Architect Labs describes AI agents jointly designing a processor, its compiler, and the software that verifies it. Frostbite runs interaction nets: programs expressed as graphs that make independent work explicit. The account explains how agents searched across the whole system while machine-checked specifications constrained their changes.

  • One checked specification: the team’s ASL language connects software models and hardware designs to Lean, a theorem-proving system. Once physical checks exposed bad early estimates, the original design document was discarded.
  • Fast tests need calibration: agents repeatedly aligned a fast simulation with slower hardware power estimates. Across four searches and a final ablation round, they tried 58 ideas and retained 37.
  • Separate tuned and unseen tests: the company reports 4.59× performance per watt across 25 workloads, but 3.83× on the 15 held out from design tuning.

The workflow makes the case for agents exploring expensive engineering alternatives under explicit checks. Its performance claims are still pre-silicon estimates against an unnamed in-order processor; measurements from a manufactured chip and broader standard tests remain outstanding.

The 271-comment thread on Hacker News adds questions about the workloads and the surrounding developer tools.

What the thread adds

  • amluto asks about programs that “build large data structures and slowly mutate them.” titzer raises a related question about stateful computation. These are requests for broader evidence, not demonstrated failures.
  • mtimpe argues that “A source level debugger for Vine is a requirement” and flags stuck computations and leaking values as debugging problems to investigate.
  • pizza234 challenges the in-order baseline; kragen replies, “They’re claiming better performance per watt, not better absolute performance.” The disagreement concerns what comparison would support the headline.
  • jkhdigital interprets the contribution as lowering the cost of productizing old research ideas. That is a different claim from inventing the underlying programming model.

HN handles are pseudonymous and expose no per-comment scores; this slice follows HN’s ranking and is not a consensus.