Augmented Intelligence

Overview

Go back to the dawn of reading and writing. Before reading and writing, the bounds on human knowledge were set by the human brain: the amount of information we could collect and use to solve problems was fundamentally limited by how much we could memorize and recall. Reading and writing externalized memory. We could suddenly use natural resources — clay tablets, scrolls, books — to capture information, and the amount of knowledge we could capture and apply to solve problems scaled far beyond any one brain.

Large language models do something similarly profound. Before the first frontier-level model — ChatGPT 3.5 — you could not ask a computer to think. After it, suddenly, you could. We have externalized thinking. It is not the same as human thought, but it is a computational approximation of thinking — one that externalizes thought and scales up the amount of thought we can apply to solve problems.

We start our exploration by taking a broad overview of this computational thought and how it can be used to solve problems.

Computational thought, up close

We begin with augmented intelligence and the idea of computational thought. The important shift is not simply that a computer can produce text. It is that we can now express goals, situations, judgment, and loosely structured problems through language and have a computer reason over them.

We will ask participants to look for capabilities that were fundamentally unavailable before LLMs. Sometimes the barrier was that traditional software could not adapt to the situation without every condition being anticipated in advance. Sometimes the barrier was that the necessary human expertise or labor could not be scaled. This distinction—between automating an existing procedure and making a previously impractical form of thought possible—will become an important test for every agent idea developed during the course.