Nov 22 (Reuters) – AI developers should adopt common limits on increasingly powerful models and use “perplexity” as a common metric for measuring AI progress across the industry, OpenAI’s chief scientist said on Sunday in a proposal backed by CEO Sam Altman.
The proposal detailed by OpenAI chief scientist Jakub Pachocki in a blog post was unveiled at a time of growing concern inside and outside the industry about the potential risks of AI’s rapidly growing capabilities.
“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,” Pachocki wrote.
Pachocki described perplexity as a model intelligence metric concerning how well a model predicts the next token — the basic unit of data — on average, and wrote that lower perplexity means better predictions.
Pachocki wrote that AI could help drive fundamental advances in AI training methods as soon as 2027, potentially accelerating progress more quickly than advances in monitoring and safety. Developers should commit to not cross agreed capability thresholds until new monitoring and safety mechanisms are in place, Pachocki said.
Pachocki advocated for an industry standard that specifies a first limit on development, and defines what milestones must be met before that limit is exceeded.
Altman backed the proposal, saying in a post on X that the industry should seek agreement on common limits for AI development. Altman described Pachocki’s essay as “the most important thing OpenAI has published in awhile,” and said such limits would help ensure safety and control.
(Reporting by Disha Mishra in Bengaluru; Editing by Will Dunham)

