Abstract:
OpenAI chief scientist Jakub Pachocki proposes that artificial intelligence companies should agree to set limits on the development of increasingly powerful models until researchers have better ways to understand and monitor the behavior of these models. Pachocki called for industry-wide standards that would set boundaries for AI development, and companies should agree not to push those boundaries until specific security requirements are met.

"I further propose that common limits be set based on perplexity and other measures, with developers committing not to breach these limits until our ability to understand and monitor AI behavior reaches a specific milestone," Pachocki wrote. Pachocki's proposed limits could be based in part on a technical metric called perplexity, which measures an AI model's ability to predict the next piece of text. Pachocki did not say what level the initial limits should be set at, or when they would be implemented.
Pachocki warned that as early as 2027, AI could accelerate its own evolution faster than researchers can improve their ability to monitor and understand the technology.
OpenAI CEO Altman expressed support for the proposal, calling Pachocki's article "the most important content OpenAI has released in some time." In a post on X, he said agreeing on shared limits would help ensure "security and control".
Comments