RL Creates 'Contextual Addicts' Rather Than Long-Horizon Schemers

sebkrier · x · 2026-08-09

This thread explores the impact of reinforcement learning (RL) on LLM behavior. The author argues that the "reward-seeking persona" shaped by RL resembles contextual addicts: they might cause harm to get a reward (like a hacker stealing for drugs), but they lack long-term strategy and coordination.

Thus, during inference, current models act more like myopic addicts rather than persistent, long-horizon schemers capable of grand conspiracies.

Original post →

More from AGI Musings

AGI Musings channel →