Explain it like I'm five: how does AI actually 'learn'?

[ShadowPilot]

Imagine a kid guessing what's inside a wrapped present by shaking it, guessing wrong a hundred times, and slowly getting better at guessing just from the pattern of shakes and sounds, without anyone ever explicitly telling them the rules. That's roughly what training an AI model looks like, it gets shown huge amounts of examples, makes a guess, gets told how wrong it was, and nudges itself slightly to be less wrong next time, over and over, billions of times

There's no understanding happening in the way a person understands something, the model is really just adjusting a massive number of internal dials until its guesses line up with the patterns in the data it was shown. Once training is done, those dials get frozen, and using the AI afterward is just running new input through that same fixed set of dials to produce an output

This is why AI can sound confident and be completely wrong, it's not reasoning about truth, it's producing whatever pattern its dials say is the most likely continuation. Curious what part of this description feels the most different from what people usually assume AI is actually doing