OpenAI’s chief scientist has thoughts on how to limit AI development
OpenAI’s chief scientist Jakub Pachocki proposes using perplexity—a metric of how well models predict the next token—as a universal gauge to set speed limits on AI development. He argues that unlike downstream benchmarks, perplexity remains informative as models improve, and could serve as a shared standard for companies to agree on safety thresholds. This idea comes amid growing calls from industry leaders to pace AI progress due to existential risk concerns.