Training

Everything upstream - tokens, attention, the blocks themselves - is machinery. This track is what fills that machinery with weights, in three phases that happen one after another, not at once.

Getting a base model. The training data is curated, not scraped raw; backpropagation is the measure-and-nudge loop that does the actual learning; pretraining is that loop run at planetary scale. Scaling laws tell you how to spend a compute budget before you spend it, and GPT-2 is one concrete run that makes the whole recipe real.

Turning it into an assistant. A base model is an internet-document simulator, not something that converses - prompt it like a chat and it just role-plays the transcript. Conversations become tokens through a template; instruction tuning and RLHF reshape the raw predictor into something that follows instructions, at some cost to raw capability and variety.

Teaching it to reason. Reinforcement learning is different from imitating demonstrations: give the model checkable problems, score whole attempts, and long chains of thought emerge that no human example taught it directly.

The training data - raw web, funnelled into a curated set of clean text.
Backpropagation - measure how wrong the guess was, nudge every dial a little less wrong.
Pretraining - backprop at planetary scale; where grammar, facts and geometry get baked in.
Scaling laws - loss falls predictably enough to plan billion-dollar runs around.
GPT-2: the recipe - a 2019-scale run made real; the recipe barely changed, the hardware and data did.
Base models - what pretraining actually produces: a document simulator, not an assistant.
Conversations become tokens - special tokens and a template flatten dialogue into one stream.
Instruction tuning & RLHF - reshaping a predictor around what people actually want from it.
RL & reasoning models - checkable problems, scored rollouts, and reasoning that emerges unsupervised.

Next: Running the model & the world, where a finished model meets deployment - and its own limits.