Skip to content

Early Evals for OpenAI O3

openai.com
30 pointsmaurycy2 comments
On HN

Comments

Wow, the demo where the user asks for untraceable payments shows some pretty sophisticated reasoning. The word "crafty" comes to mind.

New SOTA's on:

SWE-Bench - 71.7

Competition Code - 2727

ARC (Semi Private Eval) - 75.7 on low, 87.5% on high compute

Frontier Math (previous SOTA was 2%) - 25% on high compute

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.