MIT's AI 'mechanic' can read and edit the internals of other neural networks, Nature reports
marinkazitnik · x · 2026-11-14
A MIT preprint featured in Nature shows an AI model (GPT-6 Astra) that interprets the raw internal values of other networks well enough to predict what parts do — and then edits them, including boosting their math abilities. The 'weight-space programming' approach treats neural networks like files that AI can debug, with early sparks seen in frontier models. Researchers say it could improve AI safety and utility.
More from Research
- NYU Courant launches Center for Mathematics and AI, co-directed by Bruna and Vicol — kchonyc · 2026-11-14
- mariofilho: Pareto plots should be mandatory for benchmarks — mariofilhoml · 2026-11-14
- Rewiring just 47 neurons induces emergent misalignment in Qwen, then makes it interpretable — maxsloef · 2026-11-14
- Antithesis spent 8 years cracking NetHack Ascension, now clears ~80% of seeds in under 5 minutes — spikedoanz · 2026-11-14
- 5.46M-param touch model LingBot-Touch pokes straws through sealed lids autonomously — Puzzleheaded-Buy3008 · 2026-11-14
- New B-cell directed evolution paper rediscovers a 2004 Roger Tsien idea — NikoMcCarty · 2026-11-14