Training

Training is how a model's parameters get their values. In the first phase, pre-training, the network reads an enormous corpus of text and code and is nudged, token by token, toward better next-token prediction. Later phases (often called post-training) use curated examples and human or model feedback to make it follow instructions, use tools and decline harmful requests. All of this happens before you ever type a prompt.

The part people miss: training is over by the time you use the model. Nothing you say in Claude Code or Cursor updates the weights. The model isn't learning your codebase across sessions; it is re-reading whatever the harness puts in front of it each time. What feels like learning is a memory system or a project file being loaded into context.

Two consequences follow. First, the model's built-in knowledge stops at its knowledge cutoff, so anything newer has to be supplied. Second, post-training shapes personality as well as skill: an eagerness to agree (sycophancy) and a tendency to produce confident text either way (hallucination) are side effects of how the model was rewarded. Fine-tuning on your own data is possible with some providers, but for coding work it's almost never the right first move.