OpenAI chip chief: we don't fully trust AI models for chip design

Richard Ho, who leads OpenAI's hardware team, told More Than Moore that the company's own AI models modified parts of its Jalapeño chip, in one case saving over 13% of die area, sometimes without a clear reason. Ho said he still doesn't 100% trust the models, so every change goes through full validation. The second version of Jalapeño is now in the lab for qualification.

• OpenAI's AI models saved over 13% of Jalapeño's die area
• Ho: "I still don't 100% trust them"
• Every model change goes through the full validation flow
• The second version of Jalapeño is in the lab for qualification