15 exhaustive articles building a large language model up from first principles. From next-token prediction to a working GPT you code yourself, no prior ML background assumed.
0 of 15 read
Start anytime
Next-token prediction is the whole game.
Models don't see words. They see numbers.
Turning tokens into points in space.
Neurons, weights, and matrix multiplication.
How a model decides what to focus on.
Attention plus feedforward plus residuals, assembled.
Attention alone can't tell word order.
How a model actually learns from being wrong.
What a training run actually is.
Why raw pretrained models don't chat.
Why the same prompt gives different answers.
Why longer conversations cost more.
Perplexity, benchmarks, and why eval is genuinely hard.
Why bigger models just work better.
Every prior article, tied into working code.