Learn / How LLMs Actually Work
Training, generation & fine-tuning
Assemble a GPT, learn the loss, sample text, and see how assistants are aligned. A class in How LLMs Actually Work. Free to start, no card.
Curated lessons in this class
Stack blocks into a GPT (forward pass)
Assemble the full model and read out next-token scores.
- Embed and add position
- Run through the blocks
- Final norm and unembed to logits
The training objective
Cross-entropy, and watching a loss actually go down.
- Logits to probabilities
- Cross-entropy for one token
- Average over a batch
- Watch a loss go down
Sampling: how text comes out
Greedy, temperature, top-k and top-p decoding.
- Greedy: always the top token
- Temperature
- Top-k
- Top-p (nucleus)
From base model to assistant
How next-token predictors become ChatGPT: SFT and RLHF.
- The last mile: alignment
How CodeTrain teaches this
No videos and no multiple-choice quiz at the end. Each lesson is a handful of small steps in a real editor. You type the code yourself, it runs, and the tutor reviews what you actually wrote. It will not hand you the answer, and the step doesn't advance until your code passes. What you get at the end is a record of which steps you wrote unaided, not a completion percentage.
These curated courses are the starting point, not the product
Everything in this catalog is a curated course we wrote by hand. They exist so there is something to start on. What CodeTrain actually does is build a lesson on demand: name any topic you want to understand and it writes one for you on the spot, or paste any public repository and it teaches you that codebase from the code that is really in it. You are not limited to the list below, and most people stop using it within a session or two.