PARAS MADAN
All writing

Latest in AI

OpenAI's Perplexity Chart: What Log-Perplexity Can Tell You About a Model

OpenAI shared its internal pretraining trend and proposed a common metric for AI labs. Here is how to tell when two scores are actually comparable.

On November 22, OpenAI chief scientist Jakub Pachocki published internal measures of pretraining progress and proposed that AI labs use perplexity alongside other measures to inform shared safety limits. The OpenAI post is a look at how the lab tracks its base models. Its chart uses a fixed snapshot of internal OpenAI code, so the numbers are not a leaderboard you can compare directly with another lab's published scores.

What does log-perplexity measure?

Give a model a passage from a test set and, at each position, look up the probability it assigned to the actual next token. Take the negative natural logarithm of that probability and average the results across the scored tokens. That average is cross-entropy loss, also called log-perplexity in this setting. Exponentiating it gives perplexity. Lower means the model assigned more probability to the text it was asked to predict.

You can see the difference on two tokens. If a model gives the actual next token probabilities of 0.5 and 0.25, its average negative log probability is about 1.04 and its perplexity is about 2.83. A model that assigns more probability to the right tokens gets a lower loss. That example tells you exactly what improved. It doesn't tell you whether either model can finish the job that produced the text.

When are the numbers comparable?

Suppose you want to choose a base model for a code assistant. Score both checkpoints on the same held-out files, record the tokenizer and model revision, and make sure the evaluation gives each model the same input context. Check how the scorer handles file boundaries, special tokens and overlapping windows. Hugging Face's perplexity guide explains why a model can appear worse when you reset context at every chunk instead of using a sliding window.

Raw per-token scores are a poor comparison when tokenizers split the same file differently. One model might score fewer, longer tokens; another might score more, shorter ones. Normalizing the total log loss by the bytes in the same source text, and converting to bits per byte, can help with the unit mismatch. It still doesn't correct for different preprocessing, context, test-set leakage or what each model was trained to predict. Before trusting a cross-lab comparison, ask for the scoring recipe along with the number.

Does a value of 0.7 mean a model crossed a safety threshold?

No. Pachocki points to new mathematical results observed at roughly 0.76 log-perplexity on the internal measure and speculates that further capabilities could appear around 0.68 to 0.70. Those figures are tied to OpenAI's evaluation and to the author's interpretation of the trend. The proposed range is neither an agreed industry limit nor a demonstrated point at which a particular dangerous capability appears.

The proposal is to compare progress across labs and tie any future limits to concrete improvements in how model behavior is understood and monitored. The distinction matters: a measure that helps researchers notice something is changing doesn't tell them which capability to test, or whether a particular deployment is safe.

Pair prediction loss with a task you can grade

For the code assistant, use a second test with a clear outcome: take a held-out fix, give each system the same bug report and running tests, and check whether the patch passes without creating a regression. Record the model variant, tools and time budget. A model can be better at predicting the next token in a repository and still struggle to navigate the project, choose an edit or verify it. Reasoning at inference time and tools can improve the task outcome without changing the base model's pretraining loss, a limitation Pachocki explicitly notes for reasoning.

The useful part of OpenAI's proposal is the chance to standardize how progress is measured before arguing over what a particular number means. I would want the scoring protocol and the corresponding task results side by side, especially if the number were going to inform a decision about release.

I'm Paras. I work across AI, engineering and distribution.

I help AI teams explain technical results and what developers can actually do with them.

Book a 30-minute call.