'If Your Pass Rate Is 0, You Can't RL': Why Distillation Still Matters
fleetwood___ · x · 2026-09-25
A debate over RL vs distillation: one side argues SFT as the primary capability driver was a 2023-24 story, and in 2026 it's all RL environments — plus RL is a way to offload compute, especially using fraudulent subscription accounts.
fleetwood counters: if your pass rate is 0, you can't RL at all — with no initial capability there's no reward signal to learn from, so distillation remains absolutely crucial.
Related event: Distillation remains essential when RL pass rate is zero(2 posts)→
More from Models
- Claude Opus 5.5 creates a 5-minute animated UMAP explainer from a phone prompt — goodside · 2026-09-25
- User has Claude write a history of economic thought while he slept, economists impressed — soumitrashukla9 · 2026-09-25
- Podcast: Opus 5.5 Beats Fable 5.1 on GDPval at 40% Lower Cost, Tops Code Arena — thursdai_pod · 2026-09-25
- Claude Opus 5.5 and GPT-6 Sol Shipped 101 Minutes Apart, Both Cheaper — thursdai_pod · 2026-09-25
- Embedding test: RAG ranks the no-refund policy first, showing models still matter — galratner · 2026-09-25
- Meme: doing your taxes with Mistral could cost you a $20k IRS visit — Rasmic · 2026-09-25