04 · Five things to know · 05
Bright spot: OpenAI estimates its agents now finish about 46% of two-to-four-hour research tasks unaided
Know it
What happened. OpenAI estimates that the share of two-to-four-hour AI research tasks its agents complete without human help rose from 23% in March to 46% in October, and from 13% to 35% for four-to-eight-hour tasks. It calls the results “jagged”: 79% for monitoring training runs, 25% for inference and serving code. Claim
Why it matters. Longer tasks done without a person stepping in is one measure that matters for practical work, and OpenAI publishes where its agents still fail. Analysis
RED TEAM
These are OpenAI’s internal estimates on its own tasks. OpenAI says changes in what researchers choose to delegate may account for part of the gain, and its survey data are self-reported.
Sources
OpenAI, research acceleration report, 11.19 Co.
Bookmarks need a Mind Club sign-in. How sign-in works
Read it in Five things to know