LLM Essentials
A track of P10 · Generative AI.
From the Transformer to ChatGPT: pretraining, fine-tuning, RLHF and prompting — what today's LLMs are and how they were made.
Take two copies of the same model — same architecture, same size, same training text. Type an insult at the first and it may cheerfully continue the insult, because continuing text is all it knows. The second apologises, declines, and offers to help with something else. Nothing about the machinery differs; the second copy simply went through extra phases of training. What happens in those phases — how a text-continuation engine becomes something that behaves like an assistant — is the story this track tells.
The architecture track ends with an encoder block; this one explains how that blueprint became something you can talk to. The recipe has stages: pretrain on vast amounts of text until the model can continue any sentence, fine-tune it on examples of instructions being followed, align it with human preferences through RLHF — the stage that separates the two copies above — then steer the finished model at run time with prompts and context. Each stage answers a failure of the one before, which is why the order is fixed and why skipping a stage produces the strange behaviour it does.
Along the way the track faces the questions the recipe raises in practice: why models sound confident when they are wrong, what grounding and retrieval do about it, and when a better prompt beats a fine-tune — the practical decision anyone deploying these systems meets in their first week.
By the end, "ChatGPT" reads as an engineering pipeline rather than a mystery: you can name which stage produced any given behaviour — the fluency, the obedience, the refusals, and the failure modes too.