OpenAI found up to 14% PPA in AMD GPU IP. Its own benchmark shows why specs are release gates
Production chip gains and Meta's agent-driven model propagation point to the same architecture: optimizers need explicit invariants, staged evidence, and tests that can detect an incomplete objective.