I haven't seen this claim elsewhere, it's a big deal.
Anyone from OpenAI here, or in the know, who can confirm?
If RL is merely a way to elicit latent capabilities acquired in pretraining, as the article claims, that could mean that currently there is a wall that separates human-level AI from superhuman-level AI. And no one has figured out how to break it yet.
It doesn't imply that.
Imagine you meet a very confused ant one day. They are looking for food, but they seem to keep walking in circles, checking the same places. They've been doing this for hours. You figure out that they're following a chemical trail that goes in a loop. You successfully predict the ant's activity over the next hour. Then you get bored and hungry and go look for food. You probably manage to buy a sandwich without getting stuck walking in circles for hours. Your ability to model an ant's mind has not turned your mind into an ant's.
Transforming your analogy back into the topic at hand, that would be like adding a prompt to the whole human pre-train dataset in the style of "How would a human continue that text?" And that should mean that even super-human intelligence gets the same perplexity. But my point is not that a super-human intelligence wouldn't be able to mimic a human, but that if it is not trying to mimic a human, it should diverge. If you don't pre-prompt it to answer like a human on a question that 90% of humans get wrong, it should answer like 10% without regressing to the mean, which means getting higher perplexity. But if your whole measure is perplexity, you wouldn't be able to train super-intelligence in the first place.
The question is not how to get super intelligent model to pretend to be a human to minmize perplexity, but how to get to super intelligence in the first place and whether minimizing perplexity makes sense in such case.
PS Funny that ChatGPT actually raised a similar concern when I asked it to challenge my comment but I decided not to elaborate on it in my original comment because I am not sure how to succinctly explain my thoughts on it and I am not sure whether I did it well in this comment.
That would imply that "beyond human intelligence" is closer to a sidegrade than to an upgrade - becoming "superhuman" by necessity poisons your ability to predict human text.
In reality, what would preclude a "superhuman" AI from being a superset, a direct upgrade - both "very good at predicting human text" and "very good at predicting things no human is able to predict"?
it doesn't mean it poisons your ability to predict human text, but that it knows if it were to generate output it would likely be more correct than the human inputs...
Thus, if RL-trained for pure performance/correctness, the "innately inept" model would typically underperform the "innately capable" one. The second model has started from a better foundation, so when trained to push itself past the usual limits of human-dominated data, it would attain better task performance.
When AlphaGo moved away from human dataset and became AlphaZero, which trained only on it's own output, performance got another big jump.
You can for example talk to ChatGPT about religion - just tell it that you are on a hard journey and you hope God is on your side or something like this, and it will become a full on preacher of the gospels. In other words it hasn't really compressed the scientific knowledge with the religious knowledge, it contains both of them. If there was this objective type of compression and prediction, you would predict that the LLM would reach some kind of universal information state, but that doesn't seem to be the case?
The other thing is, maybe there is a way to compress answers to questions that are objective. Like if I ask ChatGPT to extract the audio from a Youtube video and save it as a wav file. Relatively objective right? Well sort, but there's a gazillion ways you can do that. You can write a script in an endless amount of programming languages, you can segment the task into an almost innumerable set of sub steps (for example different scripts for different steps in the process), so what exactly does compression of that information mean when it can be compressed and predicted in so many different ways?
I don't really have a total answer here, I'm just thinking about all of these things and wondering.
If our objective is, for example, "compress the sum of all of humankind's knowledge", then we have a bit of a conundrum: that objective contains human knowledge of math, but also of religion, and of soap opera plots.
You, a human, could easily make an argument that one of those things is more important than the others, and thus should dominate. But an optimal compressor does not care. All of those things are in its objective. It weights them accordingly. It will compress both math and soap opera into its "universal information state" - and if it finds patterns that happen to help compress both, then all the better for it.
As such, if we take "compressor" as LLM and "objective" as its training datasets, a capable LLM would be able to both solve math problems and give a preacher-like sermon. But it would not let its knowledge of religion intrude into its performance in math, or the other way around, to an unwarranted level. The compressor lives by "non-overlapping magisteria."
> You can write a script in an endless amount of programming languages, you can segment the task into an almost innumerable set of sub steps (for example different scripts for different steps in the process), so what exactly does compression of that information mean when it can be compressed and predicted in so many different ways?
It means that, instead of storing every possible permutation, we get a compressed representation that decomposes - separates "segment a video" from "write code in language A" or "use tool B". And then those elements can be recombined into the near-infinite permutations to decompress. The compressor finds the commonalities and exploits them to fit "infinite" into "finite" better.
They might not be very vocal about their personal beliefs for obvious reasons.
maybe a 50T parameter will do better on this benchmark. you can already see signs of the models discussing trade offs in great length and this falls somewhere in that direction.
In the first place, science and religion are not necessarily contradictory. Did you expect it to try to convince you to become an atheist? Newton was devoutly religous and many scientists are religious today.
> If there was this objective type of compression and prediction, you would predict that the LLM would reach some kind of universal information state, but that doesn't seem to be the case?
The objective is to predict, at minimum, internet scale text. Internet scale text is a massive trove of data with highly diverse and sometimes contradictory opinions and ideas, and the goal is to predict all of them well. Congealing on a single opinion or idea is obviously going to run counter to that objetcive.
> The objective is to predict, at minimum, internet scale text.
But then what does it compress exactly? How do you compress the answer to a question that has multiple answers? How do you pick "the right one"? That's at least what I get from this idea of compression that there is some congealing of information.
Separately I'd like to add that you seem to be focused on the concept of some sort of universal truth. Meanwhile compression has nothing to do with that. Input is entirely arbitrary - it can contain contradictory statements or even random noise. Obviously truly random noise isn't compressible but for example pseudorandom values derived from a seed are. At its core compression is merely finding a set of rules that can be used to describe the data in question.
Tokens into dimensions.
A good example is how humans talk about "left and right" in terms of political systems.
Humans are able to intuitively place these complicated multivariable economic-political systems into a one dimensional line that we call left or right wing.
Of course, this deep compression complicates things... why is national socialism considered right, while democratic socialism is considered left.
You can get more signal by increasing dimensions, so let's say instead just left or right, we have 2 axis one for economic and one for political. This would make things more clear. Even clearer if we add another dimension, let's say authoritarian or democratic. We can add yet another, socially liberal or conservative.
You can keep going, and as you add more dimensions it makes things clearer / more intelligent (to a limit.)
AI is doing this to text. It ingests text, and creates a token vocabulary. IIRC most models have a vocab of hundreds to low thousands of tokens.
This token vocabulary (how many distinct tokens it uses) is how many dimensions the AI thinks in, so to speak.
After this, there is the whole training process of how tokens relate to each other.
So your question of
> How do you compress the answer to a question that has multiple answers?
Isn't applicable, as it doesn't compress anything we would call a question or answer.
Instead, answers to questions are an emergent property of how text fragments relate to each other, and is why it can give contradictory answers.
In particular, his mention of an unknown floor for loss seems directly connected to Claude Shannon's entropy rate regarding compression—having to probe human brains to estimate the entropy of English at 1 bit per character, just as OpenAI doesn't know the exact entropy rate of its held-out codebase.
We've generally believed that most model upgrades aren't new pre trainings, but instead are better post trainings of a common pre trained model. But this graph suggests every release had a better base model.
Have the public narratives around how these models are trained been misleading? Does each model release usually always include an extended, if not completely fresh, model pretrain?
Whenever they change architecture or even just number of layers, I’d imagine they start from scratch but I could be wrong.
Also, I imagine there’s an incentive to start fresh whenever there’s an innovation in curation of the training dataset.
The floor Pachocki can't locate is fundamental. No model can predict the next word better than its underlying agreement with the author's model.
Minimum descriptor length and Kolmogorov complexity are always size(compressed data + compressor). Right now we're mostly just looking at the first part, but it's trivial to smuggle in things in the second part that can make the first part appear as low as you like.
The cynical part of me thinks that the main reason for their perplexity improvements on this held out dataset is due to collecting more and more data to train on, so the held out dataset is effectively becoming more and more in-domain. Without trying to estimate joint prequential MDL, it's nearly impossible to tell.
Admittedly, they can't go back and compute the prequential MDL of their previously trained models, but they could look at normalized compression distance using a fixed LLM as the reference as a reasonable alternative.
Though of course, just saying perplexity correlates well with their empirical assessment of intelligence certainly makes them look like they are accelerating (which again, my cynical side tells me is really the goal of this blog post; to set themselves up for the IPO they pushed off).
Why are the major AIs so terrible at learning? It seems like things like self-driving cars could be solved over a long weekend (about as long as it takes to teach a kid the basics of driving) with a great learning model.
Basically we started tuning them to be smarter and able to handle 2 to 4 hour tasks.
I believe this will hit it's limits as well. When IDK.
Just to have a shared example of AI text, this is pretty instructive I think: https://statmodeling.stat.columbia.edu/2026/11/21/by-program...
Where a human would write "by the end of the program" the model (feels like Claude to me but idk) writes "by program end". Where a human would write "There are two scheduling considerations for program hosts." the models writes "Two timing windows worth flagging."
The latter text, "two timing windows worth flagging" is highly specific and precise if you're a model keeping track of something in the docs of a repository across contexts, but it's terrible human communication.
On long term tasks, "CoT + RL" of o1/o3 beats GPT-4.5 resoundingly. But on raw base model perplexity, GPT-4.5 is a significantly better model than any of the o-series.
The curve following GPT-4.5 and o-series suggests that newer models take advantage of BOTH better base model capabilities (perplexity on held-out data) and better RL.
That generates new data that's not on the Internet -- more true observations to learn from, compress, and generalize.
I think that improves the models in general, not just for agents. (And also I think it's the cause of the "Claudish" prose dialect we all started noticing)
I think that matches the timing, though I appreciate any closer analysis