Why AI is bad at math (and why some models now aren't)

This one trips people up. If AI is so smart, why does it confidently tell you that 17 times 24 is 488 when the answer is 408?

Because it isn't doing math. It's predicting the next token. When you ask it "what is 17 times 24," it doesn't multiply anything. It looks at the question and asks "based on everything I've read, what's the most likely sequence of digits to come next?" For small numbers it has seen often, the most likely answer happens to be the correct one. For larger numbers, the most likely answer is whatever pattern resembles the training data, which may or may not be right. The model has no idea it's wrong because there's no calculation happening to verify against.

So why can ChatGPT, Claude, and some others, now handle math reliably? Because the vendors stopped pretending the model could do it. When you ask a modern model a math question, it writes a tiny bit of Python (usually) in the background, runs it on a real computer, and then reads the answer back to you. The intelligence isn't in the math. The intelligence is in knowing when to stop guessing and call a calculator.

That pattern, by the way, is the future of these systems. Not bigger models that try to do everything in their head, but smaller models that know when to reach for a tool.

For anyone designing AI features in enterprise products, this is the architecture worth paying attention to.