1. Show the model a huge amount of data
The model is fed billions of examples — text from the public web, books, code. At first its predictions are random. The 'learning' is the process of nudging it toward better predictions.
2. Make a prediction, score it, adjust
The model takes an input (say, a partial sentence) and predicts the next piece. Its prediction is compared to the real next piece. The gap is a 'loss' — a number that says how wrong it was.
3. Backpropagation: blame goes backward
An algorithm called backpropagation figures out which internal numbers (the 'weights' or 'parameters') contributed most to the wrong answer, and by how much. It's calculus, fast.
4. Gradient descent: nudge the weights
Each weight is nudged a tiny amount in the direction that reduces the loss. Do this billions of times across billions of examples and the model gets shockingly good at predicting what comes next.
5. Instruction-tune it to be useful
A raw pre-trained model is a brilliant autocomplete, not an assistant. A second, smaller training round teaches it to follow instructions, then a third (RLHF) teaches it to prefer helpful, honest, harmless answers.
6. Freeze it and serve it (inference)
Once training is done, the weights are frozen. When you chat with the model, no learning is happening — it's just running those frozen weights forward to produce its best next-token guesses very fast.