AI that uses tools: the calculator
Be able to explain why an AI computes better with a calculator than in its head.
Prerequisites
- BAI's invented answersrequired
Everyday explanation
Ask a language model what 4 871 · 396 is. It often answers something that looks right but is wrong by a couple of digits.
Why? It does not compute — it guesses the next character from how numbers usually look. Rather like guessing the answer to a multiplication problem without working it out.
The solution: give the model a calculator. It writes «compute 4871·396», a real program computes it, and the answer is sent back. Now it is right every time.
Intuition
The same principle holds for everything the model cannot do itself:
| The question | The tool needed |
|---|---|
| 4 871 · 396 | a calculator |
| What is the time? | a clock |
| What is in our price list? | searching documents |
| What will the weather be tomorrow? | a weather API |
| What does this code give? | running the code |
When are tools not needed? When the answer is in the language itself: explaining a concept, rewriting a text, giving ideas, summarising something you have pasted in.
A good AI service knows the difference and uses tools automatically when they are needed. But you should know it too — then you understand why the answer is sometimes wrong.
Mastery means
- Explains why a model computes better with a calculator
- Gives examples of when tools are needed and when they are not
Sign in to do the exercises and build your mastery up.
Sources
- arXiv — Toolformer: Language Models Can Teach Themselves to Use Tools — arXiv (open access; licence per article)
- Skolverket — About AI in school (in Swedish) — Skolverket's open terms