🧠Machine Learning (ML)

Machine learning is programming with examples instead of instructions. You choose a model: a function with many adjustable internal numbers, called parameters. You choose a loss: a formula scoring how wrong the model's outputs are. Then training adjusts the parameters to shrink the loss on your examples, and inference runs the trained function on new inputs. Every ML system you have ever heard of, from a loan scorer to GPT, fits this sentence.

Features, labels, parameters

The input variables are features: pixel brightnesses, words, transaction amounts. The desired outputs, when you have them, are labels: this photo is a cat, this email is spam. The parameters are what learning changes, and their count is the number you see in headlines. The digit reader you can train on this site has exactly 13,002 parameters; frontier language models have hundreds of billions. Same idea, different magnitude.

Training vs inference: the two lives of a model

Training is expensive and happens once (or occasionally): it needs the data, the loss, and lots of compute. Inference is cheap and happens constantly: the parameters are frozen and the function just runs. When you chat with an LLM you are doing inference; nothing you type changes its weights. What feels like memory across a conversation is engineering around a frozen model, not learning.