Wink Pings

OpenAI wants to use Perplexity to put the brakes on AI, but the brake line may be misdrawn

OpenAI has disclosed pretraining perplexity data for the first time and proposed that the industry set a common safety cap based on it. The metric has value, but as a brake line it has many problems: it is hard to compare across labs, it does not cover chain-of-thought, and its lower bound is unknown. Online commenters are more interested in when the internal model will be released.

OpenAI chief scientist Jakub Pachocki published a long article titled Perplexity and Intelligence. The same day, Sam Altman reposted it on X with the note: This is OpenAI's most important release in some time.

The most important part is that it publicly shares internal data on pretraining progress and proposes that the industry set a common safety cap based on the same metric—perplexity.

Perplexity, whose Chinese term literally means degree of confusion, measures a model's average performance in predicting the next token (text fragment). The lower the number, the better the model has grasped the patterns in the data. OpenAI's view is that pretraining is essentially training a compressor on large amounts of data, and perplexity is the core yardstick for tracking the intelligence level of that compressor.

OpenAI gave a set of data: GPT-3.5's log-perplexity was about 0.99, GPT-4 about 0.91, and by the current model it is already below 0.8. Each time a threshold is crossed, new capabilities emerge. GPT-3.5 crossed the threshold for fluent conversation; GPT-4 began to use chain-of-thought effectively; and below 0.8, the model showed transformative capabilities in cybersecurity and mathematics.

image

Pachocki argues in the article that developers should make a shared commitment: not to cross shared limits set based on perplexity and other metrics until the ability to understand and monitor AI behavior reaches a specific milestone. In other words, draw a clear brake line for AI development.

Why use perplexity as the yardstick? Because it can be compared across models and labs. But this is also the most controversial part.

Different labs use completely different datasets and data mixtures. OpenAI uses a fixed snapshot of its own internal codebase; how are others supposed to benchmark against that? Perplexity does not directly correspond to any practical task capability; it is more of an underlying metric. In addition, capability jumps from chain-of-thought reasoning will not show up in a drop in perplexity. And there is the most fundamental problem: we do not know what the intrinsic entropy of the data actually is, or where the floor on loss lies; the closer it gets to the floor, the more a tiny drop in perplexity may mean a huge jump in capability.

Pachocki also mentions an accelerating factor: AI has already begun to participate in pretraining algorithm research. He writes that when the model reaches a log-perplexity of about 0.76, it starts producing breakthrough results in mathematics, to the point that it feels surreal. He speculates that a similar avalanche in pretraining research may occur in 2027. This precisely shows that the window for humans to manually set limits may be very short.

Sam Altman said in the repost: I hope we can agree as friends on shared limits. Perhaps he typed too fast and misspelled 'friend' as 'frield'. The replies quickly latched onto this meme, but the stronger emotion came from elsewhere.

One commenter said bluntly: You will not make us stop; open-source open-weight models will win. Another asked: Who will audit these perplexity limits? Many others repeatedly asked the same question: When will OpenAI's legendary internal model be released? It seems the community's trust in self-regulation is not much higher than its confidence in this metric.

OpenAI voluntarily showing its hand this time is indeed rare. But are those cards real? Will other labs follow? Can the metric itself withstand scrutiny? All are unresolved questions. The only certainty is that, for the first time, the gap between the speed of AI development and our ability to explain it has been laid out quantitatively by a single number.

发布时间: 2026-11-23 02:42