
Data labeling is the set of right answers a model learns from. Who writes them, and why trained labelers agree only about three times out of four.

Overfitting is when a model memorised the examples instead of learning the rule. What overfitting is, why the score rises as usefulness drops, and how to spot it.

Synthetic data is examples that never existed: another model made them up. And they often work better than real ones. Here's why — and where the limit is.

A raw model from the internet is rude and useless — humans rating answers make it polite. What RLHF is in plain words, and why it makes AI a bit of a flatterer.

A neural network doesn't memorize examples or follow rules — it turns millions of knobs until it guesses right. Here's the plain-words version, no formulas.