Bitter Lesson for RLMs: harnesses may drive generalization through decomposition

viksit · x · 2026-07-21

The post argues that the *bitter lesson* points away from domain-specific structure and toward general-purpose decomposition and recombination. It draws an analogy to CPUs: once single-core clock speeds hit a heat wall, progress came from parallelizing workloads instead of just making one core faster. The author suggests LLMs may follow a similar path: as context windows and task horizons grow, scaling parameters will help, but **big-O improvements** in harnesses and decomposition may improve generalization even faster. A quoted thread adds a more concrete claim: when training RLMs on tasks with shared structure that look different on the surface, the root model can learn the same trajectory for both. In that view, the harness—not the transformer itself—induces generalization by composing calls over structurally similar trajectories.

Related event: Research: Harness Drives Generalization in Reinforcement Language Models(9 posts)→

Original post →

More from AGI Musings

AGI Musings channel →