Alignment researcher: RL-era gains are narrow and trade off other skills
dhadfieldmenell · x · 2026-10-20
Alignment researcher David Hadfield-Menell argues that RL-era model development brings much narrower improvements, often at the cost of other skills/behaviors — especially expected if RL drives genuine capability change rather than merely eliciting capabilities from pretraining.
More from AGI Musings
- ChatGPT flubs a simple crossword, shaping outsiders' skepticism about AI — iruletheworldmo · 2026-10-20
- The AI safety inversion: why alignment efforts keep increasing risk — mayfer · 2026-10-20
- AI data brokering splits: pre-training data commoditizes, RL post-training data is the real business — agihouse_org · 2026-10-20
- Google researcher: 'tradition' is the social mechanism absorbing displaced work — moultano · 2026-10-20
- Frontier labs don't want your chats — they want the world around you when you typed the prompt — marktenenholtz · 2026-10-20
- Ethan Mollick: multiplayer AI must move past group chat — its promise is working across teams — emollick · 2026-10-20