· Desk editor: Daniel Mercer · · Single source

OpenAI scientist proposes perplexity metric to limit AI speed

OpenAI chief scientist Jakub Pachocki published a blog post proposing perplexity as a standard measurement for limiting artificial intelligence development.

New benchmark for progress limits

Jakub Pachocki, OpenAI's chief scientist, released a new blog post discussing potential limits on AI development. He focuses specifically on a measurement called perplexity to establish these boundaries. This approach addresses concerns raised after a former researcher warned that companies were gambling with lives. Anthropic CEO Dario Amodei previously published a plan to pace the frontier of technology growth. OpenAI CEO Sam Altman described Pachocki's post as the most important publication in a while.

Perplexity measures prediction accuracy

This suggested measure determines the average accuracy with which a model forecasts the upcoming token. Lower perplexity values indicate better predictions and signify advanced model performance. Benchmarks focusing on downstream capabilities have quickly become outdated as technology advances rapidly. Pachocki noted that GPT-4 achieved losses around 0.91 while utilizing chain-of-thought reasoning productively. Now that losses fall below 0.8, transformative capabilities in cybersecurity and mathematics have emerged.

Agreement on safety standards

Companies would commit not to exceed a certain threshold until monitoring abilities reach a concrete milestone. This shared benchmark for general intelligence allows firms to agree on common safety standards. The goal is to prevent unchecked AI progress from reaching catastrophic or existential risks. SpaceX CEO Elon Musk has endorsed similar ideas regarding limiting the rate of advancement.

Reported by one outlet

Only one outlet has published this. Nothing here has been checked against a second report, so read it as that outlet's account and follow the link below for the original.

Reported by

1 independent outlet. Headline as published. Links open the original report.