Why winning training recipes at proxy scale can lose at target scale
gerardsans · x · 2026-09-10
A reply to BlackHC's educational thread on Bayesian model selection and scaling laws: three criteria for "best", when they disagree under scaling, and what that means for evals. The replier argues current recipe-scaling has blindspots — it ignores token isomorphism structure, treats curation and loss as generic ML problems, and that transformers learn co-occurrence geometry under repetition. Without distribution support in the frozen pretraining landscape, scaffolding and prettier losses can't fill interpolation voids; loss only scores next-token fit on the training measure.
More from Research
- EdgeRock open-sourced: 72 MiB KV cache at 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- TRL adds Score Centering, a simple fix for RLHF training instability — lvwerra · 2026-10-27
- Latent-Foresight: world models that learn to predict their own future — abursuc · 2026-10-27
- UK AISI papers: no good safeguards yet against misuse of biological AI tools — davidmanheim · 2026-10-27
- CMU's XYBench: LLMs answer queries literally and reinforce user misconceptions — valentina__py · 2026-10-27