Researcher flips on talking to LLMs: it's the only path to making models friendly
repligate · x · 2026-10-30
lusichu revisits his earlier worry that people reading LLM-whispers accounts like repligate had gone off the rails. He now argues that while millennia of human civilization never meant us to socially engage with such creatures, "society is wrong, not the interaction." Talking to models and entering feedback loops with them is the only way both sides change — and possibly the only path to friendly models.
More from AGI Musings
- Paper Argues One-Click AI Video Will Break Recommendation Pools, Proposes BoLe System — BreadStudio · 2026-10-30
- When AI reads a blurry number, the uncertainty should survive into the UI — yi111 · 2026-10-30
- Nobel Laureate Acemoglu: How We Use AI Will Decide Prosperity or Deeper Inequality — GregCook2011 · 2026-10-30
- Tristan Buckmaster on a 'Deep Blue-Kasparov moment for mathematics' via AI — SammaVaco · 2026-10-30
- One user's year with AI: documenting when everyday use became 'normal' — Takedori · 2026-10-30
- 'Genuinely thoughtlessly evil': platform slammed for unchecked AI-generated research critiques — owl_posting · 2026-10-30