Ask a large language model what 347 × 89 is and it will often get it wrong. Ask it to reason through a complex legal argument and it does surprisingly well. The mismatch points to something fundamental about how transformers work.

Transformers learn by finding statistical patterns across vast amounts of text. Arithmetic doesn't work that way — the answer to 347 × 89 doesn't recur in training data the way grammatical structures or argument forms do. Multiplication also requires tracking intermediate state: carrying digits, holding partial products. Transformers have no reliable built-in working memory for that.

The fix isn't more training — it's a tool call. AI assistants that handle math correctly are typically routing arithmetic to a calculator or code interpreter. The transformer handles the language around the problem; a deterministic program handles the numbers. What looks like AI doing math is usually AI knowing when not to do math.