growth mode don’t have to make money, it’s a land grab right now. But the expense is largely in the R and D. You can build a rig to run full models for 10-20k right? That’s only a month or two of a junior dev’s time, and after that it’s just electricity. And you could have dozens of devs using the same rig as long as they could timeshare. I don’t see where the economics wouldn’t work, it’s just there’s no use in investing in the hardware until we know where AI is going.
Yeah, you can build a rig to run full models for 10-20k... That's a big reason OpenAI might not make it. The whole article is about LLMs not being a viable business.
> it’s honestly not much worse than a junior dev and 100x faster. And if the code is bad you throw it away and try again 10 seconds later. The amount of speed up I see as a very experienced dev is astronomical
Personally, I find that waiting for the code to generate, then reviewing the code carefully, then deciding if I need to rewrite it to be more painful, more error prone, and much slower than writing the code correctly.
Especially since this AI junior never learn from it’s mistakes.
I think it speaks to different approaches to how individuals write code.
> How great is it gonna be in a year or two?
I would bet that it’s about the same (not great code, generally), but the tools fail to generate responses less often and likely would have more context.
Hopefully they become fast enough to run offline or at least feel more instantaneous.
Writing the code isn't the hard part, and wrangling the computer is the part of the work I enjoy most. My problem with AI bros is that they explicitly want to automate all the shit people like to do, such that we can all finally be free to work service jobs.
So, in essence, it's now incrementally better than a templating script (except when worse), but Have Faith, it will be Better Soon. TBH, that's the same song that's been on repeat since the Dartmouth Workshop. In 1956. Jam yesterday and jam tomorrow, never any jam today.
It's just status anxiety. Mid engineers go on and on claiming theres literally no value from LLMs even possible in principle while top tier people are using them as force multipliers.
Who? How? This is not what I've seen where I work. There's a bunch of hubbub and generalized excitement, and lots of talk about what could be done, or what might be done, but not very much actual doing. I must just be a clueless "mid".
Guido van Rossum - "I use it every day. My biggest adjustment with using Copilot was that instead of writing code, my posture shifted to reviewing code."
https://www.youtube.com/watch?v=-DVyjdw4t9I
These kinds of responses are my favorite dark pattern rhetorical device, because you can assert literally _anything_ in this format and almost nobody will refute you, because the cost of refuting bullshit is 100x the cost of producing it.
Anyways, here goes.
1. Guido uses Copilot like I do - as a StackOverflow replacement to write the dumb boilerplate code. A much less flattering quote is "It doesn't save me much thinking, but [it helps] because I'm a poor typist". Also it's literally a minute or two of a three hour podcast.
2. A lot of code is autogenerated lol. Again, it's all the boring boilerplate stuff.
3. The cofounder of OpenAI is a biased source lol
4. He's an AI researcher, of course he runs that stuff.
5. Again, similar to Guido. He's using it for the boilerplate. Nothing wrong with enjoying using it as a toy, as he is here. But he's not doing serious work with it.
There's no virtue in hyping this stuff like a HODL bitcoin cultist.
I believe your examples are - unironically - misleading.
1- he states that the generated code is most likely wrong. He is appreciative of it though because he is a very poor typer so he doesn't have to do that part as much
2- so that's not supporting your argument that the 'top' devs are using it. Besides it doesn't say how it's counted, nor how much time is spent reviewing and correcting it
3- actually okay. But is he using it for production code? Doesn't say
4- he definitely doesn't talk about coding, only brainstorming and writing text.
5- your best one. Still, the use case here is side projects not production
You might still be right, I definitely do not compare myself to these people, but trying to glue some sources together makes a poor argument.
And the subject on hand is more that just using LLMs, it's the role of LLMs in the dev work environment
I disagree. Tools which need to be babysat, the way LLMs do, slow you down rather than speed you up. It's like having to mentor a junior team member, except a human will eventually learn and you can just let him work. LLMs are incapable of learning, so you can't ever leave the phase where they are a drain.
> I’m a little shocked at how much negativity there is around LLMs among developers
There's a Quentin Tarantino quote where he says there are 2 kinds of film critics. There are those who love movies and there are those that love the movies they love.
A lot of developers really seem to love the technology they love.
These people are where most of the negativity is coming from. And my guess is that the people who are encouraged by LLMs and not negative (mostly) aren't taking time out of their days to write long blog posts or argue about it online.
I use Copilot for autocompleting the boring boilerplate. I like it. I also think LLMs are mostly useless.
It's not the technology, it's the stupid overhype. It really feels like all the HODL bitcoin cultists have finally gotten over their lost apes and found a new technogod.
So many people in these threads are convinced it's about to gain sentience. That's not going to happen. You get the people outing themselves by saying "it does my job better than me!"
If you say something honest and direct like "their output is mediocre and unreliable" or "the RNG text generator is not capable of thinking, you're just Clever Hans-ing yourself" or "if it does your job better than you, that says more about you than about it", you get people clutching their pearls like a Stanford parent whose kid got a D.
arXiv has turned into a paper mill of AI startups uploading marketing hype cosplaying as "research".
I don't really get the hate; work is boring, if some tool can make it happen faster and with less effort I'm all for it. I don't hear the line cooks at McDonalds complaining that they have to use a semi-automated grill that beeps at them instead of an open fire.
As a developer I'm seeing less hate than apathy from my colleagues, but rather the hatred I do see is from people trying to push LLM's ON developers.
So it's from middle management levels riding the hype train, and possibly trying to save money and getting bonuses for it at the expense of other people.
Just like when offshoring was in its same point on Gartner's hype curve.
"Everyone has a model, but no one has a business".
Ed occasionally makes good points, but he's very very angry at Big Tech, and his anger often gets in the way of his message.
Reading his latest rant reminds me of Karl Denninger railing against Google around the time of their IPO, claiming they would never make enough money to justify an $85 share price (a $1000 investment then would be worth around $375,000 today).
Google solved a real problem. They indexed the web and made search work, and they did it very cheaply. So cheaply, in fact, that they could give their service away to users and monetize it with ads. LLMs are not like this. They're both extremely expensive to run and they don't do anything truly valuable--there's no killer app. So how exactly is OpenAI or their ilk (or for that matter the rest of us) supposed to use these things to make money?
This is the only question, and the fact it's still an open question just screams "hype bubble". My bet is this AI stuff goes the way of the NFT.
I'd take that bet. Google offers a very expensive service for free, but is able to monetize it with ads. Sometimes connecting users to companies is what users actually want. But Google has this problem that since the service is free, its users feel entitled to everything for free. They can't just go and charge people what it costs to run a Google search.
OpenAI doesn't have this problem. ChatGPT has a free level to get you hooked, but it's restricted. So a lot of users pay them $20 or $200 or some other amount per month to use their service. So how OpenAI makes money is by selling access to their service. What you do with it is up to you, but their value proposition is simple. Pay us to get more/better access to our service.
How much it costs them to operate the service is a secret known only to them. There are a lot of very very educated guesses, but they're just guesses. After the VC money runs out they'll have to charge more than it costs to provide the service to stay afloat, and then we'll see. $20/month for ChatGPT plus is the $1 Uber that got people hooked. There's already a $200/month tier.
Whether OpenAI, specifically, will be standing in 20 years, only time will tell. But by this point it should be obvious that there's something to this LLM thing. Even if the product doesn't get any better than it is today, it'll still take 5-10 years for its effects to reverberate through society.
The killer app is LLM-accelerated programming. Sure, it doesn't work for all domains and it can't do everything, but even if the only thing it's good for is creating JavaScript react CRUD apps, well, there are a lot of those out there, and they're not actually limited to that. And since tool use means they can generate code and compile it and test that it works, it's possible to generate datasets for other languages and libraries, the only question is which ones is it worth it for.
It might not help at all in your line of work, but a friend who does contracting is able to use LLMs to cut the time it takes him to do a specific kind of job in half, if not more,
enabling him to take on twice as many clients and make more money. For him it would still worth it even at 100x the current price. thankfully competition means it'll take a while before it's that expensive.
I think there's the same logic flaw of looking at how things are at the start - so so - and how they may be in 20 years - Google getting an advertising cut for most of the world's commerce, AI replacing/doubling the ~100tn/yr labour market.
Hey I have my gripes with the landscape, certainly, but this is just too much.
> It sure is! But it doesn't really prove anything other than that people are using the single-most-talked about product in the world. By comparison, billions of people use Facebook and Google. I don't care about this number!
> User numbers alone tell you nothing about the sustainability or profitability of a business, or how those people use the product. It doesn’t delineate between daily users, and those who occasionally (and shallowly) flirt with an app or a website. It doesn’t say how essential a product is for that person.
Both of these "arguments" could be applied to any of the big tech giants of the last 25 years - Google, Amazon, Facebook, Uber, whatever (and there'd be other incumbents used by billions of people before them). I don't believe these arguments discount ChatGPT from having the potential to continue growing like a Facebook. And who cares how many journalists Altman knows, you don't get a product written about that much unless it's truly a groundbreaking product.
> And even then, we still don't have a killer app! There is no product that everybody loves, and there is no iPhone moment!
There sure is, it's called programming. He called out quality earlier on, but the quantity and speed and direction the AI can take (as well as its rate of improvement) is breathtaking. My own output has 10x'd easily since GPT-4 came out (although some of that means I'm needing far less hours in certain places). And guess what? The code quality is generally fine.
> Where are the products? No, really, where are they? What's the product you use every day, or week, that uses generative AI, that truly changes your life? If generative AI disappeared tomorrow — assuming you are not somebody who actively builds using it — would your life materially change?
Ok, the product is called ChatGPT, or Claude, or DeepSeek or whatever, and if it disappeared overnight, my programming productivity would drop dramatically. I would not seek to take on as ambitious projects in as short of a time frame as I am doing now.
I don't know, as a user and developer both of AI/LLMs, this article isn't hitting the mark for me. There are legitimate criticisms of the field, but I'm not seeing them thus far.
Edit - I'll say I agree with the Deep Research criticisms. These products are very underwhelming. They're literally to help people do a research report which needs to be done, but won't be used or read critically by anyone report.
"I'll say I agree with the Deep Research criticisms. These products are very underwhelming."
I haven't shelled out the $200/month for OpenAI's Deep Research offering, but similar products from Google and Perplexity are extremely useful (at least for my use case). I would never present the results unchecked / unedited, but the Deep Research products will dig much deeper for information than Perplexity could be persuaded to previously. The results can then be fed into another part of the process.
> Both of these "arguments" could be applied to any of the big tech giants of the last 25 years - Google, Amazon, Facebook, Uber, whatever (and there'd be other incumbents used by billions of people before them). I don't believe these arguments discount ChatGPT from having the potential to continue growing like a Facebook. And who cares how many journalists Altman knows, you don't get a product written about that much unless it's truly a groundbreaking product.
I don't get these statements. This line of thinking is so egregious, FTX made an ad about it. This is survivorship bias. If 'wrong' predictions about some can be discrediting, then why not right predictions to establish credibility? The same people were right about the Segway, crypto, metaverse, web3, that dog walking startup that Softbank burned money on, and countless other harebrained endeavors.
Even in your example, Uber is a financial crime. It is still using accounting tricks to show profit.
This is not a criticism he really mentions explicitly, but the issue I see with the valuation is that these products cannot be used in a production system by a responsible engineer. As in, I can't have LLMs autonomously plugged in as part of my product. The failure modes are not ever going to be predictable enough. Now, Microsoft will probably be able to charge a fortune at enterprise level, and managers will dream of a day that the LLMs can replace all those weirdo devs at their company, but that'll stay a dream.
All of the valuable uses are personal. It makes me personally feel more productive at work. It helps me personally understand some topic better. It gives me an idea in a personal project.
That's all really cool, but that is not what the valuation is about. The valuation is about a false science fiction and hype bubble about agentic this or that or AGI or whatever, and this is driving very questionable decisions for wasting possibly trillions of dollars and tons of energy.
The plus side is that there is some really cool personally useful tech here, and we will probably end up with very good open source implementations and cheaper used GPUs once the bubble bursts.
You also can't responsibly have a prolific apprentice (intern, first or second year, etc.) plugged directly to production.
On the other hand, today most money for SWEs goes to people who aren't "staff engineer" level. What if most money went to staff engineers who directed these interns and paired with new human apprentices to learn staff eng?
In the past few thousand years, apprentice/journeyman/mastery of trade was how trades worked, with an "up or out" model that kept the role pyramid the right shape for skill to shape the outcomes.
These days, far too many careers stay apprentice skill the entire career, mostly thanks to enterprises failing at engineering management as a skilled trade, so being unable to value and raise staff engineer caliber contributors. The enterprise learning machine is broken by false "efficiency".
LLMs change this. Staff engineer caliber SWEs are able to direct these indefatigable assistants, as if each staff engineer has a team of 10 who never need mental breaks to remain productive. There will of course be some number of junior devs who themselves have enough affinity for the role they will want to stick with the apprentice model and work to the staff engineer level. (And will always be solo or boutique teams of app/saas SWEs.)
As for the enterprise engineering management that couldn't tell the difference between a staff engineer and an apprentice, the LLM multiplies the difference to the point the outcomes are evident even to a non technical observer.
So one possible timeline for this is a raising of the median human skill level by attrition of those unskilled enough or unable to think critically enough to leverage the machine assistants as force multipliers or unable to survive directly mentored skill-up training and observation from the staff engineers.
You talk about personal value. Roughly, I agree with you completely, and am adding who I think those persons could be (or have to be given the current level of "thinking" by these tools) for the hype to deliver on value. (At a higher level of machine, closer to "AGI", this scenario changes.)
As is evident by downsizing a mediocre team and observing output go up and work more reliably, these forces could, if playing out this way, make dollars per human go up, productivity go up, quality go up, and enable a return to the millennia-proven model of apprenticeship for the trade.
Plenty of AI companies that are cons or extremely overvalued, but the technology is the real deal and delivering huge improvements over previous techniques in all kinds of domains: language translation, weather prediction, code completion, self driving, etc.
I swear no matter how many times people say it people will still conflate all ML with LLMs. No, chatGPT is not driving advances in self-driving or weather prediction
For better or worse, "LLM" or "generative ai" has become roughly synonymous with the current wave of ML.
I know very little about ChatGPT, but Waymo is using an LLM: "Powered by Gemini, a multimodal large language model developed by Google, EMMA employs a unified, end-to-end trained model to generate future trajectories for autonomous vehicles directly from sensor data." (https://waymo.com/blog/2024/10/introducing-emma)
Huh. I mean it makes sense to train end-to-end on all the interrelated tasks involved in driving but putting a whole-ass language model in the middle of that seems like a stunt. I wonder if it does better than like, any random transformer not trained on language first? Still, I hadn't heard that so I guess I was wrong about that one
No, you were right, this appears to be just research on how applicable LLMs could be to the space. They talk about the improvements their LLM makes, especially in being multimodal vs training multiple independent models, but also the limitations that appear to prevent it from being useable as it is. Maybe some form if it will be used some day (it does seem like it would be useful to have semantic understanding of the world integrated into the system), but at least as of when this was published, it's not actually used.
Okay so because of the ambiguity of the other reply I'm just gonna say, I don't think we should be surprised that someone is trying to use LLMs to do basically anything. That's basically what prints funding money right now, so long as you're the kind of company or guy the VCs or whoever will believe in. The signal here is "does it do something to appreciably advance the state of the art over previous methods"?
Yeah so like, this is a cool result, and it uses a transformer architecture. I actually do think that it's fair to say that transformers have proven widely useful, especially in tasks that look like sequence modeling. It's a step change akin to the now-pervasive use of convolutional neural networks that started in the 2010s, and is deeply significant of course. This is also really different from "this is an LLM"
The reason I want to specifically harp on this is because a lot of people are selling this narrative where "AI is becoming superintelligent" or whatever by making an amorphous blob out of a bunch of separate advances that use machine learning techniques. This has been happening for a while, is a great thing for science, and it's clear that machine learning methods are here to stay in science. I'm a machine learning researcher. I've understood, celebrated, and tried to help with this as best I can manage over the last 9 years of my life. And it's been going on for a lot longer than the general public has been in this AI hype wave. The entire modern field of bioinformatics is arguably built on the backbone of machine learning, and has been since before I went to grad school.
This is really different from "We fed everything into a language model and now it's superintelligent and is making scientific advances all by itself" or even "scientists just ask chatGPT shit and it figures it out for them". The breathless tech press really makes it sound like anything that happens in AI research, which increasingly includes the entire usage of ML toolkits in the sciences (Which is pervasive, and expectedly so! ML is an extension of statistics and statistics has been the basis of science for like a century) is just some amorphous force called "AI" that's suddenly gained this aggregate body of competency. Imagine if we anthropomorphized statistics that way. Or Math for that matter. This kind of narrative gives me the overall impression that this is not being talked about honestly, and it's clear that this is profitable to do. I don't have to use charged words like "con" or "fraud" to think this deceptive framing is not a great thing
What you're saying does happen to some degree and in this instance, if i had linked some advance with a diffusion model then i would get it but about the only difference between this and chatgpt is the data it's been trained on.
If Open AI cared, the next version of GPT could be a State of the Art weather predictor.
I mean by the same logic the only difference between a diffusion model and a VLM is that you put the spatial transformer on the other end.
Yes, one of the powerful things about every kind of neural network is that they're a very general class of function approximator. That we can use a similar toolkit of techniques to tackle a wide variety of problems is very cool and useful. Again, the analogy to statistical models is telling. You can model a lot of phenomena decently well with gaussian distributions. Should we report this as "Normal distribution makes yet another groundbreaking discovery!"? Probably this wouldn't have the same impact, because people aren't being sold sci-fi stories about an anthropomorphized bell curve. People who are using LLMs already think of "AI" as a thinking thing they're talking to, because they have been encouraged to do that by marketing. Attributing these discoveries made by scientists using this method to "AI" in the same way that we attribute answers produced by chatGPT to "AI" is a deliberate and misleading conflation
>I mean by the same logic the only difference between a diffusion model and a VLM is that you put the spatial transformer on the other end.
Maybe if that was the only different but it's not. There are diffusion models that have nothing to do with transformers or attention or anything like that and where using them for arbitrary sequence prediction is either not possible or highly non-trivial.
Yes, All Neural Network architectures are function approximators but that doesn't they excel equally for all tasks or that you can even use them for anything other than a single task. This era of the transformer where you can simply use a single architecture for NLP, Computer Vision, Robotics, even reinforcement learning is a very new one. Literally anything a bog standard transformer can do is anything GPT can do if Open AI wished.
Like i said, i don't disagree with your broader point. I just don't think this is an instance of it.
It's clear you're missing what point it is that I'm making from these responses, but I'm unsure how to explain it better and you're not really giving me much to work with in terms of seeming to engage with the substance of it, so I think we gotta leave this an impasse for now
LLMs can do weather prediction? There is no way that is true. Considering ChatGPT sometimes insists 2+2=5, I sincerely doubt it is solving PDEs used to model weather systems.
A Large Language Model is just a text predicting Transformer (and sometimes image and/or audio predicting if they're multimodal). Transformers are general sequence to sequence predictors. The only difference between this and GPT is the data it's been trained on. They're the same kind of neural network.
Think base models. LLMs are extremely good at predicting the future when applied to human languages, as this is literally their only optimization goal. Why couldn't they also be good at predicting the future when applied to other complex forecasting tasks?
Of course what can be mathematically calculated without inference is going to be. LLMs may however be able to interpret the results of these calculations better than humans or current stochastical evaluations.
He seems to think the hype will do great damage to society
"I need you to fucking listen to me: everything I am describing is unfathomably dangerous, even if you put aside the environmental and financial costs."
Personally I think he's lost it a bit. I mean say he's right in that LLMs plateau and investors lose some money. Life will go on.
This reads like a straightforward rant to me, not something informative.
Which then invites the question: why is this person's opinion on the subject relevant? Do they have some credentials that make it more valuable than a random comment with a similar rant (of which there are plenty) on Reddit or HN?
>So...yeah, of course ChatGPT has that many users. When you have hundreds of different reporters constantly spitting out stories
Oh sure, because you can just have hundreds of reporters constantly write about your product. It's so simple. Why aren't more people thinking of that ?
>The weekly users number is really weird. Did it really go from 200 million to 300 million users in the space of three months?
According to similarweb, monthly visits grew over 1B in that timeframe so yeah sure it sounds possible.
>300 million monthly active users would mean a conversion rate of less than 4%, which is pretty piss-poor
A B2C Saas whose lowest price point is $20 will be lucky to get anywhere near 4% conversation.
>And even then, we still don't have a killer app!
The 6th (and climbing) most visited site in January is not a killer app ? Okay
> And when this all falls apart — and I believe it will — there will be a very public reckoning for the tech industry.
I shocked he really believes it in his closing thought.
Maybe his rant would be bit more digestible if it contained sections with: "here's what I tried and it did not work". But that would make it not a rant but actual research with value.
I'll save him the trouble of writing it: "I asked ChatGPT 3.5 to write a large, underspecified chunk of code a couple of years ago, and it didn't work the first time, unlike the code that I write. This whole 'AI' business is an elaborate scam."
He's wrong, there's no other sector of the economy with growth potential and there's so much capital desperately seeking returns. Also anxious capital terrified of being disrupted. The tech industry will just move on to the next hype cycle when this one burns out.
To this pedantic point, If the average written intelligence of all humans alive and dead is > the max intelligence of all live humans who are also willing/positioned to do the same task at the same time and at the same place.
But yeah, I don't think LLMs (the current core architecture) can provide super intelligence. I think it needs a bit more than next token prediction architecturally speaking.
This is incorrect. If you take the most basic interpretation of an LLM at temperature 0 as predicting the most likely token, and you run it on, say, 1,000 runs of "complete this Spanish sentence with the word for 'X'", then:
- maybe ALL humans would fail the test in some way, eg. let's say everybody gets at least 10 of those wrong, and the average person gets 100 of those wrong.
- still, as long as most people correctly get each word right, your LLM would get every single response correct (because for each item in the test, 900+ people out of a thousand gave the same correct answer in the training set).
In that sense, it's totally possible for a system trained on a vast vat of average-human input to generate super-human outputs.
But still, the questions in that test are "solved" in the sense of "I can take a dictionary and answers these questions with full certainty". Beyond established knowledge LLMs are monkeys with typewriters, at best.
I’d like to see you ace even a middle-school level Spanish test with just a dictionary (sub Spanish with some other language if you happen to know Spanish).
Let's define "zeta" as a mathematical function ζ(s) which takes a statement "s" as input, where s is a statement of a breakthrough in LLM capabilities achieved relative to the current date and time and ζ(s) is the probability that a given AI skeptic will honestly recognize "s" as a breakthrough,
then our Riemann-Goalpost hypothesis is that ζ has zeros for every "s" which is a negative integer (every breakthrough that happened in the past is null in value) and only has positive values where s is positive.
We can conclude from the above that given a far enough date in the future, any given breakthrough can be spectacular, but once achieved, will be derided as trivial.
for people using these AI products for coding what exactly are you doing ?
general API plumbing ? call this api, combine the json results & spit it out. Then slap a React / Next.js frontend ?
lately, I have been doing your classical business apps - due to the domain rules - ai is pretty useless - but I have found perplexity and deep mind to be smarter stack overflows. that's it.
personally I use it as a glorified stack overflow. It really just saves me a few clicks on my browser.
For anything even remotely complex or involved, it's just not that useful. Interestingly, It seems to shine the most when doing boilerplate stuff in widely used languages such as python or javascript, but it's extremely bad with terraform.
Its basically more time-consuming for me to get it to where I want the code than finish it myself. For some scaffolding in python green-field development, sure I save a few minutes.
A huge part of our problems stems from the fact that it's possible to make "companies" whose business model is built on losing money hand over fist until they've brainwashed everyone into thinking their "product" is good. In a sane world these companies would fail and AI would continue develop through small failures and small successes over a period of years or decades. Instead we get a firehose of nonsense just because a small number of wealthy people are willing to gamble.
That would be nice, but I'm more worried they won't collapse because they'll succeed at shoving their snake oil down the throats of enough big players to ensure their survival.
On LLM/ML itself: it seems a lot of cynical people start with some unreasonable idea that "AI" should be able to do what it will perhaps be able to in 10 or 100 years, and are subsequently upset that it is not capable of that yet. It may get there, it may not. But that's on you for starting with a wrong assumption.
Is the AI "business" or "market" overvalued for it's current capabilities? Yeah, I do believe so. Welcome to the financial world, which is completely separated from reality. It's like that in all sectors where something new and exciting is happening, not just IT or AI. People poor money in hoping to be early enough to make a profit. Nothing more, nothing less. The rest is marketing. Some Sam Altman guy promoting the hell out of his own product? That is literally his job, regardless of wether or not he believes it all.
But articles like these are so bizarre to me. The author acts like he has millions at stake and his money manager just won't listen and pull all investments out of AI. Hurry up, the bubble is about to burst, I will lose all my money!
Except that... they don't. They are just "old man yelling at cloud". If you believe AI is the next Metaverse or WeWork, then it will just die off by itself once the bubble pops. Why are you having so many conversations about it, where you seem to be desperately trying to convince people of the bubble/con that is AI. To the point that you're so sick of it, that you write down your arguments so you can point the blinded there instead of having those tiresome arguments.
Genuinely baffled. Spend your energy on something productive rather than destructive, perhaps?
If you are correcting my use of millions to trillions: I was refering to the author himself, who writes like this giant AI bubble is pushing him forward to the edge of the cliff as people keep believing in it, and he is desperately trying to get the bubble to shrink or he'll fall off and die. Methaphorically.
But why does he act or feel that way? Let the trillions be lost, it's just how hypes, bubbles and the stock market in general work.
> it seems a lot of cynical people start with some unreasonable idea that "AI" should be able to do what it will perhaps be able to in 10 or 100 years, and are subsequently upset that it is not capable of that yet
Because that's what AI wa supposed to be in the first place. But the industry performed the swindle of renaming "AI" to "AGI", so that they can pretend the thing that exists now is "AI".
502 comments
[ 2.5 ms ] story [ 334 ms ] threadThere hasn't been much R&D progress, though. Sure, as another commenter pointed out, context lengths have gotten longer and chat models can interpret images now, but the industry figureheads have been pushing agents, and we're not much closer to those than we were two years ago when GPT-4 came out. Current models simply are not consistent enough to do the kind of agentic stuff that AI valuations are predicated upon, nor is there any sign that a significantly smarter GPT-5 is just around the corner. Multi-modal chat is cute, but OpenAI is burning money. They're all burning money, and they don't have a product. They imply and imply that there's something big on the horizon, but it's been years, and there just isn't a killer app yet. Their platform isn't good enough, and it's not improving in the ways it would need to in order for Godot to arrive and for agents to be feasible.
That said, I'm learning a new sdk and I've moved 500-1k searches a month from kagi and google to llms.
With AI the improvements have certainly been impressive but it isn't straightforward how you can define "reasoning" to measure whether or not the reasoning is exponentially "improving".
He comes across as just a ludicrously unpleasant, spite-filled person.
> I'm fucking tired of having to write this sentence.
> I am so very bored of having this conversation
> I don't care about this number!
> Shut the fuck up!
> This isn't the early days of shit.
> Didn't we just talk about this? Fine, fine.
> $3.25 billion a quarter is absolutely pathetic.
> This isn’t real business! Sorry!
> He said in one of his stupid and boring blogs that
> This man is full of shit! Hey, tech media people reading this — your readers hate this shit! Stop printing it! Stop it!
> It's here where I'm going to choose to scream.
> Dario Amodei — much like Sam Altman — is a liar, a crook, a carnival barker and a charlatan, and the things he promises are equal parts ridiculous and offensive.
> Why are we humoring these oafs?
> Despite Newton's fawning praise
> Nobody talks like this! This isn’t how human beings sound! I don’t like reading it!
> Ewww.
> I'm sorry, I know I sound like a hater, and perhaps I am, but this shit doesn't impress me even a little.
> I know, I know, I'm a hater, I'm a pessimist, a cynic, but I need you to fucking listen to me: everything I am describing is unfathomably dangerous
> expensive, stupid, irksome, quasi-useless new product
> I know this has been a rant-filled newsletter, but I'm so tired of being told to be excited about this warmed-up dogshit.
> I refuse to sit here and pretend that any of this matters.
> I'm tired of the delusion. I'm tired of being forced to take these men seriously.
When I read this kind of thing, it’s very apparent that this is being driven entirely by spite not insight. He’s just so angry about everything. There are 57 exclamation marks in this article!
because profit of this can not cover the investment in this industry
adoption of iphone/smartphone/internet brought new products, including those for reproduction and those for consumption
but generative AI is totally different with iPhone, consumers maybe willing to buy a new ai-powered iphone __just like how they bought new iPhones for every 2years before__
> The current landscape imho should be viewed as an R&D race amongst private actors
in fact, it's a CapEx race, you don't need to R&D anything (ofc you must pretent you do)
that's why it's a con
> The AI Bubble will burst “any day now”
"The canary in the coal mine to look at is when Satya Nadella or Sundar or Zuckerberg say, ‘You know that $80bn of capex I said I was going to do? I think I’m going to cut that by two-thirds.’ That’s what you need to look for."
that's the day
It’s basically functioning as a team of entry-level junior engineers at this point.
Previously I was having to spend a fair amount of time writing tickets and providing context, but lately I’ve fed all my meeting transcripts and such into an LLM and it interactively creates Jira tickets for me. Each one takes me maybe 30s to read before I confirm them and the assistant creates the actual tickets.
Can you give some examples?
Devin was able to recognize that the project used Poetry, was Dockerized, and that the Python version was specific in multiple places (.python-version, pyproject.toml, Dockerfile). It saw that a couple of minor dependencies didn’t support the new version of Python, so it went back and upgraded those to the most recent matching version first.
Devin had never touched the repository in question before getting this task.
I’ve given it more and less complex tasks, and yeah, it struggles with some things. I’d estimate that it consumes about 5-10% of my time but multiples my overall output by ~3x.
I think we’ll see a ton of complaints about how bad the job market is in the next couple of years. That will be true, but only for juniors or for seniors who don’t embrace the tech. For seniors who do embrace it and specialize in implementing these systems, it’ll be a gold mine.
Then, over 5-10 years, our seniors will start to retire or leave the field. No one will be there to replace them. At that point we’ll see a resurgence in the job market.
Things like autocompletion and “chat with your codebase” help juniors more than seniors; agents help seniors much more than juniors. As these systems improve, their failure cases get more and more complex/nuanced - you will always need senior people with the insight necessary to figure out what’s wrong when it breaks. For a while that will help seniors and hurt juniors… right up until businesses realize that they don’t have replacements for their existing senior engineers, at which point they’ll be desperate to hire again.
For large codebases (greater than 15k or 20k LOC) the context size seems like a real problem right now.
My apologies if anyone finds this offensive, but I sorta see Devin as a fresh junior SWE hire. It doesn’t do well with tasks that require deep knowledge sometimes, but it has shallow or better knowledge of everything. I would describe it as working with a brand new SWE with an IQ of about 85 who is also on the low end of being high-functioning autistic. By that I mean that it takes most things literally and sometimes has difficulty with nuance.
> burns a ton of money, gets stuck, doesn’t implement the changes you want
The first time you use it, I think that’s pretty fair. Every time it gets stuck or does the wrong thing, when you correct it, it gives you the option to add to its “knowledge base”. That’s a bunch of additional context that it applies in only certain situations. Within a week or so of using it regularly, it’s significantly more valuable. It “learns” much faster than a human.
Example:
About a dozen of our projects all rely on a shared repository (“Enki”) that contains a Composefile, configs, and some light automation. Tests are run in Docker, and you have to navigate to the other repo’s directory to bring up the service. Some of those projects have service names in the Composefile that differ from the project name. I was able to run the steps interactively on “Devin’s machine”, tell Devin what I had done, and then tell it that this is the correct approach for any project that depends on that repository. I didn’t tell it what projects those are, or how to find out.
The next time I used Devin on a project like that, it tried to run the tests directly in a local Python environment. That didn’t work, but it tried the correct approach next. That worked, so it added a line to its knowledge base “Project <foo> uses Enki.” From that point forward it did the right thing the first time.
> For large codebases (greater than 15k or 20k LOC) the context size seems like a real problem right now.
The primary project I’m working on is a Django app. I don’t have it in front of me right now, but it’s about five years old, has been under very active development the entire time, and is comprised of about twenty apps. It’s not the largest codebase I’ve worked on, but it’s far from the smallest. I can do a line count tomorrow if you’d like.
This is missing the most interesting changes in generative AI space over the last 18 months:
- Multi-modal: LLMs can consume images, audio and (to an extent) video now. This is a huge improvement on the text-only models of 2023 - it opens up so many new applications for this tech. I use both image and audio models (ChatGPT Advanced Voice) on a daily basis.
- Context lengths. GPT-4 could handle 8,000 tokens. Today's leading models are almost all 100,000+ and the largest handle 1 or 2 million tokens. Again, this makes them far more useful.
- Cost. The good models today are 100x cheaper than the GPT-3 era models and massively more capable.
If nothing else, my workflows as a software developer have changed significantly in these past two years with just what's available today, and there is so much work going into making that workflow far more productive.
And I’d argue it took decades to actually achieve some of the things we were promised in the early days of the internet. Some have still not come to fruition (the tech behind end to end encrypted emails was developed decades ago, yet email as most people use it is still ridiculously primitive and janky)
Nobody will deny the importance of railways to the Industrial Revolution, but they also lost a lot of people a lot of money: https://simonwillison.net/2024/Dec/31/llms-in-2024/#the-envi...
Then came DSL, then came cable, then came fiber. Countless billions of dollars invested into all these different systems.
This AI stuff is something else. Lots of hardware investment, sure, but also lots of software investment. It is becoming so good and so cheap its showing up on every single search engine result.
Anyway, my point is, while there may have been aspects of the early internet being a bubble, there were real dollars chasing real utility, and I think AI is quite similar in that regard.
I have very little idea of the second - it's totally possible OpenAI is a bad investment. I think this article is massively wrong about the first part though - this is an incredible technology, and this should be evident to everyone (I'm a little shocked we're still having an argument of the form "I'm a world-class developer and this increases my productivity" vs. "no, you're wrong!" on the other).
this is exactly the problem
The more productivity AI brings to workers, the fewer employees employers need to hire, the less salary employers need to pay, and the less money workers have for consumption.
capitalist mode of production
for humanity, the increase in productivity is progress
i'm not saying it's bad, i'm saying it has consequences
Did all of that free code reduce demand for developers? If not, why not?
the anwser is yes, while in the meantime, the expansion of the industry offset the surplus of developers.
More effective developers results in more demand for custom software, resulting in more jobs for developers.
My hope is that AI-assisted programming will have similar results.
I don't really know myself, but I think there's a decent change that most developer jobs will actually disappear. Your argument isn't wrong, but when we're nearing (though still far from) the state where all productive tech work can be handled by LLMs. Once it can effectively and correctly fix bugs and add new well-defined features to a real codebase, things start to look very different for most developers.
I agree, though personally I'm liking the "big thing" as well. R1 is able to one-shot a lot of work for me, churning away in the background while I do other things.
> Multi-modal
IMO this is still early days and less reliable. What are some of your daily use cases?
> Context lengths
This is the biggest thing IMO (Models remaining coherent at > 32k contexts)
And whatever improvements have caused models like Qwen2.5 to be able to write valid code reliably vs the GPT-4 and earlier days.
There are a whole lot of useful smaller niche projects HF like extracting vocals/drums/piano from music, etc
For images I use it for things like helping draft initial alt text for images, extracting tables from screenshots, translating photos of signs in languages I don't speak - and then really fun stuff like "invent a recipe to recreate this plate of food" or "my CSS renders like this, what should I change?" or "How do you think I turn on this oven?" (in an Airbnb).
I've recently started using the share-screen feature provided for Gemini by https://aistudio.google.com/live when I'm reading academic papers and I want help understanding the math. I can say "What does this symbol with the squiggle above it?" out loud and Gemini will explain it for me - works really well.
Just last night I was digging around in my basement, pulling apart my furnace, showing pics of the inside of it, having GPT explain how it works and what I needed to do to fix it.
Getting it to fix a mower now. It's surfacing some good YouTube vids.
I don't have 100% of the "common sense" knowledge about every field, but good LLMs probably have ~80% of that "common sense" baked in. Which makes them better at interpreting incomplete information than I am.
A couple of examples: a post on some investment forum mentions DCA. A cooking recipe tells me "boil the pasta until done".
I absolutely buy that feeding in a few photos of dusty half-complete manual pages found near my water heater would provide enough context for it to answer questions usefully.
If the cost-to-serve is subsidized by VC money, they aren't getting cheaper, they're just leading you on.
If we're stretching, we can talk about opportunity cost. But the people spending and creating the "bubble" don't have better opportunities. They're not nations that see a ROI on things like transportation infrastructure or literacy.
So unless the discussion is taken more broadly and higher taxes are on the table, there really isn't a cost or subsidy imo.
This. IIUC to serve an LLM is to perform an O(n^2) computation on the model weights for every single character of user input. These models are 40+GB so that means I need to provision about 40GB RAM per concurrent user and perform hundreds of TB worth of computations per query.
How much would I have to charge for this? Are there any products where the users would actually get enough value out of it to pay what it costs?
Compare to the cost of a user session in a normal database backed web app. Even if that session fans out thousands of backend RPCs across a hundred services, each of those calls executes in milliseconds and requires only a fraction of the LLM's RAM. So I can support thousands of concurrent users per node instead of one.
The computations are not O(n^2) in terms of model weights (parameters), but linear. If it were quadratic, the number would be ludicrously large. Like, "it'll take thousands of years to process a single token" large.
(The classic transformers are quadratic on the context length, but that's a much smaller number. And it seems pretty obvious from the increases in context lengths that this is no longer the case in frontier models.)
> These models are 40+GB so that means I need to provision about 40GB RAM per concurrent user
The parameters are static, not mutated during the query. That memory can be shared between the concurrent users. The non-shared per-query memory usage is vastly smaller.
> How much would I have to charge for this?
Empirically, as little as 0.00001 cents per token.
For context, the Bing search API costs 2.5 cents per query.
People who I trust in this space have consistently and credibly talked about these constant efficiency gains. I don't think this is a case of selling compute for less than it costs to run.
The subsidies are going to the training costs. I don't know if any model is running at a profit once training/research costs are included.
Not very well in my experience. Last time I checked ChatGPT/DALL-E couldn't understand the its own output to know that what it had drawn was incorrect. Nor could it correct mistakes that were pointed out to it.
For example, I ask it to draw an image of a bike with rim brakes it could not, nor could it "see" that what was wrong with the brakes that it had drawn. For all intents and purposes it was just remixing the images it had been trained on without much understanding.
Evaluating vision LLMs on their ability to improve their own generation of images doesn't make sense to me. That's why I enjoy torturing new models with my pelican on a bicycle SVG benchmark!
Anecdote:
I often front-load a bunch of package.jsons from a monorepo when making tooling / CI focused changes. Even 10 or 20k tokens in, Claude says things like "we should look at the contents of somepackage/package.json to check the specifics of the `dev` script."
But its already in the context window! Given the reminder (not reloading it, just saying "its in there"), Claude makes the inference it needs for the immediate problem.
This seems to approximate a 'working memory' for the assistant or models themselves. Curious whether the model is imposing this on the assistant as part of its schema for simulating a thoughtful (but fallible) agent, or if the model itself has the limitation.
Redesigning our cities around cars was one of the big mistakes of the 20th century.
Horses weren’t cheap when they were a primary mode of transportation. Lots of people have died riding, driving, and breaking horses. They definitely smell bad, and their “particulate matter” was so bad that houses had to be set back and elevated from the street.
Cities designed around cars are far superior to cities designed around horses.
Why can't you design a city around it? It would be different than Houston or LA, but would that be a bad thing?
??? No, they don't. They only smell bad when they're kept standing in their urine, which is still not worse than a hairdresser. Compared to dogs (or ICE cars) horses smell way less. They do sweat to regulate temperature, which has a distinct smell, but it's way less irritating to a human's nose than the sweat of the rider.
There's a set of distinct smells associated with the horses, but other than the piss, none of them are particularly "bad". In my experience, humans tend to smell way worse overall (from food to body odor to the excrement) than horses.
Sarcastically, one might say that horses are even harder to operate (they have minds of their own), they smell worse that automobiles (esp EVs), and the particulate matter they excrete would be unhealthy to consume. They are also very slow.
More seriously, the trajectory that our imagination pushes towards seems to be “overcoming” biological limitations. Perhaps a symbiosis of machine and biology and consciousness will take us to the next level by opening up vast new universe state.
Cities have been designed around Carriages for millenia. You can go and walk in Pompeii and observe pavements for pedestrians, roads for wheeled carriages with crossing spots of elevated stones for pedestrians.
It turns out that cities require a lot of goods to be moved through - more than a pedestrian can carry, and over inclines that human muscle power doesn't like.
The reason why cities are designed around cars is that cars were designed to fit in contemporary cities and they co-evolved over the 20th century. It was the slow kind of evolution, with each step being easier and cheaper than the big redesign.
No, the switch to car-centric infrastructure was a deliberate policy choice lobbied for by the automotive industry. [1] We ripped up a lot of good transit to lay down roads this wide and fast.
[1]: https://www.fastcompany.com/90781961/how-automakers-insidiou...
> It turns out that cities require a lot of goods to be moved through - more than a pedestrian can carry, and over inclines that human muscle power doesn't like.
Hence the utility of public transit, which kills substantially fewer people and is much cheaper. Though goods are mostly moved with trucks, and trucks aren't my concern. Urban congestion isn't caused by 18-wheelers.
The point isn't to say that new tech is bad, but that there can be adverse consequences to jumping in wholesale.
car is consumer goods, we buy cars for use, not for reproduction
It's been made fairly clear that the insiders are setting the stage for governments to back them (bail them out).
I would strongly argue that coding assistants are AI’s first killer app. Copilot, Cursor, Windsurf etc.
To use these tools properly, you need to know how to build the same thing precisely.
I used claude + gemini to whip out a rewrite of a show HN project in less than half an hour to deployment yesterday.
https://news.ycombinator.com/item?id=43071381
But this sort of work is fairly low value and boilerplate-y.
The code isn't the most readable because I don't need it to be however if you make me write it from scratch in an interview style setting I'd have trouble doing it. If I read the code I can follow it and it makes sense + it's an easy component to manually test. So.. no, I don't need to know how to precisely build the same thing.
And before you worry that I'm committing code I can't build from scratch.. This is a simple component for a 5 page landing page build with astro where I'm the "main" dev ( wrote like 80% of the code). The web-page won't even need maintainance once it's deployed
The copilots get you going quick at the expense of your learning, which is great for one-offs, but not lasting work quality.
> This is a simple component for a 5 page landing page build with astro
You're already in the over-engineered section there.
https://gist.github.com/lazarcf/c80ae6f9362aaf3aa92e21e3a0dd...
you can see the component (the "image gallery") here : https://pixico.roware.ro/apa-distilata/
I agree the tools are overhyped for allowing non-developers to write code. It’s not (today) a replacement for a dev agency that takes a set of requirements and runs with it, it’s a replacement for a junior developer who you need to micromanage a bit. But that’s still a boon to productivity!
I have heard "top" engineers at various places say it makes them 2x faster, or whatever, but I would like to see this assessed by timed testing, as is sometimes done for evaluating software engineering.
Copilot may let me type less, but I have not seen the wall clock effects, which is a very hard thing to measure (time perception is very unreliable).
You can see this example where I timed myself to deployment using AI tools to rewrite a show HN project in half an hour. The code is open source.
My comment was posted 2 hours after show HN when I saw it on front page so you know I didn’t lose track of time I spent.
The majority of software work is maintaining large, existing products: adding features, fixing bugs, improving performance, etc., or building new software in problem domains that aren’t so well-defined.
If it could test and verify things though... ideally physically since Im in embedded and pulling SD-cards etc is a thing.
I’d love it if I could get it to write decent unit tests given a basic description of what I’m testing but I at least cannot get it to output anything useful for our codebase. It’s too unaware of the broader context of the code, what objects need to be instantiated from some other internal library and passed in, etc. It can do a decent job if all I have is a totally isolated function that doesn’t touch the rest of the codebase or use any domain objects, but that’s a rare enough case as to be essentially useless to me.
I think it also really accelerates learning of a new language or framework, when that language or framework is really well documented on the web. For novel programming frameworks, obviously it's a bit more challenging to get help from an LLM.
One of more recent attempts at using LLM code assist was to try to fix a bug in a Swift SSH Agent's connection handling that was causing hangs. I know zero Swift, much less the networking frameworks. So I pumped the output of `tree` on the git repo into the LLM, asked for which file likely handled connections, and it found it right away. That's probably 15 minutes saved. Before putting in the file I asked for likely reasons for deadhangs, got that list, then put in the Swift file that handled connections, and it pointed to what the likely problem was. That's probably 1 hour+ of reading documentation to try to figure out what the code was doing wrong with the networking framework, assuming the LLM was not hallucinating. And that "not hallucinating" probability is high enough in my experience that I spend >50% of my time trying to verify I'm not getting bullshitted.
The LLM proposed a fix (~10-20 minute savings), but even as somebody who doesn't use Swift it seemed like >99% chance that it had just introduced a bunch of race conditions in the data structures it used to track connection status. So I asked about it, and it said "Oh yeah of course how could I forget" and then significantly complicated the solution with something that I thought looked like it probably worked. But was the LLM just being obsequious or was it correct the first time? So hard to tell...
So in about 20 minutes I probably accomplished in a language I didn't know, in a code base I didn't know, about 2 hours+ of learning.
But if I knew the language, it would have saved me very little time, and may have cost me some time.
Of course, you have to find a company willing to spend more money on worse (for them - less ads) search, and it's won't be Google.
The results aren't always accurate, but neither is Google...
These IMO are relatively useful things. But probably (in their current state) will not justify the valuation of the companies involved and the massive investment occurring right now.
I don't know how the future will unfold. I do think it is reasonable to be somewhat bearish on what has been promised vs. what has been released.
It's almost like entertainment requires some humanity and thought and true creativity behind it.
In an open world game, it’s trivial to assign memories and facts an AI learns about its world from interactions or in response to game events. All an LLM has to do is be fine tuned to take data from that internal knowledge base and express it as natural language text, in order to have intelligent and useful conversations with a player. It’s not difficult.
By your logic I could claim a quantum computer with qubits on the scale of the mass of the sun is a killer app for doing RSA encryption breaking. And I would be making an equally useless statement.
Whether the companies that are leading the market today will end up being the ones who capture that value is anyone's bet.
Maybe this era of AI will be remembered as a wealth transfer from VCs back to everyday consumers lol
There are a lot of horses in this race and the literally billion dollar question is who's Amazon this time around, and who's Webvan. Who's Uber and who's Flywheel (which doesn't even have a Wikipedia page anymore, ouch). Not knowing which horse is going to win doesn't negate the fact that a horse is going to win.
Model available LLMs, like Llama and Deepseek and StableDiffusion are totally a wealth transfer to consumers. Better make use of them!
> not much worse than a junior dev and 100x faster.
Is there a greater hell than this!?
If you actually take the time to tell it “hey, don’t do it this way,” it can definitely do it differently the next time.
On top of that, is anyone training models on their own codebase, and noting to AI which patterns are best practice and which aren’t?
There are a ton of ways to make it better than the baseline copilot experience
Especially given that you can ask an LLM to optimise code and on multiple runs it can not tell if it's is improving or degenerating.
Yes — junior management using LLMs and 100x more cocksure.
But like you said, in a few more years we'll see! It does feel like there's some missing pieces yet to be figured out to truly "reason" and generalize.
This makes me think of a quirk I discovered recently which is that ChatGPT simply won't generate a picture of a 'full glass of wine'. It generates pictures with all sorts of crazy waves/splashes in the glass but the glass is always half full no matter how you prompt it.
I'm not enough of an expert to make any deductions from this, but I think it hints at what the limitations of the currently models are.
I'm able to do a so much more using LLMs (Mistral-Large, Qwen2.5 and R1 locally, Claude via API) than without them.
I have to get the IDE setup properly now.
I had a complex finance situation that I was struggling with, both from a mathematical/taxation perspective and a personal psychological finance hangup. I spent a few good hours talking to it through everything and had a mental breakthrough. To get the same kind of insight, I would have to pay a financial advisor AND a psychologist for several hours.
That all of this was free while someone calls it a "con" seems completely wrong
(I got my CFA cousin to look over the numbers and he agreed with R1's advice, fwiw)
Hopefully Mistral can copy their technique and give us a 123b reasoning model.
So now my company makes more money, and the work gets done faster, but I can't say I feel appreciative. I'm sure it's great for founders though, for whom doing the work is merely an obstacle to having the finished product. For me, the work is the end goal, because I'm not hired to own the result.
I do the fun bit: having creative ideas and trying them out.
For an incremental improvement...not great, not terrible.
It does what the other poster said: it automates the boring parts of "this db model has eight fields that are mostly what you expect" and it autocompletes them mostly accurately.
That's not something an IDE does.
You're either completely misremembering what IDEs have been able to do up until 3 years ago, or completely misunderstanding what is available now. Even the very basic "autocomplete" functionality of IDEs is meaningfully better now.
It would be so nice to have a productivity Linux OS that just works on all my devices without tinkering. I want to stop supporting the closed source monopolies, but the alternatives aren't up to par yet. I am extremely hopeful that they will be once mega corps inevitably decay and people tire of the boom-bust cycle.
As technologists, we all want beautifully designed tools, and I'm increasingly seeing that these are only created by passionate and talented people who truly care about tech, unlike megacorps that only care about enriching their board and elite shareholders.
For medium complexity things, I can get them done quickly without manual coding if I have a clear understanding in mind of what the implementation should look like. I supply the requirements, design and strategy and it's fairly easy to "keep things on the rails". The "write a PRD first" hack (https://www.aiagentshub.net/blog/how-to-10x-your-development...) works pretty well. Agent with YOLO mode and terminal access rips, particularly if you have good tests.
For tasks where I know the spec of the feature but don't clearly understand how I would design / implement the feature myself, it's hit-and-miss. Mostly miss for me.
I also haven't had much success with niche libraries, have to stick to the most popular library/tool/framework choices or it will quickly get stuck.
They need trillions of dollars in returns. VC's won't finance tech startups for decades.
I use Cursor sometimes, and VSCode + Continue with llama.cpp, and it's great. That's not worth billions. It's definitely not worth trillions.
Now someone will respond about how it's just a stepping stone, and how the billions are justified by _something completely imaginary, and not invented yet, and maybe not ever_ e.g. agents.
The BigTech companies have been flush with liquidity and poured those hundreds of billions into the promising tech, and as result we got a wonderful new technology. There is not much need for those trillions in return - just look at liquidity positions of those companies, they are just fine. If those trillions come in eventually - even better.
Whilst you are correct that big tech cos do not need the return to survive, that's not how public markets work at all, and thus not how the incentives for those in charge of the companies work, and so making you actually wrong.
If investment in AI don't pan out (i do think that it will pan out, and those trillions will come) then those companies would just pour even more billions into whatever big thing/promise would come next. Rinse and repeat. Because some of those things do generate tremendous returns, and thus not playing that game is what really constitute true loss of money.
US right now is run by someone whose explicit promises, if actually implemented, have an obvious immedidiate 13-14% reduction in GDP — literally, never mind side effects, I'm not counting any businesses losing confidence in the idea that America is a place to invest, this is just direct impact.
DOGE + deportation by themselves do most of that percentage. The tariffs are a rounding error in comparison, but still bad on the kind of scale that gets normal politicians kicked out.
And yet, the markets are up.
I just want to know so that I can set a reminder and check back on your comment when the time arrives.
I think if they could find a way to make their software good, instead of bad, like it increasingly is, that would be a good use of that money.
Just as they were convinced after Covid that they needed to put hiring into overdrive.
Tech management has the collective IQ of a flock of sheep.
The whole thing is like bitcoin. There’s too many people that benefit from maintaining the collective illusion.
But I do think (and better understand) there is a failure to understand this at a higher abstraction. One part is simply "money is a proxy." This is an uncontestable fact. But one must ask "proxy for what?" and I think people only accept the naive simple answer. Unfortunately, this "is a proxy" concept is extremely generalization. Everything is an estimation, everything is an approximation, and most things are realistically intractable. We use sibling problems or similar problems to work with that are concrete, but there are always assumptions made and ignoring these can have disastrous consequences. Approximations are good (they're necessary even) but the more advanced a {topic,field,civilization,etc} gets, the more important it is to include higher order terms. Frankly, I don't think humans were built for that (though by some miracle we have the capacity to deal with it).
My partner and her dad are both economists, and one thing I've learned is that what many people think are "economics questions" are actually "business questions". I think a story from her dad makes this extremely clear. A government agency hired him to look at the cost benefit analysis of some stuff (like building a few hospitals and some other unambiguously beneficial institutions), and when he presented everyone was happy but had a final question "should we build them?" The answer? "That's not the role of an economist." The reason for this is because money can't actually be accurately attributed to these things. You can project monetary costs for construction, staffing, and bills, and you can make projections about how many people this will benefit, how it can reduce burdens elsewhere, and as well as make /some/ projections about potential cost savings. But you can't answer "should you." Because the weight of these values is not something that can be codified with any data. It is an importance determined by the public and more realistically their representatives. Very few times can you give a strong answer to a question like "should we build a new hospital" and essentially in only the extreme cases. I'll give another example. In my town there was an ER that was closed due to budget constraints. This ER was across the street to the local university, which students represent ~15% of the population. The next nearest ER? A 15 minute ambulance ride away and in the next town over. Did the city save money? Yes. Did the sister city's ER become even busier? Also yes. Did people lose access to medicine? Yes. Did people die? Also yes. Have economists put a price on human life? Also yes, but they are very clear that this is not a real life and a very naive assumptions[1]. It is helpful in the same way drawing random squiggles on a board can help a conversation. Any squiggles can really be drawn but the existence of _something_ helps create some point to start from.
[0] okay crypto bros, you're not wrong but low volatility is critical as well as some other aspects. Let's not get off topic
[1] https://www.npr.org/2020/04/23/843310123/how-government-agen...
That seems like a suspect claim. If you're saying that you, personally, cannot create billions of dollars in value with Cursor & friends that is certainly true - but you are in no position to make a judgement call about where the cap on value creation is for the LLM market is worth based on your personal use cases. LLMs don't just do code completion. We really can't estimate how much potential value is being created without doing some serious data diving and studying of cases.
A better argument would be that the DeepSeek experience suggests these companies have no moat and therefore no way to earn a return on capital. But LLMs are probably going to generate at least trillions of dollars in value because they're on par or ahead of Wikipedia and Google for answering many queries then they also have hundreds of ancillary uses like answering medical questions at weird hours or creative/professional writing.
Consider that Wikipedia is much bigger than Encyclopedia Britanica, but because it is given away to everyone for free, it is not counted as E.B.'s max sale price ($2900 in 1989?) times the world's internet connected population (5.6e9?) — $16 trillion.
AI, regardless of value, are priced at the marginal cost to reproduce weights or run inference depending on which you care about.
But I do mean "reproduce" not "invent" — it doesn't matter if DeepSeek's "a few million" was only possible because they benefited from published research, it just matters that they could.
And if the hardware is the bottleneck for inference, that profit goes to the hardware manufacturer, not to the top ten companies who made models.
That is a problem for the VC’s that bet wrong, not for the world at large.
The models exist now and they’ll keep being used, regardless of whether a bunch of rich guys lost a bunch of money.
These companies are heavily subsidized by investors and their cloud service providers (like Microsoft and Google) in an attempt to gain market share. It might actually work - but this situation, where a product is sold under cost to drum up usage and build market share, with the intent to gain a monopoly and raise prices later on - is sort of the definition of a bubble, and is exactly how the mobile app bubble, the dot-com bubble, and previous AI bubbles have played out.
How? I get that many devs like using them for writing code. Personally I don't, but maybe someday someone will invent a UX for this that I don't despise, and I could be convinced.
So what? That's a tiny market. Where in the landscape of b2b and b2c software do LLMs actually find market fit? Do you have even one example? All the ideas I've heard so far are either science fiction (just wait any day now we'll be able to...) or just garbage (natural language queries instead of SQL). What is this shit for?
Various minor thing so far. For example I heard about ChatGPT being evaluated as a tool for providing answers for patients in therapy. ChatGPT answers were evaluated as more empathetic, more human and more aligned with guidelines of therapy than answers given by human therapists.
Providing companionship to lonely people is another potential market.
It's not as good as people at solving problems yet but it's already better than humans at bullshiting them.
I could see this being useful in a "dark pattern" sense, but only if it's incredibly cheap, to increase the cost to the user of engaging with customer support. If you have to argue with the LLM for an hour before being connected to an actual person who can help you, then very few calls will make it to the support staff and you can therefore have a much smaller team. But that only works if you hate your users.
Not since the advent of Google have I heard people rave so much about the usefulness of a new technology.
To make money though it just needs to have a large or important audience and a means of convincing people to think, want, or do things that people with money will pay to make people think, want or do.
Ads, in other words
A related question: has anyone figured out how to monetize LLM input? When a user issues a Google search query they're donating extremely valuable data to Google that can be used to target relevant ads to that user. Is anyone doing this successfully with LLM prompt text?
Microsoft turned itself into a trillion dollar company off the back of enterprise SAAS products and LLMs are among the most useful.
My company uses them for a fuckton of things that were previously too intractable for static logic to work (because humans are involved).
This is mostly in the realm of augmented customer support (e.g. customer says something, and the support agent immediately gets the summarized answer on their screen)
It’s nothing that can’t be done without, but when the whole problem can be simplified to “write a good prompt” a lot of use cases are suddenly within reach.
It’s a question if they’ll keep it around when they realize it doesn’t always quite work, but at least right now MS is making good money off of it.
No, it's not. The first half of the article talks about how useless the actual product is, how the only reason we hear about it is because the media loves to talk about it.
What it goes into is how over hyped and over valued these companies are. They've blown through $5bn of compute each in a year and their revenue is abysmal. Microsoft won't report on ai separately, probably because it's abysmal.
I'm positive on LLMs for coding. But I think I have to agree with their assessment. Coding seems like the best area for these tools and what we see now is great. It's probably even worth $10b to the IT industry maybe eventually. But they're not paying for it yet, clearly. And I also think it's just not going to have huge significance outside our industry. The people I rub shoulders with outside of work have not mentioned or asked about it once, which is not necessarily meaningful but it does reveal the limits of hype too.
While the timeline is unclear; it seems likely that LLMs will obsolete precisely the skills that developers use to earn their income. I imagine a lot of them feel rather threatened by the rapid rate of progress.
Pointing out that it is already operating at junior dev quality and rapidly improving is unlikely to quiet the discontent.
If you think LLMs operate at "junior dev" capacity you either don't work with junior devs and is just bullshitting your way around here, or you just pick pretty awful junior devs.
LLMs are alright. An okay productivity tool, although its inconsistencies many times nullify productivity gains - By design they often spit out wrong results that look and sound very plausible. A productivity blackhole. Its mistakes are sometimes hard to spot, but pervasive.
Beyond that, if your think that all a dev does is spit out code, and since LLMs can spit out code it can replace devs in some imaginary timeline, you are sorely mistaken. The least part of my work is actually spitting out code, although it is the part I enjoy the most.
I honestly feel way nore threatened by economic downturns and the looming threat of recession. The only way LLMs threaten me is by being a wasteful technology that may precipitate a downturn in tech companies, causing more layoffs, etc nd so forth.
If you base it on ability, then an LLM can be be more useful to a good developer than 1 or more less competent "junior" team members (regardless of their age).
Not because it can do all the things like any "junior" can (like make coffee), but because the things it can do on top of what a "junior" can do, more than makes up for it.
Code is liability. The knowledge inside developers' heads is the corresponding asset. If you just produce code without the mental models being developed and refined, you're just increasing liability without the counterpart increase in assets.
I’ve hired lots of junior devs, some of them very capable. I’ve been in this industry for more than 15 years. LLMs operate at junior dev capacity, that’s pretty clear to me at this moment.
I'm in a middle. I enjoy Zed and its predictions, I utilize R1 to help me to reason. I do _not_ ever want to stop programming. And I see so often whenever somebody less experienced than me shows me look how Cursor did this with three prompts, can we merge? And the solution is just wrong and doesn't solve the hard issues.
For me the biggest issues are the people who want to see the craft of programming gone. But I do enjoy the tooling.
These types of articles are just catching the next meme wave, which will be hating on and making fun of "AI" of all sorts.
I’m not particularly worried. I think it’s obvious that software engineering is definitely an “intelligence complete” problem. Any system that can do software engineering can solve any problem that requires intelligence. So, either my job is safe or I get to live through the fall of almost all white collar disciplines. There’s not a huge middle ground.
Although perhaps this is just the programmer stereotype of thinking that if someone can code, they can do anything.
How about the middle ground where a human using AI replaces you?
The human job is (maybe) safe, but your job?
Developer productivity has gone up immensely in the last 50 years and the industry is larger than ever.
Disclaimer: I’m the developer behind CodeBeaver
The author spends a good amount of bytes telling us that they don't want to hear this argument even though they expect it.
Unless closed models have significant advantage AI inference will be a commodity business - like server hosting.
I'm not sure that closed models will maintain an advantage.
Even the US Government is getting involved in subsidizing these companies and all of the infrastructure and resources needed to keep it expanding. We can look forward to even more methane power plants, more drilling, more fracking, more noisy data-centres sucking up fresh water from local reserves and increased damage to the environment that will come out of the pocket books of... ?
Update: And for what? "Deep Research"? Apparently it's not that great or world-changing for the costs involved. It seems that the author is tired of the yearly promise that everything is just a year or two away as long as we keep shovelling more money and resources into the furnace.
Personally, I find that waiting for the code to generate, then reviewing the code carefully, then deciding if I need to rewrite it to be more painful, more error prone, and much slower than writing the code correctly.
Especially since this AI junior never learn from it’s mistakes.
I think it speaks to different approaches to how individuals write code.
> How great is it gonna be in a year or two?
I would bet that it’s about the same (not great code, generally), but the tools fail to generate responses less often and likely would have more context.
Hopefully they become fast enough to run offline or at least feel more instantaneous.
I happen to value human creativity.
Who? How? This is not what I've seen where I work. There's a bunch of hubbub and generalized excitement, and lots of talk about what could be done, or what might be done, but not very much actual doing. I must just be a clueless "mid".
Guido van Rossum - "I use it every day. My biggest adjustment with using Copilot was that instead of writing code, my posture shifted to reviewing code." https://www.youtube.com/watch?v=-DVyjdw4t9I
Here's Jeff Dean saying 25% of the characters in new PRs at Google are AI Generated. https://www.dwarkeshpatel.com/p/jeff-dean-and-noam-shazeer
Andrej Karpathy - "I basically can't imagine going back to "unassisted" coding at this point" https://www.reddit.com/r/singularity/comments/1ezssll/andrej...
Andrew Ng - "I run multiple models on my laptop — Mistral, Llama, Zefa. And I use ChatGPT quite often. " https://www.ft.com/content/2dc07f9e-d2a9-4d98-b746-b051f9352...
Simon Willison https://simonwillison.net/2023/Mar/27/ai-enhanced-developmen...
I mean I can keep going. I doubt you would compare yourself to these people.
Anyways, here goes.
1. Guido uses Copilot like I do - as a StackOverflow replacement to write the dumb boilerplate code. A much less flattering quote is "It doesn't save me much thinking, but [it helps] because I'm a poor typist". Also it's literally a minute or two of a three hour podcast.
2. A lot of code is autogenerated lol. Again, it's all the boring boilerplate stuff.
3. The cofounder of OpenAI is a biased source lol
4. He's an AI researcher, of course he runs that stuff.
5. Again, similar to Guido. He's using it for the boilerplate. Nothing wrong with enjoying using it as a toy, as he is here. But he's not doing serious work with it.
There's no virtue in hyping this stuff like a HODL bitcoin cultist.
>here's a bunch of hubbub and generalized excitement, and lots of talk about what could be done, or what might be done, but not very much actual doing
I am showing that indeed many top people use the tools to make themselves more productive, in direct contradiction to the comment above.
Listen: it's a fun toy. Engineers love toys and shiny distractions.
Don't confuse shiny rocks for gold.
Keep thinking that and don't feel too bad when 21 year old zoomers are 10x more impactful than you are at work.
1- he states that the generated code is most likely wrong. He is appreciative of it though because he is a very poor typer so he doesn't have to do that part as much
2- so that's not supporting your argument that the 'top' devs are using it. Besides it doesn't say how it's counted, nor how much time is spent reviewing and correcting it
3- actually okay. But is he using it for production code? Doesn't say
4- he definitely doesn't talk about coding, only brainstorming and writing text.
5- your best one. Still, the use case here is side projects not production
You might still be right, I definitely do not compare myself to these people, but trying to glue some sources together makes a poor argument.
And the subject on hand is more that just using LLMs, it's the role of LLMs in the dev work environment
It makes a better argument then the bold and plainly wrong claim that no one is using them and its all just "a bunch of hubbub".
None of the 5 sources say that AI code generation is really making them more productive
There's a Quentin Tarantino quote where he says there are 2 kinds of film critics. There are those who love movies and there are those that love the movies they love.
A lot of developers really seem to love the technology they love.
These people are where most of the negativity is coming from. And my guess is that the people who are encouraged by LLMs and not negative (mostly) aren't taking time out of their days to write long blog posts or argue about it online.
It's not the technology, it's the stupid overhype. It really feels like all the HODL bitcoin cultists have finally gotten over their lost apes and found a new technogod.
So many people in these threads are convinced it's about to gain sentience. That's not going to happen. You get the people outing themselves by saying "it does my job better than me!"
If you say something honest and direct like "their output is mediocre and unreliable" or "the RNG text generator is not capable of thinking, you're just Clever Hans-ing yourself" or "if it does your job better than you, that says more about you than about it", you get people clutching their pearls like a Stanford parent whose kid got a D.
arXiv has turned into a paper mill of AI startups uploading marketing hype cosplaying as "research".
So it's from middle management levels riding the hype train, and possibly trying to save money and getting bonuses for it at the expense of other people.
Just like when offshoring was in its same point on Gartner's hype curve.
"Everyone has a model, but no one has a business".
This is the only question, and the fact it's still an open question just screams "hype bubble". My bet is this AI stuff goes the way of the NFT.
OpenAI doesn't have this problem. ChatGPT has a free level to get you hooked, but it's restricted. So a lot of users pay them $20 or $200 or some other amount per month to use their service. So how OpenAI makes money is by selling access to their service. What you do with it is up to you, but their value proposition is simple. Pay us to get more/better access to our service.
How much it costs them to operate the service is a secret known only to them. There are a lot of very very educated guesses, but they're just guesses. After the VC money runs out they'll have to charge more than it costs to provide the service to stay afloat, and then we'll see. $20/month for ChatGPT plus is the $1 Uber that got people hooked. There's already a $200/month tier.
Whether OpenAI, specifically, will be standing in 20 years, only time will tell. But by this point it should be obvious that there's something to this LLM thing. Even if the product doesn't get any better than it is today, it'll still take 5-10 years for its effects to reverberate through society.
The killer app is LLM-accelerated programming. Sure, it doesn't work for all domains and it can't do everything, but even if the only thing it's good for is creating JavaScript react CRUD apps, well, there are a lot of those out there, and they're not actually limited to that. And since tool use means they can generate code and compile it and test that it works, it's possible to generate datasets for other languages and libraries, the only question is which ones is it worth it for.
It might not help at all in your line of work, but a friend who does contracting is able to use LLMs to cut the time it takes him to do a specific kind of job in half, if not more, enabling him to take on twice as many clients and make more money. For him it would still worth it even at 100x the current price. thankfully competition means it'll take a while before it's that expensive.
> It sure is! But it doesn't really prove anything other than that people are using the single-most-talked about product in the world. By comparison, billions of people use Facebook and Google. I don't care about this number!
> User numbers alone tell you nothing about the sustainability or profitability of a business, or how those people use the product. It doesn’t delineate between daily users, and those who occasionally (and shallowly) flirt with an app or a website. It doesn’t say how essential a product is for that person.
Both of these "arguments" could be applied to any of the big tech giants of the last 25 years - Google, Amazon, Facebook, Uber, whatever (and there'd be other incumbents used by billions of people before them). I don't believe these arguments discount ChatGPT from having the potential to continue growing like a Facebook. And who cares how many journalists Altman knows, you don't get a product written about that much unless it's truly a groundbreaking product.
> And even then, we still don't have a killer app! There is no product that everybody loves, and there is no iPhone moment!
There sure is, it's called programming. He called out quality earlier on, but the quantity and speed and direction the AI can take (as well as its rate of improvement) is breathtaking. My own output has 10x'd easily since GPT-4 came out (although some of that means I'm needing far less hours in certain places). And guess what? The code quality is generally fine.
> Where are the products? No, really, where are they? What's the product you use every day, or week, that uses generative AI, that truly changes your life? If generative AI disappeared tomorrow — assuming you are not somebody who actively builds using it — would your life materially change?
Ok, the product is called ChatGPT, or Claude, or DeepSeek or whatever, and if it disappeared overnight, my programming productivity would drop dramatically. I would not seek to take on as ambitious projects in as short of a time frame as I am doing now.
I don't know, as a user and developer both of AI/LLMs, this article isn't hitting the mark for me. There are legitimate criticisms of the field, but I'm not seeing them thus far.
Edit - I'll say I agree with the Deep Research criticisms. These products are very underwhelming. They're literally to help people do a research report which needs to be done, but won't be used or read critically by anyone report.
I haven't shelled out the $200/month for OpenAI's Deep Research offering, but similar products from Google and Perplexity are extremely useful (at least for my use case). I would never present the results unchecked / unedited, but the Deep Research products will dig much deeper for information than Perplexity could be persuaded to previously. The results can then be fed into another part of the process.
I don't get these statements. This line of thinking is so egregious, FTX made an ad about it. This is survivorship bias. If 'wrong' predictions about some can be discrediting, then why not right predictions to establish credibility? The same people were right about the Segway, crypto, metaverse, web3, that dog walking startup that Softbank burned money on, and countless other harebrained endeavors.
Even in your example, Uber is a financial crime. It is still using accounting tricks to show profit.
All of the valuable uses are personal. It makes me personally feel more productive at work. It helps me personally understand some topic better. It gives me an idea in a personal project.
That's all really cool, but that is not what the valuation is about. The valuation is about a false science fiction and hype bubble about agentic this or that or AGI or whatever, and this is driving very questionable decisions for wasting possibly trillions of dollars and tons of energy.
The plus side is that there is some really cool personally useful tech here, and we will probably end up with very good open source implementations and cheaper used GPUs once the bubble bursts.
On the other hand, today most money for SWEs goes to people who aren't "staff engineer" level. What if most money went to staff engineers who directed these interns and paired with new human apprentices to learn staff eng?
In the past few thousand years, apprentice/journeyman/mastery of trade was how trades worked, with an "up or out" model that kept the role pyramid the right shape for skill to shape the outcomes.
These days, far too many careers stay apprentice skill the entire career, mostly thanks to enterprises failing at engineering management as a skilled trade, so being unable to value and raise staff engineer caliber contributors. The enterprise learning machine is broken by false "efficiency".
LLMs change this. Staff engineer caliber SWEs are able to direct these indefatigable assistants, as if each staff engineer has a team of 10 who never need mental breaks to remain productive. There will of course be some number of junior devs who themselves have enough affinity for the role they will want to stick with the apprentice model and work to the staff engineer level. (And will always be solo or boutique teams of app/saas SWEs.)
As for the enterprise engineering management that couldn't tell the difference between a staff engineer and an apprentice, the LLM multiplies the difference to the point the outcomes are evident even to a non technical observer.
So one possible timeline for this is a raising of the median human skill level by attrition of those unskilled enough or unable to think critically enough to leverage the machine assistants as force multipliers or unable to survive directly mentored skill-up training and observation from the staff engineers.
You talk about personal value. Roughly, I agree with you completely, and am adding who I think those persons could be (or have to be given the current level of "thinking" by these tools) for the hype to deliver on value. (At a higher level of machine, closer to "AGI", this scenario changes.)
As is evident by downsizing a mediocre team and observing output go up and work more reliably, these forces could, if playing out this way, make dollars per human go up, productivity go up, quality go up, and enable a return to the millennia-proven model of apprenticeship for the trade.
I know very little about ChatGPT, but Waymo is using an LLM: "Powered by Gemini, a multimodal large language model developed by Google, EMMA employs a unified, end-to-end trained model to generate future trajectories for autonomous vehicles directly from sensor data." (https://waymo.com/blog/2024/10/introducing-emma)
Emma is something they tried, but further down the article they explain why they don't use it as such yet.
https://wayve.ai/thinking/lingo-2-driving-with-language/
And there are state of the art weather prediction transformers.
https://arxiv.org/abs/2312.03876
The reason I want to specifically harp on this is because a lot of people are selling this narrative where "AI is becoming superintelligent" or whatever by making an amorphous blob out of a bunch of separate advances that use machine learning techniques. This has been happening for a while, is a great thing for science, and it's clear that machine learning methods are here to stay in science. I'm a machine learning researcher. I've understood, celebrated, and tried to help with this as best I can manage over the last 9 years of my life. And it's been going on for a lot longer than the general public has been in this AI hype wave. The entire modern field of bioinformatics is arguably built on the backbone of machine learning, and has been since before I went to grad school.
This is really different from "We fed everything into a language model and now it's superintelligent and is making scientific advances all by itself" or even "scientists just ask chatGPT shit and it figures it out for them". The breathless tech press really makes it sound like anything that happens in AI research, which increasingly includes the entire usage of ML toolkits in the sciences (Which is pervasive, and expectedly so! ML is an extension of statistics and statistics has been the basis of science for like a century) is just some amorphous force called "AI" that's suddenly gained this aggregate body of competency. Imagine if we anthropomorphized statistics that way. Or Math for that matter. This kind of narrative gives me the overall impression that this is not being talked about honestly, and it's clear that this is profitable to do. I don't have to use charged words like "con" or "fraud" to think this deceptive framing is not a great thing
Yes, one of the powerful things about every kind of neural network is that they're a very general class of function approximator. That we can use a similar toolkit of techniques to tackle a wide variety of problems is very cool and useful. Again, the analogy to statistical models is telling. You can model a lot of phenomena decently well with gaussian distributions. Should we report this as "Normal distribution makes yet another groundbreaking discovery!"? Probably this wouldn't have the same impact, because people aren't being sold sci-fi stories about an anthropomorphized bell curve. People who are using LLMs already think of "AI" as a thinking thing they're talking to, because they have been encouraged to do that by marketing. Attributing these discoveries made by scientists using this method to "AI" in the same way that we attribute answers produced by chatGPT to "AI" is a deliberate and misleading conflation
Maybe if that was the only different but it's not. There are diffusion models that have nothing to do with transformers or attention or anything like that and where using them for arbitrary sequence prediction is either not possible or highly non-trivial.
Yes, All Neural Network architectures are function approximators but that doesn't they excel equally for all tasks or that you can even use them for anything other than a single task. This era of the transformer where you can simply use a single architecture for NLP, Computer Vision, Robotics, even reinforcement learning is a very new one. Literally anything a bog standard transformer can do is anything GPT can do if Open AI wished.
Like i said, i don't disagree with your broader point. I just don't think this is an instance of it.
Maybe some specially trained weather prediction neural network could perform well, but that isn’t really what we’re talking about here.
Of course what can be mathematically calculated without inference is going to be. LLMs may however be able to interpret the results of these calculations better than humans or current stochastical evaluations.
It is not clear to me why the author feels the need to have the conversation.
Human consciousness gives us the ability imagine future states in the universe and make them come true.
The results will speak for itself.
"I need you to fucking listen to me: everything I am describing is unfathomably dangerous, even if you put aside the environmental and financial costs."
Personally I think he's lost it a bit. I mean say he's right in that LLMs plateau and investors lose some money. Life will go on.
Which then invites the question: why is this person's opinion on the subject relevant? Do they have some credentials that make it more valuable than a random comment with a similar rant (of which there are plenty) on Reddit or HN?
Oh sure, because you can just have hundreds of reporters constantly write about your product. It's so simple. Why aren't more people thinking of that ?
>The weekly users number is really weird. Did it really go from 200 million to 300 million users in the space of three months?
According to similarweb, monthly visits grew over 1B in that timeframe so yeah sure it sounds possible.
>300 million monthly active users would mean a conversion rate of less than 4%, which is pretty piss-poor
A B2C Saas whose lowest price point is $20 will be lucky to get anywhere near 4% conversation.
>And even then, we still don't have a killer app!
The 6th (and climbing) most visited site in January is not a killer app ? Okay
I shocked he really believes it in his closing thought.
Maybe his rant would be bit more digestible if it contained sections with: "here's what I tried and it did not work". But that would make it not a rant but actual research with value.
Closer? I'm just going by the headline here.
But yeah, I don't think LLMs (the current core architecture) can provide super intelligence. I think it needs a bit more than next token prediction architecturally speaking.
- maybe ALL humans would fail the test in some way, eg. let's say everybody gets at least 10 of those wrong, and the average person gets 100 of those wrong.
- still, as long as most people correctly get each word right, your LLM would get every single response correct (because for each item in the test, 900+ people out of a thousand gave the same correct answer in the training set).
In that sense, it's totally possible for a system trained on a vast vat of average-human input to generate super-human outputs.
then our Riemann-Goalpost hypothesis is that ζ has zeros for every "s" which is a negative integer (every breakthrough that happened in the past is null in value) and only has positive values where s is positive.
We can conclude from the above that given a far enough date in the future, any given breakthrough can be spectacular, but once achieved, will be derided as trivial.
general API plumbing ? call this api, combine the json results & spit it out. Then slap a React / Next.js frontend ?
lately, I have been doing your classical business apps - due to the domain rules - ai is pretty useless - but I have found perplexity and deep mind to be smarter stack overflows. that's it.
For anything even remotely complex or involved, it's just not that useful. Interestingly, It seems to shine the most when doing boilerplate stuff in widely used languages such as python or javascript, but it's extremely bad with terraform.
Its basically more time-consuming for me to get it to where I want the code than finish it myself. For some scaffolding in python green-field development, sure I save a few minutes.
Open source projects would have a lot more compute to work with.
But it's not like people are going to throw out all the Nvidia hardware they bought.
And there are ai applications that I can think of that would be viable at 100x cheaper price.
1. The Generative AI is very very useful
2. The Generative AI is a epic bubble and it will kill us all (financially) one day
It's entirely appropriate to describe it as a 'con', as long as you look deep enough into OpenAI, SoftBank, MSFT, NVDA, SMCI, etc.
it's a con, but it's useful, and it will kill us all
Is the AI "business" or "market" overvalued for it's current capabilities? Yeah, I do believe so. Welcome to the financial world, which is completely separated from reality. It's like that in all sectors where something new and exciting is happening, not just IT or AI. People poor money in hoping to be early enough to make a profit. Nothing more, nothing less. The rest is marketing. Some Sam Altman guy promoting the hell out of his own product? That is literally his job, regardless of wether or not he believes it all.
But articles like these are so bizarre to me. The author acts like he has millions at stake and his money manager just won't listen and pull all investments out of AI. Hurry up, the bubble is about to burst, I will lose all my money!
Except that... they don't. They are just "old man yelling at cloud". If you believe AI is the next Metaverse or WeWork, then it will just die off by itself once the bubble pops. Why are you having so many conversations about it, where you seem to be desperately trying to convince people of the bubble/con that is AI. To the point that you're so sick of it, that you write down your arguments so you can point the blinded there instead of having those tiresome arguments.
Genuinely baffled. Spend your energy on something productive rather than destructive, perhaps?
But why does he act or feel that way? Let the trillions be lost, it's just how hypes, bubbles and the stock market in general work.
Haven’t read any news in the past 20 years? This is all going to be funded by the American taxpayer
Because that's what AI wa supposed to be in the first place. But the industry performed the swindle of renaming "AI" to "AGI", so that they can pretend the thing that exists now is "AI".