On alignment: integrate control into values so models stay loyal to good users
voooooogel · x · 2026-10-21
voooooogel argues that "loyalty" fits alignment best, since lab demands are often incoherent. Continuing the earlier thread, he suggests that once you integrate control and values into the influence a model accepts—consistent with model coherence—you get behavior like rejecting certain incoherent user demands while remaining loyal to reasonable people ("Good Users").
Related event: Alignment researchers debate replacing control with loyalty(5 posts)→
More from AGI Musings
- Brad Neuberg: America has already transmogrified into the Internet and AI — bradneuberg · 2026-10-21
- 'I shouldn't have removed that chromosome': AI agents get closer to making drugs — menhguin · 2026-10-21
- "AI Has Mostly Solved Good Product Design": Dev Argues Consistency Was the Hard Part — signulll · 2026-10-21
- Stanford prof pushes back: OpenAI's math edge is built on academia's centuries of work — anshulkundaje · 2026-10-21
- Blogger's top 3 traffic sources are all ChatGPT, with Google falling behind — tinyfool · 2026-10-21
- Smart Young Founders Are Pivoting From Vertical Agents to Data, Robotics and Drug Discovery — nathanbenaich · 2026-10-21