# Pretraining, Post-Training and Fine-Tuning Three stages of turning compute and data into a useful model. They differ enormously in cost, and that cost gap is what makes the cross-Pacific loop possible. ## Pretraining The model reads a very large corpus (trillions of tokens) and learns to predict the next token. This is where general knowledge and raw capability come from, and where almost all the compute goes: tens of thousands of GPUs running for weeks or months. "American-scale pretraining" means this stage, done at the size the leading labs do it. See [[Frontier Models]]. ## Post-training Everything done to a pretrained base model to make it behave: follow instructions, reason step by step, use tools, refuse harmful requests. Main techniques: - **Supervised fine-tuning (SFT):** train on curated examples of good responses, which can come from humans or from a stronger model ([[API Distillation]]). - **Preference optimisation (RLHF, DPO):** train on comparisons of better and worse answers. - **Reinforcement learning with verifiable rewards (RLVR):** let the model attempt problems whose answers can be checked automatically (maths, code that passes tests) and reward correct attempts. The quality of the checker sets the ceiling, see [[Evaluation and Verification Infrastructure]]. Post-training is a fraction of pretraining cost but increasingly drives the capability people notice. ## Fine-tuning Post-training applied by someone downstream, on a narrow domain: legal drafting, a codebase, a company's support tickets. Can be full fine-tuning or [[Parameter-Efficient Fine-Tuning (PEFT)]]. This is what vertical AI companies do when they build on [[Open Weights]] and call the result their own model. > [!important] The asymmetry > Pretraining sets the frontier and is expensive. Post-training and fine-tuning copy and specialise it cheaply. A company that "post-trained its own model" on an open base has bought the expensive part for free. Related: [[Foundational Models MOC]], [[The Cross-Pacific AI Chessboard]]