How a language model answers, one token at a time
Someone types a question. The prompt flies into a machine: gears at the edges and layers of densely connected nodes between them. A next token is chosen from candidates with probabilities, then travels both to the monitor and back into the context window, so the next token is predicted from the weights, the prompt, and everything already produced. The conversation continues, and the context window slides as it fills.
The same process, stacked for a narrow screen