General Discussion

They are both doing this a LOT less lately with Sonnet 4.6 and Gemini 3.1.

Part of my point here is they're getting better at figuring stuff out, i.e. not repeatedly giving wrong answers, promising to do better next time, then repeating the same mistake again and again. I know exactly what you're talking about with that. Less than a year ago I was still seeing it a lot, esp. on cheaper/free models. But they are moving forward from that, very rapidly with the latest, paid models.

And I'm sorry, but the question I posed required the model to do something that is not demonstrably different from "thinking". It couldn't have just derived that answer from the mess of words I posted (some of it was even redundant and unclear once I read it over again, but it was not misled by that) without understanding what I was asking. These things are thinking more and more like humans with every passing iteration. You can believe me or not, but I'm telling you, I use the shit every day. Yes, it still makes mistakes, but overall, it's getting freaking smart as hell. It's scary.