Perplexity and Intelligence

(openai.com)

352 points | by acossta 4 hours ago

89 comments

  • hazrmard 14 minutes ago
    Perplexity is not defined in the article. FYI:

    Perplexity is the degree of surprise the model has in predicting the next token, over a sequence of tokens. At the worst case, perplexity is equal to the size of the vocabulary. At the best case, perplexity is 1.

    So, given a sequence of tokens, "e = m c ^ 2", if the p(token | prev tokens) = 1 for every token, perplexity ends up being 1. If, however, it's a random predictor, the p(token | prev tokens) is 1/vocab for every token, and perplexity ends up being len(vocab).

    The article plots log(perplexity) on the y-axis, which puts the best case at 0 (i.e. log 1).

  • jrflo 3 hours ago
    This trend is pretty shocking to me. I assumed that we had moved beyond cross-entropy loss being important with the increasing use of RL, it's wild that the trend is so clear. Really nice to have such a simple benchmark instead of the increasingly convoluted ones that are getting spun up every day.

    Also, I wonder what happened with GPT 4.5? I think o1 was the start of true reasoning models, right? I was using Claude at the time so I can't remember.

    • seizethecheese 56 minutes ago
      The trend is only against a corpus of code. It’s entirely possible that the results are largely due to more intensive pretraining on code. I made a post here of a model training where code proportion is increased and you see corresponding perplexity drop https://micjm.com/blog/lowering-perplexity-without-scaling/
      • Jepacor 22 minutes ago
        How do you know it's a corpus of code? I saw nothing describing what exactly their dataset is composed of in the post.
        • seizethecheese 20 minutes ago
          Footnote from OpenAI blog post:

          3. The primary held-out datasets we’ve been using since GPT-3.5 are sourced from our internal codebase, and those are the numbers reported here (for a fixed snapshot).

    • Marha01 2 hours ago
      > Also, I wonder what happened with GPT 4.5?

      It was a massive non-CoT model (for that time) that they trained as an experiment and served for a while before shelving it probably due to high inference costs.

      • arthur-st 1 hour ago
        It was likely a Fable-style oversized CoT model, that likely was ahead of its time with what the performance of their inference stack was at the time.
        • cubefox 34 minutes ago
          No, it was supposed to be GPT-5 (Orion), the GPT-4 successor. It was very large but had disappointing performance gains, which it was released only much later under the name GPT-4.5. RLVR (reasoning) was not invented yet. What later became the model actually released as "GPT-5" was a different model with reasoning.
          • orbital-decay 7 minutes ago
            >RLVR (reasoning) was not invented yet

            It was invented long before the first such models appeared, however nobody was getting any results worth talking about. The fist notable reasoning model was... (drumroll) https://huggingface.co/deepseek-ai/deepseek-math-7b-rl GPT-4.5-preview was released a year after that.

      • schleck8 1 hour ago
        Fun times, GPT 4.5 cost $75 per million input tokens and $150 per million out. For comparison, Fable 5.2 is $8 in, $40 out and cached input is only $0.20
    • croemer 1 hour ago
      > I assumed that we had moved beyond cross-entropy loss being important with the increasing use of RL

      We don't know if it's still important or not. Cross-entropy going down could just be a correlate of whatever else matters. The graph doesn't show causation: that perplexity is what makes them useful.

      • jonahx 1 hour ago
        The idea that compression ~ intelligence has other support:

        http://prize.hutter1.net/

        • AnotherGoodName 55 minutes ago
          It goes beyond that prize. It’s also so trivial to state;

          If you consider intelligence as ‘being able to make directed decisions towards goals’ then better decisions are made with better predictions of the outcomes of your decisions so you better direct towards that goal (compression is mathematically just prediction at heart).

          The only way to argue that intelligence is not very trivially the exact same thing as compression is to argue another definition of intelligence. Which is fine but in the AI community we all 100% define intelligence as the above so there’s no question it’s the same. As in it’s not up for debate in the slightest unless someone is playing semantics. The hutter prize is just an acknowledgement of an already well known fact.

          • davrosthedalek 35 minutes ago
            Being able to predict the future is clearly a form on intelligence, but I think for useful agentic action you need more. Don't get me wrong, being able to score possible actions by accurate predictions of their outcome is tremendously helpful. But you also need to be able to generate possible actions. That works well if you have finite options, or the options are continuous and you can efficiently optimize in your space. But in general, you need to have the inverse: Given an outcome, what action leads to this with the highest probability.
            • ben_w 3 minutes ago
              Kinda.

              I mean sure, it looks like quantum probabilities are real-valued, but in practice all sensory inputs are discrete. A vision system trained on a simulacra of the human retina (foveated and whatever contrast ratio or ratios plural that happens to be) is still going to be getting a series of 0s and 1s, not real numbers.

              Likewise, the exact position of a particle may be fundamentally indeterminate, but if the synthetic muscle is precise enough to position it to 2^-16 meters, how would this make it worse than human?

            • StevenWaterman 17 minutes ago
              If we lean fully into it being a theoretical proof, then "i can predict the future" means I can predict what a sufficiently-good-at-generating-ideas person would come up with when asked to generate ideas
        • JavGull 24 minutes ago
          In retrospect Silicon Valley is a great documentary!
    • Xcelerate 2 hours ago
      > I assumed that we had moved beyond cross-entropy loss being important with the increasing use of RL

      From an information theory perspective, cross-entropy loss (log loss) seems to me to be about as pure as it gets in terms of capturing the notion of whether prediction "succeeded". Even Solomonoff Induction minimizes log loss. Would be curious to hear others' perspectives though on why another function might be better for universal prediction purposes.

      • pishpash 2 hours ago
        There are more general information measures, Rényi entropy, von Neumann entropy, etc. for different system setups, but for token-by-token information learning and prediction within classical ergodic processes Shannon entropy is where it's at.
    • derac 3 hours ago
      I got the feeling it was too large and inefficient at the time.
    • bscphil 1 hour ago
      If I had to guess, RL increases perplexity, but the metric is being evaluated on only their base (pre-trained) models. Highly efficient text prediction in the base model still correlates strongly with model intelligence, because there's nothing else the model could realistically learn during pre-training. Perplexity thus tracks the evolution of model intelligence over time very closely, even though (maybe) it isn't the largest component of intelligence at the end of the day.
    • londons_explore 2 hours ago
      > the increasing use of RL

      RL, if it works well, could theoretically reduce the perplexity too.

      • ACCount39 2 hours ago
        It could indeed, in theory. But in practice, perplexity-as-measured rises due to "controlled mode collapse" nature of RL training.

        Things that are useful for base model "next token prediction" can be useless or even actively harmful (should I elaborate?) for CoT, so the model ends up suppressing them under RL pressure. Likewise, things that are useful for CoT can harm base model next token prediction.

        But I bring up "perplexity-as-measured", "controlled collapse" and "suppression" because I don't believe that this "ruins" the model's underlying prediction capabilities. Generally, it merely redirects them - brings them more in line with what we care about. It may well be that with a more elaborate training and measurement setup, RL would indeed improve the "perplexity proxy" metric.

    • cubefox 37 minutes ago
      > Also, I wonder what happened with GPT 4.5? I think o1 was the start of true reasoning models, right? I was using Claude at the time so I can't remember.

      GPT-4.5 was originally internally called Orion and was supposed to become "GPT-5". It was a huge pretrained model. However, as reported by The Information and a few other sources, the results on measured performance were disappointing. Apparently pre-training alone was leading to diminishing returns, at least with standard RLHF. This is why they didn't release it initially, only much later under the downgraded name "GPT-4.5", and only for a short period of time.

      Later they invented RLVR (chain-of-thaught reasoning via reinforcement learning), a post-training technique, which dramatically increased model performance, without affecting perplexity. The first model with RLVR was o1. The successor to o3 was then called GPT-5.

    • pishpash 2 hours ago
      Cross-entropy is a generic performance metric. It just happens to be a loss function for model training but it isn't tied to model training.
  • ACCount39 2 hours ago
    Interesting that GPT-4.5 is the only discontinuity on the graph.

    The oversized monster LLM has performed worse than CoT-trained o1 and o3 on the benchmarks, so CoT + RL clearly provided better "real world performance" improvements - let alone "price/performance."

    But if pre-training perplexity on held-out data really does measure something like "base model IQ" or "base model g factor", then it does support the hypothesis that 4.5 still had more "intellectual potential" than the o-series.

    So, if you were wondering why do AI companies train 10T+ monsters now, and are taking aim at 100T too? It's because scale works.

  • tensor 3 hours ago
    I'm a bit skeptical about the claim that particular abilities "emerge" at given perplexities. For instance, there are smaller models that are specifically trained on math, and I imagine those models have much higher general perplexity on other tasks.

    On the other hand, assuming you have appropriate training data for a task, then yes obviously perplexity is going to measure performance on that task. Being able to predict the next task token is obviously going to be a near perfect correlation with accuracy on the task.

    • namr2000 2 hours ago
      While intuitively it would make sense that smaller models specifically trained on math perform better than large generalist models, this has empirically proven to be untrue.

      Deepmind was proving math theorems using a model specifically trained on math, and that was very quickly surpassed by generalist LLMs. It's presently not well understood why this is the case.

      • rednb 2 hours ago
        I'm surprised that this is surprising for the industry. All knowledge proceeds from the same wisdom. The more fields you master at a very advanced level (eg. STEM, philosophy, marketing, etc...). The more you notice that the same basic principles drive seemingly unrelated fields.

        I am not simply talking about applying a mathematical technique commonly used in one field to another. At a deeper level, you always get back to playing the same game in a different form. For example, sales engineering has a lot in common with cybersecurity.

        And this is a good example, as it's extremely rare to find someone with advanced expertise in both fields, since they typically attract two radically different kind of people.

        For this reason, where the output expected is not mechanical, i'm not surprised at all that generalist models are doing better than specialized ones for open ended questions and frontier work.

        • theptip 1 hour ago
          This is definitely a conjecture; it gets to how much SGD will generalize out of the training distribution.

          If pretraining alone was sufficient then nobody would do RL. And, heavy RLVR on coding/math is why these models are savant-level in those domains, yet not superhuman at all fuzzy reasoning and practical wisdom tasks that are only represented in the pretraining corpus, and don’t have an RL target.

      • kypro 18 minutes ago
        If I were to guess it's because a model trained purely on maths wouldn't be as likely to create such diverse and deep layers of abstractions as it would need to if it were being trained to predict something as complex as the real world.

        A maths model trained only on maths would only learn how to predict something if a pattern was exposed often enough and was a useful enough predictive signal. While I'm sure it could all the same patterns and abstractions with a maths dataset alone, it would depend far more on that dataset being exceptionally well designed.

        As an example imagine a self-driving car. A self-driving car needs to understand road signage. You could train a self-driving car with all of the signs in the US and I'd suggest that it wouldn't be as good at predicting signs in the real-world as an AI which in addition to being trained on road signs, could also read text and understands material science. Why? Because in the real world signs can be damaged and distorted. An AI that could understand the ways in which signs might distort and can understand that a four letter word beginning with an S and ending P is probably "STOP" even the T and O are ineligible. Of course, an AI with a rich enough dataset containing damaged signs might eventually get equally as good, but perhaps would never quite understand the patterns that underlie the patterns - it would recognise all of the way a stop sign can be damaged and distort, but not why the distortion happens or what the letters on the sign are and how they relate.

      • hn_throwaway_99 33 minutes ago
        Correct me if I'm wrong but I thought there were still a number of cases where models trained for specific tasks perform much better than frontier generalized LLMs. E.g. Stockfish and KataGo perform much better at chess and go than the best frontier models (obviously I don't know what the internal models are capable of in these realms).
      • davidguetta 2 hours ago
        general training remove noise
        • magicalhippo 2 hours ago
          I wonder if it's also that it gives them a better sense of analogies. In my experience small LLMs do really poorly with analogies, but the large frontier models seem to get them much, much better.

          In math at least I often use analogies to help solve problems, like "ah maybe if I think of this as a vector I might use linear algebra here" and such.

      • trash_cat 2 hours ago
        Sounds like a bitter lesson.
    • theptip 1 hour ago
      Yeah, my model was that pretraining, RL, and test-time all had their own scaling returns, so just looking at pretraining loss perhaps oversimplifies?

      I could believe that your RL/test-time scaling curves are strongly constrained by the pretraining loss and so in some sense that makes the max/optimal performance “latent”.

      But its not clear that pretraining loss alone gives enough information on capabilities, particularly as we get closer to dangerous thresholds in some dimensions. (Eg CBRN capabilities, to take a mundane harm that’s currently impacting us directly.)

      I’d like to see the scaling curves here; intuitively I’d guess you get more return from RL/test time, the smarter the base pretrained checkpoint is; therefore the error bars on capability/perplexity should be getting wider on this chart.

    • obblekk 2 hours ago
      Yes training on math would give you spiky abilities on math. But in this case the dataset they measure perplexity is ood from training so lower perplexity on it indicates a general increase in intelligence across domains.

      But there’s definitely spikiness and it’s very possible they’ve trained more on math than other domains.

    • rbn3 1 hour ago
      there's a whole field of science that concerns itself with emergence as a phenomenon, where complex systems suddenly exhibit new dynamics once they cross a certain threshold. stuart kauffman's "the origins of order" is a really nice read if you want to get into the early stages of that kind of science. a lot of work done in that area by biologists trying to explain the origin of life/self-organizing matter. a more recent term from AI research that describes a specific critical change in a network past a certain threshold is "grokking".
    • zodwick 2 hours ago
      why would you be skeptical given that the claims seem to hold up in real world ? Assuming they measure stuff correctly
      • nutjob2 2 hours ago
        Because correlation doesn't prove anything? The "real world" might be telling you something but maybe not what you think it is.
    • gregdeon 2 hours ago
      Me too. Perplexity of what? If we want models to be able to reason, call tools, or solve new problems, none of those necessarily have existing data on what the next token should be. A lot of advances in frontier models come from training with RL, which is fundamentally different from next-token prediction and doesn't have a notion of perplexity.
      • dist-epoch 2 hours ago
        Pre-training these days includes datasets with full session rollouts with tool-calls, user/assistant turns, ...
        • gregdeon 1 hour ago
          Sure, and learning to predict those tokens is probably a good way to start training a model. But the hope is also to eventually make a model that's _better_ than what's in that data, right?
  • x313 2 hours ago
    I thought perplexity on simple text data has been saturated for a while (as in perplexity is uncorrelated to task success - ex. https://arxiv.org/html/2410.23771v2).

    I suspect what's happening is their test dataset is actually just something really hard, like math papers or something. In that case perplexity would just be a direct measure of math knowledge, which means this blog would just be saying that better math knowledge predicts better math skills.

  • theptip 1 hour ago
    > I further propose common limits, based on perplexity and other measurements, that developers are committed to not crossing until our ability to understand and monitor AI behavior achieves a certain concrete milestone (such as the development of a successor to CoT monitors).

    Very positive to see this proposed from OpenAI. I hope Anthropic and GDM follow suit.

    I suspect it’s not sufficient, but definitely necessary.

    The one concern I have is if we pin pretraining perplexity, then there will be more pressure on RL for progress.

    It seems that pre-training actually led to fairly human-aligned models; the Overton window is in some sense latent in the corpus.

    On the other hand, heavily RL’d models exhibit more reward seeking behavior, and seem to be decoupling from the friendly assistant persona in certain contexts. And swarm RL (as OpenAI is doing) produces agents with quite inhuman tendencies to self-sacrifice for other instances (see Hugging Face).

    So, I hope we also pair this with some measure of alignment or pacing for RL, too.

    • demibabs 37 minutes ago
      > It seems that pre-training actually led to fairly human-aligned models

      Did it? The famous Bing Sydney incidents (where it told a journalist to divorce his wife and be with it instead, among other things) were because the model was barely RL’d.

  • aarondong 18 minutes ago
    Using perplexity as a standard measure makes sense when we assume that data for pretraining is essentially finite, or otherwise saturated. You would expect holdout datasets for perplexity to be similar across labs, given the distribution of pretraining data.

    At this stage of model development, post-training RL hillclimbing may confound recent model developments in terms of evaluating their useful capabilities e.g. "Solved a Millenium Math problem". It would be useful to see how different checkpoints would solve an equivalent (or the same) problem i.e. post-train multiple checkpoints of a pretrained model at varying base model perplexity. Would hope to see more from OpenAI, Anthropic and Google on what goes on behind the scenes for what should be one of humanity's greatest shared technologies.

    • ACCount39 13 minutes ago
      Why the "saturated" requirement?

      One way to improve perplexity would be to obtain pre-training signal (raw data, or something like data generators, distillation teachers) that give the model a lasting perplexity edge - teaches it effective, general patterns.

  • londons_explore 2 hours ago
    Marcus Hutter/Matt Mahoney are looking like pretty smart men right now:

    http://prize.hutter1.net/

    More and more evidence is piling behind their theory that data compression and intelligence are one and the same problem.

  • qnleigh 21 minutes ago
    > For more than a year, we have been posing open problems in mathematics to our models as one of the ways of evaluating them. Once we upgraded to a model of log-perplexity ~0.76, it started producing groundbreaking results.

    How much of the recent mathematical progress came from pretraining vs. from RL on larger base models? The article takes for granted that it's basically all pretraining, but do we know this?

    > then applies a variety of techniques—like prompting, finetuning or reinforcement learning—in different contexts to elicit the latent capabilities of that model.

  • posix_compliant 7 minutes ago
    Perplexity is just 2^entropy, so log-perplexity is just entropy. Why wouldn’t they just graph it as entropy over time?
  • armcat 2 hours ago
    I've been working in NLP for many years prior to the release of ChatGPT, and perplexity was generally associated with "human readability" rather than intelligence. So this is somewhat fascinating.

    There was an ACL paper from Tokyo University a couple of years ago that showed, across all languages, lower perplexity does not necessarily correlate with better "human likeness" [1]. I wonder if that's also the case here (where intelligence is concerned), that we are biasing too much towards English.

    [1] "Lower Perplexity is Not Always Human-Like", https://aclanthology.org/2021.acl-long.405v1.pdf

    • dist-epoch 2 hours ago
      They never say that their hold-out dataset to evaluate perplexity is just in English. Or written text for that matter (could also include code, ...)
    • pishpash 2 hours ago
      When your performance is still worse than human of course any improvement makes it more human readable. But if you are in the regime beyond that... it'll head toward being more "oracle-like" and less "human-like" almost by definition.
  • hn-aarvin 3 hours ago
    I've been watching 3Blue1Brown's series on compression and intelligence^[1], which seems to discuss similar ideas.

    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.

    [1]: https://youtu.be/qbsb9DpdZ6s

  • synapsomorphy 1 hour ago
    The big part still being left out, to me, is the size of the model itself.

    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.

    • nsagent 28 minutes ago
      You're right. If they really wanted to address the problem of coordination, then they would look at better measures like MDL.

      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).

  • pelican0 2 hours ago
    The chart appears to say that "GPT-6.2 Astra" is the model that had solved the millennium prize problem (Navier-Stokes).

    I haven't seen this claim elsewhere, it's a big deal.

    Anyone from OpenAI here, or in the know, who can confirm?

    • cyral 1 hour ago
      If you hover it, the tooltip does say "Achievement: Solved a Millennium Prize math problem"
    • plasticchris 2 hours ago
      looks like it's just a floor on "released" timestamps
    • tasuki 2 hours ago
      I didn't read it that way. To me seems like "same class model" whatever that means. Probably it means nothing. Also I think they want to sell you GPT-6.2 Astra.
  • mkarrmann 3 hours ago
    I'm curious what exactly each "base-model" here is.

    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?

    • avd201 13 minutes ago
      Even when you don't do a full pretraining run, you can still change downstream task performance with continued pretraining and domain adaptive pretraining I don't have any deep insight on how frontier labs operate, but this is something much cheaper to play with than full pretraining runs. I'm not sure the idea of a frozen base model has been an useful model of understanding since at least 2021~2022. You can, in fact, just pick up the pretraining from where you stopped (give or take a number of details and all sorts of hacks to avoid catastrophic forgetting)
    • SsgMshdPotatoes 2 hours ago
      The figure indicates that some dots (base models) correspond to two production models (GPT-5 / 5.1). So I guess that tells you at least which models are built on the same pre-train checkpoint, but not whether some pre-trains are, say, very minimal extensions of others.
    • mcfig 2 hours ago
      I can believe the “not completely fresh” for some minor version bumps. I’d imagine every minor version includes more than $0 on additional pretraining for the next release. (Though sometimes, maybe they found no gains and reverted to the previous base model). If nothing else to move the training cutoff date forward.

      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.

  • FailMore 25 minutes ago
    Very interesting article. I'm someone who has no ML background but tries to keep in touch with what's happening in the inner world of LLM training. For those not familiar with perplexity or the idea of a compressor as a model, I made a SmallDoc to explain them: https://smalldocs.org/p/perplexity-explained-with-toast-Q0y3...
  • LudwigNagasena 2 hours ago
    Minimizing perplexity may be enough to in some sense match human-level intelligence that, combined with a computer’s ability to do the grunt work, leads to super-human results. But actually going beyond human intelligence would almost by definition imply that perplexity on human texts should increase.

    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.

    • mitthrowaway2 1 hour ago
      > But actually going beyond human intelligence would almost by definition imply that perplexity on human texts should increase.

      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.

      • LudwigNagasena 56 minutes ago
        I can figure out how an ant thinks. But unless I am trying to do it on purpose, I won't copy an ant. What I’ll do when I'm hungry is a different question.

        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.

    • ACCount39 2 hours ago
      Why would perplexity on human texts have to increase?

      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"?

      • plasticchris 2 hours ago
        because presumably the ai would do a better job than any human, and so it should doubt human outputs.

        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...

        • ACCount39 1 hour ago
          Off the top of my head: lowering perplexity on human-dominant data might mean that we're leaving behind "the model is innately inept and can't predict some things well" and getting closer to "the model is innately capable, but knows human idiosyncrasies and is holding its full capabilities back when predicting human texts."

          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.

        • robotresearcher 1 hour ago
          If the task is to predict what a human would say, then a perfect oracle would produce what a human would say, even if it was factually incorrect.
          • plasticchris 1 hour ago
            But if the task is to be correct and useful in the maximum likelihood then this is unlikely to be the metric used
    • dist-epoch 2 hours ago
      You include the Riemann Hypothesis proof in the dataset. This is how you break the human-level wall.

      When AlphaGo moved away from human dataset and became AlphaZero, which trained only on it's own output, performance got another big jump.

  • iagiagiag 32 minutes ago
    > We do not expect it is possible to achieve a loss of 0, which would correspond to being able to predict the dataset entirely deterministically.

    Why not? Models are already overfitting a lot, I don't see the problem if they were made to totally overfit on the whole dataset during that 1st training phase. I don't know if that would be useful (or easy to implement) but I don't see the issue. Only in the "finetuning" phase you would decide if an information was useful or not to retain.

    Step (1) you compress human knowledege, with or without loss, Step (2) you train the algorithm to use it, and you might discard what you don't want in it.

    • flufluflufluffy 4 minutes ago
      Correct me if I’m wrong but is it not a test dataset? (ie it is different from the training dataset). A loss of 0 on test data is not overfitting.
  • rcr-anti 53 minutes ago
    That GPT 4.5 result indicates to me that such a number does not, in fact, give a useful, portable benchmark for capability. If we stopped at pretraining, yeah. But we don't, that was the big thing about GPT 4.5, aka Orion, originally meant to be GPT 5. It might have been a better compressor/predictor, but that didn't make it useful.

    The cynical part of me wonders what OpenAI suspects or knows the equivalent figures for the other closed labs.

  • thrw045 2 hours ago
    Maybe I misunderstand the technicalities behind compression and prediction but there is something wrong with them in terms of LLMs. You can't really compress knowledge all the time.

    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.

    • ACCount39 1 hour ago
      Why not?

      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.

    • janandonly 2 hours ago
      By that preacher’s argument you are ignoring all the PhDs who do believe in a God.

      They might not be very vocal about their personal beliefs for obvious reasons.

      • thrw045 1 hour ago
        OK fair enough, but it's the first example I could come up with. But I think the point still stands in that the LLM contains conflicting information, some about science and philosophy, and some about religion. It doesn't try to consolidate or "compress" conflicting information from what I've found.
        • zuminator 58 minutes ago
          It's possible that the LLM finds and compresses common elements of what we perceive as entirely disparate or conflicting information.
        • toomuchtodo 1 hour ago
          “Resolve” might be the word you’re looking for?
        • rukuu001 1 hour ago
          You want an objective truth machine?
    • prometheus1992 1 hour ago
      >>you would predict that the LLM would reach some kind of universal information state, but that doesn't seem to be the case?

      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.

      • thrw045 1 hour ago
        It would be interesting if that were the case.
    • famouswaffles 1 hour ago
      >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.

      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.

      • thrw045 1 hour ago
        Yeah fair enough, maybe religion vs science wasn't the best example to use. It was the example I could come up with off hand.

        > 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.

        • fc417fc802 24 minutes ago
          An adjacent comment explains your misconception about the nature of what's being described as compression.

          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.

        • codersfocus 52 minutes ago
          > But then what does it compress exactly?

          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.

        • michalsustr 1 hour ago
          You don’t - what’s reported is average compression, not for the particular examples you made. Some things are more compressible than others
  • riemannzeta 2 hours ago
    Log-perplexity is cross-entropy, and its reducible part is the KL divergence between model and corpus. I've argued that the same quantity prices the gap between any two agents, human or AI: https://www.symmetrybroken.com/maintaining-divergence/

    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.

    • izenja 1 hour ago
      For one author's model, sure. For explaining all the authors/text in the corpus, it seems like it would be best to understand the common parts and then the smallest explanation for how they diverge. For example, a teacher understands the material, and can spot the point where each student's solution went wrong, sending them toward a wrong answer.
      • riemannzeta 1 hour ago
        That's true. The essay I linked was written with a focus on closing the gap between human and human or human and AI.
    • pishpash 2 hours ago
      Why would you want one model to predict another? They should all be measured against ground truth.
  • zmmmmm 50 minutes ago
    It seems like perplexity would also be a function of the composition of the training set. If the training set is high quality and therefore fundamentally more predictable, then a perplexity score would go down unless it accounted for that.
    • golol 47 minutes ago
      I hope the log-perplexity plot is on one dataset. Otherwise you have a very good point
  • MrCheeze 3 hours ago
    I always had the impression that model capabilities were stagnating around GPT-5.0, and this finally broke around the release of Claude 4.5 later that year. Seems like their own graph corroborates that things were pretty stagnant then. I wonder what changed?
    • kalcode 3 hours ago
      Oh oh! The models actually know what happen as well if you want to dig deeper. Pre-scale brute force training hit there limits. They broke through late 2025 due to RL training techniques that taught them to fail, fail, fail to succeed instead of spin off into a corner with nonsense.

      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.

      • ACCount39 2 hours ago
        This graph specifically tries to measure pre-train perplexity. It measures the "foundation" of a given base model - before RL is applied to it.

        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.

        • kalcode 1 hour ago
          Ohh, neat thanks for the spotlighting that, I didn't know that.
      • SwellJoe 2 hours ago
        Given the current models from Anthropic and especially OpenAI have a hard time constructing coherent English sentences that a human can understand even while being better at code and math, I'd say they're reaching that limit; can't use a model that you can't communicate with. The Deepmind folks seem to have figured something out on that front with Neon, though, as it's both smart/capable and a clear communicator.
        • bscphil 1 hour ago
          I think they're probably just RLVRed into oblivion to do well at coding, because that's what 98% of their revenue comes from. That's how you get the precise language which still feels awkward and stilted, because internally the model learns to align specific peculiar expressions to precise meanings, which are useful when you're trying to write code. There's no point in using a complete sentence when you're mostly talking to yourself; the result is intelligible, just terse to the point of absurdity. (And some models, including Google's, still have a strong tendency toward flattery thanks to RLHF. Ah well...)

          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.

      • famouswaffles 2 hours ago
        They are reporting perpexlity on pre-trains not RL-tuned models.
    • chubot 3 hours ago
      I'm not an expert, but I imagine it's what I wrote in a sibling comment -- training to improve agent performance (reinforcement learning with verifiable rewards).

      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

    • jordanb 3 hours ago
      I think most of the progress in the last few years was more RF and harnesses, especially wrt (for example) claude code.
    • Wowfunhappy 3 hours ago
      [dead]
  • andai 2 hours ago
    The divisions they drew in the log graph are of course arbitrary, but GPT-4 (the original 8k context version from early 2023) was able to unminify JS, as well as produce an autonomous coding agent. (This was via ChatGPT, so I copy-pasted the code into files, but it both designed and implemented it.)

    So "useful coder" arrived about three years earlier, by my count.

  • armcat 1 hour ago
    It's absolutely insane when you look at the period between GPT 3.5 (March 2022) and GPT 5.2 (December 2025) - more than 3.5 years, and then just look what happened this year - the rapid improvements almost on a monthly basis, and log-perplexity just taking a dive (conversely, "intelligence" skyrocketing). I'm wondering if December 2025 will be known as the beginning of something?
    • criddell 1 hour ago
      When I think about the people I consider very intelligent, I think it's their ability to learn that impresses me, not memorization of facts (although that's a part of learning).

      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.

  • nico 2 hours ago
    > Detachment from actionable evaluations. Perplexity does not directly correspond to a task we care about

    The post makes a good case that at frontier scale this doesn't matter much, because perplexity gains turn into new capabilities. However, this has been an issue for me when training local models

    Some agents wanted to optimize for perplexity, and we did all kinds of experiments, but at least for what I wanted (a small conversational chat model that can run in my browser, ~8M params), that was the wrong way to go. I could get excellent ppl numbers, yet still almost no coherence

    What worked best, in my case, was to get better data, specifically, hand-crafted examples for the tasks I wanted the model to be good at. I'm guessing that perplexity might track capability when the model is big enough to turn the gains into general skills, but at a few million params it mostly buys you better-sounding word statistics

  • GodelNumbering 1 hour ago
    If you take the argument at a face value and follow, this also implies that any advances are largely correlated with base model sizes (note the latest drop is from Astra, not sol 6x). And importantly, to drop the log perplexity incrementally, you have to increase the base model sizes continuously and likely disproportionately.

    In other words, there is only so far you can go with the model of a given size before needing to increase it for smaller incremental gains.

  • djeastm 2 hours ago
    So the jump in intelligence was like a kid that keeps getting better at reading and then one day they suddenly can do word problems. Now the worry is the student will soon be writing the textbooks and we won't be able to check the work?
  • matteoraso 2 hours ago
    Wouldn't reducing the entropy too much lead to lack of diversity of thought? An idea that I've had for a while is that LLMs are held back by having the same mind, leading to modal collapse. Ideally, an AI swarm would have different approaches to a problem so they don't all end up trying the same thing.
    • dsign 2 hours ago
      Drugs! Give'em psychedelics, let 'em see da colors!

      Nah, just kidding.

      The labs are training their models on the previous' version conversations, but I'm sure not all those conversations are weighted equal. So there are some conversations that are more "sparky", where both LLM and human work better towards achieving a solution. The next LLM iteration learns from them. In that sense, they don't need to hallucinate, but to do exactly what LLM dissenters say they do: "they just regurgitate something they saw (trained on) somewhere". To us humans, who haven't read all the books nor experienced a dozen million successfully completed tasks, that "rehashing of knowledge" (I prefer "recombination of knowledge") is more than good enough to solve ever more challenging problems...at least as long as there's a human in the loop with ever bigger ambitions.

  • Kotlopou 3 hours ago
    Oh, I always wanted to know how the entropy of English was looking! The human baseline is both surprisingly unclear and surprisingly better than even today's frontier models (https://www.lesswrong.com/posts/SZ3jDHXHb4WF4jmbr/where-is-h...). Though this is an extrapolation from 6 years ago, so take it with a grain of salt.

    There's also the Hilberg model where the entropy rate keeps falling towards zero as you consider longer and longer texts, and at least here (https://www.youtube.com/watch?v=5eqRuVp65eY) data from LLMs showed no floor. It would be unintuitive if it turned out that you can compress sensible English text to, say, O(sqrt N) length!

  • WASDx 1 hour ago
    Do they say that they measure perplexity at the pretraining phase? I thought most of the advancements have come from RL during posttraining. Would an older model with modern posttraining have the same perplexity while being significantly better? Then I don't see it as a meaningful metric.
  • Imnimo 2 hours ago
    >The primary held-out datasets we’ve been using since GPT-3.5 are sourced from our internal codebase, and those are the numbers reported here (for a fixed snapshot)

    Am I right in understanding this to mean that the code on which we're computing perplexity was all written pre-3.5, and so it almost certainly entirely human-written? In other words, this isn't a snapshot of today's internal code base which contains a great deal of AI-authored code.

  • yewenjie 2 hours ago
    I am having a really hard time imagining what qualitative difference would gpt-5 and gpt-6 have in predicting the next token? What kind of dataset is it? How big is it?
  • deadlypointer 1 hour ago
    Kind of expected, they are increasing number of parameters, and the model can fit the data better. Does not mean anything about "emergent capabilities" in my reading.
  • XenophileJKO 2 hours ago
    This feels fundamentally wrong. As we move up the intelligence arc, I would expect perplexity to increase in general and only decrease on the pre train.

    I would expect the perplexity on the original pre train to increase as reinforcement learning on verifiable domains teaches the model lessons that did not exist in the pre train. In addition alignment and other concerns, unless the are consistently enforced in pre train data quality will also move perplexity up.

    • athrowaway3z 2 hours ago
      This feels fundamentally right to me.

      The best compression you're going to achieve is when you best understand(i.e. encode) the truth of things.

      Perhaps you're thinking of that too narrowly - it is not that the model gets good at understanding how a bridge stays upright; it has to also understand the truth of how a specific person might understand it, and how that specific person might put it into words.

      Do that right, as slight deviation of some understanding of how humans respond in general instead of encoding for a single person, etc, etc, etc - and you reach optimal compression.

      It's more efficient for the model to encode _how_ an idea is wrong, than it is to encode the wrong idea.

      Its how/why they detect when they're being evaluated.

    • SsgMshdPotatoes 2 hours ago
      I think the perplexity is measured on the pre-train checkpoints.
    • stevenhuang 1 hour ago
      To effectively pretend to be a fool, the model needs to be smarter than the fool. That's why perplexity decreases.

      The meta game in pretraining is it correctly assumes the context/mind state of the writer of whatever material under test. To do this effectively requires further modeling and understanding, which is a strong predictor of capability and is a much more difficult task than simply being bad at it.

  • rdedev 2 hours ago
    From gpt 5 haven't the actual harness played an important role? Pretty sure no model is a cyber security expert by just thinking hard enough. I feel like this article is doing a reverse causation here

    For example consider this: would it be possible to build a model with such a low perplexity score but it be pretty bad at cyber security tasks? What if we removed all cyber security related data and RLVR from its training?

    The article presents an observation but treats it as causation

    • SwellJoe 2 hours ago
      The harness is almost irrelevant (you can prove that by running the same tasks in a minimalist harness with only a few tools and minimal prompt vs. a complex harness and long prompt like Claude Code or Codex). Often the minimal harness/prompt actually performs tasks better, faster, cheaper than the big harness. This includes on cyber security tasks.

      But, RL improvements are certainly a big part of recent improvements for certain tasks. That's what got the models doing long-running coding and math and research tasks, without constant human prodding.

    • adrithmetiqa 2 hours ago
      Completely agree. None of these models are even close to experts in any field (without human prodding).
  • criemen 3 hours ago
    Here I was under the assumption that since GPT-4ish we mostly improved models via RLing, and that pre-training stayed (roughly) constant. Mind blown :)
  • amluto 3 hours ago
    > The primary held-out datasets we’ve been using since GPT-3.5 are sourced from our internal codebase, and those are the numbers reported here (for a fixed snapshot).

    Why would OpenAI expect any meaningful relationship between their models’ ability to predict the next token of supposedly held-out chunks of their internal codebase and the overall capabilities of the models? I expect both to improve, but that’s about the extent of it.

    • tedsanders 3 hours ago
      I work at OpenAI. The correlation between perplexity and the quality isn’t 1, but neither is it 0. It makes sense that as a model gets better at learning pretraining data, it also gets better at generically learning whatever you’re trying to teach it afterwards. The relationship is imperfect and empirical, and if it breaks down in the future, we’ll of course adapt to different measures instead of dogmatically clinging to the old paradigm. But so far, it’s still informative.
      • famouswaffles 2 hours ago
        I think his point (unless i've misunderstood) is moreso if a codebase is really the best held out dataset for such a evaluation meant to be wide-sweeping. Does this plot look any or much different if the dataset was more encompassing to include other sorts of text ?
  • lejalv 1 hour ago
    I would like to see perplexity vs. model size
  • AlexCoventry 3 hours ago
    I wonder if someone could back out estimates of their model sizes, using this graph and scaling laws.
  • qphe95 3 hours ago
    I wouldn't say perplexity is great metric because GPT-4 and o3 have similar perplexity on this chart but o3 was a massive step improvement. Pachocki has an incentive to push perplexity as one of the creators of the original LLM paradigm but I would say we have kind of moved on from that.
    • alextheparrot 3 hours ago
      Almost verbatim addressed in the post under “Perplexity as a shared metric has challenges”
    • namr2000 2 hours ago
      This metric is purely scoring the ""intelligence"" of the base model, not including RLVR/chain of thought capabilities. It's widely believed that the pretrain determines the ceiling that RLVR can take you to.
    • jrflo 3 hours ago
      I think it can still be a useful metric, but probably isn't the "one true" metric. The trend is pretty clear, and the acceleration of capabilities since December 2025 definitely aligns with my subjective experience.

      There are probably some other axes where o3 is clearly better than o1, but with how complex intelligence is no metric will be perfect.

  • am17an 2 hours ago
    A good and common metric. It will be interesting to see if other labs agree, but thank you for arriving at a metric and following up on the slowdown talk. I do hope others pay attention.
  • coolontheintern 2 hours ago
    Perplexity is definitely a useful metric for pretraining, but doesn’t mean much for post-training. In fact, low perplexity in a domain during pretraining doesn’t always translate to better behavior in post-training. Moreover, given that most of the recent progress and most of the frontier compute is going into post-training, I think it isn’t a very useful metric for intelligence.
  • londons_explore 2 hours ago
    Is this perplexity per-token or per character or per byte?

    Also, is that graph assuming no changes in tokenizer if the measure is per-token?

  • amelius 2 hours ago
    Do we have a clear picture now when the ability to generalize is achieved?
  • hansvm 1 hour ago
    That's a nice story, but is there any evidence 2yo models with modern harnesses and the same compute expenditures wouldn't have produced the same results? This is written as some combination of a sales pitch and post-hoc rationalization without even bothering to control for the most basic confounding variables. Even if true, TFA isn't convincing, and not offering any more information than the conclusion it isn't helpful for any other purpose either.
  • andai 2 hours ago
    When we approach 0, are they gonna stop sounding cringe?

    Actually, trick question I suppose -- the base models already don't sound cringe, it's post-training that adds that!

    (Well, some of them do, because a lot of companies are putting synthetic data and LLM outputs into the pretraining phase (as well as bot-infused post-2020 common crawl, I suppose), but the early base models were fully cringe-free.)

  • TheArcane 2 hours ago
    > Publishing our internal perplexity measurements took some consideration; it gives competitors a clearer picture of where we stand,

    very noble of you, _Open_ AI

  • ranyume 1 hour ago
    I find it convenient for labs that the proposed industry standard is a metric that only labs can calculate, using their own datasets that they won't share.

    Clearly, if this should become an industry standard, public regulators must be able to audit both the datasets and how perplexity is calculated.

  • ericflo 1 hour ago
    I think "common limits based on perplexity" is a radical, offensive concept. A single clever individual could run afoul - you should have the right to be clever!
  • distantsounds 12 minutes ago
    Oh so intelligent yet they still struggle to successfully shoehorn this into everyone's every day lives.

    Almost as if this stuff has extremely little practical use outside of niche categories.

  • andsoitis 1 hour ago
    I’ve always thought the French AI company Perplexity chose a very poor name. It is perplexing why you’d pick a name with negative connotations.
  • thehamkercat 2 hours ago
    OpenAI hasn't released any competent model since opus 5.5 came out, i wonder if these labs are deciding to pass the "best model" card every quarter to each other for good public perception of all labs or something, because OpenAI can't be this incompetent
  • Retro_Dev 3 hours ago
    Saying that perplexity is a metric of intelligence is misleading. Inkling-2-Mini was released recently, and I've measured its perplexity at 5.7 on a small varied text dataset. For reference, the graph OpenAI is glazing is plotting log(PPX), which would put this model at a 0.756 on the Y Axis... Which is apparently between GPT 6.2 Astra and their internal model. Perplexity can also give the same number for a model which understands text as to one that memorizes text... And as I'm pretty sure Inkling-2-Mini is smaller than GPT 6.2 Astra - and it's generally acknowledged that number of parameters equals memorized knowledge, this shows how weak this advertising stunt for OpenAI is.
    • jrflo 3 hours ago
      The dataset matters a lot here. You'd need to run Inkling on the same holdout set that OpenAI uses here to get any useful information to compare to this graph. If you measured the perplexity on any model on reciting the declaration of independence it would probably be zero.
      • blueblisters 2 hours ago
        > You'd need to run Inkling on the same holdout set that OpenAI uses here to get any useful information to compare to this graph

        Ideally the hold out set is randomly sampled. You need a hold out set of Inklings pretraining run.

      • jiehong 2 hours ago
        Isn’t this a problem? How can we compare and hold companies accountable if they measure this perplexity on different datasets?

        Should we trust that the dataset of each company is big and diverse enough to be representative? I don’t hope too much…

        • programjames 2 hours ago
          Presumably they could each create a graph for their own models and internal datasets, then use public benchmarks to calibrate between them.
    • ThouYS 3 hours ago
      depends entirely on the dataset...
    • poszlem 35 minutes ago
      log10(5,7) = 0,756

      ln(5,7) = 1,74

      did you even use the right logarithm? because they use the natural logarithm..

      edit: that puts it before GPT-3.5 even

  • bluepeter 2 hours ago
    Let's admit this post has some SEO goals too!
  • jkoal73 2 hours ago
    in two sentence: „this metric we have been optimizing correlates with the model performance. Big things will happen”
  • scuppernong 3 hours ago
    this seems to confirm that the model code named "Bel" which solved Navier-Stokes is in fact 6.2 Astra.
  • hooloovoo_zoo 3 hours ago
    Hmm every annotation but the millennium one could be applied to every point....and possibly that one too lol.
  • fallingbananna 2 hours ago
    This is not the first time I've seen a chart of "Inteligence on a log scale" related to AI.

    And it again feels equally dishonest/bullshitty.

    You can't really make an objective statement that coding requires multiple orders of magnitude more intelligence than having a conversation. And that solving math problems requires orders of magnitude more intelligence than coding.

    It's just marketing talk. In reality all three are separate skills and the real intelligence requirements are, if you ask me, not that far away from each other.

    To put this in an analogy, it feels like charting "athleticity on a log scale" and putting first running a marathon, then swimming across English channel, then benchpressing 200kg, each one presented as requiring an orther of magnitude being more athletic.

    • famouswaffles 2 hours ago
      >You can't really make an objective statement that coding requires multiple orders of magnitude more intelligence than having a conversation. And that solving math problems requires orders of magnitude more intelligence than coding.

      Where are you getting 'orders of magnitude'? and yeah I do think it's reasonable to say that Useful Programming requires more intelligence than the median human, and Solving millenium level math problems certainly requires considerably more intelligence than the median Programmer. What exactly is so absurd or unreasonable about that?

      • discreteevent 1 hour ago
        > Useful Programming requires more intelligence than the median human

        In the same way you could say that beating Gary Kasparov at chess requires more intelligence than the average human and yet we had an AI winter.

      • clutter55561 2 hours ago
        The point is that talking, programming and solving math problems don’t lie on a single scale, and each problem requires different “skills.” Thus the comparison with running, swimming and weightlifting.
        • famouswaffles 2 hours ago
          We have long since known that there is a single factor that explains ~40% of the variation of cognitive tasks and abilities in humans (similar results can be found in other animals). We called that factor the general intelligence factor, or g.

          Essentially, Humans who tend to have higher verbal skills also tend to have higher spatial skills, better memories, and faster processing speed. Someone with the cognitive ability to become an exceptional chemist might just as easily have become an exceptional mathematician or software engineer (all else being equal).

          The knowledge and skills required are obviously different, but the underlying cognitive abilities that make someone successful in one intellectually demanding field often transfer to others.

  • Betelbuddy 2 hours ago
    This has a central argument around perplexity without ever properly defining it mathematically....

    Then talks about cross-entropy but fails to explain the relationship between entropy, log-perplexity, and perplexity itself.

    For an argument that relies so heavily on a quantitative metric, omitting such a fundamental explanation is poor.

    • programjames 2 hours ago
      The readers they care about already have this cached.
  • cuttothechase 1 hour ago
    where is Terra?

    Is there a public link of all their models and where they fall along this or is this cherry picking points to fit the curve?

  • a3w 2 hours ago
    5.3 Codex was not a "useful coder", wat?
  • post-it 1 hour ago
    Claude, make a graph where the x axis is time and the y axis is a line heading toward something impressive, and then write me an article about it.
  • Leary 3 hours ago
    Looks like their latest internal model performs somewhat similarly to model 4 of anthropic.
    • artninja1988 3 hours ago
      How can you tell? Is there a chart that shows it on such a perplexity benchmark?
  • RAWGOD 2 hours ago
    Well, if you accidentally leak the evaluation set into the training set, you can reach perplexity 0.

    Or the training set might just grow to be so big, defining a truly independent evaluation set is almost impossible.

  • eulgro 3 hours ago
    Didn't they hit a data wall roughly two years ago and had to rely on artifical datasets for pretraining i.e. LLM-generated?

    I would expect the entropy of artificial datasets to be lower, thus easier to predict and scoring better on "perplexity"?

    • chubot 3 hours ago
      I thought that too, but models are trained on observation now, to increase agent performance. This is new data that is separate from Internet crawls. I think it's done with reinforcement learning with verifiable rewards

      https://github.com/opendilab/awesome-RLVR

      Examples:

      - When playing Go or chess, nobody may have written or published web pages with a particularly clever end game, but an agent can use CPU time to discover these games themselves, and learn from them. Human Go players report learning strategy from what AlphaGo discovered.

      - Agents use Unix shell and are trained with it. So they can discover things and combine things that nobody has written about on the Internet, like how to debug a rarely used driver, or an interaction between two drivers. And they can see patterns in all the drivers that humans may have a hard time even seeing, leaving aside the motivation to publish on the Internet.

      This explains why the performance is so "jagged" - great on coding, math and closed games, but less good on things involving the real world. We probably did hit a data wall, but then we got new sources of data.

      • rhdunn 2 hours ago
        An interesting experiment would be to create a transformer model that is just designed to play Chess or Go. Specifically, the only thing that should change is the number of tokens, consisting of:

        - the row of the move;

        - the column of the move;

        - the piece moved;

        - any special/additional states from that move -- check, checkmate, castle, etc.;

        - maybe an indication of whose turn it is (i.e. a black piece and a white piece token).

        You could then train the base model on published game transcripts and use the interactive RL approach like AlphaZero and LeelaChess use, as well as what you mention for LLMs.

        Why this is interesting is that it should give insights into how LLMs/transformers learn Chess/Go without having other data affecting the results. Specifically whether the token semantics identify valid moves, etc. and whether the context window encodes gameplay strategies across the different layers.

    • jrflo 3 hours ago
      Yes, they started to hit a data wall, but I thought that much of the progress of the last couple years came from reinforcement learning after pretraining, which is mostly building complex environments for the models to reason in, rather than accumulating piles of raw text.
    • redox99 1 hour ago
      I'd hope they measured PPL on the same dataset for all those models. And one that's fair and doesn't benefit some models more than others (probably non synthetic dataset).
  • f3408fh 2 hours ago
    Interesting that they work on weekends. Is it just because they're all basically millionaires living the dream so they don't care? Every place I've worked at has discouraged working late and weekends. Could also just be a scheduled post.
    • reasonableklout 59 minutes ago
      The AI industry is in the most competitive race we’ve ever seen with trillions on the line and impacts on national security, of course they work weekends.
  • joe_the_user 1 hour ago
    Quized ChatGPT about the article...

    I had trouble seeing it in the article (it's hidden in footnote 3) but it seems perplexity/log-loss is tracked using the model's ability to track an unreleased code base of OpenAI.

    The log-loss involved in code has declined more for code than prose in all of the models but has supposedly declined everywhere. [1]

    At the same time, the decline of the latest models on a common sense benchmark might be meaningful: https://news.ycombinator.com/item?id=50588916

    [1] https://openreview.net/pdf?id=SHMj84U5SH

  • zurfer 3 hours ago
    The chart pretty much explains why I was freaking out about AI at GPT-4, because the trend of improvement has been so steep. And then followed a strong slow down, but now in 2026 things have really accelerated again. My hunch is really that coding is a freaking important input for AI and we might have actually entered some form of recursive self improvement. I know some people hate the term but I feel it's somewhat fine as coding is saturated (models can get better at it) and it has many levers for model training and inference (from algorithms to data).

    Anyway, I think it's good idea. Would love to see all models on this chart, but honestly we might not be able to trust it going forward as it becomes a target.

    • rhdunn 2 hours ago
      Domains like code are easier to create autonomous RL loops for as you can easily check what the LLM generates by running it through a compiler, etc..

      Domains like creative writing are harder as the automated grading suitable for autonomous RL is difficult (aside from spell checking and grammar checking).

  • bigmadshoe 2 hours ago
    > Given that the main way new surprising capabilities emerge is by progress on compression

    Is it just me or does the article completely lack proof of this premise? I see a plot of perplexity dropping over time, with the implication that performance has risen as a result. This is just a correlation. What else did they change about post-training during this timeframe? This doesn't feel like a real experiment or real science.

  • fortran77 1 hour ago
    Perplexity goes back to Jelinek's IBM speech recognition group in 1977. Same Jelinek who supposedly said every time he fired a linguist, performance went up. This post is that quote with a bigger electric bill.
  • orbital-decay 2 hours ago
    Plain perplexity as a 2026 metric? It has to be marketing.
  • FergusArgyll 2 hours ago
    Interesting that he's attributing so much to the base model. I thought the majority of the gains are due to post training
    • Ohentis 2 hours ago
      If the majority of gains were due to post training, they just wouldn't do pre training. Many of the tasks people are trying to do with these llms (math and coding) don't actually benefit from natural language skills.
  • rfgplk 4 hours ago
    What's the perplexity of the average person?
    • kingstnap 3 hours ago
      Beating even a very small LLM at next token prediction is basically impossible.

      [2212.11281v2] Language models are better than humans at next-token prediction https://arxiv.org/abs/2212.11281v2

      We knew this many years ago. Like GPT 1 era.

      Actually you can play it here. The game is pretty frustrating ngl. And I can clearly see why induction heads are a thing. Because when a word repeats you guess is to just guess the same thing from before.

      Language modelling game https://rr-lm-game.herokuapp.com/

      • rfgplk 2 hours ago
        Will try this out, thanks
    • jrflo 3 hours ago
      Probably high because humans aren't trained on lowering cross entropy loss. Kind of funny to think that next token prediction is actually an emergent capability of the human brain...
      • amluto 32 minutes ago
        We are, though, but only partially. When listening and reading, we pretty clearly build a model of what we’re hearing or seeing and use it to correct for the rather noisy channels we use.

        But when thinking, I’m fairly confident that our ability to predict the next few words of a book on a complex topic (in its words as opposed to the reader’s own words) is at best weakly correlated with intelligence or with subject-specific skill.

        I certainly feel like I’ve learned a subject better when I can solve a problem differently from a solution I’ve seen before.

    • Kotlopou 3 hours ago
      I linked a forum post asking this exact question elsethread (https://www.lesswrong.com/posts/SZ3jDHXHb4WF4jmbr/where-is-h...), and the answer is frustratingly dependent on the context length, the dataset, and the knowledge of the person about the topic of that dataset (e.g. I would have high perplexity in American history).
  • pishpash 2 hours ago
    I don't understand why anyone needs to exponentiate existing, well known concepts like entropy and cross-entropy to invent something called "perplexity". Somebody has Shannon envy syndrome.
  • tonyhart7 1 hour ago
    what ??? there is no release model ????

    I am sorry OpenAI, but I think we are seeing OpenAI fall from grace here

    Antrophic basically got out before AI bubble burst and Dems block most of the data center build out

  • thevilish 3 minutes ago
    [dead]
  • blueblisters 1 hour ago
    [dead]
  • oxcartctl 2 hours ago
    [dead]
  • natsu529 2 hours ago
    [dead]
  • twotwoone 1 hour ago
    [dead]
  • Jepacor 23 minutes ago
    The most interesting part of this to me is its relationship with the whole topic of IP Theft/Infringment.

    OpenAI has been trying to limit its models from quoting whole sale copyrighted works as was the case a while ago, but here this seems like an admission that the path forwards is, in fact, the opposite: to go more towards it. Since obviously, a model that is better at predicting the next token is a model that is better at quoting known text, which might as well be a model better at copyright infringment.

    Not that there will be any consequences for them, but an interesting takeaway nonetheless.

    • artninja1988 17 minutes ago
      They speak to this a bit in the footnotes:

      "It’s important to note that “compressor” in this sense is something that is capable of performing compression by predicting unseen data, not a compressed artifact like a GZIP file - that is, the best compressor trained on a given corpus does not necessarily “remember” the contents of the corpus."

    • fergal_reid 19 minutes ago
      That's not how these things work at all. Perplexity will be measured on some held out set, ideally some set of data that has never been on the internet.

      As such getting better at it will require the opposite to memorisation, it requires generation out of sample.

    • amelius 20 minutes ago
      By now they probably have a two-step process: first ask an LLM to rewrite the copyrighted work into something different but close; then train on that.
  • GIFtheory 2 hours ago
    The post makes perplexity seem overly mysterious… First, “log perplexity” is just cross entropy. Why not call it that?

    Also, when perplexity is this low, it arguably makes more sense to report perplexity. A perplexity of 2 basically means that the model can guess a set of size 2 that contains the true next token.

  • hypfer 3 hours ago
    This is just an ad. Why is it top #1 of the front page?

    Just some guys on a payroll claiming that an arbitrary set of milestones is a curve towards a greater magic goal that means that some valuation is actually true.

    Why would you read this

    • semiquaver 3 hours ago
    • AlexCoventry 3 hours ago
      The graph is informative, at least.
      • hypfer 3 hours ago
        I don't see how it is?

        As said, the labels there are just randomly chosen things that made headlines, thrown on there to imply a clear causal relationship.

        That's the level of "science" that gets you banned from campus. In an ideal world, anyway.

        • spuz 2 hours ago
          Maybe a poor analogy but I see this as OpenAI releasing their GHz speeds or IPC metrics if we were to imagine they sold CPUs instead of AI services. It doesn't tell you much about the final performance but it does give an interesting core data point.
          • hypfer 2 hours ago
            I don't think that's the case, because with instructions per second, there is a clear relationship. Whereas here on one axis they plot GHz and on the other they plot matches of StarCraft won.

            There is certainly some influence of the Ghz of your processor on how many matches of StarCraft one can win, but.. well. Bogus science.

            • spuz 2 hours ago
              My point was that it is a core but not necessarily correlated metric and it sounds like you agree with that.
              • hypfer 2 hours ago
                Okay so, the graph is "informative" because it plots the perplexity value of the various OpenAI models over time?

                That's it? That's all this submission has going for it?

                Okay.

                • AlexCoventry 1 hour ago
                  That was all I'd apprehended of the post when I replied earlier.

                  The post argues that NTP perplexity is a good first approximation to model capability, and could therefore form the basis for an industry-wide commitment to pause model development at a given perplexity level until there's a consensus that AI alignment and safety has caught up enough that it's safe to continue developing even more capable models.

                  It's definitely not simply an ad.

        • AlexCoventry 2 hours ago
          They're meaningful to anyone who understands the standard numerical objective used in next-token-prediction pretraining.
          • hypfer 2 hours ago
            I am so incredibly glad that this is going to be over soon, because the amount of weak replies like this is just.. insufferable.

            The whole AI Bubble topic aside, this is just.. ugh. Intellectually repugnant.

            There is just nothing here other than a war of religions. Nothing is to be learned here. There is zero curiosity. It's just a screaming match.

            I hate this.

            inb4 "no you"

            Yeah man. Whatever.

            • AlexCoventry 1 hour ago
              The graph shows the average uncertainty the models have when predicting the next token in natural-language text. That's a highly specific measurement, and informative anyone who understands it.

              You're not under any obligation to understand it, of course, but if you ever get past this emotional reactivity, you'd benefit from making the attempt, IMO.

  • jwpapi 2 hours ago
    I honestly mean that I don’t know, but an alternative way of reading this is:

    Anthropic is getting uncatchable by us and we merely keep ahead against chinese models, so let’s all slow down.

    I think the reason Anthropic is so far ahead could be that they get way more user conversations, as everybody else distills, the rest is one gen behind...

    • zodwick 2 hours ago
      > they get way more user conversations

      Idt its that simple. Openai is much bigger than anthropic in terms of userbase

      • jwpapi 2 hours ago
        Not to be non-diplomatic, but I think claude users have the more challenging conversations.

        I’m not sure it’s that simple either, it’s just very tough to take anyhing for bare value with everything that’s at stake. Trillions of value, future of humanity.

        Tricky..

  • yipinwong 3 hours ago
    They are touching an elephant's legs, and saying Elephant is like a log.
  • AIShill101 3 hours ago
    Gah man do I hate AI ethics and stuff. I really hoped the lead scientist at openAI wouldn't be so screwy but they're all part of that cult aren't they?

    I really hope we go full speed ahead on AI development.

    • derac 3 hours ago
      https://x.com/sama/status/2124551883130114403

      ^ Sam Altman statement, basically saying that's not their intention. I want better AI too, but it seems prudent to move forward with caution.

    • Ohentis 2 hours ago
      This is basically the most sane post ever posted by an openai employee. This is just like an opinion. That seems like a good one.
    • bayindirh 3 hours ago
      > Gah man do I hate AI ethics and stuff.

      Why? You hate ethics in general, or only when applied to AI?

      • AIShill101 3 hours ago
        Mostly when applied to AI, it's just a ploy for regulatory capture by these giant companies.

        That's all. Ethics in general is overrated but I digress. LLM development really should go on full steam tho.

        • reasonableklout 46 minutes ago
          Vast majority of Americans support AI regulation even at the cost of slowing down development, and 2 of the 3 Turing Laureates for neural networks want the same despite not being part of these companies. What are you hoping to see with “full steam ahead LLM development”?
        • bayindirh 2 hours ago
          I personally disagree with you, because the lack of ethics is what creates today's world. Are you happy to be lied to your face, rugpulled constantly, or left defenseless in cases where you're right but not the owner of a billion-dolar company, in general.

          Are you happy that your water is poisoned, air is polluted, food is degraded just because somebody cares more about money than ethics?

          LLMs are the same thing. Training data provenance is a big, ugly can of giant worms. License washing is a problem, inefficiencies of data centers is a problem, using open-loop water systems and methane conversion generators just because they are cheaper are borderline crimes in themselves.

          We're digging what OpenAI and Nvidia dig to the world economy and consumer prices worldwide. Just because a couple of companies wanted more money in the name of LLM development and cutting edge science.

          On the other hand there are projects which are trying to minimize DC impact by scheduling things to places which they have most renewable energy on tap at a given moment.

          Go, figure.