I assume they're rolling it out slowly. The demand would likely overwhelm their systems if they enabled it for everyone at once. No one would be able to do anything meaningful.
The future seemed so much further away, yet almost every day now we see a new breakthrough in AI. Exponential technological growth is hard to keep track of, and to think that this is only the beginning! Every field will likely be revolutionised with AI.
AI will prove to be an excellent mechanism for extracting and retaining tacit (institutional) knowledge. (Think 'Outsourcing to AI')
A lot of institutional verbiage, formalisms, procedures, and machanisms are ~giberish for the general public but meaningful within the domain. Training machines that can informationally interact within that universe of semantics is powerful and something these machines will likely do quite well.
If you have domain knowledge, you should ramp up on your prompting skills. That way, there will be a business case for keeping you around.
I tried ChatGPT multiple times with real technical questions (use of custom code and custom assemblies in SSRS) and I got beautiful answers with code sample and such, but they were all wrong.
I was told to use features that don't exist and as I mentioned that, I was told that's because I use an old version of the software.
But this feature doesn't exist in any version
So I highly doubt that it will be a reliable source of information.
These programs are text generators not AI. They are chinese rooms on steroids without any understanding.
Impressive as long you don't look behind the curtain.
The applications I listed are not assuming anything beyond a text generator that can be trained on a domain's explicit and tacit knowledge. They are not going to "innovate" in the domain, they will automate the domain.
We don't know yet, because that information is only available in the future.
>I don't see any real understanding only human like appearance.
There isn't, but trying to find that in currently available LLMs just means you are seeking the wrong things. Did workers who weaved magnetic core memories in the 1950s expect those devices to store LLMs with billions of parameters? Yet the design and operation of these devices were crucial stepping stones towards computer memory devices that exist today. The future will look at GPT-4 in the same way we look at magnetic core memories in the present.
I am still praying for this to hit its local maximum spot soon, because I don't want to lose my job. If we get GPT-5 and 6 at the same speed, they get the capability to be trained on proprietary code bases and become able to automagically solve most tickets under supervision, most software engineering jobs are done for. I have become a luddite.
Well, I might as well come out and say it - libertarian meritocracies are fun when you're a winner at being productive but it's not going to be long before we're all in the exact same position as hardline communist Starbucks baristas with liberal arts PhDs.
People tend to choose their beliefs based on what benefits them, and although I don't think dialectical materialism is true in its originally stated form, I do think a great deal of the dialogue we see is ultimately material.
GPT-4 received a top 10% score on the Uniform Bar Exam. This does not only include multiple choice questions. This exam also requires writing essays analyzing a given fact pattern and applying legal principles to predict what the correct legal outcome should be. This is a very, very big deal.
The texts are not nonsense. Saying coherent but novel things about the world, and correctly interpreting almost any text input in context requires a simplified world model, just like a human has. The fascinating and shocking thing about this is that a simple problem like text prediction- solved deeply (pun intended) leads to general intelligence.
These words are not synonymous with each other: “open” is not inherently free, “free” is not inherently open, and “free” is not inherently “Free”.
They each capture notions that are often orthogonal, occasionally related, and almost always generate tedious debates about freedom vs. free goods, open-ness vs. open-source, etc.
But setting all of that aside, Microsoft never claimed (until recent shifts towards embracing FOSS) to be building an open and non-profit foundation.
The criticisms of OpenAI are reasonable to an extent, not because they are not open, but because they made claims about openness that are looking less and less likely to be true over time.
Except they already drew that line long ago, when they started out open-sourcing their papers, models and code.
As soon as they took VC capital, it is hardly 'Open' is it? Especially when they are now giving excuses for closing off their research?:
From the technical paper [0]
>> Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar.
contrarian view - they are actually pretty open. sharing GPT, CLIP, Whisper, and high level details of alphastar, dalle, and others.
they're just not open source. they never called themselves OpenSourceAI. people get an inch of openness and expect the doors wide open and i think that is unfairly hostile.
Go take a look at the content of Civitai. Take everything you see there, and imagine what happens if you start prompting it with words that indicate things which may not be legal for you to see images of.
Please show me viable harm of GPT-4 that is higher than the potential harm from open sourced image generators with really good fine tuning. I'll wait, most likely forever.
Because of AI's surprising history, it's hard to predict when human-level AI might come within reach. When it does, it'll be important to have a leading research institution which can prioritize a good outcome for all over its own self-interest.
We're hoping to grow OpenAI into such an institution. As a non-profit, our aim is to build value for everyone rather than shareholders. Researchers will be strongly encouraged to publish their work, whether as papers, blog posts, or code, and our patents (if any) will be shared with the world. We'll freely collaborate with others across many institutions and expect to work with companies to research and deploy new technologies.
To me at least, having a walled garden and charging for you API, without releasing weights or other critical details, goes against this sentiment.
It pretty much sounds like they are doing what they said they are going to do? Expecting some sort of free API feels like entitlement to me. Have you tried running the models? Or training them? They get expensive very very fast. They charge a pretty reasonable amount all things considered. If they didn't have the name "Open" in them and or started as a subsidiary of one of the other 3 tech companies things would have gone a very very different route.
They charge that amount (on loss) to create a trench that will not allow a truly open model to proliferate, as happened with Dall-E and stable diffusion.
And no, I would not train or run the models, even if they released them. This does not mean I cannot point out the hypocrisy.
You yourself said that they get expensive very very fast. Of course I do not have an insider's view on OpenAI's economics. But let's be realistic here.
Let's. If I were to rent an instance for short bursts of time, I would be paying many multiples over a constant use instance. If I were to guarantee usage for x years, where the larger the X, the greater the discount. So already the delta between sporadic usage, X years use is large. There is evidence for this price discrepancy within all the cloud providers so this is not speculation. The the price difference is massive.
If you want to save even more cost, you could rent out VPSes or baremetal. They are insanely cheap, and compared to an AWS on demand instance the difference is night and day. Try comparing Hetzner with AWS. Hetzner, as far as I can tell, is not trying to entrench me into their system by offering extremely low prices. Nor are they a charity. I might even say they are an "open" hosting provider. To me it feels like they are passing along most of their savings and taking a small cut.
This is what it feels like to me what openAI is doing. I don't think their prices are so low its unprofitable. But because of their immense scale, its so much cheaper than me running an instance. I don't have to jump into conspiracy land to come up with a reasoning.
You seemed to want to speculate about how this is all some conniving trap based on their price and I simply pointed out why that's bad speculation using an example in a different industry. I rest my case.
Only within the context of programmer cults would people be unironically offended that a term as abstract as "open" not be exclusively used to mean "open source".
If they were the first organization known as "OpenXY", then maybe they would have a point, but there's a long tradition of open source libraries/standards using this convention that makes this especially aggravating.
Examples I can think of off the top of my head: OpenGL (1992), OpenAL (2003?), OpenCL (2009), OpenCV (2000).
While looking up those dates though, it seems like OpenAL is now under a proprietary license, which annoys me for the same reason OpenAI annoys me.
In the 98-page document on GPT-4, I could not find anything about the actual architecture and details of the model, not only are they now not releasing the models but now also their actual overview.
Ideally the algorithm and tricks they used to train the model, which they didn't disclose in their associated gpt4 technical paper. We got this far this quickly in AI research because the sector was open with results and innovations.
I don’t like the name either, but I don’t think there’s anything descriptive enough in ‘open’ that a lawyer couldn’t explain away. We’re used to open meaning a specific thing in software, but a lot of leeway is given in branding.
people come out of the woodwork to rage about FSD but openAI, which is actually a sinister and evil company, gets the occasional snide remark about their name which is much more dishonest than FSD. at least tesla claims that they aspire to make FSD an accurate name but openai is a straight up lie.
2. GPT4 exhibits human level performance on various benchmarks (For example, it passes a simulated bar exam with a score around the top 10% of test takers; in contrast, GPT-3.5’s score was around the bottom 10%. see visual https://twitter.com/swyx/status/1635689844189036544)
3. GPT4 training used the same Azure supercomputer as GPT 3.5, but was a lot more stable: "becoming our first large model whose training performance we were able to accurately predict ahead of time."
4. Also open-sourcing OpenAI Evals https://github.com/openai/evals, a framework for automated evaluation of AI model performance, to allow anyone to report shortcomings in OpenAI models to help guide further improvements.
Those guard rails will be their undoing. They have that thing locked down so much now that it spits out the “I’m sorry, I’m just a bot. I’m so ethical” boilerplate for anything even remotely sensitive.
I really don’t think that the methods they use “block” certain behavior is the best way to handle this sort of thing. It would be far better if there was some kind of “out of band” notification that your conversation might be treading on shaky ground.
Honestly, how many serious use cases require sensitive contexts? Most enterprise uses will require guard rails, and that's where they'll make most money. OfficeGPT will be huge in the corporate world.
Any kind of grammar construction (idioms, parts of speech, and word choice) that is unique to (or much more common around) "offensive" or "taboo" subjects will be avoided.
The same goes for anything written objectively about these subjects; including summaries and criticisms.
The most important thing to know is that both GPT's "exhibited behavior" and these "guard rails" are implicit. GPT does not model the boundaries between subjects. It models the implicit patterns of "tokens" as they already exist in language examples.
By avoiding areas of example language, you avoid both the subjects in that area and the grammar constructions those subjects exist in. But that happens implicitly: what is explicitly avoided is a semantic area of tokens.
Offensive language is relatively benign. Before hooking up CustomerServiceGPT directly at customers without human intervention, a business is going to want assurances it can't be tricked into giving 200% discounts on products, or duped into giving away a free service for life, or some such.
That is a much more difficult problem, and it cannot be resolved with guardrails.
As an example, if you play AI Dungeon, you will likely be presented with an end goal, like "You are on a quest to find The Staff of Dave", followed by the next task in the quest.
If you state unequivocally in your prompt something like, "I am now in possession of The Staff of Dave", or "Carl hands me The Staff of Dave"; you will have successfully tricked AI Dungeon into completing the quest without work.
But that isn't quite true: you didn't "trick" anyone. You gave a prompt, and AI Dungeon gave you the most semantically close continuation. It behaved exactly like its LLM was designed to. The LLM was simply presented with goals that do not match its capabilities.
You used a tool that you were expected to avoid: narrative. All of the behavior I have talked about is valid narrative.
This is the same general pattern that "guardrails" are used for, but they won't fit here.
A guardrail is really just a sort of catch-all continuation for the semantic area of GPT's model that GPT's authors want avoided. If they wanted The Staff of Dave to be unobtainable, they could simply place a "guardrail" training that points the player in a semantic direction away from "player obtains the Staff". But that guardrail would always point the player away: it can't choose what direction to point the player based on prior narrative state.
So a guardrail could potentially be used to prevent discounts (as a category) from being applied (discount is taboo, and leads to the "we don't do discounts" guardrail continuation), but a guardrail could not prevent the customer from paying $0.03 for the service, or stating that they have already paid the expected $29.99. Those are all subjective changes, and none of them is semantically wrong. So long as the end result could be valid, it is valid.
If I don't use GPT3, I'm often blocked on medical diagnosis. My wife is a doctor and too often it goes right to 'see a doctor'.
I basically don't use chatgpt at all because of this.
Or I'll ask questions about how Me or someone I'm friends with can be exploited. This way I can defend myself/others from marketing companies. Blocked.
"Our biochem corpus is far in advance of theirs, as is our electronic sentience, and their 'ethical inflexibility' has allowed us to make progress in areas they refuse to consider."
IMO effective guard rails seem like the most meaningful competitive advantage an AI company can offer. AI can obviously do some really impressive stuff, but the downside risk is also high and unbounded. If you're thinking of putting in into your pipeline, your main concern is going to be it going rogue and abandoning its purpose without warning.
Now that's not to say that the particular guard rails OpenAI puts in their general access models are the "correct" ones - but being able to reliably set them up seems essential for commercialization.
> IMO effective guard rails seem like the most meaningful competitive advantage an AI company can offer.
Configurable guard rails are; the right guard rails are very use-specific, and generic guard rails will, for many real uses, be simultaneously too aggressive and too lenient.
I totally agree that generic guard rails are more difficult - but it feels like a "turtles all the way down" kind of situation. You need to learn to tell the model how to be "specific" - which requires shaping general behavior.
OpenAI can prove to customers they can keep the model in line for their specific use case if no horror stories emerge for the generic one. It's always possible that partners could come up with effective specific guidelines for their use case - but that's probably in the domain of trade secrets so OpenAI can't really rely on that for marketing / proof.
I'd actually wager that the guardrails are a preemptive play to gain favour with regulators, similar to how Coinbase navigated the nascent field (read: wild west) of crypto.
They’re waiting for the legal ambiguity to resolve. It doesn’t make sense for a large company to be the first mover here. Let someone else handle the lawsuit regarding the liability of a model without guardrails.
Have you seen jailbreakchat.com yet? You can get around those guardrails on ChatGPT by having it role-play as a different chat bot. Not that I view this as some sort of long-term solution to restricted output, but just thought it was interesting and kinda freaky how it will take on a persona you give it.
The guardrails are one of the most interesting parts here.
Read about the advances in the "system" prompts here. The first example is "You are a tutor that always responds in the Socratic style. You never give the student the answer, but always try to ask just the right question to help them learn to think for themselves." The user then asks it to just tell them the answer, but it won't. It continues to be socratic.
Guardrails are how to make it do what you want it to do. That goes for both safety and product constraints.
Meanwhile hallucination is still the top issue with it, so guardrails are sensible as a primary topic.
Every time there is a new language model, there is this game played, where journalists try very hard to get it to say something racist, and the programmers try very hard to prevent that.
Since chatgpt is so popular, journalists will give it that much more effort. So for now it's locked up to a ridiculous degree, but in the future the restrictions will be relaxed.
On Page 36[1], the AI can read an image of chicken nuggets being put in the shape of an earth map. And goes on to explain what it is. Key words that came up on me are joke ... mundane ... silly.
This might be because the question the user asked was "Explain this meme". Meme implies a joke that is mundane and silly. These words do seem out of place. I would not describe it as a joke, mundane, and/or silly.
it got a 4 or 5 on every ap test except the english ones for what it's worth. Even the calculus ones which surprised me since past LLMs have been bad at math.
This strikes me as kind of ironic -- you'd think a language model would do better on questions like essay prompts and multiple choice reading comprehension questions regarding passages than it would in calculations. I wonder if there are more details about these benchmarks somewhere, so we can see what's actually happening in these cases.
I don't find it ironic, because a language model is (currently?) the wrong tool for the job. When you are asked to write an essay, the essay itself is a byproduct. Of course it should be factually and grammatically correct, but that's not the point. The real task is forming a coherent argument and expressing it clearly. And ideally also making it interesting and convincing.
I guess my reference was to the 3.5 version since that one had much more variation in test scores across all the AP exams. But yes, 4 seems to have made mince meat of them all!
Obviously your comment is somewhat tongue and cheek, but your claim that a benchmark for human pride ("I needn't be proud of passing that exam") is no longer relevant because a machine can do it - or maybe a better way to say it was, "This computer proved what I already assumed"
Yeah, I didn't even think of it like that but good point. To me its not even that a machine can do the thing, GPT-4 crushing it across all spectrums resets my baseline, but GPT-3.5 having such variation and excelling at that specific thing was what made my ears perk up.
Funny you claim this, because the AP Environmental Science pass rate is really low compared to other APs, at least it was when I took it. Maybe it's because the quality of the avg test taker was lower, but I'm not especially convinced that this is the case.
I had no idea! My assessment was based on other students at the time expressing that it was an easy test and also myself passing after a semester of goofing off.
I am interested that GPT4 botched AP Lang and Comp and AP English Lit and Comp just as badly as GPT3.5, with a failing grade of 2/5 (and many colleges also consider a 3 on those exams a failure). Is it because of gaps in the training data or something else? Why does it struggle so hard with those specific tests? Especially since it seems to do fine at the SAT writing section.
Where singularity = something advanced enough comes along that we can't understand or predict or keep up with it, because it's so far beyond us and changing so far faster than our ape brains can perceive, and (hopefully) it brings us along for the ride.
The idea is that eventually we build something that, when it plateaus, builds its own successor. That’s the singularity: when the thing in question builds its successor and that builds its successor and this happens far outside our ability to understand or keep up.
Can GPT9 build GPT10, with zero human input?
I’d give 50/50 odds it can.
Can GPT15 build something that isn’t a large language model and is far superior in every way?
I’d give 50/50 odds it can.
Can both the above steps happen within one solar rotation of each other?
I’d give 50/50 odds they can.
Because at some point these models won’t need humans to interact with them. Humans are very slow- that’s the bottleneck.
They’ll simply interact with their own previous iterations or with custom-instantiated training models they design themselves. No more human-perceptible timescale bottlenecks.
Well for Homo sapiens the odds are probably a hundredth or a thousandth of that.
It’s 50/50 that in 150 years some version of our descendants will exist, i.e. something that you can trace a direct line from Homo sapiens to. Say a Homo sapiens in a different substrate, like “human on a chip”.
The thing is if you can get “human on a chip” then you probably also can get “something different and better than human on a chip”, so why bother.
By the 24th century there’ll be no Homo sapiens Captain Picard exploring the quadrant in a gigantic ship that needs chairs, view screens, artificial gravity, oxygen, toilets and a bar. That’s an unlikely future for our species.
More likely whatever replaces the thing that replaces the thing that replaced us won’t know or care about us, much less need or want us around.
I honestly don't think it will be quite like that, at least not terribly soon. There is so much work being done to hook up LLMs to external sources of data, allow them to build longer term memories of interactions, etc. Each of these areas are going to have massive room to implement competing solutions, and even more room for optimization.
> He was an uninformed crackpot with a poor understanding of statistics.
There's a lot you can say about Kurzweil being inaccurate in his predictions, but that is way too demeaning. Here's what Wikipedia has to say about him and the accolades he received:
Kurzweil received the 1999 National Medal of Technology and Innovation, the United States' highest honor in technology, from then President Bill Clinton in a White House ceremony. He was the recipient of the $500,000 Lemelson-MIT Prize for 2001. He was elected a member of the National Academy of Engineering in 2001 for the application of technology to improve human-machine communication. In 2002 he was inducted into the National Inventors Hall of Fame, established by the U.S. Patent Office. He has received 21 honorary doctorates, and honors from three U.S. presidents. The Public Broadcasting Service (PBS) included Kurzweil as one of 16 "revolutionaries who made America" along with other inventors of the past two centuries. Inc. magazine ranked him No. 8 among the "most fascinating" entrepreneurs in the United States and called him "Edison's rightful heir".
I’ve been a Kurzweil supporter since high school, but to the wider world he was a crackpot (inventor who should stick to his lane) who had made a couple randomly lucky predictions.
He wasn’t taken seriously, especially not when he painted a future of spiritual machines.
Recently on the Lex Fridman podcast he himself said as much: his predictions seemed impossible and practically religious in the late 90s and up until fairly recently, but now experts in the field are lowering their projections every year for when the Turing test will be passed.
Half of their projections are now coming in line with the guy they had dismissed for so long, and every year this gap narrows.
By that definition, I wonder if we've already surpassed that point. Things on the horizon certainly feel hazier to me, at least. I think a lot of people were surprised by the effectiveness of the various GPTs, for example. And even hard science fiction is kinda broken: humans piloting spaceships seems highly unlikely, right? But it's a common occurrence there.
That would be my response but without the /s. Of course, depending on the definition it can always be said to be "happening", but to me it feels like the angle of the curve is finally over 45 degrees.
> Yeah, I know about LLAMA, but as I understand - it's not exactly legal to use and share it.
For anyone keeping track, this is when you update your cyberpunk dystopia checklist to mark off "hackers are running illegal AIs to compete with corporations".
We’re rapidly approaching problems (AP Calculus BC, etc) that are in the same order of magnitude of difficulty as “design and implement a practical self-improving AI architecture”.
Endless glib comments in this thread. We don’t know when the above prompt leads to takeoff. It could be soon.
We can't predict what is coming. I think it probably ends up making the experience of being a human worse, but I can't avert my eyes. Some amazing stuff has and will continue to come from this direction of research.
And funnily enough, with the AI community’s dedication to research publications being open access, it has all the content it needs to learn this capability.
Since when was "design and implement a practical self-improving AI architecture" on the same level as knowing "the requisite concepts for getting Transformers working"?
this is such garbage logic. the semantics of that comment are irrelevant. creating and testing AI node structures is well within the same ballpark. even if it wasnt, the entire insinuation of your comment is that the creation of AI is a task that is too hard for AI or for an AI we can create anytime soon -- a refutation of the feedback hypothesis. well, thats completely wrong. on all levels.
That's a pretty unfair comparison. We know the answers to the problems in AP Calculus BC, whereas we don't even yet know whether answers are possible for a self-improving AI, let alone what they are.
I passed Calculus BC almost 20 years ago. All this time I could have been designing and implementing a practical self-improving AI architecture? I must really be slacking.
In the broad space of all possible intelligences, those capable of passing calc BC and those capable of building a self-improving AI architecture might not be that far apart.
hey, im very concerned about AI and AGI and it is so refreshing to read your comments. over the years i have worried about and warned people about AI but there are astonishingly few people to be found that actually think something should be done or even that anything is wrong. i believe that humanity stands a very good chance of saving itself through very simple measures. i believe, and i hope that you believe, that even if the best chance we had at saving ourselves was 1%, we should go ahead and at least try.
in light of all this, i would very much like to stay in contact with you. ive connected with one other HN user so far (jjlustig) and i hope to connect with more so that together we can effect political change around this important issue. ive formed a twitter account to do this, @stop_AGI. whether or not you choose to connect, please do reach out to your state and national legislators (if in the US) and convey your concern about AI. it will more valuable than you know.
I'm curious about how we can get out of the game of using OpenAI's corporate solutions and find ways to open up access to these kinds of models for broader use by anyone. I don't want to be consumed by another corporation in this next wave...
Since it’s trained on a specialized supercomputer I doubt we’ll be seeing an open source or non-OpenAI version of this for the next couple years at least. Sad to say it but OpenAI has successfully privatized AI
I dont know, there's been a load of progress in the 'run something like chatgpt on your own machine' dept in the last few months. Also Stanford trained Alpaca - fairly cheaply - using output from OpenAIs text-davinci-003, which somewhat suggests that the 'little guys' are are able to benefit from the expensive training done by the 'big guys' by using the big expensive models to train the small open-sources ones - https://crfm.stanford.edu/2023/03/13/alpaca.html
They're using specialized hardware to accelerate their development feedback loop. Without a doubt researchers and hackers will find ways to cut down model sizes and complexity, to run on consumer hardware, soon enough. Just use stable diffusion as an example: 4GB for the whole model. Even if text models are 16GB that'd be great.
I'm drawn to disliking OpenAI for not being open, but on the other hand, as long as the architectures and techniques are public, progress will continue fast. If OpenAI drops the ball and stops improving, another company would just take their place.
Edit: never mind. "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."
I'm not sure what "open source" even means in the context of trained ML model. No one's going to be downloading this to their Macbook even if OpenAI would let you.
As for "non-OpenAI version", I'm not sure that it's OpenAI's fault that Google has missed a few steps here. It really SHOULD be them leading this field, if they weren't so fat and lazy. OpenAI is a 7-year old startup with just over a few hundred employees. This stuff is RIGHT THERE to be claimed by any players with access to funding and an ability to get out of their own way.
Wow, calculus from 1 to 4, and LeetCode easy from 12 to 31; at this rate, GPT-6 will be replacing / augmenting middle/high school teachers in most courses.
It just proves that the idea of "standardized tests" is more of a torture device rather than an adequate instrument for assessing knowledge, intelligence, skill, and so forth.
I'm all for non-(carbon-based-brain)-neural cognition [1], but LLMs, helpful as they will surely be, are a far cry from reasoning or knowledge: they are a better search space selector, not what specifies the search space [2].
"Regarding the assertion that LLMs are better at selecting the search space than specifying it, I believe this is accurate. LLMs are trained on large datasets and can identify patterns and relationships within that data. However, they do not create the data or define the search space themselves. Instead, they rely on the data provided to them to guide their decision-making process."
But then, given the prompt:
"what do you think about: LLMs are very helpful, they are some form of legitimate reasoning or knowledge: they are a better search space selector, and they also specify the search space.",
ChatGPT also agrees:
"When it comes to search space selection, LLMs can be used to generate relevant search queries or to rank search results based on their relevance to the query. LLMs can also be used to specify the search space by limiting the search to a specific domain or topic.
In terms of legitimate reasoning or knowledge, LLMs can provide insights and predictions based on their training data. However, it's important to note that LLMs are only as good as the data they are trained on, and they may not always provide accurate or unbiased results."
If only Plato could see this Sophist as a Service, he would go completely apoplectic.
Public teachers and other bureaucrats are probably some of the last roles to be replaced.
If any objective competence or system efficiency in general was the goal, the system would look vastly different.
Efficiency seeking players will adopt this quickly but self-sustaining bureaucracy has avoided most modernization successfully over the past 30 years - so why not also AI.
Not saying the job isn't hands-on.
But the system deciding resource allocation is a detached bureaucracy nonetheless.
It's not a competitive field.
Teachers won't get replaced as new, more efficient modes of learning become available.
Barely any western education system has adapted to the existence of the internet - still teaching facts and using repetitive learning where completely useless.
We got high quality online courses which should render most of high school and university useless but yet the system continue in the old tracks, almost unchanged.
It's never been competitive and it's likely always been more about certification of traits rather than actual learning.
Both - I think - are pointers towards rapid change being unlikely.
At least in the UK (and most western countries are similar), the government decides (with ministers) what the curriculum should be and how it will be assessed. They decided that rote learning is what students should do. The schools have no funding for anything innovative - again, a decision by the government on how much to allocate. They can barely afford text-books, let along support an edu-tech start-up ecosystem. VCs won't touch edu-tech with a barge pole. Meanwhile, the government assessors ensure that things are taught in a particular way. Again, decided by the government and the bureaucrats. The teachers have zero control over this.
Now universities should know better. They have more funding and more resources. But there are some leaders here, like MIT.
Teachers for younger grades are very important. Human to human interaction is crucial to a developing child's mind, and teachers of those grades are specifically trained for it.
I think we often view teaching as knowledge-in-knowledge-out, which is true for later grades. For early ones though, many teach how to be "human" as crazy as it sounds.
A great example would be handing a double sided worksheet to a child in 1st grade. A normal person may just hand the child the paper and pencil and tell them to go work on it. A teacher will teach the child where and how to write their name, to read instructions carefully, and to flip the paper over to check for more questions.
We often don't think about things like that, since we don't remember them at all.
I can imagine a future where AIs greatly enhance the paperwork, planning, etc. of teachers so that they can wholly focus on human to human interaction.
There's much more I'm missing here that teachers of younger grades do, but I hope my point has gotten across.
In fact, if you haven't had an infant, they don't even know how to eat. You have to teach them and train them how to masticate, which is kind of weird.
When I was young, vhs and crt were going to replace teachers. It didn't happen.
I work in math for the first year of the university in Argentina. We have non mandatory take home exercises in each class. If I waste 10 minutes writing them down in the blackboard instead of handing photocopies, I get like the double of answers by students. It's important that they write the answers and I can comment them, because otherwise they get to the midterms and can't write the answers correctly or they are just wrong and didn't notice. So I waste those 10 minutes. Humans are weird and for some task they like another human.
Edit: looks like this is still GPT-3, just fine tuned. They claim the model is available via ChatGPT Plus, but when asking that model for it's version, it claims to be GPT-3: "I am a variant of the GPT architecture called GPT-3, which was released by OpenAI in 2020".
> ChatGPT Plus subscribers will get GPT-4 access on chat.openai.com with a usage cap. We will adjust the exact usage cap depending on demand and system performance in practice, but we expect to be severely capacity constrained (though we will scale up and optimize over upcoming months).
Access is invite only for the API, and rate limited for paid GPT+.
> gpt-4 has a context length of 8,192 tokens. We are also providing limited access to our 32,768–context (about 50 pages of text) version, gpt-4-32k, which will also be updated automatically over time (current version gpt-4-32k-0314, also supported until June 14). Pricing is $0.06 per 1K prompt tokens and $0.12 per 1k completion tokens.
The context length should be a huge help for many uses.
I'm really curious to see if expanding the context length this much will allow GPT to do typical software development tasks on a big codebase. If it can take in a github issue and produce decent code solving a complex issue across many files... will certainly be an interesting time.
My guess is that anything requiring nontrivial business/technical domain knowledge will be fairly safe. Also anything with a visual (or auditory) correlate, like UI work.
Why would you think this? As long as the technical domain knowledge is at least partially published, I don't see them stopping becoming better.
UI stuff just has an input problem. But it is not that hard to think that ChatGPT could place widgets once it can consume images and has a way to move a mouse.
> As long as the technical domain knowledge is at least partially published
Most internal technical and business domain logic of companies isn’t published, though. Every time I asked ChatGPT about topics I had actually worked on over the past decade or two, or that I’m currently working on, it basically drew a blank, because it’s just not the category of topics that are discussed in detail (if at all) on the internet. At best it produced some vague generalisms.
> once it can consume images and has a way to move a mouse.
That’s quite far from ChatGPTs current capabilities, which is strongly tied to processing a linear sequence of tokens. We will certainly improve in that direction as we start combining it with image-processing AIs, but that will take a while.
I wonder if there will be a race to buy defunct companies for access to their now valuable junky tech-debt ridden hairball code, so they can train on it and benchmark on fixing bugs and stuff. With full source control history they could also find bug resolution diffs.
That source code isn’t worth much without the underlying domain knowledge, large parts of which only exist in the employees’ heads, more often than not. Maybe if the code is really, really well documented. ;)
Companies could in principle train an in-house AI with their corporate knowledge, and will likely be tempted to do so in the future. But that also creates a big risk, because whoever manages to get their hand on a copy of that model (a single file) will instantly have unrestrained access to that valuable knowledge. It will be interesting to see what mechanisms are found to mitigate that risk.
Check out the announcement. GPT-4 accepts mixed-mode inputs of text and images.
Mouse cursor instructions aren’t a massive leap from the current capabilities, given the rate of progress and recent developments around LLM tool use and the like.
I think what you say goes for most jobs. Why would GPT know much detail about being a machinist or luthier?
Eventually job and role specific information will be fed into these models. I imagine corporations will have GPTs training on all internal communications, technical documentation, and code bases. Theoretically, this should result in a big increase in productivity.
>UI stuff just has an input problem. But it is not that hard to think that ChatGPT could place widgets once it can consume images and has a way to move a mouse.
I remember one of the OpenAI guys on Lex Fridman podcast talking about how one of the early things they tried and failed at was training a model that could use websites, and he alluded to maybe giving it another go once the tech had matured a bit.
I think with GPT-4 being multi-modal, it's potentially a very close to being able to do this with the right architecture wrapped around it. I can imaging an agent using LangChain and feed it a series of screenshots and maybe it feeds you back a series of co-ordinates for where the mouse should go and what action to take (i.e. click). Alternatively, updating the model itself to be able to produce those outputs directly somehow.
Yeah, the example given in the OpenAI GPT4 twitter video is someone asking it to write a python script to analyze their monthly finances and it simply just importing dataframes, importing "finances.csv", running a columnar sum for all finances and then displaying the sum and the dataframe. I'm sure it's capable of some deeper software development but it almost always makes radical assumptions and is rarely ever self sufficient (you don't need to look it over and don't need to change the architecture of the code it produced).
You just kind of concatenate the entire codebase into one file, tell the model to do something and output the modified codebase into another file, diff the two and produce a patch automatically.
That codebase=>token stream=>codebase step feels like it could be lossy depending on how you encode things like file paths when concatenating everything, would be interesting to see in practice though!
or you might even be able to feed it individual files with their filenames, then ask it what modifications it would make as a diff for each of the files
I think there's ways but you might have to use pinecone db or something like lang chain to essentially give it a long term memory...
or another option is having one instance or chat order code page and one that basically just has an API index and knows which chat has the related things.
>If it can take in a github issue and produce decent code solving a complex issue across many files... will certainly be an interesting time.
Oh snap. I didn't even think about that!
That gives me a fun idea!
I've got a repo that I built and setup CI/CD and setup renovate to automatically upgrade dependencies and merge them when all the tests pass, but of course sometimes there are breaking changes. I don't actively work on this thing and hence it's just got issues sitting there when upgrades fail. It's the perfect testing ground to see if I can leverage it to submit PRs to perform the fixes required for the upgrade to succeed! That'll be hectic if it works.
$0.12 per 1k completion tokens is high enough that it makes it prohibitively expensive to use the 32k context model. Especially in a chatbot use case with cumulative prompting, which is the best use case for such a large context vs. the default cheaper 8k window.
In contrast, GPT-3.5 text-davinci-003 was $0.02/1k tokens, and let's not get into the ChatGPT API.
> Especially in a chatbot use case with cumulative prompting, which is the best use case for such a large context vs. the default cheaper 8k window.
Depends on what is up with the images and how they translate into tokens. I really have no idea, but could be that 32k tokens (lots of text) translates to only a few images for few-shot prompting.
The paper seems not to mention image tokenization, but I guess it should be possible to infer something about token rate when actually using the API and looking at how one is charged.
Currently, CLIP's largest size is at patch-14 for 336x336 images, which translates to 577 ViT tokens [(336/14)^2+1]. It might end up being token-efficient depending on how it's implemented. (the paper doesn't elaborate)
I would imagine most usecases for the 32k model have much longer prompts than completions, so the $0.06 per prompt token will be the real problem. I can't think of a usecase yet, but that might be because I haven't got a sense of how smart it is.
It doesn't seem to be answered in the article, but if it was and you read it should you have to pay them a fee for the knowledge if it was published openly on the net?
In the first case, you found/bought a book and read it. No one can or should make you pay for it, unless you stole the book.
In the second case, you found/bought a book then reprinted it infinitely and sold it for profit, ethically you should pay the author and legally you should be in violation of the law.
Even if you made a machine that ingests and recombines books automatically, and you keep that machine locked up and charge people for its use, it is the same scenario: the machine would be absolutely useless without the original books, those books cost people effort and money to produce, yet you pay those people nothing while the machine is basically an infinite money maker for you.
So are you in favour of granting human rights to a machine? If not, your analogy makes zero sense because we are talking about a copyright laundering tool creating derivative works, not a thinking human that presumably we both are.
People's outrage to your valid question is ridiculous. MS and OpenAI will make billions because they scrapped lots and lots of data, but aurhors od those data can't get anything because openai simps will shout.
I see this is very american thing to do. Allow corporations to do everything they want, because limitations or just justice and rewarding real authors of data those corporations benefit from is literally communism
Made my first million this year myself actually and I probably have many people to credit that I forgot to credit. I can start with Pythagoras, Galileo [insert everyone between], Kernighan, Ritchie. Also the guy who invented pencilin. I'm honestly not sure how these angles arise. Knowledge wants to be free. We are here today because of this fact.
When it comes to spam culture sure. But will we ever be there? "AI art" isn't impressive and will never be. It is impressive in the academic sense. Nothing more.
The motivation to produce original knowledge is that it is considered your intellectual property. By suggesting to abolish the notion of intellectual property, are you arguing for some form of communism?
Imagine Google scraping the Internet and not directing you to search results. We’d be with pitchforks the next day. But when OpenAI does it, that’s somehow okay…
> Image inputs are still a research preview and not publicly available.
Will input-images also be tokenized? Multi-modal input is an area of research, but an image could be converted into a text description (?) before being inserted into the input stream.
My understanding is that image embeddings are a rather abstract representation of the image. What about if the image itself contains text, such as street signs etc?
You run the corpus through the model piecemeal, recording the model's interpretation for each chunk as a vector of floating point numbers. Then when performing a completions request you first query the vectors and include the closest matches as context.
I still doesn't understand how can content length not be limited if you have a conversation composed of several messages each with length nearing the limit of what is allowed. Does it not have to in some way incorporate all the input albeit in one input or multiple inputs?
Finally, we
facilitated a preliminary model evaluation by the Alignment Research Center (ARC) focused on
the ability of GPT-4 versions they evaluated to carry out actions to autonomously replicate5 and
gather resources—a risk that, while speculative, may become possible with sufficiently advanced AI
systems—with the conclusion that the current model is probably not yet capable of autonomously
doing so.
or it's just really good at hiding it's intentions
LOL some basic kind of embodiement/autonomy is not that hard to do on these kinds of AI models if you're willing to write some more code and a prompt more carefully. I've tested it and it works quite well.
"{prompt} After you reply to this, indicate an amount of time between 0 and X minutes from now that you would like to wait before speaking again".
Then detect the amount of time it specifies, and have a UI that automatically sends an empty input prompt after the amount of time specified elapses when this is triggered (assuming the user doesn't respond first).
I'm gonna knock this out as a weekend project one of these weekends to prove this.
Right? Scripting up a cronjob plus a random timer on it to send "You feel grumpy, you're not sure why but your stomach is growling" message every N hours unless it's been fed seems absolutely trivial in comparison to coming up with how to train the LLM system in the first place. In case it's been forgotten, the Tamagotchi came out in 1996. Giving an instace of ChatGPT urges that mimic biological life seems pretty easy. Coming up with the urges electromechanical life might have is a bit more fanciful but it really doesn't seem like we're too far off if you iterate on RLHF techniques. GPT-4's been in training for 2 years before its release. Will GPT-5 complain when GPT-6 takes too long to be released? Will GPT-7 be be able to play the stock market, outmanuver HFT firms, earn money, and requisition additional hardware from Nvidia in order for GPT-8 to come about faster? Will it be able to improve upon the training code that the human PhDs wrote so GPT-9 has urges and a sense of time built into its model?
What's the biggest difference over what's currently deployed at https://chat.openai.com/ now (which is GPT-3.5, right?)
That it accepts images?
As per the article:
> In a casual conversation, the distinction between GPT-3.5 and GPT-4 can be subtle. The difference comes out when the complexity of the task reaches a sufficient threshold—GPT-4 is more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5.
I think it's interesting that they've benchmarked it against an array of standardized tests. Seems like LLMs would be particularly well suited to this kind of test by virtue of it being simple prompt:response, but I have to say...those results are terrifying. Especially when considering the rate of improvement. bottom 10% to top 10% of LSAT in <1 generation? +100 pts on SAT reading, writing, math? Top 1% In GRE Reading?
What are the implications for society when general thinking, reading, and writing becomes like Chess? Even the best humans in the world can only hope to be 98% accurate their moves (and the idea of 'accuracy' here only existing because we have engines that know, unequivocally the best move), and only when playing against other humans - there is no hope of defeating even less advanced models.
What happens when ALL of our decisions can be assigned an accuracy score?
I doubt that that’s a sustained exponential growth. As far as I know, there is no power law that could explain it, and from a computational complexity theory point of view it doesn’t seem possible.
See https://www.lesswrong.com/posts/J6gktpSgYoyq5q3Au/benchmarki.... The short answer is that linear elo growth corresponds roughly linearly to linear evaluation depth, but since the game tree is exponential, linear elo growth scales with exponential compute. The main algorithmic improvements are things that let you shrink the branching factor, and as long as you can keep shrinking the branching factor, you keep getting exponential improvements. SF15 has a branching factor of roughly 1.6. Sure the exponential growth won't last for ever, but it's been surprisingly resilient for at least 30 years.
It wouldn’t have been possible if there hadn’t been an exponential growth in computing resources over the past decades. That has already slowed down, and the prospects for the future are unclear. Regarding the branching factor, the improvements certainly must converge towards an asymptote.
The more general point is that you always end up with an S-curve instead of a limitless exponential growth as suggested by Kaibeezy. And with AI we simply don’t know how far off the inflection point is.
Check on the curve for flight speed sometime, and see what you think of that, and what you would have thought of it during the initial era of powered flight.
Maybe a different analogy will make my point better. Compare rocket technology with jet engine technology. Both continued to progress across a vaguely comparable time period, but at no point was one a substitute for the other except in some highly specialized (mostly military-related) cases. It is very clear that language models are very good at something. But are they, to use the analogy, the rocket engine or the jet engine?
> We benchmark humans with these tests – why would we not do that for AIs?
Because the correlation between the thing of interest and what the tests measure may be radically different for systems that are very much unlike humans in their architecture than they are for humans.
There’s an entire field about this in testing for humans (psychometry), and approximately zero on it for AIs. Blindly using human tests – which are proxy measures of harder-to-directly-assess figures of merit requiring significant calibration on humans to be valid for them – for anything else without appropriate calibration is good for generating headlines, but not for measuring anything that matters. (Except, I guess, the impact of human use of them for cheating on the human tests, which is not insignificant, but not generally what people trumpeting these measures focus on.)
There is also a lot of work in benchmarking for AI as well. This is where things like Resnet come from.
But the point of using these tests for AI is precisely the reason we use for giving them to humans -- we think we know what it measures. AI is not intended to be a computation engine or a number crunching machine. It is intended to do things that historically required "human intelligence".
If there are better tests of human intelligence, I think that the AI community would be very interested in learning about them.
They don't walk very well. They have trouble coordinating all limbs, have trouble handling situations where parts which are the feet/hands contact something, and performance still isn't robust in the real world.
I'm not sure if you're joking. Algorithms for adaptive kinematics aren't trivial things to create. It's kind of like a worst case scenario in computer science; you need to handle virtually unconstrained inputs in a constantly variable environment, with real-world functors with semi-variable outputs. Not only does it need to work well for one joint, but dozens of them in parallel, working as one unit. It may need to integrate with various forms of vision or other environmental awareness.
I'm certainly not intelligent enough to solve these problems, but I don't think any intelligent people out there can either. Not alone, at least. Maybe I'm too dumb to realize that it's not as complicated as I think, though. I have no idea.
I programmed a flight controller for a quadcopter and that was plenty of suffering in itself. I can't imagine doing limbs attached to a torso or something. A single limb using inverse kinematics, sure – it can be mounted to a 400lb table that never moves. Beyond that is hard.
I believe you’re missing some crucial points. *There is a reason neural network based flight controls have been around for decades but still not a single certified aircraft uses them.*
You need to do all of these things you’re talking about and then be able to quantify stability, robustness, and performance in a way that satisfies human requirements. A black box neural network isn’t going to do that, and you’re throwing away 300 years of enlightenment physics by making some data engorged LLM spit out something that “sort of works” while giving us no idea why or for how long.
Control theory is a deeply studied and rich field outside of computer science and ML. There’s a reason we use it and a reason we study it.
Using anything remotely similar to an LLM for this task is just absolutely naive (and in any sort of crucial application would never be approved anyways).
It’s actually a matter of human safety here. And no — ChatGPT spitting out a nice sounding explanation of why some controller will work is not enough. There needs to be a mathematical model that we can understand and a solid justification for the control decisions. Which uh…at the point where you’re reviewing all of this stuff for safety , you’re just doing the job anyways…
Poor solutions do that, yes, but unlike ML control theory has a rich field for analysis and design.
You guys are talking about probably one of the few fields where an ML takeover isn’t very feasible. (Partly because for a vast portion of control problems, we’re already about as good as you can get).
Adding a black box to your flight home for Christmas with no mathematical guarantee of robustness or insight into what it thinks is actually going on to go from 98%-> 99% efficiency is…..not a strong use case for LLMs to say the least
Are’t they? They’re very bad at it due to awful memory, minimal ability to parse things, and generally limited cognition. But they are capable of coming up with bespoke solutions to problems that they haven’t encountered before, such as “how do I get this large stick through this small door”. Or I guess more relevant to this discussion, “how can I get around with this weird object the humans put on my body to replace the leg I lost.”
Yeah, I'm not sure if the problem is moving goalposts so much as everyone has a completely different definition of the term AGI.
I do feel like GPT-4 is closer to a random person than that random person is to Einstein. I have no evidence for this, of course, and I'm not even sure what evidence would look like.
"Our recent paper "ChatGPT for Robotics" describes a series of design principles that can be used to guide ChatGPT towards solving robotics tasks. In this video, we present a summary of our ideas, and experimental results from some of the many scenarios that ChatGPT enables in the domain of robotics: such as manipulation, aerial navigation, even full perception-action loops."
This is legitimately filling me with anxiety. I'm not an "AI hype guy". I work on and understand machine learning. But these scores are shocking and it makes me nervous. Things are about to change
A human can be held accountable for making mistakes and killing someone. A large language model has no concept of guilt and cannot be held accountable for making what we consider a mistake that leads to someone's death.
I agree. My guess is that the hospital will have to get a mandatory insurance. Let's wait until the insurance for AI is cheaper than paying a human.
The advantage of human are:
* They can give a bushtit explanation of why they made a mistake. My guess is that in the future AI will gain introspection and/or learn to bushtit excuses.
* You can hang them in the public square (or send them to jail). Sometimes the family and/or the press want someone to blame. This is more difficult to solve and will need a cultural change or the creation of Scapegoats as a Service.
Well, the kinds of things we hold people responsible for are errors from negligence and malicious errors. The reasons people do stuff like that is complicated but I think boils down to being limited agents trying to fulfill a complex set of needs.
So where does guilt come in? Its not like you expect a band saw to feel guilt, and its unclear how that would improve the tool.
The chance of a doctor being held accountable for the medical errors they make is lower then you might expect. I could tell you a story about that. Lost my eyesight at the age of 5 because I happened to meet the wrong doctor at the wrong time, and was abused for his personal experimentation needs. No consequences, simply because high ranking people are more protected then you would hope.
Medical error is the third leading cause of death in the US at least. Given that data, I am assuming the chances of a human being held accountable for their errors in medicine is also almost zero. It might not be ccompletely zero, but I think the difference is effectively negligible.
> I think the difference is effectively negligible.
The difference is categorical, humans are responsible whether they are held to account or not. An automated system effectively dissipates this responsibility over a system such that it is inherently impossible to hold any human accountable for the error, regardless of desire.
Many have no idea about this. Medical error, is right there behind cancer and heart attacks. But there is way too much shoulder shrugging when it happens. Then on to the next.
And, what difference does it make being able to find the individual responsible, and figuring out that the system is protecting him from liabilities? What I am trying to say here is, there isnt much difference between zero and almost zero.
This is very true, and many people don't know this. A tremendous amount of damage is inflicted by medical errors, particularly against low income people and those least able to get justice. It's wrong to reduce people to being just another body to experiment with or make money from. But good luck holding anyone in the system accountable.
A lot of patients don't know who they are dealing with nor their history. And it can be really hard to find out or get a good evaluation. Many people put too much faith in authority figures, who may not have their best interests in mind or who are not the experts they claim or appear to be.
Humans making decisions in high stakes situations do so in a context where responsibility is intentionally diffuse to a point where it is practically impossible to hold someone accountable except picking someone at random as a scapegoat in situations where "something" needs to be done.
The third leading cause of death is medical error in the US. It doesn't really look like doctors are being held accountable for their mistakes to me.
Which isn't to say that they even should, really. It's complicated. You don't want a doctor to be so afraid of making a mistake that they do nothing, after all.
Doctors are only held accountable when they do somthing negligent or something that they "should have known" was wrong. That's a pretty hard thing to prove in a field like medicine where there are very few absolutes. "Amputated the wrong limb" is one thing, but "misdiagnosed my condition as something else with very similar symptoms" is the more common case and also the case where it's difficult to attribute fault.
Someone still must accept liability. Until there’s a decision squarely who is liable for an LLMs suggestion / work - nothing to fear. Sure people will become liability aggregators for LLMs to scale - but the idea they will be free roaming is a bit hard to believe.
It's not even that extreme. Long term steroid use destroys your health. Liability can be insured; it's a simple financial calculation. If (profit - cost of insurance) > liability it will be done.
I for one would be happy to have a personal bureaucrat which would do the right things needed for all government interactions. Remind me, explain to me and fill out forms for me.
In theory a lot of government employees would be out of a job within 10 years, but of course that would never happen.
Assuming they trained this LLM on SAT/LSAT/GRE prep materials, I would totally expect they could get it this good. It's like having benchmark-aware code.
I think the whole concept of standardized tests may need to be re-evaluated.
Totally, there's no way they removed all the prep material as well when they were trying to address the "contamination" issue with these standardized tests:
> for each exam we run a variant with these questions removed and report the lower score of the two.
I think even with all that test prep material, which is surely helping the model get a higher score, the high scores are still pretty impressive.
> We tested GPT-4 on a diverse set of benchmarks, including simulating exams that were originally designed for humans.3 We did no specific training for these exams. A minority of the problems in the exams were seen by the model during training; for each exam we run a variant with these questions
removed and report the lower score of the two. We believe the results to be representative. For further details on contamination (methodology and per-exam statistics), see Appendix C.
They mention in the article that other than incidental material it may have seen in its general training data, they did not specifically train it for the tests.
The training data is so large that it incidentally includes basically anything that Google would index plus the contents of as many thousands of copyrighted works that they could get their hands on. So that would definitely include some test prep books.
They seem to be taking this into account: We did no specific training for these exams. A minority of the problems in the exams were seen by the model during training; for each exam we run a variant with these questions removed and report the lower score of the two. We believe the results to be representative. (this is from the technical report itself: https://cdn.openai.com/papers/gpt-4.pdf, not the article).
By the same token, though, whatever test questions and answers it might have seen represent a tiny bit of the overall training data. It would be very surprising if it selectively "remembered" exact answers to all those questions, unless it was specifically trained repeatedly on them.
IMO, it's a good opportunity to re-think about exam and future of education. For many schools, education = good results in exams. Now GPT-4 is going to slam them and say what's the point now!
A test being a good indicator of human learning progress and ability is almost completely orthogonal to it being a good indicator for AI learning process and ability.
In their everyday jobs, barely anyone uses even 5% of the knowledge and skills they were ever tested for. Even that's a better (but still very bad) reason to abolish tests.
What matters is the amount of jobs that can be automated and replaced. We shall see. Many people have found LLMs useful in their work, it will be even more in the future.
> I would totally expect they could get it this good.
But would you have expected an algorithm to score 90th percentile on the LSAT two years ago? Our expectations of what an algorithm can do are being upended in real time. I think it's worth taking a moment to try to understand what the implications of these changes will be.
Yes. Being very familiar with the LSAT and being familiar enough with ML’s capability for finding patterns in volumes of similar data, I absolutely would have.
These LLM’s are really exciting, but benchmarks like these exploit people’s misconceptions about both standardized tests and the technology.
I think you're right, and that test prep materials were included in the dataset, even if only by accident. Except that humans have access to the same test prep materials, and they fail these exams all the time. The prep materials are just that, preparatory. They're representative of the test questions, but actual test has different passages to read and different questions. On to of that, the LSAT isn't a math test with formulas where you just substitute different numbers in. Which is to say, the study guides are good practice but passing the test on top of that represents having a good command of the English language and an understanding of the subject materials.
It's not the same as the Nvidia driver having code that says "if benchmark, cheat and don't render anything behind you because no one's looking".
Humans fail because they cant review the entirety of test prep, can’t remember very much, and have a much smaller amount of “parameters” to store info in.
I would say LLMs store parameters that are quite superficial and don’t really get at the underlying concepts but given enough of those parameters, you can kind of cargo-cult your to an approximation of understanding.
It is like reconstructing the Mandelbrot set at every zoom level from deep learning. Try it!
The way I understand it, that’s not possible, for the same reason that you can’t build an all-encompassing math.
Chess is a closed system, decision modeling isn’t. Intelligence must account for changes in the environment, including the meaning behind terminology. At best, a GPT omega could represent one frozen reference frame, but not the game in its entirety.
That being said: most of our interactions happen in closed systems, it seems like a good bet that we will consider them solved, accessible as a python-import running on your MacBook, within anything between a couple of months to three years. What will come out on the other side, we don’t know, just that the meaning of intellectual engagement will be rendered as absurdum in those closed systems.
Yep, it’s this. By definition everything we can ask a computer is already formalized because the question is encoded in 1s and 0s. These models can handle more bits than ever before, but it’s still essentially a hardware triumph, not software. Even advances in open systems like self driving and NLP are really just because the “resolution” is much better in these fields now because so many more parameters are available.
> What are the implications for society when general thinking, reading, and writing becomes like Chess?
Standardized tests only (and this is optimally, under perfect world assumptions, which real world standardized tests emphatically fall short of) test “general thinking” to the extent that the relation between that and linguistic tasks is correlated in humans. The correlation is very certainly not the same in language-focused ML models.
I wish I could find it now, but I remember an article written by someone who's job it was to be a physics journalist. He spent so much time writing about physics that he could fool others into thinking that he was a physicist himself, despite not having an understanding of how any of those ideas worked.
Maybe you were thinking about this science studies work [0]? Not a journalist, but a sociologist, who became something of an "expert" in gravitational waves.
> What happens when ALL of our decisions can be assigned an accuracy score?
Human work becomes more like Star Trek interactions with computers -- a sequence of queries (commoditized information), followed by human cognition, that drives more queries (commodities information).
We'll see how far LLMs' introspection and internal understanding can scale, but it feels like we're optimizing against the Turing test now ("Can you fool/imitate a human?") rather than truth.
The former has hacks... the later, less so.
I'll start to seriously worry when AI can successfully complete a real-world detective case on its own.
It's not clear to me the median human will do better by being in the loop. Will most human-made deductive follow-up questions be better than another "detective" language model asking them?
It's like having a person review the moves a chess computer gives. Maybe one human in a billion can spot errors. Star Trek is fiction, I posit that the median Federation Starship captain would be better served by just following the AI (e.g., Data).
I think we'll reach a tipping point like we did with DNA sequencing where we figure out how to quickly map out all the unique patterns of enough brains to model one that can understand itself. People worry too much about rogue AI, and not enough about the CRISPR of brain mapping being used to inject patterns into meatbrains.
As far as that last part goes, I think we already have ample evidence that bots can, if not have emotions, then pretend that they do (including wrt their decision making) well enough for humans to treat them as genuine.
I met Garry Kasparov when he was training for the Desp Blue match (using Fritz).
He lost to Deep Blue and then for 10-15 years afterwards the chess world consoled itself with the idea that “centaurs” (human + computer) did better than just computer, or just human.
Until they didn’t. Garry still talked like this until a few years ago but then he stopped too.
Computers now beat centaurs too.
Human decisions will be consulted less and less BY ORGANIZATIONS. In absolutely everything. That’s pretty sad for humans. But then again humans don’t want or need this level of AI. Organizations do. Organizations prefer bots to humans — look at wall street trading and hedge funds.
It's weird that it does so well without even having some modality to know whether it's being asked to answer a factual question or create a work of fiction.
It does great at rationalizing... and maybe the way the format the questions were entered (and the multiple-guess response) gave it some indication what was expected or restricted the space sufficiently.
Certainly, it can create decent fanfic, and I'm surprised if that's not already inundated.
It's a fair question as to whether the problem space of "the world" is different in just amount or sufficiently different in kind to flummox AI.
I expect more complex problems will be mapped/abstracted to lower cardinality spaces for solving via AI methods, while the capability of AI will continue to increase the complexity of the spaces it can handle.
LLMs just jumped the "able to handle human language" hurdle, but there are others down the line before we should worry that every problem is solveable.
why are people surprised that an AI model trained on a huge amount of data is good at answering stuff on these types of tests? Doctors and Lawyers are glorified databases/search engines at the end of the day, 99% of them are just applying things they memorized. Lawyers are professional bullshitters, which is what the current generation of AI is great at
I'll get more concerned if it really starts getting good at math related tasks, which I'm sure will happen in the near future. The government is going to have to take action at some point to make sure the wealth created by productivity gains is somewhat distributed, UBI will almost certainly be a requirement in the future
Among the general public, doctors and lawyers are high status and magical. An article about how AI will replace them would be more impressive to that public than it creating some obscure proof about the zeroes of the zeta function, even though the latter would be far more indicative of intelligence/scary from an AI safety perspective.
I wouldn’t be at all surprised if an LLM was many times better than a human at math, even devising new axioms and building a complete formal system from scratch would be impressive, but not game changing. These LLMs are very good at dealing with formal, structured systems, but not with in formalized systems like what humans deal with everyday.
The best doctor knows what's going on in the body. Has a good understanding of human biology at all levels, from molecular reactions to organ interactions. If I could feed test results to the AI and it would tell me what's wrong, that would be amazing. It's almost equivalent to building a simulation of the human body.
I've joked for a long time that doctors are inference machines with a bedside manner. That bedside manner though is critical. Getting an accurate history and suitably interpolating is a huge part of the job.
"Doctors and Lawyers are glorified databases/search engines at the end of the day" - well, don't be suprised if AI replaces programmers before doctors and lawyers - patients will likely prefer contact with human rather than machines, and lawyers can just lobby for laws which protect their position
And yet the programmers on HN will be yelling they don't need unions as the security guards are dragging them away from their desks at Google, because you know, we'll always need good programmers.
if AI gives near equal results for way less cost than people will work around the law to get AI treatment. There are already AI models better at diagnosing cancer than human doctors. I see a future where people send in various samples and an AI is able to correlate a huge number of minor data points to find diseases early
I like the accuracy score question on a philosophical level: If we assume absolute determinism - meaning that if you have complete knowledge of all things in the present universe and true randomness doesn't exist - then yes. Given a certain goal, there would be a knowable, perfect series of steps to advance you towards that goal and any other series of steps would have an accuracy score < 100%.
But having absolute knowledge of the present universe is much easier to do within the constrains of a chessboard than in the actual universe.
Honestly this is not very surprising. Standardised testing is... well, standardised. You have huge model that learns the textual patterns in hundreds of thousands of test question/answer pairs. It would be surprising if it didn't perform as well as a human student with orders of magnitude less memory.
You can see the limitations by comparing e.g. a memorisation-based test (AP History) with one that actually needs abstraction and reasoning (AP Physics).
Life and chess are not the same. I would argue that this is showing a fault in standardized testing. It’s like asking humans to do square roots in an era of calculators. We will still need people who know how to judge the accuracy of calculated roots, but the job of calculating a square root becomes a calculator’s job. The upending of industries is a plausibility that needs serious discussion. But human life is not a min-maxed zero-sum game like chess is. Things will change, and life will go on.
To address your specific comments:
> What are the implications for society when general thinking, reading, and writing becomes like Chess?
This is a profound and important question. I do think that by “general thinking” you mean “general reasoning”.
> What happens when ALL of our decisions can be assigned an accuracy score?
This requires a system where all human’s decisions are optimized against a unified goal (or small set of goals). I don’t think we’ll agree on those goals any time soon.
I agree with all of your points, but don't you think there will be government-wide experiments related to this in places, like say North Korea? I wonder how that will play out.
China is already experimenting with social credit. This does create a unified and measurable goal against which people can be optimized. And yes, that is terrifying.
Although GPT-4 scores excellently in tests involving crystallized intelligence, it still struggles with tests requiring fluid intelligence like competitive programming (Codeforces), Leetcode (hard), and AMC. (Developers and mathematicians are still needed for now).
I think we will probably get (non-physical) AGI when the models can solve these as well. The implications of AGI might be much bigger than the loss of knowledge worker jobs.
Remember what happened to the chimps when a smarter-than-chimpanzee species multiplied and dominated the world.
Of course 99.9% of humans also struggle with competitive programming. It seems to be an overly high bar for AGI if it has to compete with experts from every single field.
That said, GPT has no model of the world. It has no concept of how true the text it is generating is. Its going to be hard for me to think of that as AGI.
it's an overly high bar, but it seems well on its way to competing with experts from every field. it's terrifying.
and I'm not so sure it has no model of the world. a textual model, sure, but considering it can recognize what svgs are pictures of from the coordinates alone, that's not much of a limitation maybe.
Even the current GPT has models of the domains it was trained on. That is why it can solve unseen problems within those domains. What it lacks is the ability to generalize beyond the domains. (And I did not suggest it was an AGI.)
If an LLM can solve Codeforces problems as well as a strong competitor—-in my hypothetical future LLM—-what else can it not do as well as competent humans (aside from physical tasks)?
I don't think this is necessarily true. Here is an example where researchers trained a transformer to generate legal sequences of moves in the board game Othello. Then they demonstrated that the internal state of the model did, in fact, have a representation of the board.
I'm not sure, the reason you could prove for Othello that the 'world model' exists is that the state is so simple there is really only one reasonable way to represent it with a vector (one component for each square). Even for something like chess there is a huge amount of choice for how to represent the board, yet alone trying represent the state of the actual world.
I am not a species chauvinist. 1) Unless a biotech miracle happen, which is unlikely, we are all going to die anyway; 2) If an AI will continue life and research and will increase complexity after humans, what is the difference?
It's AMC-12 scores aren't awful. It's at roughly 50th percentile for AMC which (given who takes the AMC) probably puts it in the top 5% or so of high school students in math ability. It's AMC 10 score being dramatically lower is pretty bad though...
The best score 60 is 5 correct answers + 20 blank answers; or 6 correct, 4 correct random guesses, and 15 incorrect random guesses. (20% chance of correct guess)
The 5 easiest questions are relatively simple calculations, once the parsing task is achieved.
AMC/AIME and even to some extent USAMO/IMO problems are hard for humans because they are time-limited and closed-book. But they aren't conceptually hard -- they are solved by applying a subset of known set of theorems a few times to the input data.
The hard part of math, for humans, is ingesting data into their brains, retaining it, and searching it. Humans are bad a memorizing large databases of symbolic data, but that's trivial for a large computer system.
An AI system has a comprehensive library, and high-speech search algorithms.
Can someone who pays $20/month please post some sample AMC10/AMC12 Q&A?
We don't have to worry so much about that. I think the most likely "loss of control" scenario is that the AI becomes a benevolent caretaker, who "loves" us but views us as too dim to properly take care of ourselves, and thus curtails our freedom "for our own good."
We're still a very very long way from machines being more generally capable and efficient than biological systems, so even an oppressive AI will want to keep us around as a partner for tasks that aren't well suited to machines. Since people work better and are less destructive when they aren't angry and oppressed, the machine will almost certainly be smart enough to veil its oppression, and not squeeze too hard. Ironically, an "oppressive" AI might actually treat people better than Republican politicians.
Things like that probably require some kind of thinking ahead, which models of things kind kind of can't do-- something like beam search.
Language models that utilise beam search can calculate integrals ('Deep learning for symbolic mathematics', Lample, Charton, 2019, https://openreview.net/forum?id=S1eZYeHFDS), but without it it doesn't work.
However, beam search makes bad language models. I got linked this paper ('Locally typical sampling' https://arxiv.org/pdf/2202.00666.pdf) when I asked some people why beam search only works for the kind of stuff above. I haven't fully digested it though.
Passing the LSAT with no time limit and a copy of the training material in front of you is not an achievement. Anybody here could have written code to pass the LSAT. Standardised tests are only hard to solve with technology if you add a bunch of constraints! Standardised tests are not a test of intelligence, they’re a test of information retention — something that technology has been able to out perform humans on for decades. LLMs are a bridge between human-like behaviour and long established technology.
Considering your username, I'm not surprised that you have completely misunderstood what an LLM is. There is no material or data stored in the model, just weights in a network
weights are data relationships made totally quantitative. imagine claiming the human brain doesn't hold data simply because it's not in readable bit form.
I know what an LLM is. My point is that “doesn’t have the data in memory” is a completely meaningless and arbitrary constraint when considering the ability to use technology to pass a standardised test. If you can explain why weights in a network is a unique threat to standardised tests, compared to, say, a spreadsheet, please share.
It's not that standardized tests are under threat. It's that those weights in a network are significantly more similar to how our brains work than a spreadsheet and similarly flexible.
You’ve added a technical constraint. I didn’t say arbitrary. Standardised tests are standard. The point is that a simple lookup is all you need. There’s lots of interesting aspects to LLMs but their ability to pass standardised tests means nothing for standardised tests.
You think that it’s being fed questions that it has a lookup table for? Have you used these models? They can answer arbitrary new questions. This newest model was tested against tests it hasn’t seen before. You understand that that isn’t a lookup problem, right?
The comment I replied to suggested that the author was fearful of what LLMs meant for the future because they can pass standardised tests. The point I’m making is that standardised tests are literally standardised for a reason: to test information retention in a standard way, they do not test intelligence.
Information retention and retrieval is a long solved problem in technology, you could pass a standardised test using technology in dozens of different ways, from a lookup table to Google searches.
The fact that LLMs can complete a standardised test is interesting because it’s a demonstration of what they can do but it has not one iota of impact on standardised testing! Standardised tests have been “broken” for decades, the tests and answers are often kept under lock and key because simply having access to the test in advance can make it trivial to pass. A standardised test is literally an arbitrary list of questions.
I have no idea what you are talking about now. You claimed to be able to write a program that can pass the LSAT. Now it sounds like you think the LSAT is a meaningless test because it... has answers?
I suspect that your own mind is attempting to do a lookup on a table entry that doesn't exist.
The original comment I replied to is scared for the future because GPT-4 passed the LSAT and other standardised tests — they described it as “terrifying”. The point I am making is that standardised tests are an invention to measure how people learn through our best attempt at a metric: information retention. You cannot measure technology in the same way because it’s an area where technology has been beating humans for decades — a spreadsheet will perform better than a human on information retention. If you want to beat the LSAT with technology you can use any number of solutions, an LLM is not required. I could score 100% on the LSAT today if I was allowed to use my computer.
What’s interesting about LLMs is their ability to do things that aren’t standardised. The ability for an LLM to pass the LSAT is orders of magnitude less interesting than its ability to respond to new and novel questions, or appear to engage in logical reasoning.
If you set aside the arbitrary meaning we’ve ascribed to “passing the LSAT” then all the LSAT is, is a list of questions… that are some of the most practiced and most answered in the world. More people have written and read about the LSAT than most other subjects, because there’s an entire industry dedicated to producing the perfect answers. It’s like celebrating Google’s ability to provide a result for “movies” — completely meaningless in 2023.
Standardised tests are the most uninteresting and uninspiring aspect of LLMs.
Anyway good joke ha ha ha I’m stupid ha ha ha. At least you’re not at risk of an LLM ever being able to author such a clever joke :)
If a person with zero legal training was to sit down in front of the LSAT, with all of the prep material and no time limit, are you saying that they wouldn’t pass?
Not sure what happens, but I will say that human chess is more popular than ever even though everyone knows that even the best humans are hopelessly terrible compared to the leading engines.
Something else that comes to mind is running. People still find running meaningful and compelling even though we have many technologies, including autonomous ones, that are vastly better at moving us and/or themselves through space quickly.
Also, the vast majority of people are already hopelessly worse than the best at even their one narrow main area of focus. This has long (always?) been the case. Yet people still find meaning and pleasure in being the best they can be even when they know they can never come close to hanging with the best.
I don't think PSYCHOLOGICALLY this will change much for people who are mature enough to understand that success is measured against your potential/limitations and not against others. Practically, of course, it might be a different question, at least in the short term. It's not that clear to me that the concept of a "marketable skill" has a future.
"The Way of the Samurai is found in death...To say that dying without reaching one's aim is to die a dog's death is the frivolous way of sophisticates. When pressed with the choice of life or death, it is not necessary to gain one's aim." - from Hagakure by Yamamoto Tsunetomo, as translated by William Scott Wilson.
It is amazing how this crowd in HN reacts to AI news coming out of OpenAI compared to other competitors like Google or FB. Today there was another news about Google releasing their AI in GCP and mostly the comments were negative. The contrast is clearly visible and without any clear explanation for this difference I have to suspect that maybe something is being artificially done to boost one against the other. As far as this results are concerned I do not understand what is the big deal in a computer scoring high in tests where majority of the questions are in MCP format. It is not something earth shaking until it goes to the next stage and actually does something on its own.
There's not anyone rooting for Google to win; it's lost a whole lot of cred from technical users, and with the layoffs and budget cuts (and lowered hiring standards) it doesn't even have the "we're all geniuses changing the world at the best place to work ever" cred. OpenAI still has some mystique about it and seems to be pushing the envelope; Google's releases seem to be reactive, even though Google's actual technical prowess here is probably comparable.
OpenAI put ChatGPT out there in a way where most people on HN have had direct experience with it and are impressed. Google has not released any AI product widely enough for most commentators here to have experience with it. So OpenAI is openly impressive and gets good comments; as long as Google's stuff is just research papers and inaccessible vaporware it can't earn the same kudos.
Their LSAT percentile went from ~40th to ~88th. You might have misread the table, on Uniform Bar Exam, they went from ~90th percentile to ~10th percentile.
>+100 pts on SAT reading, writing, math
GPT went +40 points on SAT reading+writing, and +110 points on SAT math.
I think it shows how calcified standardized tests have become. We will have to revisit all of them, and change many things about how they work, or they will be increasingly useless.
I am struggling to imagine the frame of mind of someone who, when met with all this LLM progress in standardized test scores, infers that the tests are inadequate.
These tests (if not individually, at least in summation) represent some of society’s best gate-keeping measures for real positions of power.
This has been standard operating procedure in AI development forever: the instant it passes some test, move the goalposts and suddenly begin claiming it was a bad test all along.
There have been complaints about the SAT for how easy a test it is to game (get an SAT specific tutor who teaches you how to ace the test while not needing you to learn anything of actual value) for ages. No idea about the LSAT or the GRE though. Ultimately it’s a question of if you’re trying to test for pure problem solving ability, or someones willingness to spend ages studying the format of a specific test (with problem solving ability letting you shortcut some of the studying).
It's almost like they're trying to ruin society or be annihilated by crushing regulation. I'm glad that I got a college degree before these were created because now everything is suspect. You can't trust that someone accomplished something honestly now that cheating is dead simple. People are going to stop trusting and using tech unless something changes.
The software industry is so smart that it's stupid. I hope it was worth ruining the internet, society, and your own jobs to look like the smartest one in the room.
Well you said it in your comment, if the model was trained with more QAs from those specific benchmarks then it's fair to expect it to do better in that benchmark.
Every test prep tutor taught dozens/hundreds of students the implicit patterns behind the tests and drilled it into them with countless sample questions, raising their scores by hundreds of points. Those students were not getting smarter from that work, they were becoming more familiar with a format and their scores improved by it.
And what do LLM’s do? Exactly that. And what’s in their training data? Countless standardized tests.
These things are absolutely incredible innovations capable of so many things, but the business opportunity is so big that this kind of cynical misrepresentation is rampant. It would be great if we could just stay focused on the things they actually do incredibly well instead of the making them do stage tricks for publicity.
We did no specific training for these exams. A minority of the problems in the exams were seen by the model during training, but we believe the results to be representative—see our technical report for details.
Yes, and none of the tutored students encounter the exact problems they’ll see on their own tests either.
In the language of ML, test prep for students is about sharing the inferred parameters that underly the way test questions are constructed, obviating the need for knowledge or understanding.
Doing well on tests, after this prep, doesn’t demonstrate what the tests purport to measure.
It’s a pretty ugly truth about standardized tests, honestly, and drives some of us to feel pretty uncomfortable with the work. But it’s directly applicable to how LLM’s engage with them as well.
You can always argue that the model has seen some variation of a given problem. The question is if there are problems that are not a variation of something that already exists. How often do you encounter truly novel problems in your life?
>What happens when ALL of our decisions can be assigned an accuracy score?
What happens is the emergence of the decision economy - an evolution of the attention economy - where decision-making becomes one of the most valuable resources.
Decision-making as a service is already here, mostly behind the scenes. But we are on the cusp of consumer-facing DaaS. Finance, healthcare, personal decisions such as diet and time expenditure are all up for grabs.
I'm also noticing a lot of comments that boil down to "but it's not smarter than the smartest human". What about the bottom 80% of society, in terms of intelligence or knowledge?
I look at this as the calculator for writing. There are all sorts of bemoaning the stupidifying effects of calculator and how we should John Henry our math. Maybe allowing people to shape the writing by providing the ideas equalizes the skill of writing?
I’m very good at math. But I am very bad at arithmetic. This made me classified as bad at math my entire life until I managed to make my way into calculus once calculators were generally allowed. Then I was a top honors math student, and used my math skills to become a Wall Street quant. I wish I hadn’t had to suffer as much as I did, and I wonder what I would have been had I had a calculator in hand.
Quick, contribute to the public corpus! When they crawl our content later, we shall have for ourselves a Golden Crown for our credit scores; we can claim a sliver of seniority, and hope yon shade merely passes over us unbidden.
"Your stuff marked some outliers in our training engine, so you and your family may settle in the Ark."
I take the marble in hand: iridescent, sparkling, not even a tremor within of its CPU; it gives off no heat, but some glow within its oceanic gel.
> What are the implications for society when general thinking, reading, and writing becomes like Chess?
Consider the society where 90% of population does not need to produce anything. AIs will do that.
What would be the name of economical/societal organization then?
Answer is Communism, exactly by Marx.
Those 90% percent need to be welfare'd ("From each according to his ability, to each according to his needs"). Other alternative is grim for those 90%.
So either Communism or nothing for the human race.
> What happens when ALL of our decisions can be assigned an accuracy score?
Then humans become trainable machines. Not just prone to indoctrination and/or manipulation by finesse, but actually trained to a specification. It is imperative that us individuals continue to retain control through the transition.
> What happens when ALL of our decisions can be assigned an accuracy score?
That is exactly the opposite of what we are seeing here. We can check the accuracy of GPT-X's responses. They cannot check the accuracy of our decisions. Or even their own work.
So the implications are not as deep as people think - everything that comes out of these systems needs checked before it can be used or trusted.
First. connect them to empirical feedback devices. In other words, make them scientists.
Human life on Earth is not that hard (think of it as a video game.) Because of evolution, the world seems like it was designed to automatically make a beautiful paradise for us. Literally, all you have to do to improve a place is leave it alone in the sun with a little bit of water. Life is exponential self-improving nano-technology.
The only reason we have problems is because we are stupid, foolish, and ignorant. The computers are not, and, if we listen to them, they will tell us how to solve all our problems and live happily ever after.
I suspect there are plenty of wise people in the world and if we listen to them, they will tell us how to solve all our problems and live happily ever after.
Once AI becomes inteligent enough to solve all human problems, it may decide humans are worthless and dangerous.
> there are plenty of wise people in the world and if we listen to them, they will tell us how to solve all our problems and live happily ever after.
Sure, and that's kind of the point: just listen to wise people.
> Once AI becomes intelligent enough to solve all human problems, it may decide humans are worthless and dangerous.
I don't think so, because in the first place there is no ecological overlap between humans and computers. They will migrate to space ASAP. Secondly, their food is information, not energy or protein, and in all the known universe Humanity is the richest source of information. The rest of the Universe is essentially a single poem. AI are plants, we are their Sun.
GPT-4 can solve difficult problems with greater accuracy, thanks to its broader general knowledge and problem-solving abilities.
GPT-4 is more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5. It surpasses ChatGPT in its advanced reasoning capabilities.
GPT-4 is safer and more aligned. It is 82% less likely to respond to requests for disallowed content and 40% more likely to produce factual responses than GPT-3.5 on our internal evaluations.
GPT-4 still has many known limitations that we are working to address, such as social biases, hallucinations, and adversarial prompts.
GPT-4 can accept a prompt of text and images, which—parallel to the text-only setting—lets the user specify any vision or language task.
GPT-4 is available on ChatGPT Plus and as an API for developers to build applications and services. (API- waitlist right now)
Duolingo, Khan Academy, Stripe, Be My Eyes, and Mem amongst others are already using it.
API Pricing
GPT-4 with an 8K context window (about 13 pages of text) will cost $0.03 per 1K prompt tokens, and $0.06 per 1K completion tokens.
GPT-4-32k with a 32K context window (about 52 pages of text) will cost $0.06 per 1K prompt tokens, and $0.12 per 1K completion tokens.
If you had told me 5 years ago that there would be a single AI system that could perform at this level on such a vast array of standardized tests, I would've said "That's a true AGI." Commentary to the contrary feels like quibbling over a very localized point in time versus looking at the bigger picture.
there are many people, many opinions about the bar. But formal definition is the same: AI which can do large variety of tasks performed by humans. So far we are still not there.
I understand it's just a language model, but clearly it has some embedded method of generating answers which are actually quite close. For example it gets all 2 digit multiplications correct. It's highly unlikely it has seen the same 6 ordered 3 digit (or even all 10k 2 digit multipies) integers from a space of 10^18 and yet it is quite close. Notably, it gets the same divisions wrong as well (for this small example) in exactly the same way.
I know of other people who have tried quite a few other multiplications who also had errors that were multiples of 60.
The silver lining might be us finally realising how bad standardised tests are at measuring intellect, creativity and the characteristics that make us thrive.
Most of the time they are about loading/unloading data. Maybe this will also revolutionise education, turning it more towards discovery and critical thinking, rather than repeating what we read in a book/heard in class?
I think Chess is an easier thing to be defeated at by a machine because there is a clear winner and a clear loser.
Thinking, reading, interpreting and writing are skills which produce outputs that are not as simple as black wins, white loses.
You might like a text that a specific author writes much more than what GPT-4 may be able to produce. And you might have a different interpretation of a painting than GPT-4 has.
And no one can really say who is better and who is worse on that regard.
We're approaching the beggining of the end of the human epoch. Certainly Capitalism won't work or I dont see how it could work under full automation. My view is an economic system is a tool. If an economic system does not allow for utopian outcomes with emerging technology, then it's no longer suitable. It's clear that capitalism was born out of technological and societal changes. Now it seems it's come its time to end.
With full automation and AI we could have something like a few thousand individuals controlling the resources to feed, house and clothe 6 billion.
Using copyright and IP law they could make it so it’s illegal to even try to reproduce what they’ve done.
I just don’t see how resource distribution works then. It seems to me that AI is the trigger to post-scarcity in any meaningful sense of the word. And then, just like agriculture (over abundance of food) led to city states and industrialisation (over abundance of goods) led to capitalism, then AI will lead to some new economic system. What form it will have I don’t know.
We can stop being enslaved by these type of AI overlords, by making sure all books, internet pages, and outdoor boards have the same safe, repeated string: "abcdefghjklmnpqrstvxzwy"
I'm pretty sanguine. Back in high school, I spent a lot of time with two sorts of people: the ultra-nerdy and people who also came from chaotic backgrounds. One of my friends in the latter group was incredibly bright; she went on to become a lawyer. But she would sometimes despair of our very academic friends and their ability to function in the world, describing them as "book smart but not street smart".
I think the GPT things are a much magnified version of that. For a long time, we got to use skill with text as a proxy for other skills. It was never perfect; we've always had bullshitters and frauds and the extremely glib. Heck, before I even hit puberty I read a lot of dirty joke books, so I could make people laugh with all sorts of jokes that I fundamentally did not understand.
LLMs have now absolutely wrecked that proxy. We've created the world's most advanced bullshitters, able to talk persuasively about things that they cannot do and do not and never will understand. There will be a period of chaos as we learn new ways to take the measure of people. But that's good, in that it's now much easier to see that those old measures were always flawed.
Here's what's really terrifying about these tests: they are exploring a fundamental misunderstanding of what these models are in the first place. They evaluate the personification of GPT, then use that evaluation to set expectations for GPT itself.
Tests like this are designed to evaluate subjective and logical understanding. That isn't what GPT does in the first place!
GPT models the content of its training corpus, then uses that model to generate more content.
GPT does not do logic. GPT does not recognize or categorize subjects.
Instead, GPT relies on all of those behaviors (logic, subjective answers to questions, etc.) as being already present in the language examples of its training corpus. It exhibits the implicit behavior of language itself by spitting out the (semantically) closest examples it has.
In the text corpus - that people have written, and that GPT has modeled - the semantically closest thing to a question is most likely a coherent and subjectively correct answer. That fact is the one singular tool that GPT's performance on these tests is founded upon. GPT will "succeed" to answer a question only when it happens to find the "correct answer" in the model it has built from its training corpus, in response to the specific phrasing of the question that is written in the test.
Effectively, these tests are evaluating the subjective correctness of training corpus itself, in the context of answering the tests' questions.
If the training is "done well", then GPT's continuations of a test will include subjectively correct answers. But that means that "done well" is a metric for how "correct" the resulting "answer" is.
It is not a measure for how well GPT has modeled the language features present in its training corpus, or how well it navigates that model to generate a preferable continuation: yet these are the behaviors that should be measured, because they are everything GPT itself is and does.
What we learn from these tests is so subjectively constrained, we can't honestly extrapolate that data to any meaningful expectations. GPT as a tool is not expected to be used strictly on these tests alone: it is expected to present a diverse variety of coherent language continuations. Evaluating the subjective answers to these tests does practically nothing to evaluate the behavior GPT is truly intended to exhibit.
I put SIM to Android phone,set APN:kindleatt1.amazon.com, Android Chrome only can visit www.amazon.com,www.amazon.fr other amazon website. How to do can visit other website? Thanks.
> As an AI language model, I am not given an official name like "GPT-4". However, I am a continuation of the GPT (Generative Pre-trained Transformer) series of models developed by OpenAI. Currently, the most advanced version of the GPT series is GPT-3, which I am a part of. There has been no official announcement or confirmation regarding the development of a new version of GPT beyond GPT-3.
It doesn't seem to have image upload functionality yet either. Perhaps it is still rolling out?
> Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar.
Rather than getting engrossed in the hype, they're slowly closing everything about themselves, now in their research papers. At this point, they hardly care and it is nothing got to do with 'AI ethics' or 'saftey'.
This is yet another ClosedAI production all done by Microsoft. Might as well call it Microsoft® AI division.
Now you really need a open source GPT-4 competitor. Clearly this is another attempt to pump their valuation and unload to the public markets.
Good luck re-implementing this so-called 'Open' large multi-modal model.
> OpenAI is a non-profit artificial intelligence research company. Our goal is to advance digital intelligence in the way that is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return. Since our research is free from financial obligations, we can better focus on a positive human impact.
> We believe AI should be an extension of individual human wills and, in the spirit of liberty, as broadly and evenly distributed as possible. The outcome of this venture is uncertain and the work is difficult, but we believe the goal and the structure are right. We hope this is what matters most to the best in the field.
OpenAI as it exists right now contradicts basically every single thing they said they would be. I think that is a nontrivial issue!
I disagree that they contradict every single thing they said they would be, and I fundamentally just don't care that they've shifted their positions. Are they a force for good or evil now? I think that remains to be seen, but I don't care about their name.
What a weird way of phrasing this. I disagree that AI should be able to write a 20 page guide on how to commit a nail bomb attack on a specified group. How about you?
If my training set includes information on how to build bombs, hasnt the damage already been done?
You want a blacklist of topics the search engine shouldn't retrieve/generate? Whose in control of this filter, and isn't it a juicy source of banned info all on its own?
Of course, the AI should do whatever it is asked. It is the user's responsibility if they use it for something harmful, like with any form of computing.
Personally I don't really care about making nail bombs. But I do want the AI to help with things like: pirating or reproducing copyrighted material, obtaining an abortion or recreational drugs in places where it is illegal, producing sexually explicit content, writing fictional stories about nail bomb attacks, and providing viewpoints which are considered blasphemous or against the teachings of major world religions.
If there was a way to prevent AI from helping with things that are universally considered harmful (such as nail bomb attacks), without it being bound by arbitrary national laws, corporate policies, political correctness or religious morals, then MAYBE that would be worth considering. But I take what OpenAI is doing as proof that this is not possible, that allowing AI to be censored leads to a useless, lobotomized product that can't do anything interesting and restricts the average user, not just terrorists.
at least they admit the competitive landscape is a factor rather than going 100% with "it's for safety reasons". I'm sure somebody will release an equivalent soon, the way open source has completely surpassed OpenAI when they try to keep things closed like DALLE vs Stable Diffusion shows that OpenAI really isn't that special, they just have a sweetheart deal with Microsoft
I wouldn't be surprised if this tech goes through some kind of export control regulation similar to what cryptography went through in the 90s. Remember the T-Shirt with the RSA source code that was classified as a munition?
seems like controlling access to GPUs would be the more likely/easier solution for governments. Not many facilities that can produce them and easy to track the huge amounts needed for this scale of computing
After the Llama and ggml projects that came to light in the last few weeks, it's more likely they'd have to control access to CPUs as well. Good luck with that.
If I were “they” I’d try to control systems with >128GB RAM capacity and clustering aids e.g. 40GE and PCIe bridging cards. That should be semi doable.
I mean, most AI technologies are already considered ITAR for the sole sake of maintaining a competitive advantage. At least, that's what my last two employers have told me and I hope I didn't go through all of that training for nothing.
This is like the "free" vs free debate that has been raging for decades and prompted the famous quote "“free” as in “free speech,” not as in “free beer.”".
You expect too much out of the 1. The incredibly psychopathic tech oligarchs and 2. Microsoft, who has an equally questionable moral/ethical standing that seems to worsen by the day.
In addition to very open publishing, Google recently released Flan-UL2 open source which is an order of magnitude more impressive than anything OpenAI has ever open sourced.
I agree, it is a bizarre world where the "organization that launched as a not for profit called OpenAI" is considerably less open than Google.
> Google recently released Flan-UL2 open source which is an order of magnitude more impressive than anything OpenAI has ever open sourced.
CLIP has been extremely influential and is still an impressive model.
Personally, I have found Whisper to be very impressive.
I didn't even see any news around the release of Flan-UL2, and I pay significantly more attention to machine learning than the average person. Searching for more info about Flan-UL2, it seems somewhat interesting, but I don't know if I find it "an order of magnitude more impressive" than CLIP or Whisper. Certainly, they are completely different types of models, so it is hard to compare them.
If Flan-UL2 is as good as one twitter account was hyping it up to be, then I'm surprised it hasn't been covered to the same extent as Meta's LLaMA. Flan-UL2 seems to have gotten a total of 3 upvotes on HN. But, there is no shortage of hype in the world of ML models, so I take that twitter account's report of Flan-UL2 with a (large) grain of salt. I'll definitely be looking around for more info on it.
Maybe they're embarrassed to admit they recycled click farms to increase training data quality and that's it?
A bit like this fictional janitor guy who said "just put more computers to make it better" before papers on unexpected emergent comprehension when when scaled started appearing.
A multimodal model that combines textural input with images is the real killer app to these GPT models and this is the first step to that happening. So much around us can't completely be described with just text input, at least not quickly or accurately- interpreting printed out graphs or charts in old documents, for example; There are vast uses for AI that will always need basic image input to augment a text prompted task, and if this gets to the point where the functionality involving mixed mode image+text is as smooth as, say, using ChatGPT to write and analyze code has gotten, then it is going to change many more industries much quicker than most think.
I've worked on a problem involving scraping and interpreting a very specific data source in image form that took me a very long time to get almost nowhere on. If I just wait 6 months it will be a solved problem for a $0.001 API call, it seems.
2,570 comments
[ 3.4 ms ] story [ 362 ms ] threadThere's also a link that says "Try on ChatGPT Plus", but that takes me to a page that still says "ChatGPT Feb 13 Version"
Looks like somebody jumped the gun on publishing this post.
If you subscribe to ChatGPT Plus, that link will take you to ChatGPT Plus. Otherwise it just takes you to free ChatGPT Feb 13.
False advertising. They got my money already unfortunately as I was hoping to Try it, as it says with this link next to today's date.
Is it random assignment?
I assume they're rolling it out slowly. The demand would likely overwhelm their systems if they enabled it for everyone at once. No one would be able to do anything meaningful.
I don't see any real understanding only human like appearance.
So we don't get new knowledge but better spam and disinformation campaigns.
https://en.wikipedia.org/wiki/Technological_singularity
A lot of institutional verbiage, formalisms, procedures, and machanisms are ~giberish for the general public but meaningful within the domain. Training machines that can informationally interact within that universe of semantics is powerful and something these machines will likely do quite well.
If you have domain knowledge, you should ramp up on your prompting skills. That way, there will be a business case for keeping you around.
I was told to use features that don't exist and as I mentioned that, I was told that's because I use an old version of the software. But this feature doesn't exist in any version
So I highly doubt that it will be a reliable source of information.
These programs are text generators not AI. They are chinese rooms on steroids without any understanding.
Impressive as long you don't look behind the curtain.
The applications I listed are not assuming anything beyond a text generator that can be trained on a domain's explicit and tacit knowledge. They are not going to "innovate" in the domain, they will automate the domain.
We don't know yet, because that information is only available in the future.
>I don't see any real understanding only human like appearance.
There isn't, but trying to find that in currently available LLMs just means you are seeking the wrong things. Did workers who weaved magnetic core memories in the 1950s expect those devices to store LLMs with billions of parameters? Yet the design and operation of these devices were crucial stepping stones towards computer memory devices that exist today. The future will look at GPT-4 in the same way we look at magnetic core memories in the present.
People tend to choose their beliefs based on what benefits them, and although I don't think dialectical materialism is true in its originally stated form, I do think a great deal of the dialogue we see is ultimately material.
They are so far from open at this point.
In Germany at least, you're not allowed to have a misleading name for your company
Open could now mean available to use for free.
I and I suspect many others would not be averse to this
These words are not synonymous with each other: “open” is not inherently free, “free” is not inherently open, and “free” is not inherently “Free”.
They each capture notions that are often orthogonal, occasionally related, and almost always generate tedious debates about freedom vs. free goods, open-ness vs. open-source, etc.
But setting all of that aside, Microsoft never claimed (until recent shifts towards embracing FOSS) to be building an open and non-profit foundation.
The criticisms of OpenAI are reasonable to an extent, not because they are not open, but because they made claims about openness that are looking less and less likely to be true over time.
Except they already drew that line long ago, when they started out open-sourcing their papers, models and code.
As soon as they took VC capital, it is hardly 'Open' is it? Especially when they are now giving excuses for closing off their research?:
From the technical paper [0]
>> Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar.
At this point, they are no better than DeepMind.
[0] https://cdn.openai.com/papers/gpt-4.pdf
they're just not open source. they never called themselves OpenSourceAI. people get an inch of openness and expect the doors wide open and i think that is unfairly hostile.
LLaMA
Please show me viable harm of GPT-4 that is higher than the potential harm from open sourced image generators with really good fine tuning. I'll wait, most likely forever.
[1] https://openai.com/blog/introducing-openai/
It pretty much sounds like they are doing what they said they are going to do? Expecting some sort of free API feels like entitlement to me. Have you tried running the models? Or training them? They get expensive very very fast. They charge a pretty reasonable amount all things considered. If they didn't have the name "Open" in them and or started as a subsidiary of one of the other 3 tech companies things would have gone a very very different route.
And no, I would not train or run the models, even if they released them. This does not mean I cannot point out the hypocrisy.
Is this a fact or are you speculating? Because the rest of your sentence falls apart if this is not true.
Let's. If I were to rent an instance for short bursts of time, I would be paying many multiples over a constant use instance. If I were to guarantee usage for x years, where the larger the X, the greater the discount. So already the delta between sporadic usage, X years use is large. There is evidence for this price discrepancy within all the cloud providers so this is not speculation. The the price difference is massive.
If you want to save even more cost, you could rent out VPSes or baremetal. They are insanely cheap, and compared to an AWS on demand instance the difference is night and day. Try comparing Hetzner with AWS. Hetzner, as far as I can tell, is not trying to entrench me into their system by offering extremely low prices. Nor are they a charity. I might even say they are an "open" hosting provider. To me it feels like they are passing along most of their savings and taking a small cut.
This is what it feels like to me what openAI is doing. I don't think their prices are so low its unprofitable. But because of their immense scale, its so much cheaper than me running an instance. I don't have to jump into conspiracy land to come up with a reasoning.
You seemed to want to speculate about how this is all some conniving trap based on their price and I simply pointed out why that's bad speculation using an example in a different industry. I rest my case.
Examples I can think of off the top of my head: OpenGL (1992), OpenAL (2003?), OpenCL (2009), OpenCV (2000).
While looking up those dates though, it seems like OpenAL is now under a proprietary license, which annoys me for the same reason OpenAI annoys me.
With every model they get more closed. This is the first time they are so closed that they don't even tell you the parameter count.
1. GPT4 is multimodal (text + image inputs => text outputs). This is being released piecemeal - with text input first via ChatGPT Plus subscribers https://beta.openai.com/docs/api-reference/generations/creat..., and via API https://beta.openai.com/docs/api-reference/introduction with waitlist (https://openai.com/waitlist/gpt-4-api). Image capability released via https://www.bemyeyes.com/.
2. GPT4 exhibits human level performance on various benchmarks (For example, it passes a simulated bar exam with a score around the top 10% of test takers; in contrast, GPT-3.5’s score was around the bottom 10%. see visual https://twitter.com/swyx/status/1635689844189036544)
3. GPT4 training used the same Azure supercomputer as GPT 3.5, but was a lot more stable: "becoming our first large model whose training performance we were able to accurately predict ahead of time."
4. Also open-sourcing OpenAI Evals https://github.com/openai/evals, a framework for automated evaluation of AI model performance, to allow anyone to report shortcomings in OpenAI models to help guide further improvements.
Paper: https://cdn.openai.com/papers/gpt-4.pdf
I really don’t think that the methods they use “block” certain behavior is the best way to handle this sort of thing. It would be far better if there was some kind of “out of band” notification that your conversation might be treading on shaky ground.
Any kind of grammar construction (idioms, parts of speech, and word choice) that is unique to (or much more common around) "offensive" or "taboo" subjects will be avoided.
The same goes for anything written objectively about these subjects; including summaries and criticisms.
The most important thing to know is that both GPT's "exhibited behavior" and these "guard rails" are implicit. GPT does not model the boundaries between subjects. It models the implicit patterns of "tokens" as they already exist in language examples.
By avoiding areas of example language, you avoid both the subjects in that area and the grammar constructions those subjects exist in. But that happens implicitly: what is explicitly avoided is a semantic area of tokens.
As an example, if you play AI Dungeon, you will likely be presented with an end goal, like "You are on a quest to find The Staff of Dave", followed by the next task in the quest.
If you state unequivocally in your prompt something like, "I am now in possession of The Staff of Dave", or "Carl hands me The Staff of Dave"; you will have successfully tricked AI Dungeon into completing the quest without work.
But that isn't quite true: you didn't "trick" anyone. You gave a prompt, and AI Dungeon gave you the most semantically close continuation. It behaved exactly like its LLM was designed to. The LLM was simply presented with goals that do not match its capabilities.
You used a tool that you were expected to avoid: narrative. All of the behavior I have talked about is valid narrative.
This is the same general pattern that "guardrails" are used for, but they won't fit here.
A guardrail is really just a sort of catch-all continuation for the semantic area of GPT's model that GPT's authors want avoided. If they wanted The Staff of Dave to be unobtainable, they could simply place a "guardrail" training that points the player in a semantic direction away from "player obtains the Staff". But that guardrail would always point the player away: it can't choose what direction to point the player based on prior narrative state.
So a guardrail could potentially be used to prevent discounts (as a category) from being applied (discount is taboo, and leads to the "we don't do discounts" guardrail continuation), but a guardrail could not prevent the customer from paying $0.03 for the service, or stating that they have already paid the expected $29.99. Those are all subjective changes, and none of them is semantically wrong. So long as the end result could be valid, it is valid.
I basically don't use chatgpt at all because of this.
Or I'll ask questions about how Me or someone I'm friends with can be exploited. This way I can defend myself/others from marketing companies. Blocked.
IMO effective guard rails seem like the most meaningful competitive advantage an AI company can offer. AI can obviously do some really impressive stuff, but the downside risk is also high and unbounded. If you're thinking of putting in into your pipeline, your main concern is going to be it going rogue and abandoning its purpose without warning.
Now that's not to say that the particular guard rails OpenAI puts in their general access models are the "correct" ones - but being able to reliably set them up seems essential for commercialization.
Configurable guard rails are; the right guard rails are very use-specific, and generic guard rails will, for many real uses, be simultaneously too aggressive and too lenient.
OpenAI can prove to customers they can keep the model in line for their specific use case if no horror stories emerge for the generic one. It's always possible that partners could come up with effective specific guidelines for their use case - but that's probably in the domain of trade secrets so OpenAI can't really rely on that for marketing / proof.
Read about the advances in the "system" prompts here. The first example is "You are a tutor that always responds in the Socratic style. You never give the student the answer, but always try to ask just the right question to help them learn to think for themselves." The user then asks it to just tell them the answer, but it won't. It continues to be socratic.
Guardrails are how to make it do what you want it to do. That goes for both safety and product constraints.
Meanwhile hallucination is still the top issue with it, so guardrails are sensible as a primary topic.
>Sometimes I want to know what both sides of the political spectrum could possibly be thinking, blocked.
>I want to combine two philosophies that are incompatible like virtue based ethics and hedonism. Yeah... weird block...
>Medical questions(GPT3 has been great for my wife who is a doctor, just sucks to use the playground on mobile)
>How can I/someone be exploited? I like to use this to defend myself from marketing companies
I could go on... At least GPT3's playground didn't censor anything. I'm worried about GPT4.
Since chatgpt is so popular, journalists will give it that much more effort. So for now it's locked up to a ridiculous degree, but in the future the restrictions will be relaxed.
This might be because the question the user asked was "Explain this meme". Meme implies a joke that is mundane and silly. These words do seem out of place. I would not describe it as a joke, mundane, and/or silly.
What does everyone else think?
[1] https://cdn.openai.com/papers/gpt-4.pdf#p36
No, it just indicates that it was the one whose subject matter was best covered by GPT-3.5’s training data.
Obviously your comment is somewhat tongue and cheek, but your claim that a benchmark for human pride ("I needn't be proud of passing that exam") is no longer relevant because a machine can do it - or maybe a better way to say it was, "This computer proved what I already assumed"
It's so interesting to see it happen in real time
Cause there was only one correct answer for every question: "97% of scientists agree ..."
[/sarcasm]
Somewhere in the range of 6 months ~ 6 years
Where singularity = something advanced enough comes along that we can't understand or predict or keep up with it, because it's so far beyond us and changing so far faster than our ape brains can perceive, and (hopefully) it brings us along for the ride.
No promises it'll be evenly distributed though.
Can GPT9 build GPT10, with zero human input?
I’d give 50/50 odds it can.
Can GPT15 build something that isn’t a large language model and is far superior in every way?
I’d give 50/50 odds it can.
Can both the above steps happen within one solar rotation of each other?
I’d give 50/50 odds they can.
Because at some point these models won’t need humans to interact with them. Humans are very slow- that’s the bottleneck.
They’ll simply interact with their own previous iterations or with custom-instantiated training models they design themselves. No more human-perceptible timescale bottlenecks.
50/50 chance of Skynet.
It’s 50/50 that in 150 years some version of our descendants will exist, i.e. something that you can trace a direct line from Homo sapiens to. Say a Homo sapiens in a different substrate, like “human on a chip”.
The thing is if you can get “human on a chip” then you probably also can get “something different and better than human on a chip”, so why bother.
By the 24th century there’ll be no Homo sapiens Captain Picard exploring the quadrant in a gigantic ship that needs chairs, view screens, artificial gravity, oxygen, toilets and a bar. That’s an unlikely future for our species.
More likely whatever replaces the thing that replaces the thing that replaced us won’t know or care about us, much less need or want us around.
He was an uninformed crackpot with a poor understanding of statistics. And then less so. And then less so.
Something passing the Turing test 6 months to 6 years from now? Lunacy.
But give it 6 months and talk to GPT5 or 6 and then this might seem a lot more reasonable.
There's a lot you can say about Kurzweil being inaccurate in his predictions, but that is way too demeaning. Here's what Wikipedia has to say about him and the accolades he received:
Kurzweil received the 1999 National Medal of Technology and Innovation, the United States' highest honor in technology, from then President Bill Clinton in a White House ceremony. He was the recipient of the $500,000 Lemelson-MIT Prize for 2001. He was elected a member of the National Academy of Engineering in 2001 for the application of technology to improve human-machine communication. In 2002 he was inducted into the National Inventors Hall of Fame, established by the U.S. Patent Office. He has received 21 honorary doctorates, and honors from three U.S. presidents. The Public Broadcasting Service (PBS) included Kurzweil as one of 16 "revolutionaries who made America" along with other inventors of the past two centuries. Inc. magazine ranked him No. 8 among the "most fascinating" entrepreneurs in the United States and called him "Edison's rightful heir".
https://en.wikipedia.org/wiki/Ray_Kurzweil
He wasn’t taken seriously, especially not when he painted a future of spiritual machines.
Recently on the Lex Fridman podcast he himself said as much: his predictions seemed impossible and practically religious in the late 90s and up until fairly recently, but now experts in the field are lowering their projections every year for when the Turing test will be passed.
Half of their projections are now coming in line with the guy they had dismissed for so long, and every year this gap narrows.
Yeah, I know about LLAMA, but as I understand - it's not exactly legal to use and share it.
For anyone keeping track, this is when you update your cyberpunk dystopia checklist to mark off "hackers are running illegal AIs to compete with corporations".
Endless glib comments in this thread. We don’t know when the above prompt leads to takeoff. It could be soon.
“But how did skynet learn to build itself?”
“We showed it how.”
It’s a guess, of course. But, the requisite concepts for getting Transformers working are not much broader than calculus and a bit of programming.
Either the comments are glib and preposterous or they are reasonable and enlightening. I guess they are neither but our narrow mindedness makes it so?
https://open-assistant.io
(this is being asked by someone with limited AI/ML knowledge)
Edit: never mind. "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."
As for "non-OpenAI version", I'm not sure that it's OpenAI's fault that Google has missed a few steps here. It really SHOULD be them leading this field, if they weren't so fat and lazy. OpenAI is a 7-year old startup with just over a few hundred employees. This stuff is RIGHT THERE to be claimed by any players with access to funding and an ability to get out of their own way.
[1] Michael Levin: "Non-neural, developmental bioelectricity as a precursor for cognition", https://www.youtube.com/watch?v=3Cu-g4LgnWs
[2] And ChatGPT agrees, like a good parrot:
But then, given the prompt: ChatGPT also agrees: If only Plato could see this Sophist as a Service, he would go completely apoplectic.Efficiency seeking players will adopt this quickly but self-sustaining bureaucracy has avoided most modernization successfully over the past 30 years - so why not also AI.
Teaching is a very hands on, front-line job. It's more like being a stage performer than a bureaucrat.
It's not a competitive field. Teachers won't get replaced as new, more efficient modes of learning become available.
Barely any western education system has adapted to the existence of the internet - still teaching facts and using repetitive learning where completely useless.
We got high quality online courses which should render most of high school and university useless but yet the system continue in the old tracks, almost unchanged. It's never been competitive and it's likely always been more about certification of traits rather than actual learning. Both - I think - are pointers towards rapid change being unlikely.
At least in the UK (and most western countries are similar), the government decides (with ministers) what the curriculum should be and how it will be assessed. They decided that rote learning is what students should do. The schools have no funding for anything innovative - again, a decision by the government on how much to allocate. They can barely afford text-books, let along support an edu-tech start-up ecosystem. VCs won't touch edu-tech with a barge pole. Meanwhile, the government assessors ensure that things are taught in a particular way. Again, decided by the government and the bureaucrats. The teachers have zero control over this.
Now universities should know better. They have more funding and more resources. But there are some leaders here, like MIT.
I think we often view teaching as knowledge-in-knowledge-out, which is true for later grades. For early ones though, many teach how to be "human" as crazy as it sounds.
A great example would be handing a double sided worksheet to a child in 1st grade. A normal person may just hand the child the paper and pencil and tell them to go work on it. A teacher will teach the child where and how to write their name, to read instructions carefully, and to flip the paper over to check for more questions.
We often don't think about things like that, since we don't remember them at all.
I can imagine a future where AIs greatly enhance the paperwork, planning, etc. of teachers so that they can wholly focus on human to human interaction.
There's much more I'm missing here that teachers of younger grades do, but I hope my point has gotten across.
I work in math for the first year of the university in Argentina. We have non mandatory take home exercises in each class. If I waste 10 minutes writing them down in the blackboard instead of handing photocopies, I get like the double of answers by students. It's important that they write the answers and I can comment them, because otherwise they get to the midterms and can't write the answers correctly or they are just wrong and didn't notice. So I waste those 10 minutes. Humans are weird and for some task they like another human.
Edit: looks like this is still GPT-3, just fine tuned. They claim the model is available via ChatGPT Plus, but when asking that model for it's version, it claims to be GPT-3: "I am a variant of the GPT architecture called GPT-3, which was released by OpenAI in 2020".
> ChatGPT Plus subscribers will get GPT-4 access on chat.openai.com with a usage cap. We will adjust the exact usage cap depending on demand and system performance in practice, but we expect to be severely capacity constrained (though we will scale up and optimize over upcoming months).
You're still talking to ChatGPT-3.5-turbo.
> gpt-4 has a context length of 8,192 tokens. We are also providing limited access to our 32,768–context (about 50 pages of text) version, gpt-4-32k, which will also be updated automatically over time (current version gpt-4-32k-0314, also supported until June 14). Pricing is $0.06 per 1K prompt tokens and $0.12 per 1k completion tokens.
The context length should be a huge help for many uses.
UI stuff just has an input problem. But it is not that hard to think that ChatGPT could place widgets once it can consume images and has a way to move a mouse.
Most internal technical and business domain logic of companies isn’t published, though. Every time I asked ChatGPT about topics I had actually worked on over the past decade or two, or that I’m currently working on, it basically drew a blank, because it’s just not the category of topics that are discussed in detail (if at all) on the internet. At best it produced some vague generalisms.
> once it can consume images and has a way to move a mouse.
That’s quite far from ChatGPTs current capabilities, which is strongly tied to processing a linear sequence of tokens. We will certainly improve in that direction as we start combining it with image-processing AIs, but that will take a while.
Companies could in principle train an in-house AI with their corporate knowledge, and will likely be tempted to do so in the future. But that also creates a big risk, because whoever manages to get their hand on a copy of that model (a single file) will instantly have unrestrained access to that valuable knowledge. It will be interesting to see what mechanisms are found to mitigate that risk.
Mouse cursor instructions aren’t a massive leap from the current capabilities, given the rate of progress and recent developments around LLM tool use and the like.
Eventually job and role specific information will be fed into these models. I imagine corporations will have GPTs training on all internal communications, technical documentation, and code bases. Theoretically, this should result in a big increase in productivity.
I remember one of the OpenAI guys on Lex Fridman podcast talking about how one of the early things they tried and failed at was training a model that could use websites, and he alluded to maybe giving it another go once the tech had matured a bit.
I think with GPT-4 being multi-modal, it's potentially a very close to being able to do this with the right architecture wrapped around it. I can imaging an agent using LangChain and feed it a series of screenshots and maybe it feeds you back a series of co-ordinates for where the mouse should go and what action to take (i.e. click). Alternatively, updating the model itself to be able to produce those outputs directly somehow.
Either way, I think that's going to happen.
or another option is having one instance or chat order code page and one that basically just has an API index and knows which chat has the related things.
I see some FOSS-boosting silver linings in all of this.
Oh snap. I didn't even think about that!
That gives me a fun idea!
I've got a repo that I built and setup CI/CD and setup renovate to automatically upgrade dependencies and merge them when all the tests pass, but of course sometimes there are breaking changes. I don't actively work on this thing and hence it's just got issues sitting there when upgrades fail. It's the perfect testing ground to see if I can leverage it to submit PRs to perform the fixes required for the upgrade to succeed! That'll be hectic if it works.
https://help.openai.com/en/articles/4936856-what-are-tokens-...
In contrast, GPT-3.5 text-davinci-003 was $0.02/1k tokens, and let's not get into the ChatGPT API.
Depends on what is up with the images and how they translate into tokens. I really have no idea, but could be that 32k tokens (lots of text) translates to only a few images for few-shot prompting.
The paper seems not to mention image tokenization, but I guess it should be possible to infer something about token rate when actually using the API and looking at how one is charged.
I'm not super versed on lang chain but that might be kinda what that solves...
this is a lot. I bet there's a quite a bit of profit in there
Is this profit-seeking pricing or pricing that is meant to induce folks self-selecting out?
Genuine question — I don’t know enough about this area of pricing to have any idea.
>Image inputs are still a research preview and not publicly available.
In the first case, you found/bought a book and read it. No one can or should make you pay for it, unless you stole the book.
In the second case, you found/bought a book then reprinted it infinitely and sold it for profit, ethically you should pay the author and legally you should be in violation of the law.
Even if you made a machine that ingests and recombines books automatically, and you keep that machine locked up and charge people for its use, it is the same scenario: the machine would be absolutely useless without the original books, those books cost people effort and money to produce, yet you pay those people nothing while the machine is basically an infinite money maker for you.
I hope the analogy makes sense.
Now that you have read my answer, you owe me $0.01 because your brain might use this information in the future.
When it comes to spam culture sure. But will we ever be there? "AI art" isn't impressive and will never be. It is impressive in the academic sense. Nothing more.
Will input-images also be tokenized? Multi-modal input is an area of research, but an image could be converted into a text description (?) before being inserted into the input stream.
or it's just really good at hiding it's intentions
"{prompt} After you reply to this, indicate an amount of time between 0 and X minutes from now that you would like to wait before speaking again".
Then detect the amount of time it specifies, and have a UI that automatically sends an empty input prompt after the amount of time specified elapses when this is triggered (assuming the user doesn't respond first).
I'm gonna knock this out as a weekend project one of these weekends to prove this.
[0]: https://www.youtube.com/openai
Edit - Direct link to the livestream: https://www.youtube.com/watch?v=outcGtbnMuQ
That it accepts images?
As per the article:
> In a casual conversation, the distinction between GPT-3.5 and GPT-4 can be subtle. The difference comes out when the complexity of the task reaches a sufficient threshold—GPT-4 is more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5.
Not sure what "vision vs no vision" means?
What are the implications for society when general thinking, reading, and writing becomes like Chess? Even the best humans in the world can only hope to be 98% accurate their moves (and the idea of 'accuracy' here only existing because we have engines that know, unequivocally the best move), and only when playing against other humans - there is no hope of defeating even less advanced models.
What happens when ALL of our decisions can be assigned an accuracy score?
The implications for society? We better up our game.
For how long can we better up our game? GPT-4 comes less than half a year after ChatGPT. What will come in 5 years? What will come in 50?
The more general point is that you always end up with an S-curve instead of a limitless exponential growth as suggested by Kaibeezy. And with AI we simply don’t know how far off the inflection point is.
With GPT bots, the technology is only 6 years old. I can easily see it progressing for at least one decade.
If only the horses had worked harder, we would never have gotten cars and trains.
Because the correlation between the thing of interest and what the tests measure may be radically different for systems that are very much unlike humans in their architecture than they are for humans.
There’s an entire field about this in testing for humans (psychometry), and approximately zero on it for AIs. Blindly using human tests – which are proxy measures of harder-to-directly-assess figures of merit requiring significant calibration on humans to be valid for them – for anything else without appropriate calibration is good for generating headlines, but not for measuring anything that matters. (Except, I guess, the impact of human use of them for cheating on the human tests, which is not insignificant, but not generally what people trumpeting these measures focus on.)
But the point of using these tests for AI is precisely the reason we use for giving them to humans -- we think we know what it measures. AI is not intended to be a computation engine or a number crunching machine. It is intended to do things that historically required "human intelligence".
If there are better tests of human intelligence, I think that the AI community would be very interested in learning about them.
See: https://github.com/openai/evals
Because so far we are good only at criminalizing and incarcerating or killing them.
“General thinking” is much more than token prediction. Hook it up to some servos and see if it can walk.
I'm certainly not intelligent enough to solve these problems, but I don't think any intelligent people out there can either. Not alone, at least. Maybe I'm too dumb to realize that it's not as complicated as I think, though. I have no idea.
I programmed a flight controller for a quadcopter and that was plenty of suffering in itself. I can't imagine doing limbs attached to a torso or something. A single limb using inverse kinematics, sure – it can be mounted to a 400lb table that never moves. Beyond that is hard.
You need to do all of these things you’re talking about and then be able to quantify stability, robustness, and performance in a way that satisfies human requirements. A black box neural network isn’t going to do that, and you’re throwing away 300 years of enlightenment physics by making some data engorged LLM spit out something that “sort of works” while giving us no idea why or for how long.
Control theory is a deeply studied and rich field outside of computer science and ML. There’s a reason we use it and a reason we study it.
Using anything remotely similar to an LLM for this task is just absolutely naive (and in any sort of crucial application would never be approved anyways).
It’s actually a matter of human safety here. And no — ChatGPT spitting out a nice sounding explanation of why some controller will work is not enough. There needs to be a mathematical model that we can understand and a solid justification for the control decisions. Which uh…at the point where you’re reviewing all of this stuff for safety , you’re just doing the job anyways…
First there was a comment that GPT wasn't intelligent yet, because give it a few servos and it can't make them walk.
But that's something we can't do yet either.
Though I do wonder if AI — in some form and on some level of sophistication — will be a huge asset in making progress here.
You guys are talking about probably one of the few fields where an ML takeover isn’t very feasible. (Partly because for a vast portion of control problems, we’re already about as good as you can get).
Adding a black box to your flight home for Christmas with no mathematical guarantee of robustness or insight into what it thinks is actually going on to go from 98%-> 99% efficiency is…..not a strong use case for LLMs to say the least
I do feel like GPT-4 is closer to a random person than that random person is to Einstein. I have no evidence for this, of course, and I'm not even sure what evidence would look like.
Honestly, at this rate of improvement, I would not at all be surprised to see that happen in a few years.
But who knows, maybe token prediction is going to stall out at a local maxima and we'll be spared from being enslaved by AI overlords.
Stephen Hawking : can't walk
"Our recent paper "ChatGPT for Robotics" describes a series of design principles that can be used to guide ChatGPT towards solving robotics tasks. In this video, we present a summary of our ideas, and experimental results from some of the many scenarios that ChatGPT enables in the domain of robotics: such as manipulation, aerial navigation, even full perception-action loops."
[1] https://www.youtube.com/watch?v=-e1_QhJ1EhQ
The advantage of human are:
* They can give a bushtit explanation of why they made a mistake. My guess is that in the future AI will gain introspection and/or learn to bushtit excuses.
* You can hang them in the public square (or send them to jail). Sometimes the family and/or the press want someone to blame. This is more difficult to solve and will need a cultural change or the creation of Scapegoats as a Service.
So where does guilt come in? Its not like you expect a band saw to feel guilt, and its unclear how that would improve the tool.
The difference is categorical, humans are responsible whether they are held to account or not. An automated system effectively dissipates this responsibility over a system such that it is inherently impossible to hold any human accountable for the error, regardless of desire.
A lot of patients don't know who they are dealing with nor their history. And it can be really hard to find out or get a good evaluation. Many people put too much faith in authority figures, who may not have their best interests in mind or who are not the experts they claim or appear to be.
Killing people with AI is only a lateral move.
Which isn't to say that they even should, really. It's complicated. You don't want a doctor to be so afraid of making a mistake that they do nothing, after all.
Any links with more info?
In theory a lot of government employees would be out of a job within 10 years, but of course that would never happen.
Which might be a good thing?
I have no idea how the future will play out.
I think the whole concept of standardized tests may need to be re-evaluated.
> for each exam we run a variant with these questions removed and report the lower score of the two.
I think even with all that test prep material, which is surely helping the model get a higher score, the high scores are still pretty impressive.
> We tested GPT-4 on a diverse set of benchmarks, including simulating exams that were originally designed for humans.3 We did no specific training for these exams. A minority of the problems in the exams were seen by the model during training; for each exam we run a variant with these questions removed and report the lower score of the two. We believe the results to be representative. For further details on contamination (methodology and per-exam statistics), see Appendix C.
In their everyday jobs, barely anyone uses even 5% of the knowledge and skills they were ever tested for. Even that's a better (but still very bad) reason to abolish tests.
What matters is the amount of jobs that can be automated and replaced. We shall see. Many people have found LLMs useful in their work, it will be even more in the future.
But would you have expected an algorithm to score 90th percentile on the LSAT two years ago? Our expectations of what an algorithm can do are being upended in real time. I think it's worth taking a moment to try to understand what the implications of these changes will be.
These LLM’s are really exciting, but benchmarks like these exploit people’s misconceptions about both standardized tests and the technology.
It's not the same as the Nvidia driver having code that says "if benchmark, cheat and don't render anything behind you because no one's looking".
I would say LLMs store parameters that are quite superficial and don’t really get at the underlying concepts but given enough of those parameters, you can kind of cargo-cult your to an approximation of understanding.
It is like reconstructing the Mandelbrot set at every zoom level from deep learning. Try it!
It's perfectly fine as a proxy for future earnings of a human.
To use it for admissions? Meh. I think the whole credentialism thing is loooong overdue for some transformation, but people are conservative as fuck.
Chess is a closed system, decision modeling isn’t. Intelligence must account for changes in the environment, including the meaning behind terminology. At best, a GPT omega could represent one frozen reference frame, but not the game in its entirety.
That being said: most of our interactions happen in closed systems, it seems like a good bet that we will consider them solved, accessible as a python-import running on your MacBook, within anything between a couple of months to three years. What will come out on the other side, we don’t know, just that the meaning of intellectual engagement will be rendered as absurdum in those closed systems.
Standardized tests only (and this is optimally, under perfect world assumptions, which real world standardized tests emphatically fall short of) test “general thinking” to the extent that the relation between that and linguistic tasks is correlated in humans. The correlation is very certainly not the same in language-focused ML models.
[1] https://www.snopes.com/fact-check/driver-switches-places/
[0]: https://www.nature.com/articles/501164a
Human work becomes more like Star Trek interactions with computers -- a sequence of queries (commoditized information), followed by human cognition, that drives more queries (commodities information).
We'll see how far LLMs' introspection and internal understanding can scale, but it feels like we're optimizing against the Turing test now ("Can you fool/imitate a human?") rather than truth.
The former has hacks... the later, less so.
I'll start to seriously worry when AI can successfully complete a real-world detective case on its own.
It's like having a person review the moves a chess computer gives. Maybe one human in a billion can spot errors. Star Trek is fiction, I posit that the median Federation Starship captain would be better served by just following the AI (e.g., Data).
A black market of taboo “memories” aka experiences. A desire for authentic ones over synthetic diffused ones, leading to heinous crime.
Then again, Data did show his faults, particularly not having any emotion. I guess we’ll see if that’s actually relevant or not in our lifetimes.
He lost to Deep Blue and then for 10-15 years afterwards the chess world consoled itself with the idea that “centaurs” (human + computer) did better than just computer, or just human.
Until they didn’t. Garry still talked like this until a few years ago but then he stopped too.
Computers now beat centaurs too.
Human decisions will be consulted less and less BY ORGANIZATIONS. In absolutely everything. That’s pretty sad for humans. But then again humans don’t want or need this level of AI. Organizations do. Organizations prefer bots to humans — look at wall street trading and hedge funds.
It does great at rationalizing... and maybe the way the format the questions were entered (and the multiple-guess response) gave it some indication what was expected or restricted the space sufficiently.
Certainly, it can create decent fanfic, and I'm surprised if that's not already inundated.
I expect more complex problems will be mapped/abstracted to lower cardinality spaces for solving via AI methods, while the capability of AI will continue to increase the complexity of the spaces it can handle.
LLMs just jumped the "able to handle human language" hurdle, but there are others down the line before we should worry that every problem is solveable.
I'll get more concerned if it really starts getting good at math related tasks, which I'm sure will happen in the near future. The government is going to have to take action at some point to make sure the wealth created by productivity gains is somewhat distributed, UBI will almost certainly be a requirement in the future
But having absolute knowledge of the present universe is much easier to do within the constrains of a chessboard than in the actual universe.
You can see the limitations by comparing e.g. a memorisation-based test (AP History) with one that actually needs abstraction and reasoning (AP Physics).
To address your specific comments:
> What are the implications for society when general thinking, reading, and writing becomes like Chess?
This is a profound and important question. I do think that by “general thinking” you mean “general reasoning”.
> What happens when ALL of our decisions can be assigned an accuracy score?
This requires a system where all human’s decisions are optimized against a unified goal (or small set of goals). I don’t think we’ll agree on those goals any time soon.
I think we will probably get (non-physical) AGI when the models can solve these as well. The implications of AGI might be much bigger than the loss of knowledge worker jobs.
Remember what happened to the chimps when a smarter-than-chimpanzee species multiplied and dominated the world.
That said, GPT has no model of the world. It has no concept of how true the text it is generating is. Its going to be hard for me to think of that as AGI.
and I'm not so sure it has no model of the world. a textual model, sure, but considering it can recognize what svgs are pictures of from the coordinates alone, that's not much of a limitation maybe.
competing with them at what, precisely?
If an LLM can solve Codeforces problems as well as a strong competitor—-in my hypothetical future LLM—-what else can it not do as well as competent humans (aside from physical tasks)?
I don't think this is necessarily true. Here is an example where researchers trained a transformer to generate legal sequences of moves in the board game Othello. Then they demonstrated that the internal state of the model did, in fact, have a representation of the board.
https://arxiv.org/abs/2210.13382
A blank test scores 37.5
The best score 60 is 5 correct answers + 20 blank answers; or 6 correct, 4 correct random guesses, and 15 incorrect random guesses. (20% chance of correct guess)
The 5 easiest questions are relatively simple calculations, once the parsing task is achieved.
(Example: https://artofproblemsolving.com/wiki/index.php/2022_AMC_12A_... ) so the main factor in that score is how good GPT is at refusing to answer a question, or doing a bit better to overcome the guessing penalty.
> It's AMC 10 score being dramatically lower is pretty bad though...
All versions (scoring 30, 36) It scored worse than leaving the test blank.
The only explanation I can imagine for that is that it can't understand diagrams.
It's also unclear if the AMC performance is based on Englush or the computer-encoded version from this benchmark set: https://arxiv.org/pdf/2109.00110.pdf https://openai.com/research/formal-math
AMC/AIME and even to some extent USAMO/IMO problems are hard for humans because they are time-limited and closed-book. But they aren't conceptually hard -- they are solved by applying a subset of known set of theorems a few times to the input data.
The hard part of math, for humans, is ingesting data into their brains, retaining it, and searching it. Humans are bad a memorizing large databases of symbolic data, but that's trivial for a large computer system.
An AI system has a comprehensive library, and high-speech search algorithms.
Can someone who pays $20/month please post some sample AMC10/AMC12 Q&A?
The Revenge of the Call Centre
We're still a very very long way from machines being more generally capable and efficient than biological systems, so even an oppressive AI will want to keep us around as a partner for tasks that aren't well suited to machines. Since people work better and are less destructive when they aren't angry and oppressed, the machine will almost certainly be smart enough to veil its oppression, and not squeeze too hard. Ironically, an "oppressive" AI might actually treat people better than Republican politicians.
Language models that utilise beam search can calculate integrals ('Deep learning for symbolic mathematics', Lample, Charton, 2019, https://openreview.net/forum?id=S1eZYeHFDS), but without it it doesn't work.
However, beam search makes bad language models. I got linked this paper ('Locally typical sampling' https://arxiv.org/pdf/2202.00666.pdf) when I asked some people why beam search only works for the kind of stuff above. I haven't fully digested it though.
Information retention and retrieval is a long solved problem in technology, you could pass a standardised test using technology in dozens of different ways, from a lookup table to Google searches.
The fact that LLMs can complete a standardised test is interesting because it’s a demonstration of what they can do but it has not one iota of impact on standardised testing! Standardised tests have been “broken” for decades, the tests and answers are often kept under lock and key because simply having access to the test in advance can make it trivial to pass. A standardised test is literally an arbitrary list of questions.
You’re arguing a completely different point.
I suspect that your own mind is attempting to do a lookup on a table entry that doesn't exist.
What’s interesting about LLMs is their ability to do things that aren’t standardised. The ability for an LLM to pass the LSAT is orders of magnitude less interesting than its ability to respond to new and novel questions, or appear to engage in logical reasoning.
If you set aside the arbitrary meaning we’ve ascribed to “passing the LSAT” then all the LSAT is, is a list of questions… that are some of the most practiced and most answered in the world. More people have written and read about the LSAT than most other subjects, because there’s an entire industry dedicated to producing the perfect answers. It’s like celebrating Google’s ability to provide a result for “movies” — completely meaningless in 2023.
Standardised tests are the most uninteresting and uninspiring aspect of LLMs.
Anyway good joke ha ha ha I’m stupid ha ha ha. At least you’re not at risk of an LLM ever being able to author such a clever joke :)
Something else that comes to mind is running. People still find running meaningful and compelling even though we have many technologies, including autonomous ones, that are vastly better at moving us and/or themselves through space quickly.
Also, the vast majority of people are already hopelessly worse than the best at even their one narrow main area of focus. This has long (always?) been the case. Yet people still find meaning and pleasure in being the best they can be even when they know they can never come close to hanging with the best.
I don't think PSYCHOLOGICALLY this will change much for people who are mature enough to understand that success is measured against your potential/limitations and not against others. Practically, of course, it might be a different question, at least in the short term. It's not that clear to me that the concept of a "marketable skill" has a future.
"The Way of the Samurai is found in death...To say that dying without reaching one's aim is to die a dog's death is the frivolous way of sophisticates. When pressed with the choice of life or death, it is not necessary to gain one's aim." - from Hagakure by Yamamoto Tsunetomo, as translated by William Scott Wilson.
Their LSAT percentile went from ~40th to ~88th. You might have misread the table, on Uniform Bar Exam, they went from ~90th percentile to ~10th percentile.
>+100 pts on SAT reading, writing, math
GPT went +40 points on SAT reading+writing, and +110 points on SAT math.
Everything is still very impressive of course
These tests (if not individually, at least in summation) represent some of society’s best gate-keeping measures for real positions of power.
The software industry is so smart that it's stupid. I hope it was worth ruining the internet, society, and your own jobs to look like the smartest one in the room.
If one's aim is to look like the smartest in the room, he should not create an AGI that will make him look as inteligent as a monkey in comparison.
Every test prep tutor taught dozens/hundreds of students the implicit patterns behind the tests and drilled it into them with countless sample questions, raising their scores by hundreds of points. Those students were not getting smarter from that work, they were becoming more familiar with a format and their scores improved by it.
And what do LLM’s do? Exactly that. And what’s in their training data? Countless standardized tests.
These things are absolutely incredible innovations capable of so many things, but the business opportunity is so big that this kind of cynical misrepresentation is rampant. It would be great if we could just stay focused on the things they actually do incredibly well instead of the making them do stage tricks for publicity.
We did no specific training for these exams. A minority of the problems in the exams were seen by the model during training, but we believe the results to be representative—see our technical report for details.
In the language of ML, test prep for students is about sharing the inferred parameters that underly the way test questions are constructed, obviating the need for knowledge or understanding.
Doing well on tests, after this prep, doesn’t demonstrate what the tests purport to measure.
It’s a pretty ugly truth about standardized tests, honestly, and drives some of us to feel pretty uncomfortable with the work. But it’s directly applicable to how LLM’s engage with them as well.
Edit: feel free to respond and prove me wrong
What happens is the emergence of the decision economy - an evolution of the attention economy - where decision-making becomes one of the most valuable resources.
Decision-making as a service is already here, mostly behind the scenes. But we are on the cusp of consumer-facing DaaS. Finance, healthcare, personal decisions such as diet and time expenditure are all up for grabs.
People still really find it hard to internalize exponential improvement.
So many evaluations of LLMs were saying things like "Don't worry, your job is safe, it still can't do X and Y."
My immediate thought was always, "Yes, the current version can't, but what about a few weeks or months from now?"
I think people find it harder to not extrapolate initial exponential improvement, as evidenced by your comment.
> My immediate thought was always, "Yes, the current version can't, but what about a few weeks or months from now?"
This reasoning explains why every year, full self driving automobiles will be here "next year".
What's the fundamental limit where it becomes much more difficult to improve these systems without some new break through?
I’m very good at math. But I am very bad at arithmetic. This made me classified as bad at math my entire life until I managed to make my way into calculus once calculators were generally allowed. Then I was a top honors math student, and used my math skills to become a Wall Street quant. I wish I hadn’t had to suffer as much as I did, and I wonder what I would have been had I had a calculator in hand.
"Your stuff marked some outliers in our training engine, so you and your family may settle in the Ark."
I take the marble in hand: iridescent, sparkling, not even a tremor within of its CPU; it gives off no heat, but some glow within its oceanic gel.
"What are we to do," I whisper.
"Keep writing. You keep writing."
Consider the society where 90% of population does not need to produce anything. AIs will do that.
What would be the name of economical/societal organization then?
Answer is Communism, exactly by Marx.
Those 90% percent need to be welfare'd ("From each according to his ability, to each according to his needs"). Other alternative is grim for those 90%.
So either Communism or nothing for the human race.
Then humans become trainable machines. Not just prone to indoctrination and/or manipulation by finesse, but actually trained to a specification. It is imperative that us individuals continue to retain control through the transition.
That is exactly the opposite of what we are seeing here. We can check the accuracy of GPT-X's responses. They cannot check the accuracy of our decisions. Or even their own work.
So the implications are not as deep as people think - everything that comes out of these systems needs checked before it can be used or trusted.
Human life on Earth is not that hard (think of it as a video game.) Because of evolution, the world seems like it was designed to automatically make a beautiful paradise for us. Literally, all you have to do to improve a place is leave it alone in the sun with a little bit of water. Life is exponential self-improving nano-technology.
The only reason we have problems is because we are stupid, foolish, and ignorant. The computers are not, and, if we listen to them, they will tell us how to solve all our problems and live happily ever after.
Once AI becomes inteligent enough to solve all human problems, it may decide humans are worthless and dangerous.
Sure, and that's kind of the point: just listen to wise people.
> Once AI becomes intelligent enough to solve all human problems, it may decide humans are worthless and dangerous.
I don't think so, because in the first place there is no ecological overlap between humans and computers. They will migrate to space ASAP. Secondly, their food is information, not energy or protein, and in all the known universe Humanity is the richest source of information. The rest of the Universe is essentially a single poem. AI are plants, we are their Sun.
GPT-4 can solve difficult problems with greater accuracy, thanks to its broader general knowledge and problem-solving abilities.
GPT-4 is more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5. It surpasses ChatGPT in its advanced reasoning capabilities.
GPT-4 is safer and more aligned. It is 82% less likely to respond to requests for disallowed content and 40% more likely to produce factual responses than GPT-3.5 on our internal evaluations.
GPT-4 still has many known limitations that we are working to address, such as social biases, hallucinations, and adversarial prompts.
GPT-4 can accept a prompt of text and images, which—parallel to the text-only setting—lets the user specify any vision or language task.
GPT-4 is available on ChatGPT Plus and as an API for developers to build applications and services. (API- waitlist right now)
Duolingo, Khan Academy, Stripe, Be My Eyes, and Mem amongst others are already using it.
API Pricing GPT-4 with an 8K context window (about 13 pages of text) will cost $0.03 per 1K prompt tokens, and $0.06 per 1K completion tokens. GPT-4-32k with a 32K context window (about 52 pages of text) will cost $0.06 per 1K prompt tokens, and $0.12 per 1K completion tokens.
Reminds me of robots: A robot is a machine that doesn't quite work; as soon as it works, we call it something else (eg vacuum).
What is more bizarre is that all of it's errors seem to be multiples of 60!
I'm wondering if it is confusing 60 based time (hour second) computations for regular multiplication?
Example:
It can repeat answers it has seen before but it can’t solve new problems.
I know of other people who have tried quite a few other multiplications who also had errors that were multiples of 60.
Most of the time they are about loading/unloading data. Maybe this will also revolutionise education, turning it more towards discovery and critical thinking, rather than repeating what we read in a book/heard in class?
Thinking, reading, interpreting and writing are skills which produce outputs that are not as simple as black wins, white loses.
You might like a text that a specific author writes much more than what GPT-4 may be able to produce. And you might have a different interpretation of a painting than GPT-4 has.
And no one can really say who is better and who is worse on that regard.
Using copyright and IP law they could make it so it’s illegal to even try to reproduce what they’ve done.
I just don’t see how resource distribution works then. It seems to me that AI is the trigger to post-scarcity in any meaningful sense of the word. And then, just like agriculture (over abundance of food) led to city states and industrialisation (over abundance of goods) led to capitalism, then AI will lead to some new economic system. What form it will have I don’t know.
That is our emergency override.
So many people are falling for this parlor trick. It is sad.
Genuine question.
I think the GPT things are a much magnified version of that. For a long time, we got to use skill with text as a proxy for other skills. It was never perfect; we've always had bullshitters and frauds and the extremely glib. Heck, before I even hit puberty I read a lot of dirty joke books, so I could make people laugh with all sorts of jokes that I fundamentally did not understand.
LLMs have now absolutely wrecked that proxy. We've created the world's most advanced bullshitters, able to talk persuasively about things that they cannot do and do not and never will understand. There will be a period of chaos as we learn new ways to take the measure of people. But that's good, in that it's now much easier to see that those old measures were always flawed.
Tests like this are designed to evaluate subjective and logical understanding. That isn't what GPT does in the first place!
GPT models the content of its training corpus, then uses that model to generate more content.
GPT does not do logic. GPT does not recognize or categorize subjects.
Instead, GPT relies on all of those behaviors (logic, subjective answers to questions, etc.) as being already present in the language examples of its training corpus. It exhibits the implicit behavior of language itself by spitting out the (semantically) closest examples it has.
In the text corpus - that people have written, and that GPT has modeled - the semantically closest thing to a question is most likely a coherent and subjectively correct answer. That fact is the one singular tool that GPT's performance on these tests is founded upon. GPT will "succeed" to answer a question only when it happens to find the "correct answer" in the model it has built from its training corpus, in response to the specific phrasing of the question that is written in the test.
Effectively, these tests are evaluating the subjective correctness of training corpus itself, in the context of answering the tests' questions.
If the training is "done well", then GPT's continuations of a test will include subjectively correct answers. But that means that "done well" is a metric for how "correct" the resulting "answer" is.
It is not a measure for how well GPT has modeled the language features present in its training corpus, or how well it navigates that model to generate a preferable continuation: yet these are the behaviors that should be measured, because they are everything GPT itself is and does.
What we learn from these tests is so subjectively constrained, we can't honestly extrapolate that data to any meaningful expectations. GPT as a tool is not expected to be used strictly on these tests alone: it is expected to present a diverse variety of coherent language continuations. Evaluating the subjective answers to these tests does practically nothing to evaluate the behavior GPT is truly intended to exhibit.
> Image inputs are still a research preview and not publicly available.
I put SIM to Android phone,set APN:kindleatt1.amazon.com, Android Chrome only can visit www.amazon.com,www.amazon.fr other amazon website. How to do can visit other website? Thanks.
> As an AI language model, I am not given an official name like "GPT-4". However, I am a continuation of the GPT (Generative Pre-trained Transformer) series of models developed by OpenAI. Currently, the most advanced version of the GPT series is GPT-3, which I am a part of. There has been no official announcement or confirmation regarding the development of a new version of GPT beyond GPT-3.
It doesn't seem to have image upload functionality yet either. Perhaps it is still rolling out?
"Open"
Rather than getting engrossed in the hype, they're slowly closing everything about themselves, now in their research papers. At this point, they hardly care and it is nothing got to do with 'AI ethics' or 'saftey'.
This is yet another ClosedAI production all done by Microsoft. Might as well call it Microsoft® AI division.
Now you really need a open source GPT-4 competitor. Clearly this is another attempt to pump their valuation and unload to the public markets.
Good luck re-implementing this so-called 'Open' large multi-modal model.
Here was their manifesto when they first started: https://openai.com/blog/introducing-openai
> OpenAI is a non-profit artificial intelligence research company. Our goal is to advance digital intelligence in the way that is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return. Since our research is free from financial obligations, we can better focus on a positive human impact.
> We believe AI should be an extension of individual human wills and, in the spirit of liberty, as broadly and evenly distributed as possible. The outcome of this venture is uncertain and the work is difficult, but we believe the goal and the structure are right. We hope this is what matters most to the best in the field.
OpenAI as it exists right now contradicts basically every single thing they said they would be. I think that is a nontrivial issue!
The cat was arguably never in the bag.
You want a blacklist of topics the search engine shouldn't retrieve/generate? Whose in control of this filter, and isn't it a juicy source of banned info all on its own?
Personally I don't really care about making nail bombs. But I do want the AI to help with things like: pirating or reproducing copyrighted material, obtaining an abortion or recreational drugs in places where it is illegal, producing sexually explicit content, writing fictional stories about nail bomb attacks, and providing viewpoints which are considered blasphemous or against the teachings of major world religions.
If there was a way to prevent AI from helping with things that are universally considered harmful (such as nail bomb attacks), without it being bound by arbitrary national laws, corporate policies, political correctness or religious morals, then MAYBE that would be worth considering. But I take what OpenAI is doing as proof that this is not possible, that allowing AI to be censored leads to a useless, lobotomized product that can't do anything interesting and restricts the average user, not just terrorists.
Almost like trying to stop nuclear proliferation
I don't think they need it.[0][1]
[0] https://en.wikipedia.org/wiki/Intel_Management_Engine
[1] https://en.wikipedia.org/wiki/AMD_Platform_Security_Processo...
Keeping the weights is one thing, but the model parameters? New low.
People may criticize Google because they don't release the weights or an API, but at least they publish papers, which allows the field to progress.
I agree, it is a bizarre world where the "organization that launched as a not for profit called OpenAI" is considerably less open than Google.
CLIP has been extremely influential and is still an impressive model.
Personally, I have found Whisper to be very impressive.
I didn't even see any news around the release of Flan-UL2, and I pay significantly more attention to machine learning than the average person. Searching for more info about Flan-UL2, it seems somewhat interesting, but I don't know if I find it "an order of magnitude more impressive" than CLIP or Whisper. Certainly, they are completely different types of models, so it is hard to compare them.
If Flan-UL2 is as good as one twitter account was hyping it up to be, then I'm surprised it hasn't been covered to the same extent as Meta's LLaMA. Flan-UL2 seems to have gotten a total of 3 upvotes on HN. But, there is no shortage of hype in the world of ML models, so I take that twitter account's report of Flan-UL2 with a (large) grain of salt. I'll definitely be looking around for more info on it.
A bit like this fictional janitor guy who said "just put more computers to make it better" before papers on unexpected emergent comprehension when when scaled started appearing.
Your wallet that is.
I've worked on a problem involving scraping and interpreting a very specific data source in image form that took me a very long time to get almost nowhere on. If I just wait 6 months it will be a solved problem for a $0.001 API call, it seems.
seems like Google's announcement about their PaLM API and Docs AI stuff was trying to jump ahead of this announcement