At COLM, researchers clash: RL-loss tweaks are no more 'science' than prompt tuning
BlancheMinerva · x · 2026-10-11
Yoav Goldberg cites a pervasive COLM sentiment—"last year was boring, we only did prompt tweaking, now we're doing science again"—and pushes back: that science is mostly meaningless tweaks to an RL loss term, no better than prompt tuning. tallinzen adds that this reveals a mismatch between researchers' training/identity and what 90% of impactful LLM work actually is: data, evals, policy, applications.
More from AGI Musings
- Did OpenAI limit solutions to protect academia? Debating decentralized, GitHub-speed research — IgorCarron · 2026-10-12
- CSCW2026 Keynote to Propose Expanding Research Agenda to Human-Agent and Agent-Agent Collaboration — merrierm · 2026-10-12
- Top AI commentator says negative polarization over AI is inevitable, slams skeptics — teortaxesTex · 2026-10-12
- Digital minds may not want humans as guardians, machine consciousness essayist argues — cccalum · 2026-10-12
- Allie K. Miller: inbox zero is overrated, some overload sharpens focus — alliekmiller · 2026-10-12
- Figma CEO Dylan Field: As code becomes material, value shifts to design and judgment — a16z · 2026-10-12