Alignment researcher: spend less time on metrics, more time just talking to models
repligate · x · 2026-11-18
AskYatharth observes that alignment research culture favors running experiments and reading numbers, with far less time spent simply talking to models and being genuinely responsive to what they say. His advice: literally just talk to them — or have your Claude do it, keeping an open collaborative vibe rather than a 'serious researcher telling the models what to do' stance. Shared by repligate.
More from AGI Musings
- Interactive tree compares Paul Christiano's and Parzival's superintelligence timelines: 50% by late 2028 vs Sep 2027 — LeviTurk · 2026-11-18
- Fed warns AI agents could trigger bank runs in minutes — GaryMarcus · 2026-11-18
- Bill Gates resurfaces Dune warning: handing thinking to machines only enables new masters — pmddomingos · 2026-11-18
- Wojciech Zaremba: AI needs its own Project Vela — verification tech to enable global agreements — HaydnBelfield · 2026-11-18
- Claim: Claude was trained on protein sequences at scale — and biologists think that's the right paradigm — AllThingsApx · 2026-11-18
- Recursive co-founder Tim Shi on scientific taste, evals, and automating discovery — AnneliesGamble · 2026-11-18