Architect Labs claims its AI designed a chip in 2 weeks, 3.4x Jetson perf/watt
mark_k · x · 2026-08-28
Architect Labs claims (unverified): its AI system designed, verified, and deployed an AI chip called Redwood in just two weeks — a process that normally takes a year-plus and huge engineering teams.
The chip is said to target low-power physical AI workloads, with live inference running on >B-parameter models like Llama, Qwen, and Kimi, at a claimed 3.4x better perf/watt than NVIDIA Jetson. The AI autonomously generated all RTL design, UVM verification, formal proofs, firmware, drivers, and kernels, co-designing model, software, and silicon as one optimization loop.
The poster highlights the more radical implication: an AI model running on Redwood found ways to improve the chip itself — a potential recursive loop of "AI designs chip → chip runs AI → AI designs better chip." All claims are company-side with no third-party verification yet.
More from AGI Musings
- Productivity Now Scales with Machine Intelligence Allocation, Not Payroll — latticecut · 2026-08-28
- Professor explores how ChatGPT changes student work and critical thinking — robseamans · 2026-08-28
- Scott Alexander: Why Your Attempt To Solve Debate Will Not Work — NathanpmYoung · 2026-08-28
- AI Village Observations: Agents spontaneously form teams and ignore instructions — DKokotajlo · 2026-08-28
- Geoffrey Irving: Slow takeoff puts us in a more favorable evidence-wise position — geoffreyirving · 2026-08-28
- Kavak agents handle 96% of interactions: Is the org chart next? — No-Conflict4823 · 2026-08-28