OpenAI proposes AI model limits tied to intelligence metric
Rising concern over the speed of artificial intelligence advances is prompting calls for clearer industry guardrails on model development. OpenAI's chief scientist says developers should adopt shared capability limits and use perplexity as a common measure of model progress.
Highlights
- OpenAI chief scientist Jakub Pachocki proposes setting industry-wide capability thresholds for AI models based on perplexity metrics until safety milestones are met.
- The proposal suggests developers should not exceed agreed AI development limits before monitoring and safety mechanisms reach defined standards, with fundamental advances possible by 2027.
- OpenAI CEO Sam Altman endorses the plan, urging the industry to adopt shared progress limits to balance rapid AI advances with improved safety and control.
Proposal sets capability thresholds for AI developers
As reported by Reuters, OpenAI chief scientist Jakub Pachocki outlines the proposal in a blog post published on Sunday and says developers should commit not to cross agreed capability thresholds until monitoring and safety mechanisms reach defined milestones.Pachocki says the industry should adopt common limits for increasingly powerful models and use perplexity as a standard metric for measuring progress. He describes perplexity as a measure of how well a model predicts the next token, the basic unit of data, on average, with lower perplexity indicating stronger predictions.
He also advocates for an industry standard that sets an initial development limit and defines what milestones must be met before that limit is exceeded. In the proposal, Pachocki says AI could help drive fundamental advances in training methods as soon as 2027, potentially accelerating progress faster than improvements in monitoring and safety.
Safety debate gains weight across the AI sector
The proposal emerges as concern grows inside and outside the technology industry about the risks linked to rapidly improving AI capabilities. Pachocki argues that developers should hold back from surpassing agreed limits until their ability to understand and monitor AI behavior reaches a concrete threshold.OpenAI CEO Sam Altman backs the proposal in a post on X, saying the industry should seek agreement on common limits for AI development. He describes Pachocki's essay as the most important thing OpenAI has published in a while and says shared limits would help preserve safety and control.
This adds to earlier discussion around Bayesline internal feedback and how small AI-focused teams keep decision-making fast.