a16z partner: AI editing other models' weights is an underrated breakthrough for continuous learning
illscience · x · 2026-11-07
a16z General Partner Anish Acharya argues that models editing other models' weights is a bigger deal than people realize and could unlock continuous learning. He cites a recent weight-editing paper that drew enthusiasm then criticism, says the smartest people he knows have replicated and extended it, while noting possible constraints like models holding opposing ideas.
The underlying work (cited by @rikiyatakehi) let GPT-6 Astra edit Qwen3.5-0.8B weights to improve math: changing only 0.016% of weights, with no gradients or training, it fixed the 0.8B model's reasoning loops, improved math up to 6x and instruction following 3x. They call this Weight-Space Programming: frontier LLMs natively reading/writing neural network weights.
More from AGI Musings
- 'The End of AI Frontiers': compute scarcity pushes Chinese labs toward efficiency — DimitrisPapail · 2026-11-07
- Every cloned its Slack with 51 AI coworkers trained on chat history via DSPy GEPA — lateinteraction · 2026-11-07
- Ranking sci-fi's AI predictions: Ted Chiang's 'Lifecycle of Software Objects' comes first — _akpiper · 2026-11-07
- Gladstone AI CEO: Washington won't slow AI until an 'AI Chernobyl' strikes — AaronBergman18 · 2026-11-07
- Bloomington, Indiana passes first law in history banning superintelligence, 7-0 — DavidSKrueger · 2026-11-07
- The six-step power-seeking AI x-risk argument, broken down for credence calibration — AndyMasley · 2026-11-07