This is a deliberately tiny educational model — — parameters. Real LLMs use billions of parameters and far more sophisticated architectures. ·
The mechanism you just watched — forward pass, loss, backpropagation, gradient descent — is the same one. The difference is scale, architecture and data, not the underlying idea. These figures are a general illustration of the gap, not a description of any particular commercial model.
ChatGPT and Claude start out exactly like the model you just trained. Three stages turn a next-word guesser into an assistant.
1 · READ
What you just did, on a vast amount of text.
YOUR MODEL, RIGHT NOW
ASK A QUESTION AT THIS STAGE
What is the capital of France?
What is the capital of Spain? What is the capital of Italy?
It carries on the text the way text usually goes, and online a question is often followed by more questions.
2 · PRACTISE
The same next-word training, on conversations people wrote to show how a good assistant replies.
ONE PRACTICE QUESTION
User: What is the capital of France?
Assistant: The capital is ? → Paris
Nothing new inside. It still guesses the next word, but the text it learns to continue is now a helpful reply.
3 · GET RATED
People are shown two replies and pick the better one.
WHICH REPLY IS BETTER?
✓ The capital of France is Paris.
✕ Why do you want to know?
The dials are nudged to make replies like the picked one more likely. The same nudging you watched, aimed at helpful instead of next word.
All three stages run the loop you watched: guess, measure how wrong, nudge the dials.
The example text is invented to show the idea. Researchers call the three stages pre-training, fine-tuning, and reinforcement learning from human feedback (RLHF). Real assistants also get further training for safety, and differ in the details.