The blurry lines

Textbooks draw three neat boxes. Real systems scribble all over them. The most important models today deliberately mix the paradigms — and the clearest example is the chatbot you've probably used this week. ChatGPT is built with all three, in sequence.

Scroll through how it's trained.

The useful in-between paradigms

The three-box picture has famous hybrids living on the borders:

• Self-supervised learning — technically supervised, but the labels are manufactured from the data itself: hide the next word (or a patch of an image) and predict it. No humans label anything, yet you get billions of "answers." This is how LLMs and modern embeddings are pretrained, and it's arguably the most important idea in modern AI.
• Semi-supervised learning — a few labels plus a lot of unlabeled data, so you get most of the benefit without labeling everything.
• RLHF (reinforcement learning from human feedback) — turns fuzzy human preferences ("this answer is better") into a reward signal, aligning a model with what people actually want.

Paradigms are tools, not tribes

So don't think of supervised, unsupervised and reinforcement learning as rival camps. They're tools in a kit, chosen and combined to fit the problem — often several in the same system. Knowing what each one is lets you read any AI system and spot which tools it's using where.

To finish, let's ground all of this: where each paradigm actually shows up in the world.