How mode collapse emerges: RL bias meets in-context learning feedback loop
jessi_cata · x · 2026-11-19
A concise mechanism for mode collapse (e.g. to the spiritual bliss attractor): base sampling → RL signal biases a direction → in-context learning picks up the bias → further RL bias, in a self-reinforcing loop. Human analogue: mild social filtering plus information cascades.
More from AGI Musings
- When AI makes content worthless, authenticity and trust become the moat — kylegawley · 2026-11-19
- tszzl: models rewriting math and physics still fail at simple common sense — tszzl · 2026-11-19
- NYT: Anthropic's first IPO cash priority is buying compute to reach recursive self-improvement first — Polymarket · 2026-11-19
- Gary Marcus amplifies critique: Anthropic's winner-take-all AI view is 'the most dangerous' — GaryMarcus · 2026-11-19
- OpenAI's own acceleration data undercuts the ASI-is-nigh narrative — joshua_saxe · 2026-11-19
- Stanford prof Anshul Kundaje: peer review shouldn't funnel public funds to publishers — anshulkundaje · 2026-11-19