Researchers debate: does post-training stop a transformer from being a language model?
burny_tech · x · 2026-09-27
Responding to Melanie Mitchell's critique of LLMs, researcher francip argues that pretraining, SFT, preference optimization and RL all modify the same weights and the same inference process producing a conditional distribution over tokens. Post-training can radically change that distribution and behavior, but that doesn't make the result "not a language model" — just one optimized for a different objective. Part of the ongoing debate over whether modern LLMs are still next-token predictors.
Related event: Melanie Mitchell Sparks Debate: Today's Systems Are Not LLMs(31 posts)→
More from Models
- Running DeepSeek 4.1 locally on an $800 rig with a used RTX 3090 at 2-3 tok/s — LankyGuitar6528 · 2026-11-11
- Google open-sources Harness², an agent that rewrites its own setup per task — Arindam_1729 · 2026-11-11
- Dev dumps Codex desktop app over quality issues, falls back to codex-cli — tekbog · 2026-11-11
- Why do old models get dumber before launches? Lab party, says joking X post — glcst · 2026-11-11
- Grok Bot now detects hidden prompt injections, warns users and refuses to follow them — XFreeze · 2026-11-11
- Shanghai Institute Open-Sources Nex-N2.6, a 321B Multimodal Agentic Model — victormustar · 2026-11-11