The free lunch is mostly over. Single core performance is not increasing by much. And computers nowadays are so fast that all applications should be really snappy. There is a "Better Software Conference" that talks about this kind of software. I'm totally in that camp. For cross platform development there is currently nothing better (to my knowledge) than Qt. I have yet to see an imGUI application that respects my display scaling (everything is really tiny). On the other hand, Qt is quite dated in its design. I guess we need to wait for reflection in future standards to see some changes (unfortunately, C++26 has been delayed). Still, there is nothing better available. Electron is never a good alternative for our use case because the overhead is too big.

We are still using QWidgets. But that is mostly for legacy reasons. It seems to me that QML is also quite fast today and thus a viable alternative. If you have a dedicated designer for your software, you should probably go with QML.

At our company we don't use any coding agents (but may ask dedicated questions to AI without revealing the actual code). Even though you can select not to make your code available for training, there is no guarantee that it won't be used. The risk is then that our only competitor might accidentally "steal" our code through use of AI. Even local models now seem to talk to the cloud, so that even in this case your code is not safe. And local models might still be too small to be really useful (and appropriate hardware is expensive). With the advent of AI reverse engineering I am not sure if there is any advantage of not (possibly) sharing your code with AI, if it can just extract the logic from the final executable. The question is also if you can't just move a lot faster that potential losses because of AI stealing your work for a competitor is offset.

What should you do today? Everyone is now able to write crappy clones of existing software. Have a look at PhotoCraft: I guess it is quite easy to get the first 80% working. But, it will be really hard to get the last 20% working correctly and fast. One of the biggest problems (currently) is that AI will not respect the DRY principle. Bugs will be multiplied easily, but only be fixed in a single place. Also, with the goal to get "correct" software, AI sometimes likes to cheat by adapting the test cases to get all tests to pass (instead of fixing the code).

With PhotoCraft and the likes this means that AI written software needs to be provided free of charge (that's the user expectation). Because otherwise somebody else will make an AI clone free of charge. This is not a sustainable buisiness model. I also expect that AI will only be able to basically write clones (or remix existing software). It cannot invent something new. So, if you want to earn money, you have to come up with a new idea and develop entirely new algorithms for your product. You can stand out from the crowd if your software is always snappy and responsive, uses a lot less than half a GB of disk space, and uses a lot less RAM. You might have to hand code these kinds of applications. You might know the AI 'jokes' of "and make no mistakes" and I forgot to say "make it safe" and "no bugs". What do you expect AI to do? It was trained on all available source code which might mean that it represents the average programmer. You should strive to become an above-average programmer which means you cannot use AI to write above-average code. To me, that is the only selling point (as in: it's the only way to actually make money) if your software is much better than AI-written software. Because users will expect AI software to be basically free (they could've done it themselves).

Legacy software will always be treated differently and make different choices than an entirely new software project.