Dev finds bugs in RNN training env, splits runs into optimizer, curriculum and Funsearch-style tasks
cephaloform · x · 2026-11-12
A developer who planned to keep a training run going for a few more days stopped early after discovering issues in the RNN training environment. He is splitting the work to be safe, covering optimizer engineering, curriculum ordering, and aux loss design, plus more Funsearch-style tasks like bin packing, MaxCut, and sorting networks — the ones that weren't broken "worked sorta."
More from Research
- Amateurs with Claude + Codex claim discovery of a new dwarf planet in public telescope data — willdepue · 2026-11-12
- Paper applies statistical mechanics to 16 literary works, finding word frequencies follow Rayleigh-Jeans thermalization — burny_tech · 2026-11-12
- Embodied First Exam: 1,156 tasks on 47 robots test coding agents, best agent solves only 34% in standard setting — burny_tech · 2026-11-12
- LeapQuant: training-free quantization makes linear attention up to 4.25x faster — tianyin_xu · 2026-11-12
- K-means for supervised learning: self-supervised learning as a special case — caglar_ee · 2026-11-12
- SPoT boosts calibration: AUROC 0.61 to 0.84 while learning to abstain — polkirichenko · 2026-11-12