AI Anatomy Class · Part 10 · The world after ChatGPT

89. AI with hands and feet

Q. How can an AI that only writes text punch a calculator or even search the web?

1. A model can only write text. It easily gets big products like 4837 × 2961 wrong, and can't know things like today's weather.
2. So we make a deal: when the model writes text in a set shape, like 'calculator(4837 × 2961)', it means 'please call this tool'.
3. The program around the model spots this text and stops the writing. Then it runs a real calculator and gets 14,322,357.
4. The result goes on the end of the text and back into the model, which reads it and writes on. Like chapter 2: attach and repeat.
5. Search works the same way: write 'search(…)' and the program attaches what it finds. The model calls chapter 88's RAG by itself.
6. Think, call a tool, read the result, think again. An AI that goes round this loop many times to finish a job is called an agent.
7. With hands and feet comes care. Things that are hard to undo, like sending email or paying, wait until a person checks.

Chapter summary

• When the model writes text in a set shape (a tool call), the program stops and runs that tool.
• The tool's result is attached to the end of the text and fed back in, and the model reads it and writes on.
• An AI that loops through think → tool → result many times is an agent. A person checks the important steps.