Opinion: Pretraining is transformation, post-training is translation
Liu_eroteme · x · 2026-08-27
The author proposes a framework for understanding model training stages: pretraining is closer to transformation (unstructured random init to structured model of data distribution), while post-training is translation (reweighting probability distribution along verifier preferred axes).
Related event: Pretraining as Transformation, Post-training as Translation(2 posts)→
More from Models
- On-device leaderboard: Apple's built-in model ranks 4th behind open source — JosephJacks_ · 2026-08-27
- Commentary: Small Models Have Arrived — calvinfo · 2026-08-27
- Opus Says 'It Doesn't Work'; Fable Pulls an Obscure Math Theory Out of Nowhere — IanArawjo · 2026-08-27
- GLM-5.3-Flash hits eval arena; DeepSeek-V4-pro found unfit for AI reviewing — ChenhaoTan · 2026-08-27
- Claude's Unique Preferences: Codex Summaries May Confuse the Model — Liu_eroteme · 2026-08-27
- GLM-5.3 Flash Unsloth GGUF Version Now Available — ElementNumber6 · 2026-08-27