CONCEPT
Post-Training Primacy
The claim — increasingly hard to argue with — that post-training matters as much as pretraining for what a frontier model ends up being. The most visible trend of 2023–2025 frontier releases.
The post-training-primacy thesis is the contention that, for a given base model, the post-training pipeline determines more of the deployed assistant's character, capability profile, and safety properties than the pretraining run that produced the base model. Through 2021 the field treated pretraining as the dominant investment and post-training as a cheap finishing step. The pattern reversed quickly. Releases from OpenAI (o1 reasoning, GPT-4o, GPT-5), Anthropic (Claude 3 families, extended thinking), Google (Gemini's thinking), and DeepSeek have separated through their post-training strategies while sharing broadly comparable pretraining. The capability frontier is now substantially a post-training frontier.
In The You On AI Field Guide
Several developments pushed post-training into primacy. First, the base-model capability gap between frontier labs narrowed; all of them could produce a strong GPT-4-class base model by 2023–2024. Second, post-training techniques multiplied — RLHF, DPO, constitutional AI, reasoning-specific tuning, tool-use training — each opening capability-gain dimensions that were independent of base-model quality. Third, compute efficiency for post-training improved faster than for pretraining; a post-training
Keep reading with YOU ON AI
Unlock the full book, 10,000+ field-guide entries, and a 1000+ thinker library. If you have a book code, register now — it takes a minute.