Mollick: Small Accuracy Gains Compound in Long AI Tasks

Ethan Mollick highlights an ICLR 2026 paper showing that small per-step accuracy advantages compound exponentially over long tasks, and notes that cheap small models save money early but degrade quality at the tail end.

2026-10-16 ~ 2026-10-16 · 2 related posts

• Mollick points to ICLR paper showing small accuracy gains compound exponentially over long-horizon LLM tasks — emollick · 2026-10-16
• Mollick: small models save money short-term but need larger models orchestrating and error-checking — emollick · 2026-10-16