Perplexity and Intelligence
Summary
The article argues that modern AI development fundamentally relies on training large-scale compressors (models) on vast datasets, with perplexity (or cross-entropy loss) serving as a key metric for measuring progress on this "intelligence substrate." The author traces how decreasing perplexity has correlated with emergent capabilities: GPT-2 (~coherent paragraphs), GPT-3 (~in-context learning), GPT-3.5 (~fluent conversation), GPT-4 (~chain-of-thought reasoning), and more recent models with log-perplexity below 0.8 showing transformative abilities in cybersecurity and mathematics. The piece emphasizes that these capabilities were not predictable a priori from perplexity thresholds alone. The author proposes that frontier model developers should establish shared norms for evaluating models and red-teaming safety cases, including transparency standards and common limits on development measured by perplexity and other metrics. These limits would remain until concrete milestones in understanding and monitoring AI behavior are achieved, such as a successor to chain-of-thought monitors. The article acknowledges challenges with perplexity as a shared metric—difficulty of cross-lab comparison due to different datasets, detachment from actionable evaluations, failure to account for advances outside pretraining (e.g., chain-of-thought reasoning), and an unknown floor—but maintains that perplexity has stood the test of time and can play an important role in standardizing discussions of AI progress.
(Source:OpenAI)