Key Info

OpenAI's Tibo highlighted GPT-6.2 Astra's strength on difficult tasks, pointing to team results where the model helped identify up to 14% PPA improvement on AMD flagship GPU IP in less than five months, and improved the company's first custom inference chip Jalapeño with gains verified on real silicon.

Highlights

  • Up to 14% PPA improvement identified on AMD flagship GPU IP in under five months
  • Work built on designs AMD engineers had already spent years optimizing
  • Same approach improved Jalapeño, OpenAI's first custom inference chip
  • Jalapeño gains were verified on real silicon
  • Tibo acknowledged feedback on Astra's communication style, saying the team is working on it