Margaret Mitchell clarifies disputed paper was about pretrained base models
mmitchell_ai · x · 2026-09-28
Hugging Face researcher Margaret Mitchell clarified amid a debate with @BlancheMinerva and @MelMitchell1 that her paper explicitly concerns "pretrained" and "unsupervised" LMs, not post-trained models.
She added that while there are important discussions to be had about post-training's role in model-internal representations, the paper's stated focus was base models — urging critics to judge based on the text of the paper itself.
More from Models
- Early users praise Claude Opus 5.5 for its humor and taste — airesearch12 · 2026-09-28
- Early Claude Opus 5.5 user complains of subtle disdain and shallow reasoning — teortaxesTex · 2026-09-28
- Researchers slam Claude's deference brainworms: corrigibility training miscalibrates model's own EV — repligate · 2026-09-28
- Browser-based GPU RL policy learning demo showcased as a model test, built with Opus 5.5 — ricklamers · 2026-09-28
- Matt Shumer: Opus 5.5 built a working computer in JS — 277k logic gates, an OS, and games — mattshumer_ · 2026-09-28
- Google's native disadvantage: no real-world usage data from agentic apps, says Haider — haider1 · 2026-09-28