Early Evals for OpenAI O3openai.com 30 pointsmaurycy1 year ago2 commentsSaveHideCopy link On HNComments−blurbleblurble1yWow, the demo where the user asks for untraceable payments shows some pretty sophisticated reasoning. The word "crafty" comes to mind.−famouswaffles1yNew SOTA's on:SWE-Bench - 71.7Competition Code - 2727ARC (Semi Private Eval) - 75.7 on low, 87.5% on high computeFrontier Math (previous SOTA was 2%) - 25% on high compute
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