I just spent multiple evenings talking through a small data model. With the massive number of misses the model had, I can’t imagine how bad it would have been if I had let it yolo the first design of its choice, to build on.

The requirements didn’t really change, opus just couldn’t see all the real world potential issues with its various suggestions.

But it was helpful for me to talk it out. So I guess it fell in the (very) slow category of productive use.

The rubber ducking potential is the first thing that got me interested in AI as a useful tool.

You do have to be very careful of when the AI might be confidently wrong, or leading you down some kind of weird narrow path.

When searching the web to gather info on a topic, all have a tendency to find 2-4 search results and fixate on those as if they represented the entire scope of the topic. No surprise about the fixation because that's kind of how LLMs work (X, Y, and Z are in context -> more tokens about X, Y, and Z). But you have to be aware of it, and I often have to prompt it to do another search and keep turning up more options.

I've found that my requirements and acceptance criteria are much less clear than I think they are. AI reveals when you are being too vague, almost instantly sometimes.