[csomar]

> It looks fine to the outsiders, it might even function correctly MOST of the time.

Most code (even written by good programmers) rarely functions correctly most of the time. Most code is broken. This is not the problem with AI. Unless I am using the tools wrong, LLMs can generate fully functioning scripts (and some of them are good) but they break after the 50k token context and start doing insane things that not even juniors will do (like randomly removing code).

If you want to see a shit-show, go to Bolt Discord channel. Some users are able to get a very simple and rough kinda single script app running. Everything else breaks once they start making simple amendments. This is not fixed by Claude 3.7 or O1 Pro or whatever. This is a fundamental issue in all of the LLM and a local maxima of the current tech.

Not that the current tech is not amazing. It is and there is a lot of value to be extracted from it. But everyone and his investor think they are about to reach nirvana and want to replace everything with "AI" where "AI" is a 100k context LLM.