OpenAI's Chief Scientist, Jakub Pachocki, has proposed that AI labs adopt shared "pacing limits" for developing more powerful models, warning of a potential "avalanche of fundamental pretraining advancements" as early as 2027. In a blog post, Pachocki suggested using "perplexity"—a metric for how well a model predicts the next piece of data—as a common yardstick to track progress and set safety thresholds across the industry.

As a step toward this, OpenAI publicly shared its internal perplexity measurements for the first time, from GPT-3.5 to its latest models. The company argued that the competitive risk of revealing this data is outweighed by the need for shared safety protocols. Pachocki stated that progress in AI safety and monitoring is unlikely to keep pace with the potential acceleration in pretraining capabilities, necessitating a more controlled development process. CEO Sam Altman endorsed the post, calling it a crucial step towards ensuring future AI systems remain safe and controllable.