METR's Time Horizon 2.0 report ships too few appendices, researchers complain — leaked topic list
dfrsrchtwts · x · 2026-11-21
Researcher dfrsrchtwts complained that METR included disappointingly few appendices in its Time Horizon 2.0 report. A screenshot of the appendix table of contents reveals heavyweight topics: difficulty estimation and baselining, time horizon diagnostics, MCI modeling methodology, limitations of IRT, robustness checks, data integrity and reward hacking, changes in capability growth rates, inference scaling, swarm experiments, Budget NanoGPT, continual learning and more — hinting at the analytical depth behind this AI capability evaluation report.
More from Models
- DeepSeek-V4.1-Flash takes 90 minutes and half a credit balance on same coding task — HeyToha · 2026-11-21
- GPT-6.2 Astra scores 41.2% on robotics benchmark Ember-0, double September's mark — gdb · 2026-11-21
- Head-to-head coding test: Tencent Hy4 Flash finishes in 7:38 vs DeepSeek's 90 minutes — HeyToha · 2026-11-21
- Inkling-2-Mini builds a Sims-style town with 20 AI villagers running offline on a MacBook — simonguozirui · 2026-11-21
- Inkling 2 Mini ties Claude Opus 5.5 on CyberBench at less than a quarter of the cost — LiTianleli · 2026-11-21
- llama.cpp adds llama serve: run Inkling 2 Mini GGUF locally with one command — simonguozirui · 2026-11-21