Researcher Argues AI Accelerates Faster on Verifiable Tasks; "Squishy" Tasks Face Delayed-Reward Trap
davidmanheim · x · 2026-11-11
Researcher David Manheim lays out a structural argument about AI acceleration:
- Acceleration speeds differ by task type: generalization to verifiable tasks like math has occurred far faster than to "squishy" tasks like writing style.
- Economic value implies delayed feedback: economically valuable squishy tasks must be embedded in the real world with longer delays — a trilemma: if delays were short enough, tasks would be verifiable; with no feedback at all, they couldn't have economic value.
- Delayed-reward RL is fundamentally harder: model-free RL propagates value exponentially slowly with more steps, and simulation accuracy for complex systems is fundamentally compute-limited, making sparse-reward traps and spurious solutions hard to avoid.
He stresses this doesn't mean models can't do such tasks, but "verifiable" isn't binary — unverifiable tasks are sparse-reward problems, which may limit real-world impact.
More from AGI Musings
- François Fleuret to biologists: look how cocksure other fields were before AI — francoisfleuret · 2026-11-11
- Could Recursive AI Improvement Compress Millennia of Progress into One Decade? — VraserX · 2026-11-11
- AI 2027 authors say reality is running 5 months ahead of their scenario — LeviTurk · 2026-11-11
- teortaxesTex: RL may give coding models 'full fluency' with no superhuman headroom — teortaxesTex · 2026-11-11
- Frontier AI Still Lacks 'Heresy': Breakthroughs Come from Defying Consensus — JeremyNguyenPhD · 2026-11-11
- Dev argues optimism is the only trait that matters as intelligence gets commoditized — zack_overflow · 2026-11-11