On Aug 25, 2026, in Palo Alto, we got together for Cool Chips - Hot Takes, an independent community gathering organized by AiNEKKO, AIFoundry, OpenHW Foundation, and 10xEngineers.
Saad demoed for the first time the "Agentic compiler". People who actually listened to OpenAI talk on Jalapeno at Hot Chips or had time to catch up on their blog post, got to appreciate the similarities of the that to what Saad’s team did on the ET-SOC-1 edge accelerator. The similarity does not end at the trajectory of the graph, the system has to be fast (velocity is the differentiation and harness with domain knowledge helps), cost-efficient (don't just throw LLMs on a large search space; time and $ are still of importance to many), closed-loop ( input --> optimization (explore or exploit) --> expert approval --> implementation --> validation on real hardware --> documentation --> feedback --> repeat), and human (for now) and machine verifiable (expert approval goes a long way; some human still has to own it).
Without much more spoilers watch the demo - the numbers and the time and cost of gettin go those numbers is really impressive and that’s just the beginning.
