Is RL generalization a myth? Skeptic argues AGI via RL needs a bespoke environment for every task
akbirthko · x · 2026-10-12
A debate on whether RL leads to AGI. Quoting @waterloointern: LLMs' jagged capabilities and the fact they emerge from RL on revenue-curated environments suggest RL generalization doesn't exist — reaching AGI this way would mean mobilizing the world economy to build an environment for every task, making 'environment/data sellers' the most valuable companies. Main poster akbirthko reframes: if AGI is the godlike ability to materialize anything, the Last Task is defining what's worth wanting, which in practice we only approximate.
More from AGI Musings
- 'The most important chart for the AI economy': models commoditize, compute wins — FinanceYF5 · 2026-10-12
- AI Has Come For Call Centers: Job Losses Now Spread to Lower-Middle-Income Countries — FinanceYF5 · 2026-10-12
- Re-evaluating Steve Jobs: iPhone's net impact on society — granawkins · 2026-10-12
- From nuclear safety to AI: containing agents with signals they can't produce — WeldPond · 2026-10-12
- Personal Agents Nail First Impressions but Lack Daily Use Cases for Retention — sujingshen · 2026-10-12
- willcb: every non-RL algorithm is a proxy — policy gradients are the minimum-variance unbiased estimator — willcb · 2026-10-12