Station 1
How does an LLM learn?
Use one student character to separate training from generation: broad reading, teacher guided examples, and practice with feedback.
Pretraining
The student reads a huge library and gradually discovers patterns in language.
Large amounts of text -> Predict missing or next tokens -> Learn language patternsStation 2
Words become tokens
Begin with ordinary text. A language model does not receive the sentence as a single human-like thought; it receives a sequence of token pieces.
Press Tokenize to break the sentence into model-sized pieces.
Station 3
Tokens enter the context
The context is the text currently available to the model while it predicts the next token. It is working information, not human memory.
Context
The text currently available to the model when generating its next token.
Station 4
A large neural network processes the context
The neural network idea from the previous lab now operates at enormous scale. Tokens flow through learned parameters and become probabilities.
Parameters
Values learned during training that encode patterns discovered from data.
The earlier learning station explains how pretraining, supervised fine tuning, and feedback shape these learned parameters.
Station 5
Predict the next token
This is the central LLM move: given the current context, the model estimates which token is most likely to come next.
Station 6
One token becomes many
A response appears because the selected token is added back into the context. Then the model predicts again. The loop repeats.
AI
Station 7
The surprise: fluent text can still be wrong
If the model is predicting likely text, does that guarantee the answer is true?
Hallucination
A plausible sounding but incorrect or unsupported response generated by an AI model.
"The solar-powered pencil was invented by Eleanor Marsh in 1882 after early experiments with miniature photovoltaic cells."
Factually unsupportedStation 8
The whole LLM mental model
Knowledge in parameters
Patterns learned during training. These may be useful, outdated, incomplete, or wrong.
Knowledge in context
Information currently provided to the model, such as your prompt, examples, or uploaded material.
Response
Learned patterns plus current context, expressed through repeated next-token prediction.