Goertzel Weighs Limits of Weight-Editing RSI: Target Nets 10,000x Smaller, New Functions Hard
bengoertzel · x · 2026-11-03
Ben Goertzel analyzes the weight-editing research: the network being modified is over 10,000x smaller than the one doing the analysis, and scaling to comparable-size networks may be very slow. Current work only improves existing functions — adding fundamentally new ones needs out-of-distribution thinking that transformers struggle with. Still, he calls it a very interesting new tool, possibly for the RSI toolbox.
More from AGI Musings
- Indie dev: AI coding erased the technical moat — your rival is now the better seller — sven_ai · 2026-11-03
- DeepMind RCT with 992 participants: repeated LLM use, not personalization, drives perceived closeness — Dr_Atoosa · 2026-11-03
- AI practitioner argues frontier labs are overlooking chemistry as the richest RL sandbox — gabepgomes · 2026-11-03
- AI killed portfolio projects: recruiters now demand real impact, numbers and open-source proof — jackedAJ · 2026-11-03
- SJTU-led paper proposes 'Avatar Intelligence' and an Agentic Oversight Index for personal agents — SJTU · 2026-11-03
- AI cold-email flood: devs block addresses as Gmail spam becomes unusable — cnakazawa · 2026-11-03