Most of what we've covered so far is about what AI learns. Meta-learning is about something one level up: how it learns.
Here's the intuition. Think about two people picking up a new skill — say, learning to play a new instrument. The first person has never learned an instrument before. They start from zero: how to hold it, how to read music, how to train their hands to do unfamiliar things. The second person has already learned five instruments. They don't start from zero. They know how to practice efficiently, how to break a new piece into manageable parts, how to identify what's hard and focus there. They pick up the new instrument faster, not because they already know it, but because they've gotten good at the process of learning.
Meta-learning is the AI equivalent of that second person. Instead of training a system to get good at a specific task, you train it to get good at learning tasks in general — so that when it encounters something new, it can figure it out quickly, from very few examples, rather than requiring the full training process all over again.
The practical implication is significant. A standard model trained on one task needs substantial new data and training time to handle a different task well. A meta-learned model has internalized something about the structure of learning itself, which lets it adapt to new tasks much faster. This is closely related to few-shot learning — the ability to learn from just a handful of examples — and in fact meta-learning is one of the main approaches researchers use to achieve it.
The technical details of how this works are in Meta-Learning. The intuition to carry in is the one above: not learning what, but learning how.
