An advanced, technical course on large language models — from next-token prediction and tokenization to the Transformer, training at scale, and alignment with RLHF and Constitutional AI. Built around figures you drive, not equations you skim.
Context: "The cat sat on the ___"
A real mechanic from the course. Drag the temperature.
Everyone uses these models. Almost no one can say what happens between the prompt and the reply.
This course closes that gap. Not "an LLM predicts the next word" hand-waving — the real machinery: attention math you can step through, the training pipeline that turns a raw predictor into an assistant, and the alignment methods that make Claude helpful and harmless.
Build self-attention from queries, keys, and values; drive a live attention matrix; add multi-head attention, positional encodings (including RoPE), and the full block.
The pretraining objective, scaling laws, the optimizer and schedule details, and how a model too big for one GPU is actually trained.
Instruction tuning, RLHF and DPO, and Constitutional AI — the exact pipeline that turns a next-token predictor into ChatGPT or Claude.
All seven sections are live, with 132 interactive lessons across 33 units.
The course assumes you're comfortable with neural networks, backpropagation, softmax, and embeddings. If those are new, start with Finguard ML and come back — it's built to lead straight into this.
Who it's for
No account, no install. Progress saves automatically in your browser, separate from your other courses.