- Training an LLM uses the same loop as any neural network: run examples, measure the error, adjust every weight a little. Only the answer key, the scale and the goal change.
- No human labels are needed at first: the text supplies its own answer key. Each wrong guess moves every parameter towards the right answer and away from what the model expected.
- A model is only as good, balanced and current as its reading list. Collecting and cleaning the data is most of the work, and it shapes strengths, gaps and biases.
- Training stops by observation, not proof. Reliability is measured on text set aside: mostly right, never guaranteed right.
- Assistants are shaped in later phases, with human judgments and automatic checks. That is where vendors set tone, helpfulness and safety.
- Everything is "multiply and add" on huge tables: identical, independent operations, exactly what GPUs are built for. Training is limited by computing power; answering by memory bandwidth.
- GPUs are expensive for structural reasons: specialised memory, one dominant supplier, demand above supply, power. Small models run fine on ordinary servers; large models for many users need GPUs.
- Using a model does not change it. Whether your data feeds future training is a contractual question.
- Don't rebuild: libraries and APIs cover every layer. Any real project should still start from a business case.