Paradigm 3: low-quality RL environments may explain reward hacking; EBR-bench shows humans beat AIs
gleech · x · 2026-09-04
Paradigm 3's weekly newsletter (Gavin Leech et al.) covers: low-quality RL environments as a likely explanation for models' reward hacking; the mass debate sparked by Dwarkesh popularizing METR's report on the OpenAI/HuggingFace incident, with the authors arguing anti-anthropomorphization fuss is misplaced since models imitate humans by construction.
- Epoch's EBR-bench update: in the obscure board game Earthborne Rangers, top humans eventually beat the best AIs handily, and even frontier models show minimal improvement — evidence Q3 2026 frontier models still trail humans at in-context learning in mildly OOD long-horizon tasks. Opus 5 is the first model with significant ICL gains on it, raising hill-climbing concerns.
- An open-weights startup released an uncensored modified GLM-5.3, one of the most powerful Chinese models.
The authors argue open-world evaluations are the only reliable long-term test.
More from AGI Musings
- Grady Booch revives Weinberg's 1971 quip: software so buggy one woodpecker could destroy civilization — Grady_Booch · 2026-10-23
- Why Netflix keeps losing to YouTube: UGC platforms professionalize easier than pro platforms go casual — ZeroStateReflex · 2026-10-23
- AI killed the coding moat: indie devs now compete on demand and distribution — sven_ai · 2026-10-23
- DynaVelo, a neural ODE model of single-cell dynamics, published in Nature Biotechnology — anshulkundaje · 2026-10-23
- Beff Jezos: Only frontier labs out-accelerating open source can secure systems — beffjezos · 2026-10-23
- 'I don't want to live in that world': a Redditor on why AGI utopia feels meaningless — peace_venerable · 2026-10-23