Humans scale superlinearly on 14-day tasks while agents hit log-linear limits
AlexGDimakis · x · 2026-09-16
- New paper compares expert human coders vs coding agents on the same AtCoder Heuristic Contest tasks using Elo-per-token analysis.
- Agents initially outpace repeated sampling but converge to a log-linear scaling curve over long horizons; humans improve superlinearly, evidence they do continual learning while solving problems.
- Practical budget rule: 5M tokens → 1 Claude Code session; 30M → 2 independent 15M sessions; 100M → 3. Splitting at the scaling inflection point yields significant gains over a single long run.
More from coding & agent
- Hugo Bowne on the biggest agent failure mode: it's the harness, not the model — hugobowne · 2026-09-16
- Lean Verified Transformers: human-written text, AI-written proofs of core invariants — srush_nlp · 2026-09-16
- One prompt turns any book into a chapter-by-chapter podcast site with parallel subagents — armand_ruiz · 2026-09-16
- Image-to-WebDev Arena: GPT-6 Astra tops at 1733, GLM-5.3-Flash shines on price — arena · 2026-09-16
- Probability-only model tested as LLM judge: 25x faster, 600x cheaper than Fable-class — danshipper · 2026-09-16
- Two years with an AI career coach: my agent now knows my job better than I do — danshipper · 2026-09-16