This reminded me of an interview I listened to by a startup founder talking about how his company integrates AI into all of its workflows. During the Q&A, he said that they could tackle any challenge simply by iteratively constructing better contexts for the AI. At first this sounded optimistic, but then it struck me that it was actually the ultimate pessimistic view of what current AI can do. His assumption seemed to be that software engineers have already implemented all the primitives humans will ever need. If that’s true, then the only task left is to phrase our instructions in the right way so the model can stitch those primitives together into a production system.