We're certainly seeing a lot of speedup for certain kinds of products, and I expect to see far more. But some of these speedups are dramatically overstated because people are measuring only the time to produce the prototype or the cool demo, not actually the time to get through the full build/measure/learn feedback loop. In fact, if you think about it, I think it's pretty clear that it's the learn step of the feedback loop that has always been and will always be the bottleneck.

In terms of good uses for AI to figure out what people want, I really think the best advice is to focus on your own skills and agency. In other words, never ask the AI to make an artifact for you. Instead, ask it to teach you how to make the same artifact.

If you do this over and over again, if you do it step-by-step, having the AI as a teaching companion, critiquing mistakes that you make, breaking the problem down into smaller sub-problems, making sure you've mastered each step before moving on to the next, pretty soon you'll be able to create some really awesome artifacts.

Is this approach, in some sense, "slower" compared to vibe coding everything? Yes. But I think, at least in my experience, this is a go slow to go fast kind of situation. Now I had the benefit of being able to use a completely custom AI rig made by the company I help co-found, Answer.ai (with Jeremy Howard). If you want to learn more about that, you can access it at solve.it.com.