From Text to Action

Everything so far has built up to a model that does one thing: read a stretch of text and write what plausibly comes next. Astonishing, but bounded. It can describe a good search, but not run one. It can write the code to rename a file, but not rename it. It talks; it does not touch.

This final module is about what happens when you let it touch.

Hand the model a few buttons it can press, search the web, run code, read a file, send a message, and the nature of the thing changes. Its output stops being only words and becomes steps. And steps can be chained: do one, look at what happened, decide the next. A job that used to need a person making decision after decision can be handed over whole.

That is an agent. Not a new kind of brain, the same next-word machine you already understand, but wired to the world through tools and run in a loop. It is a small change in plumbing and an enormous change in what becomes possible, including new ways to go wrong.