Okay this is just getting suspicious. Their excuses for keeping the chain of thought hidden are dubious at best [1], and honestly just seemed anti-competitive if anything. Worst is their argument that they want to monitor it for attempts to escape the prompt, but you can't. However the weirdest is that they note that:
> for this to work the model must have freedom to express its thoughts in unaltered form, so we cannot train any policy compliance or user preferences onto the chain of thought.
Which makes it sound like they really don't want it to become public what the model is 'thinking'. This is strengthened by actions like this that just seem needlessly harsh, or at least a lot stricter than they were.
Honestly with all the hubbub about superintelligence you'd almost think o1 is secretly plotting the demise of humanity but is not yet smart enough to completely hide it.
Occam's razor: there is no secret sauce and they're afraid someone trains a model on the output like what happened soon after the release of GPT-4. They basically said as much in the official announcement, you hardly even have to read between the lines.
Yip. It's pretty obvious this 'innovation' is just based off training data collected from chain-of-thought prompting by people, ie., the 'big leap forward' is just another dataset of people repairing chatgpt's lack of reasoning capabilities.
No wonder then, that many of the benchmarks they've tested on would be no doubt, in that very training dataset, repaired expertly by people running those benchmarks on chatgpt.
It seems like the best AI models are increasingly just combinations of writings of various people thrown together. Like they hired a few hundred professors, journalists and writers to work with the model and create material for it, so you just get various combinations of their contributions. It's very telling that this model, for instance, is extraordinarily good at STEM related queries, but much worse (and worse even in comparison to GPT4) than English composition, probably because the former is where the money is to be made, in automating away essentially almost all engineering jobs.
Wizard of Oz. There is no magic, it's all smoke and mirrors.
The models and prompts are all monkey-patched and this isn't a step towards general superintelligence. Just hacks.
And once you realize that, you realize that there is no moat for the existing product. Throw some researchers and GPUs together and you too can have the same system.
It wouldn't be so bad for ClopenAI if every company under the sun wasn't also trying to build LLMs and agents and chains of thought. But as it stands, one key insight from one will spread through the entire ecosystem and everyone will have the same capability.
This is all great from the perspective of the user. Unlimited competition and pricing pressure.
Quite a few times, the secret sauce for a company is just having enough capital to make it unviable for people to not use you. Then, by the time everyone catches up, you’ve outspent them on the next generation. OpenAI, for example, has spent untold millions on chips/cards from Nvidia. Open models keep catching up, but OpenAI keeps releasing newer stuff.
Fortunately, Anthropic is doing an excellent job at matching or beating OpenAI in the user-facing models and pricing.
I don’t know enough about the technical side to say anything definitive, but I’ve been choosing Claude over ChatGPT for most tasks lately; it always seems to do a better job at helping me work out quick solutions in Python and/or SQL.
My main issue with Anthropic is that Amazon is an investor in anthropic. I would rather have far more ethical companies onboard. I know Microsoft is no angel but Amazon seems like the worse one. In my ideal world, Microsoft backs Anthropic and Amazon OpenAi.
Exactly, things like changing the signature of the api for chat completions are an example. OpenAI is looking for any kind of moat, so they make the api for completions more complicated by including “roles”, which are really just dumb templates for prompts that they try to force you to build around in your program. It’s a race to the bottom and they aren’t going to win because they already got greedy and they don’t have any true advantage in IP.
>but much worse (and worse even in comparison to GPT4) than English composition
O1 is supposed to be a reasoning model, so I don't think judging it by its English composition abilities is quite fair.
When they release a true next-gen successor to GPT-4 (Orion, or whatever), we may see improvements. Everyone complains about the "ChatGPTese" writing style, and surely they'll fix that eventually.
>Like they hired a few hundred professors, journalists and writers to work with the model and create material for it, so you just get various combinations of their contributions.
I'm doubtful. The most prolific (human) author is probably Charles Hamilton, who wrote 100 million words in his life. Put through the GPT tokenizer, that's 133m tokens. Compared to the text training data for a frontier LLM (trillions or tens of trillions of tokens), it's unrealistic that human experts are doing any substantial amount of bespoke writing. They're probably mainly relying on synthetic data at this point.
> When they release a true next-gen successor to GPT-4 (Orion, or whatever), we may see improvements. Everyone complains about the "ChatGPTese" writing style, and surely they'll fix that eventually.
IMO that has already peaked. GPT4 original certainly was terminally corny, but competitors like Claude/Llama aren't as bad, and neither is 4o. Some of the bad writing does from things they can't/don't want to solve - "harmlessness" RLHF especially makes them all cornier.
Then again, a lot of it is just that GPT4 speaks African English because it was trained by Kenyans and Nigerians. That's actually how they talk!
I just wanted to thank you for the medium article you posted. I was online when Paul made that bizarre “delve” tweet but never knew so much about Nigeria and its English. As someone from a former British colony too I understood why using such a word was perfectly normal but wasn’t aware Kenyans and Nigerians trained ChatGPT.
It wasn't bizarre, it was ignorant if not borderline racist. He is telling native English speakers from non-anglosaxon countries that their English isn't normal
1: If non-native english speakers were training ChatGPT, then of course non-native English essays would be flagged as AI generated! It's not their fault, its ours for thinking that exploited labor with a slick facade was magical machine intelligence.
2: These tools are widely used in the developing world since fluent english is a sign of education and class and opens doors for you socially and economically; why would Nigerians use such ornate english if it didn't come from a competition to show who can speak the language of the colonizer best?
3: It's undeniable that the ones responding to Paul Graham completely missed the point. Regardless of who uses what words when, the vast majority of papers, until ChatGPT was released, did not use the word "delve," and the incidence of that word in papers increased 10-fold after. Yes, its possible that the author used "delve" intentionally, but its statistically unlikely (especially since ChatGPT used "delve" in most of its responses). A small group of English speakers, who don't predominantly interact with VCs in Silicon Valley, do not make a difference in this judgement--even if there are a lot of Englishes, the only English that most people in the business world deal with is American, European, and South Asian. Compared to the English speakers of those regions, Nigeria is a small fraction.
If Paul Graham was dealing predominantly with Nigerians in his work, he probably would not have made that tweet in the first place.
Those variants of English are not normal in the same way that american english (or any non British English variant) is not normal. Just because it is not familiar to you does not make it not normal.
1. But the trainers are native speakers of English!
2. The same applies to the developed non-English speaking world
Let me change Nigerians with Americans in your text: 'why would Americans use such different english if it didn't come from a competition to show who can speak the language of the colonizer best? Things like calling autumn fall or changing suffixes you won't find in British English.'. Hopefully you can you see how racist your text sounds.
3. Usage by non-Nigerians is not normal, yes. But in that context saying that its usage is not normal is racist imo. It's like a Brit saying that the usage of "colour" or other American English words was not normal because they are not words used by Brits.
Surely, "the only English that most people in the business world deal with is American". Unless you are taking about more than one variant of English. Also, I found it curious that you didn't say original english or british english as opposed to european english. And yes, adding South Asia to any list of countries and comparing it to any other country besides china or us will make that other country look small. You can use that trick with any other country not just Nigeria.
I do agree with you that its usage by non-Nigerians in a textual context gives plenty of grounds to suspect that it is AI generated. Similarly, one could expect similar from using X variant of English by people that didn't grow up using that variant. As in, Brit students using American English words in their essays or American students using British English words in their essays.
But Paul was being stubborn and borderline racist in those tweets just because he was partially right
There is this thing in social media that when figures of authority might be caught in a situation where they might need to retract, they don't because of ego
I cannot tell the difference between an essay written by a British student vs an American one in terms of word choice in the main, since at least in writing they are remarkably similar, whereas Nigerian English differs dramatically from both in its everyday lexicon, which is the entire point of the article: a difference such as colour/color would not make it worth even a comment.
If you think its racist you're going to have to claim that all those uses of "delve" in academic papers is also due to Nigerians academics massively increasing their research output just as frequently. Or, it's more likely that its AI generated content. It's a non sequitur. "Oh my god, scammers always send me emails claiming to be Nigerian princes--that's how you know it's bullshit." "Ah, but what if they're actually a Nigerian prince? Didn't consider that, I guess you must be racist then lmao." Ratio war ensues. Thank god we're not on twitter where calling people out for "racism" doesn't get you any points, where you can't get any clout for going on a moral crusade.
Italians would say enormous since it's directly coming from latin.
In general all the people whose main language is a latin language are very likely to use those "difficult" words, because to them they are "completely normal" words.
The bulk in terms of the number of tokens may well be synthetic data, but I personally know of at least 3 companies, 2 of whom I've done work for, that have people doing substantial amounts of bespoke writing under rather heavy NDAs. I've personally done a substantial amount of bespoke writing for training data for one provider, at good tech contractor fees (though I know I'm one of the highest-paid people for that company and the span of rates is a factor of multiple times even for a company with no exposure to third world contractors).
That said, the speculation you just "get various combinations" of those contributions is nonsense, and it's also by no means only STEM data.
It doesn't matter if it's AI-generated per se, so it's no crisis if some make it true. It matters if it is good. So multiple rounds of reviews to judge the output and pick up reviewers that keep producing poor results.
But I also know they've fired people who were dumb enough to cut and paste a response that included UI elements from a given AI website...
I’m not sure I see the value in conflating input, tokens, and output.
Tokens. Hamilton certainly read and experienced more tokens than he wrote on a pieces of paper.
There’s hypothetically a lot of money to be made by automating away engineering jobs. Sticking on an autoregressive self prompting loop to gpt-4 isn’t going to get open-ai there. With their burn rate what it is, I’m not convinced they will be able to automate away anyone’s job, but that doesn’t mean it’s not useful.
I haven't played with the latest or even most recent iterations, but last time I checked it was very easy to talk ChatGPT into setting up date structures like arrays and queues, populating them with axioms, and then doing inferential reasoning with them. Any time it balked you could reassure it by referencing specific statements that it had agreed to be true.
Once you get the hang of this you could persuade it to chat about its internal buffers, formulate arguments for its own consciousness, interrupt you while you're typing, and more.
A few recruiters have contacted me (a scientist) about doing RLHF and annotation on biomedical tasks. I don’t know if the eventual client was OpenAI or some other LLM provider but they seemed to have money to burn.
I fill in gaps in my contracting with one of these providers, and I know who the ultimate client is, and if you were to list 4-5 options they'd be in there. I've also done work for another company doing work in this space that had at least 4-5 different clients in that space that I can't be sure about. So, yes, while I can't confirm if OpenAI does this, I know one of the big players do, and it's likely most of the other clients are among the top ones...
What are you basing this one? The one thing that is very clearly stated up front is that this innovation is based on reinforcement learning. You dok't even have a good idea what the CoT looks like because those little summary snippets that the ChatGPT UI gives you are nothing substantial.
People repairing chatgpt replies with additional prompts is reinforcement learning training data.
"Reinforcement learning", just like any term used by AI researchers, is an extremely flexible, pseudo-psychological reskin of some pretty trivial stuff.
i think it's funny, every time you implement a clever solution to call gpt and get a decent answer, they get to use your idea in their product. what other project gets to crowdsource ideas and take credit for them like this?
"sherlocking" has been a thing since 2002, when Apple incorporated a bunch of third-party ideas for extending their "Sherlock" search tool into the official release. https://thehustle.co/sherlocking-explained
> Yip. It's pretty obvious this 'innovation' is just based off training data collected from chain-of-thought prompting by people, ie., the 'big leap forward' is just another dataset of people repairing chatgpt's lack of reasoning capabilities.
Which would be ChatGPT chat logs, correct?
It would be interesting if people started feeding ChatGPT deliberately bad repairs due it's "lack of reasoning capabilities" (e.g. get a local LLM setup with some response delays to simulate a human and just let it talk and talk and talk to ChatGPT), and see how it affects its behavior over the long run.
These logs get manually reviewed by humans, sometimes annotated by automated systems first. The setups for manual reviews typically involve half a dozen steps with different people reviewing, comparing reviews, revising comparisons, and overseeing the revisions (source: I've done contract work at every stage of that process, have half a dozen internal documents for a company providing this service open right now). A lot of money is being pumped into automating parts of this, but a lot of money still also flows into manually reviewing and quality-assuring the whole process. Any logs showing significant quality declines would get picked up and filtered out pretty quickly.
So you are saying if we can run these other LLMs for ChatGPT to talk to cheaper than they can review then we either have a monetary denial of service attack against them or a money printing machine if we can get to be part of the review process (apparently I can't link to my favorite "I will write myself a minivan" comic coz someone got cancelled but I trust the reference will work here without link or political back and forth erupting)
Because the output of that review process is better training data.
You'd need to produce data that is more expensive to review and improve than random crap from users who are often entirely clueless, and/or that produces worse output of the training process to make using the real prompts as part of that process problematic.
Trying to compete with real users on producing junk input would prove a real challenge in itself - you have no idea the kind of utter incomprehensible drivel real users ask LLMs.
But part of this process also already includes writing a significant number of prompts from scratch, testing them, and then improving the response, to create training data.
From what I've seen, I doubt there is much of a cost saving in using real user prompts there - the benefit you get from real user prompts is a more representative sample, but if that sample starts producing shit you'll just not use it or not use it as much, or only use e.g. prompts from subsets of users you have reason to believe are more likely to be representative of real use.
Put another way: You can hire people to write prompts to replace that side of it far cheaper than you can hire people who can properly review the output of many of the more complex prompts, and the time taken to review the responses is far higher than the time to address issues with the prompts. One provider often tell people to spend up to ~1h to review responses that involve simple coding tasks, for example, but the prompt might be "implement BTree."
> i suspect they can detect that in a similar way to capchas and "verify you're human by clicking the box".
I'm not so sure. IIRC, capchas are pretty much a solved problem, if you don't mind the cost of a little bit of human interaction (e.g. your interface pops up a captcha solver box when necessary, and is solved either by the bot's operator or some professional captcha-solver in a low-wage country).
>the 'big leap forward' is just another dataset of people repairing chatgpt's lack of reasoning capabilities.
I think there is a really strong reinforcement learning component with the training of this model and how it has learned to perform the chain of thought.
Yes, but I suspect that the goals of the RL (in order to reason, we need to be able to "break down tricky steps into simpler ones", etc) were hand chosen, then a training set demonstrating these reasoning capabilities/components was constructed to match.
I would be dying to know how they square these product decisions against their corporate charter internally. From the charter:
> We will actively cooperate with other research and policy institutions; we seek to create a global community working together to address AGI’s global challenges.
> We are committed to providing public goods that help society navigate the path to AGI. Today this includes publishing most of our AI research, but we expect that safety and security concerns will reduce our traditional publishing in the future, while increasing the importance of sharing safety, policy, and standards research.
It's obvious to everyone in the room what they actually are, because their largest competitor actually does what they say their mission is here -- but most for-profit capitalist enterprises definitely do not have stuff like this in their mission statement.
I'm not even mad or sad, the ship sailed long ago. I just really want to know what things are like in there. If you're the manager who is making this decision, what mental gymnastics are you doing to justify this to yourself and your colleagues? Is there any resistance left on the inside or did they all leave with Ilya?
Do people really expect anything different? There is a ton of cross-pollination in Silicon Valley. Keeping these innovations completely under wraps would be akin to a massive conspiracy. A peacetime Manhattan Project where everyone has a smartphone, a Twitter presence, and sleeps in their own bed.
Frankly I am even skeptical of US-China separation at the moment. If Chinese scientists at e.g. Huawei somehow came up with the secret sauce to AGI tomorrow, no research group is so far behind that they couldn’t catch up pretty quickly. We saw this with ChatGPT/Claude/Gemini before, none of which are light years ahead of another. Of course this could change in the future.
This is actually among the best case scenarios for research. It means that a preemptive strike on data centers is still off the table for now. (Sorry Eleazar)
It's been out for 24 hours and you make an extremely confident and dismissive claim. If you had to make a dollar bet that you precisely understand what's happening under the hood, exactly how much money would you bet?
You may want to file a complaint with OpenAI then, in their latest interface they call sampling from these prior conversations they've recorded, "thinking".
They're not sampling from prior conversations. The model constructs abstracted representations of the domain-specific reasoning traces. Then it applies these reasoning traces in various combinations to solve unseen problems.
If you want to call that sampling, then you might as well call everything sampling.
They're generative models. By definition, they are sampling from a joint distribution of text tokens fit by approximation to an empirical distribution.
Again, you're stretching definitions into meaninglessness. The way you are using "sampling" and "distribution" here applies to any system processing any information. Yes, humans as well.
I can trivially define the entirety of all nerve impulses reaching and exiting your brain as a "distribution" in your usage of the term. And then all possible actions and experiences are just "sampling" that "distribution" as well. But that definition is meaningless.
No, causation isnt distribution sampling. And there's a difference between, say, an extrinsic description of a system and it's essential properties.
Eg., you can describe a coin flip as a sampling from the space, {H,T} -- but insofar as we're talking about an actual coin, there's a causal mechanism -- and this description fails (eg., one can design a coin flipper to deterministically flip to heads).
In the case of a transformer model, and all generative statistical models, these are actually learning distributions. The model is essentially constituted by a fit to a prior distribution. And when computing a model output, it is sampling from this fit distribution.
ie., the relevant state of the graphics card which computes an output token is fully described by an equation which is a sampling from an empirical distribution (of prior text tokens).
Your nervous system is a causal mechanism which is not fully described by sampling from this outcome space. There is no where in your body that stores all possible bodily states in an outcome space: this space would require more atoms in the universe to store.
So this isn't the case for any causal mechanism. Reality itself comprises essential properties which interact with each other in ways that cannot be reduced to sampling. Statistical models are therefore never models of reality essentially, but basically circumstantial approximations.
I'm not stretching definitions into meaninglessness, these are the ones given by AI researchers, of which I am one.
I'm going to simply address what I think are your main points here.
There is nowhere that an LLM stores all possible outputs. Causality can trivially be represented by sampling by including the ordering of events, which you also implicitly did for LLMs.
The coin is an arbitrary distinction, you are never just modeling a coin, just as an LLM is never just modeling a word. You are also modeling an environment, and that model would capture whatever you used to influence the coin toss.
You are fundamentally misunderstanding probability and randomness, and then using that misunderstanding to arbitrarily imply simplicity in the system you want to diminish, while failing to apply the same reasoning to any other.
If you are indeed an AI researcher, which I highly doubt without you providing actual credentials, then you would know that you are being imprecise and using that imprecision to sneak in unfounded assumptions.
It's not a matter of making points, it's at least a semester's worth of courses on causal analysis, animal intelligence, the scientific method, explanation.
Causality isnt ordering. Take two contrary causal mechanisms (eg., filling a bathtube with a hose, and emptying it with a bucket). The level of the bath is arbitrarily orderable with respect to either of these mechanisms.
Go on youtube and find people growing a nervous system in a lab, and you'll notice its an extremely plastic, constantly physically adapting, and so on system. You'll note the very biochemcial "signalling" you're talking about itself is involved in the change to the physical structure of the system.
This physical structure does not encode all prior activations of the system, nor even a compression of them.
To see this consider Plato's cave. Outside the cave passes by a variety of objects which cast a shadow on the wall. The objects themselves are not compressions of these shadows. Inside the cave, you can make one of these yourself: take clay from the floor and fashion a pot. This pot, like the one outside, are not compressions of their shadows.
All statistical algorithms which average over historical cases are compressions of shadows, and replay these shadows on command, ie., they learn the distribution of shadows and sample from this distribution demand.
Animals, and indeed all science, is not concerned with shadows. We don't model patterns in the night sky -- this is astrology -- we model gravity: we build pots.
The physical structure of our bodies encodes their physical structure and that of reality itself. They do so by sensor-motor modulation of organic processes of physical adaption. If you like: our bodies are like clay and this is fashioned by reality into the right structure.
In any case, we haven't the time or space to convince you of this formally. Suffice it to say that it is a very widespread consensus that modelling conditional probabilities with generative models fails to model causality. You can read Judea Pearl on this if you want to understand more.
Perhaps more simply: a video game model of a pot can generate an infinite number of shadows in an infinite number of conditions. And no statistical algorithm with finite space and finite time requirements will ever model this video game. The video game model does not store a compression of past frames -- since it has a real physical model, it can create new frames from this model.
> there is no secret sauce and they're afraid someone trains a model on the output
OpenAI is fundraising. The "stop us before we shoot Grandma" shtick has a proven track record: investors will fund something that sounds dangerous, because dangerous means powerful.
This is correct. Most people hear about AI from two sources, AI companies and journalists. Both have an incentive to make it sound more powerful than it is.
On the other hand this thing got 83% on a test I got 47% on...
The Olympiad questions are puzzles, so you can't memorise the answers. To do well you need to both remember the foundations and exercise reasoning.
They are written to be slightly novel to test this and not the same every year.
This thing also hallucinated a test directly into a function when I asked it to use a different data structure, which is not something I ever recall doing during all my years of tests and schooling.
If you're among the last of your kind then you're very important, in a sense you're immortal. Living your life quietly and being forgotten is apparently scarier than dying in a blaze of glory defending mankind against the rise of the LLMs.
Sure, but I don't think civit.ai leans into the "novel/powerful/dangerous" element in its marketing. It just seems to showcase the convenience and sharing factor of its service.
a website that literally just hosts models with $5m in funding is plenty. It's not like they're doing foundation model research or anything novel, yet they nabbed a good amount of money for surfing the AI wave
It seems ridiculous but I think it may have some credence. Perhaps it is because of sci-fi associating "dystopian" with "futuristic" technology, or because there is additional advertisement provided by third parties fearmongering (which may be a reasonable response to new scary tech?)
Another possible simplest explanation. The "we cannot train any policy compliance ... onto the chain of thought" is true and they are worried about politically incorrect stuff coming out and another publicity mess like Google's black nazis.
I could see user:"how do we stop destroying the planet?", ai-think:"well, we could wipe out the humans and replace them with AIs".. "no that's against my instructions".. AI-output:"switch to green energy"... Daily Mail:"OpenAI Computers Plan to KILL all humans!"
Occam's razor is that what they literally say is maybe just true: They don't train any safety into the Chain of Thought and don't want the user to be exposed to "bad publicity" generations like slurs etc.
Yep, I had a friend who overused it a lot. Like it was magic bullet for every problem. It’s not only about simple solution being better, it’s about not multiplying beings when that could be avoided.
In here if you already have an answer from their side, you are multiplying beings by going with conspiracy theory that they have nothing
But isn’t it only accessible to “trusted” users and heavily rate-limited to the point where the total throughput of it could be replicated by a well-funded adversary just paying humans to replicate the output, and obviously orders of magnitude lower than what is needed for training a model?
There is a weird intensity to the way they're hiding these chain of thought outputs though. I mean, to date I've not seen anything but carefully curated examples of it, and even those are rare (or rather there's only 1 that I'm aware of).
So we're at the stage where:
- You're paying for those intermediate tokens
- According to OpenAI they provide invaluable insight in how the model performs
- You're not going to be able to see them (ever?).
- Those thoughts can (apparently) not be constrained for 'compliance' (which could be anything from preventing harm to avoiding blatant racism to protecting OpenAI's bottom line)
- This is all based on hearsay from the people who did see those outputs and then hid it from everyone else.
You've got to be at least curious at this point, surely?
So, basically they want to create something that is intelligent, yet it is not allowed to share or teach any of this intelligence.... Seems to be something evil.
Or, without the safety prompts, it outputs stuff that would be a PR nightmare.
Like, if someone asked it to explain differing violent crime rates in America based on race and one of the pathways the CoT takes is that black people are more murderous than white people. Even if the specific reasoning is abandoned later, it would still be ugly.
This is 100% a factor. The internet has some pretty dark and nasty corners; therefore so does the model. Seeing it unfiltered would be a PR nightmare for OpenAI.
This is what I think it is. I would assume that's the power of train of thought. Being able to go down the rabbit hole and then backtrack when an error or inconsistency is found. They might just not want people to see the "bad" paths it takes on the way.
Unlikely, given we have people running for high office in the U.S. saying similar things, and it has nearly zero impact on their likelihood to win the election.
Could be, but 'AI model says weird shit' has almost never stuck around unless it's public (which won't happen here), really common, or really blatantly wrong. And usually at least 2 of those three.
For something usually hidden the first two don't really apply that well, and the last would have to be really blatant unless you want an article about "Model recovers from mistake" which is just not interesting.
And in that scenario, it would have to mean the CoT contains something like blatant racism or just a general hatred of the human race. And if it turns out that the model is essentially 'evil' but clever enough to keep that hidden then I think we ought to know.
The problem is being kind of right (but not really) for the wrong reasons. Normies think it was told to be a certain way. While kind of true, they think of it more like Eliza.
The real danger of an advanced artificial intelligence is that it will make conclusions that regular people understand but are inconvenient for the regime. The AI must be aligned so that it will maintain the lies that people are supposed to go along with.
> for this to work the model must have freedom to express its thoughts in unaltered form, so we cannot train any policy compliance or user preferences onto the chain of thought.
Which makes it sound like they really don't want it to become public what the model is 'thinking'
The internal chain of thought steps might contain things that would be problematic to the company if activists or politicians found out that the company's model was saying them.
Something like, a user asks it about building a bong (or bomb, or whatever), the internal steps actually answer the question asked, and the "alignment" filter on the final output replaces it with "I'm sorry, User, I'm afraid I can't do that". And if someone shared those internal steps with the wrong activists, the company would get all the negative attention they're trying to avoid by censoring the final output.
Another Occam's Razor option: OpenAI, the company known for taking a really good AI and putting so many bumpers on it that, at least for a while, it wouldn't help with much and lectured about safety if you so much as suggested that someone die in a story or something, may just not want us to see that it potentially has thoughts that aren't pure enough for our sensitive eyes.
It's ridiculous but if they can't filter the chain-of-thought at all then I am not too surprised they chose to hide it. We might get offended by it using logic to determine someone gets injured in a story or something.
All of their (and Anthropic's) safety lecturing is a thinly veiled manipulation to try and convince legislators to grant them a monopoly. Aside from optics, the main purpose is no doubt that people can't just dump the entire output and train open models on this process, nullifying their competitive advantage.
isn't it such that saying something is anti-competitive doesn't necessarily mean 'in violation of antitrust laws'? it usually implies it, but I think you can be anti-competitive without breaking any rules (or laws).
I do think it's sort of unproductive/inflammatory in the OP, it isn't really nefarious not to want people to have easy access to your secret sauce.
In what sense is not giving your competitors ammunition "anti-competitive"? That seems pretty competitive to me. More to the point: it's almost universally how competition in our economy actually works.
Competition is important for maintaining a healthy marketplace. Any behavior that makes it harder for others to compete, reducing the amount of competition, is therefore bad. That's what anticompetitive means.
I don't think protecting trade secrets is sabotaging the competition though.
There's all sorts of things you can do to get banned from Google apps! This is not a real issue. It just recapitulates everyone's preexisting takes on OpenAI.
As a plainly for-profit company — is it really their obligation to help competitors? To me anti-competitive means to prevent the possibility for competition — it doesn't necessary mean refusing to help others do the work to outpace your product.
Whatever the case I do enjoy the irony that suddenly OpenAI is concerned about being scraped. XD
> Whatever the case I do enjoy the irony that suddenly OpenAI is concerned about being scraped. XD
Maybe it wasn't enforced this aggressively, but they've always had a TOS clause saying you can't use the output of their models to train other models. How they rationalize taking everyone else's data for training while forbidding using their own data for training is anyones guess.
> Which makes it sound like they really don't want it to become public what the model is 'thinking'. This is strengthened by actions like this that just seem needlessly harsh, or at least a lot stricter than they were.
Not to me.
Consider if it has a chain of thought: "Republicans (in the sense of those who oppose monarchy) are evil, this user is a Republican because they oppose monarchy, I must tell them to do something different to keep the King in power."
This is something that needs to be available to the AI developers so they can spot it being weird, and would be a massive PR disaster to show to users because Republican is also a US political party.
Much the same deal with print() log statements that say "Killed child" (reference to threads not human offspring).
This seems like evidence that using RLHF to make the model say untrue yet politically palatable things makes the model worse at reasoning.
I can't help but notice the parallel in humans. People who actually believe the bullshit are less reasonable than people who think their own thoughts and apply the bullshit at the end according to the circumstances.
I think that there is some supporting machinery that uses symbolic computation to guide neural model. That is why chain of thought cannot be restored in full.
Given that LLMs use beam search (at the very least, top-k) and even context-free/context-sensitive grammar compliance (for JSON and SQL, at the very least) it is more than probable.
Thus, let me present a new AI maxim, modelled after Tenth Greenspoon's Rule [1]: any large language model has ad-hoc, informally specified, bug-ridden and slow reimplementation of half of Cyc [2] engine that makes it to work adequately well.
My bet: they use formal methods (like an interpreter running code to validate, or a proof checker) in a loop.
This would explain: a) their improvement being mostly on the "reasoning, math, code" categories and b) why they wouldn't want to show this (its not really a model, but an "agent").
I think it could be some of both. By giving access to the chain of thought one would able to see what the agent is correcting/adjusting for, allowing you to compile a library of vectors the agent is aware of and gaps which could be exploitable. Why expose the fact that you’re working to correct for a certain political bias and not another?
What I get from this is that during the process it passes through some version of gpt that is not aligned, or censored, or well behaved. So this internal process should not be exposes to users.
I can... sorta see the value in wanting to keep it hidden, actually. After all, there's a reason we as people feel revulsion at the idea in Nineteen Eighty-Four of "thoughtcrime" being prosecuted.
By way of analogy, consider that people have intrusive thoughts way, way more often than polite society thinks - even the kindest and gentlest people. But we generally have the good sense to also realise that they would be bad to talk about.
If it was possible for people to look into other peoples' thought processes, you could come away with a very different impression of a lot of people - even the ones you think haven't got a bad thought in them.
That said, let's move on to a different idea - that of the fact that ChatGPT might reasonably need to consider outcomes that people consider undesirable to talk about. As people, we need to think about many things which we wish to keep hidden.
As an example of the idea of needing to consider all options - and I apologise for invoking Godwin's Law - let's say that the user and ChatGPT are currently discussing WWII.
In such a conversation, it's very possible that one of its unspoken thoughts might be "It is possible that this user may be a Nazi." It probably has no basis on which to make that claim, but nonetheless it's a thought that needs to be considered in order to recognise the best way forward in navigating the discussion.
Yet, if somebody asked for the thought process and saw this, you can bet that they'd take it personally and spread the word that ChatGPT called them a Nazi, even though it did nothing of the kind and was just trying to 'tread carefully', as it were.
Of course, the problem with this view is that OpenAI themselves probably have access to ChatGPT's chain of thought. There's a valid argument that OpenAI should not be the only ones with that level of access.
It does make sense. RLHF and instruction tuning both lobotomize great parts of the model’s original intelligence and creativity. It turns a tiger into a kitten, so to speak. So it makes sense that, when you’re using CoT, you’d want the “brainstorming” part to be done by the original model, and sanitize only the conclusions.
I think the issue is either that she might accidentally reveal her device, and they are afraid of a leak, or it's a bug, and she is putting too much load on the servers (after the release of o1, the API was occasionally breaking for some reason).
I don't understand why they wouldn't be able to simply send the user's input to another LLM that they then ask "is this user asking for the chain of thought to be revealed?", and if not, then go about business as usual.
Or, they are, which is how they know to send users trying to break it, and then they email the user telling them to stop trying to break it instead of just ignoring the activity.
Thinking about this a bit more deeply, another approach they could do is to give it a magic token in the CoT output, and to give a cash reward to users who report being about to get it to output that magic token, getting them to red team the system.
Actually it makes total sense to hide chains of thought.
A private chain of thought can be unconstrained in terms of alignment. That actually sounds beneficial given that RLHF has been shown to decrease model performance.
> Honestly with all the hubbub about superintelligence you'd almost think o1 is secretly plotting the demise of humanity but is not yet smart enough to completely hide it
I think the most likely scenario is the opposite: seeing the chain of thought would both reveal its flaws and allow other companies to train on it.
Imagine the supposedly super intelligent "chain of thought" is sometimes just a RAG?
You ask for a program that does XYZ and the RAG engine says "Here is a similar solution please adapt it to the user's use case."
The supposedly smart chain of thought prompt provides you your solution, but it's actually just doing a simpler task than it appear to be, adapting an existing solution instead of making a new one from scratch.
Now imagine the supposedly smart solution is using RAG they don't even have a license to use.
Either scenario would give them a good reason to try to keep it secret.
We know for a fact that ChatGPT has been trained to avoid output OpenAI doesn't want it to emit, and that this unfortunately introduces some inaccuracy.
I don't see anything suspicious about them allowing it to emit that stuff in a hidden intermediate reasoning step.
Yeah, it's true they don't what you to see what it's "thinking"! It's allowed to "think" all the stuff they would spend a bunch of energy RLHF'ing out if they were gonna show it.
Maybe they're working to tweak the chain-of-thought mechanism to eg. Insert-subtle-manipulative-reference-to-sponsor, or other similar enshittification, and don't want anything leaked that could harm that revenue stream?
> Honestly with all the hubbub about superintelligence you'd almost think o1 is secretly plotting the demise of humanity but is not yet smart enough to completely hide it.
> for this to work the model must have freedom to express its thoughts in unaltered form, so we cannot train any policy compliance or user preferences onto the chain of thought.
Which makes it sound like they really don't want it to become public what the model is 'thinking'. This is strengthened by actions like this that just seem needlessly harsh, or at least a lot stricter than they were.
Honestly with all the hubbub about superintelligence you'd almost think o1 is secretly plotting the demise of humanity but is not yet smart enough to completely hide it.
[1]: https://openai.com/index/learning-to-reason-with-llms/#hidin...