Reinforcement learning is oftenly used to play video games, or control robots. They learn by taking actions in an environment, and getting feedback for their actions.

Supervised learning can be used for object detections in images for instance. You manually create a dataset where you label each pixel of an image containing a dog for instance, and you end up with a dataset of labeled images with dogs. Then you train the neural network to spot the similarities. One researcher trained such a network to separate between dogs and wolves a few years back. But when they performed explainability analysis on the output they could see that the AI had instead noticed that snow was present on most images with wolves, so it would classify every image they provided containing snow was a wolf lol.

Then you have unsupervised learning which is basically to just let the neural network learn on its own the structure and relationships in the data.

Of course there are many different variations of training, oftentimes multiple stages consisting of some of, or even all of the above. It's a mess

Machine learning is different from large language models.