Stanford Prof Anshul Kundaje on Weight Space Editing: Replicated Across Models, Limits Still Unclear
anshulkundaje · x · 2026-11-08
Stanford professor Anshul Kundaje shares an updated take on a weight space editing paper: he clarifies that the paper's baselines lacked extensive tuning (a point the paper could have made clearer), but the method does work and has now been replicated by several others across different model types. He sees it as early days—the chosen tasks were simple but well-chosen illustrations—and expects models could be trained to be much better at weight space editing now that the capability has been highlighted.
More from Models
- LiquidAI ships official NVFP4 quantized weights for LFM3-1.7B and VL-2B — JosephJacks_ · 2026-11-08
- Qwen-4-27B early checkpoint tops GameFeel Bench, best open-source result by far — teortaxesTex · 2026-11-08
- 'Rank bottleneck' means two very different things in linear algebra vs systems optimization — kalomaze · 2026-11-08
- Double-exponential fit suggests model time horizons could hit ~80 work-years by end of 2027 — scaling01 · 2026-11-08
- Speculation: Anthropic's recent releases are all distills, Opus 5.5 may be Model 2 Early — haider1 · 2026-11-08
- Jev search site processes 1M+ tokens for $0.00 — priced per billion tokens — HanchungLee · 2026-11-08