#LLMs can't do arithmetic

LLMs can’t count. They’re built to handle words, not numbers.

Proof: a researcher tested ChatGPT (4o and o1-mini) on multiplications with bigger and bigger numbers. Past 10 digits by 10 digits, the answer is almost always wrong. The charts speak for themselves.

Which makes sense, because these models aren’t designed to calculate. They’re trained to predict the next word from the semantic context, not to carry out exact operations. For math, we’ve had very good software for decades.

It’s like blaming a hammer for being bad at driving screws: it’s just not the right tool.

In the end, that’s the whole point with AI: know its strengths and weaknesses so you use it where it belongs. That’s how you know when to trust it and, above all, when you need to take over so you don’t get caught out.

Source: Yuntian Deng on X