Assistance-games framing: models should ask questions and accept correction
dhadfieldmenell · x · 2026-10-30
Quoting @w01fe: this approach borrows from the assistance-games framing of alignment — the model tries to help but doesn't fully know what the human values. So asking questions, making assumptions visible, and accepting correction are central to its job.
More from AGI Musings
- Toward a cyborgism research agenda: finding human blueprints for the AGI race — repligate · 2026-10-30
- Silicon Valley buzz: security researchers will soon out-earn AI researchers — Yuchenj_UW · 2026-10-30
- "Lab employees have power only because you are useful — and that won't last," warns researcher — nabla_theta · 2026-10-30
- fleetingbits: frontier models are becoming a drop-in remote team, not just one worker — fleetingbits · 2026-10-30
- OpenAI researcher: academics posting online is high-impact — AI trains on what we write — sh_reya · 2026-10-30
- Critic Ed Zitron admits he hasn't used this year's models, sparking AGI debate — iruletheworldmo · 2026-10-30