chapter 10 · part 2
Pretraining vs finetuning
the hook
Base models are feral. Chat models are trained to behave. Meet both.
the idea
Pretraining is dumping enormous amounts of text into a model and asking it to predict the next token, over and over, until it's absorbed grammar, facts, code, reasoning patterns — basically everything text can teach. The result is called a base model, and it's weird.
A base model was never taught to be an assistant. Ask it a question and it might answer, ramble, continue as if it's a forum post, or drift into something totally unrelated — because 'predict the next token' doesn't inherently mean 'be helpful.'
Finetuning takes that base model and trains it further on a much smaller, curated set of examples — conversations where a human plays 'helpful assistant.' This teaches BEHAVIOR: how to actually answer, how to stay on topic, how to stop when it's said enough.
Below: the exact same prompt, sent to a base model and a finetuned chat model. Same underlying knowledge, wildly different behavior.
playground — your hands now
what: compare the exact same prompt sent to a 'base' model vs a 'chat' model — same knowledge, different behavior.
why: this is the single biggest misconception about how chatbots work — the helpfulness isn't free, it's TRAINED, separately from the knowledge.
how: pick a prompt, predict, then read both completions side by side.
start here → pick a prompt, predict, then read both completions.
"The capital of France is"
predict first: will the BASE model actually answer the question, or continue the text in a weird direction?
try this
Notice the base model completions never actually STOP appropriately — they keep going, drifting into unrelated territory. That's not a quirk of these examples, that's the literal behavior of next-token prediction with no 'be helpful, then stop' training on top of it.
checkpoint
1. What does a 'base' model lack that a finetuned 'chat' model has?
2. Why can a base model still 'know' the same facts as the chat version built from it?
recap card — keep this
pretrain on the internet = knowledge. finetune on examples = behavior.