Bias-variance decomposition of frontier-model errors. Define error incoherence as the fraction of error attributable to variance (incoherence = variance / error; 0 = systematic, 1 = random). Four findings across frontier reasoning models (Claude Sonnet 4, o3-mini, o4-mini, Qwen3) plus self-trained synthetic optimizers: (1) longer reasoning → more incoherent errors across GPQA, MMLU, SWE-Bench, safety evaluations, and synthetic optimization; (2) inconsistent scale–incoherence relationship (synthetic tasks: incoherence rises with size; benchmarks: easy tasks become more coherent, hard tasks more incoherent or flat); (3) natural overthinking spikes incoherence more than deliberate reasoning-budget increases modestly reduce it; (4) ensembling reduces variance. Synthetic-optimizer experiment: train transformers to emulate steepest descent; larger models reduce bias faster than variance — "knowing what to do" outpaces "reliably doing it." Authors frame LLMs as "natively dynamical systems, not optimizers" — making them coherent optimizers requires training, and that training doesn't automatically scale with model size. Implication for safety priorities: future failures may look like industrial accidents rather than coherent pursuit of misaligned goals, raising the relative importance of reward-hacking and goal-misspecification research over constraining a perfect optimizer. Affiliations: Anthropic + Anthropic Fellows Program + EPFL (Hägele) + U. Edinburgh (Gema) + Constellation (Sleight). Builds on Sohl-Dickstein's 2023 hot mess theory of misalignment survey post. Companion Anthropic blog: alignment.anthropic.com/2026/hot-mess-of-ai/. Code: github.com/haeggee/hot-mess-of-ai. ICLR 2026 accepted; v1 2026-01-30, v2 2026-04-10.