Zhejiang University's Mechanist: AI studies its own mechanisms like a neuroscientist
jiqizhixin · x · 2026-09-15
Zhejiang University and collaborators present Mechanist, an AI system that studies itself: proposing hypotheses about its internal representations, locating mechanisms, designing experiments, verifying causality, and even intervening on specific circuits to improve its own capabilities. Unlike AI Scientists aimed at drugs or materials, Mechanist targets how models acquire knowledge and compute internally, organizing mechanistic research into a causal chain: discover behavior → locate mechanism → causal verification → active intervention → capability improvement. Representative result: in Pythia-1B, only 0.05% of parameters are highly correlated with belief-state reasoning.
More from Research
- Probabilistic programming advocates: engineer AI by design, don't grow it — xuanalogue · 2026-09-15
- FDA-approved pancreatic cancer drug daraxonrasib extends median survival by 6.5 months — EricTopol · 2026-09-15
- SoK: systematizing 58 cryptographic private transformer inference frameworks — chaumian · 2026-09-15
- ReactHuman: a physics-grounded benchmark exposes reactive safety failures in embodied multimodal LLMs — Yizhan Li · 2026-09-15
- OpenAI says ~10k concurrent agents solved the Navier-Stokes Millennium Prize Problem — DKokotajlo · 2026-09-15
- OSS harness boosts OS World 2.0 score ~20%, jumping Sol/Max from 4th to 1st — demeyer1 · 2026-09-15