Matthew Honnibal: LLMs can 'see' weight matrices — but only one layer at a time
honnibal · x · 2026-11-05
Explosion AI co-founder Matthew Honnibal offers a fresh take on weight reading: image recognition already means reading a matrix of numbers, so it's plausible a model can genuinely "see" a weight matrix — an angle he hadn't considered and thinks could be useful. He draws a clear limit, though: this likely works only for single layers; asking the model to mentally apply non-linearities across layers "seems silly."
More from Research
- ScreamingFace launches leaderboards showing multi-model systems can beat the best single model inside them — iamtrask · 2026-11-05
- OpenMined launches ScreamingFace: open-source model fusions beat best single models on live leaderboards — iamtrask · 2026-11-05
- New Paper Proposes Cryptographic "Replication Certificates" Verifiable in Seconds — Kaiping_Chen · 2026-11-05
- Vanilla ES lifts Qwen3.5-4B from 12.1% to 21.1% on Terminal-Bench 2.1, beating RL checkpoint — RulinShao · 2026-11-05
- MIT/Stanford Paper: Single-Feature Behavioral Profiles Best for LLM Human Simulation — sebkrier · 2026-11-05
- Stanford's ParallelKernelBench: best LLM solves only 32% of multi-GPU kernel tasks — burny_tech · 2026-11-05