Kernaut's discovered kernels stay interpretable: 16 scalar functions, inner-product form
sirbayes · x · 2026-10-08
Another follow-up on Kernaut: a key virtue of the approach is interpretability. For glucose prediction, the best discovered kernel takes the form k(x,x') = φ(x)^T φ(x'), where x is the 5-d input (meal/bolus size/time) and φ(x) is a discovered set of 16 scalar functions — letting researchers inspect what the model learned.
Related event: Kernaut: Coding Agents Autodiscover Interpretable GP Kernels(4 posts)→
More from Research
- 0 of 722 AI bio papers cured cancer, as biologists tell tech bros it's brutally hard — iScienceLuvr · 2026-10-08
- New multi-agent swarm writeup: smarter models benefit more from collaborative work — repligate · 2026-10-08
- COLM adds missing panel title: Math, Science, and Knowledge Production in the Era of LLMs — gregd_nlp · 2026-10-08
- Actionable Interpretability workshop at COLM 2026: four keynotes from Goodfire, Mila, Stanford — megamor2 · 2026-10-08
- COLM Paper: Meaningful Neuron Directions Extracted from MLP Weights Alone — megamor2 · 2026-10-08
- CoBPE tokenization cuts sequences 30% vs BPE, better LLMs at same compute — alisawuffles · 2026-10-08