NavList:

John Clements, you wrote:
"It might be interesting to try using a prompt that asks copilot to "show its work", by describing the intermediate steps at each step of the algorithm. It's possible that this will suddenly start generating the right answer, or at a minimum give you some insight into the issue."

A risk with asking this one single chatbot (MS Copilot) to explain its steps and fix its error is that, rather than going back to square one in its analysis (and "trashing" its original response as a human might do), the chatbot may invent overly creative ways around the problem, building a little fortress on the swamp of its original answer. They do that, too... :) Another solution is to try two reasonably independent bots, separately, on the same problem... Like in this case: You get some simple code from Copilot which you find suspect. So pick it up, carry it over (copy and paste!) to Google Gemini, and ask that chatbot if it can see any issues with the code. It's a bit like asking two chronometers to serve as checks on each other. Extending that analogy, we have no idea what's going on with all the "gears" inside the "heads" of thse AI chatbots any more than a navigator two centuries ago could understand the "gears" inside a chronometer if he tried to take one apart. But we don't need to if we can compare the work of two or more.

Frank Reed