Kenneth Stanley spotlights Kernel Autoresearch: agents discovering transferable inductive biases
kenneth0stanley · x · 2026-11-13
Kenneth Stanley boosted Richard Suwandi's upcoming ML Collective talk, Kernel Autoresearch for Open-Ended Model Discovery, exploring whether coding agents can help discover new models for scientific discovery.
- Core idea: let agents discover reusable inductive biases expressed as programs that transfer across unseen tasks.
- Per the talk slides, the system combines LLM-guided kernel search (propose→revise→interpret), a deterministic verification backend (output shape checks, PSD tests, consistency checks), a MAP-Elites archive preserving the best candidate per behavior niche, and certificate contracts (feature maps, spectral, input transforms, PSD closure).
- The talk streams at 10am PT via ML Collective's long-running paper reading group, which has met weekly since 2018 and doesn't record sessions.
More from coding & agent
- Streaming output and visible progress, not faster models, keep users from abandoning slow AI features — XxFreakhawkxX · 2026-11-13
- Prototyping a prediction-accuracy tracker on HN comments with an LLM — Darpinian · 2026-11-13
- ArtCraft clones Photoshop with AI: 1.7M daily downloads, 4,000 devs join — Latent Space · 2026-11-13
- Dev builds ComfyUI nodes to add batch generation to MiniMax H3 workflows — an80sPWNstar · 2026-11-13
- The Agent Did the Task Perfectly — It Was Just the Wrong Task: Reddit Debates Intent Mismatch — djgoel123 · 2026-11-13
- Heavy Codex Users Struggle to Track Old Sessions and Scattered Outputs — grahamhart_ · 2026-11-13