Letting it run code and count

The same AI that writes a flawless essay will cheerfully tell you 9.11 is bigger than 9.9. How can something so capable be so bad at a sum a child can do?

It's guessing the answer, not calculating it

Here's the surprise: by default, the AI doesn't calculate at all. It predicts its answer the same way it predicts words — a quick, intuitive next-guess. For "9.11 or 9.9?", the ".11 looks longer than .9" pattern is strong, so it blurts the wrong one before any real working. Force it to slow down and lay the steps out, though — line the numbers up, compare digit by digit — and it gets there. Reasoning takes room to happen, and a blurted one-shot answer gives it none.

Or just let it use a calculator

There's an even more reliable fix, and it's the same trick from the last lesson: a tool. When precision really matters, a good assistant doesn't do the sum in its head at all — it writes a tiny bit of code, actually runs it, and reads back the exact result. The computer does what computers are good at; the AI does the talking. For anything that must be exact — totals, dates, converting units, sorting a list — the surest path is to have it run the numbers, not recall them.

The gist

• —By default the AI guesses answers in one shot — fine for words, unreliable for exact maths
• —Made to work step by step, it has room to actually reason and often gets it right
• —For answers that must be exact, let it write and run code — a calculator beats a guess