RL-trained Kimi base model designs power transformers, hitting 93% of unseen specs in minutes
simonguozirui · x · 2026-09-15
Startup GenTrajectory demonstrates RLVR (RL with verifiable rewards) applied beyond math and code — to electrical engineering:
- Method: decades of physics-based verifiers built by engineers make electromagnetic systems a natural fit for RLVR; the team used Tinker to RL-train a Kimi base model to design medium-power transformers
- Results: 93% of unseen specs met, compressing multi-week engineering iterations into minutes of inference
- Motivation: AI is already improving its own architectures and chips; the power infrastructure it runs on deserves the same attention
- Tinker API's official account amplified the takeaway: RLVR isn't just for math and code — off-the-shelf physics verifiers are natural reward functions
A concrete example of the "physics verifier as reward function" route, relevant to AI for Science and industrial design.
More from AGI Musings
- Article argues intelligence has a speed limit, curbing recursive self-improvement hype — docmilanfar · 2026-09-15
- Banning data centers to save the world? A 500-year history lesson says otherwise — thursdai_pod · 2026-09-15
- Why Is LeCun So Unconcerned About AI Loss-of-Control Risks? — IndependentFresh628 · 2026-09-15
- Kapoor and Narayanan's 13,000-word essay reframes AI loss-of-control incidents — sayashk · 2026-09-15
- New 13,000-word essay: AI safety should bet on control and governance over alignment — sayashk · 2026-09-15
- Stop The AI Race pauses Occupy OpenAI protest while pushing international AI treaty open letter — DavidSKrueger · 2026-09-15