TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 16—18 · FACTUAL · EVOLVING
Pretraining and post-training
A model's knowledge comes mostly from pretraining and its behaviour mostly from post-training.
The idea
A language model is built in at least two stages. Pretraining comes first: the model reads an enormous amount of text and learns to predict what comes next. What comes out is a base model. It can continue almost anything, and it does not know you want an answer rather than more text. Post-training comes after: the builders show it good replies, then have people rank its answers and train it to prefer the higher-ranked ones. This is where it learns to answer a question, follow a format, hedge, and refuse. Knowledge is mostly pretraining. Behaviour is mostly post-training.
Why it matters
Three things students blame on "the AI" — that it invents facts, that it refuses, that it has a temperament — are each partly a post-training story. Once you know there are two stages, you know which stage to ask about.
See it in the studio
Two tools built on similar base models. One gives you a bare bullet list of options for a façade; the other asks what the client wants first and pushes back on a weak brief. The difference is not what they know. It is what they were trained to do with it, after pretraining.
Watch for this
Treating the chatbot's politeness as the model's nature. A base model is not polite or rude; it continues text. The helpful tone, the hedging and the apology were trained in later by people ranking answers — and people tend to rank agreeable answers higher. That is one place a flattering streak can come from.
Try it
Give a chatbot a weak design idea and ask for an honest assessment. Note how it softens the verdict. Now ask it to reply as a harsh jury member. The knowledge did not change; the instruction did. You are pushing against post-training with context — and you can feel how far it bends.
Prove it
Explain what a base model is and what post-training adds, and name one behaviour of a chatbot you use that most likely came from each stage.
How it works
Pretraining is self-supervised: hide the next token, predict it, adjust, across a corpus of billions of words. The result predicts text well and does nothing else. The most widely cited recipe for post-training has two steps. First, supervised fine-tuning on examples written by people — a question and a good answer. Second, reinforcement learning from human feedback, RLHF: people rank several model answers, a reward model learns those rankings, and the language model is trained to score well on it. The 2022 InstructGPT paper used exactly this recipe and found that a 1.3-billion-parameter model trained this way was preferred by people over a 175-billion-parameter base model. A 2022 method called Constitutional AI replaced some of the human ranking with a written list of principles and AI feedback. Post-training is also where refusals are trained in, which is why GUARDRAILS on this strand says a refusal can come from the weights themselves. And WHY IT MAKES THINGS UP depends on it: pretraining has no truth check, and post-training reduces invention without removing it. EVOLVING because the recipe keeps changing — newer models add reasoning training, tool-use training and other stages on top — and because what a given product did in post-training is rarely disclosed in full. The two-stage shape has held since 2022; the details of the second stage have not.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.