rw-book-cover

Metadata

Highlights

Foundational models and query writing are fundamentally mismatched. The former is probabilistic, creative, and inductive—it’s best used to generate ideas from short prompts. But analysts have to do the opposite when answering questions. They need to be precise, rigorous, and deductive—they need to know all of the nuanced laws governing how data is used, and apply them to questions that don’t specify those details. (View Highlight)

o far, one of the most striking things about Foundational models is that they’re much better at the creative parts of analysis than they are at the mechanical parts. 8 Ask ChatGPT to write a SQL query against an artificially simple schema; it’s a junior analyst, at best. But ask it to come up with possible hypotheses to explain why there’s some anomaly in a metric, and it does better than I would. (View Highlight)

In other words, it inverted the workflow between human and computer. It was better at coming up with ideas than I was; I was better at doing the work. I was its assistant, its agent, its copilot. (View Highlight)