Programming

I seriously feel like I’m going crazy since it seems to me like I’m the only one who isn’t satisfied with the “capabilities” of LLMs.

This is an unsurprising impression to form. But it's false.

You are seeing a lot of people boosting the LLM's capabilities, for many reasons.

Some of them are sincere, to a degree: they feel like they are doing more, because they're churning out code so much faster! But they're discounting or ignoring the downsides: They have far less comprehension of that code, they can't possibly review it effectively, they don't perceive the lost time trying to work the extruded output into something they can use; etc.

Some of them are insincere, because they're grifting: hoping to hype others into committing more to this technology and to their apparent experience, for clout or money. Those people can die in a hole.

A lot of them are insincere, because they perceive it will hurt their short-term career to not spout enthusiastically about using LLMs for coding. Their bosses demand it be used, measure its use and rank employees based on token burn, and job adverts overwhelmingly seem to require use of LLMs. People pretend to love this stuff just to stay employed.

Don't believe it just because people say it. Believe your own experience: you have seen how godawful is the code extruded from these slop machines. You have seen how voluminous and ignorant of context it all is, making it worse than useless to try integrating into an existing project. You have seen how the LLM can't effectively maintain its own code medium-, let along long-term. It's all and a colossal waste of effort and resources, a dead end.

Let people use the LLM for their own code if it makes them happy. But don't accept it into your project if you don't want it; and definitely don't trust their self-reported productivity unless they bring the code, and you independently assess it. You'll find very few claimed successes survive that simple scrutiny.