That said, once the project goes beyond a certain threshold, LLMs offer little more than a highly capable autocomplete. In larger codebases, the current context window is too small for the tools to even answer questions properly, let alone make any non-trivial changes.

Productivity starts going down once you try to use LLMs in a large codebase and expect them to be as helpful as they were in a smaller one. Often, these authoritative little shits will whirl you around in a doom loop and still won’t find any useful solution. And suddenly you find yourself having to clean up the mess.

I’d still say the productivity benefit is positive, but at the same time, the hype around these is bonkers. Employers are holding onto the hype cycle to bring down wages through FUD.

Exactly this. I've really tried to find use for LLMs in my big tech company SWE job and I just can't. The context is just too large, and not just the code context. In the time that I can "explain" everything to the LLM, keep iterating until it spits out something semi-useful and massage that into something I can merge, I'd rather just do the whole thing myself.

But it's amazing for greenfield personal projects.

I find LLM autocomplete extremely annoying compared to traditional intellisense

It is wrong way more and I don't want multi-line autocomplete, it's too intrusive

It's more or less the same as intellisense most of the time, but occasionally it tries to guess an entire multi-line function and it throws me way off

I don't know if it's a matter of just sticking with it to learn like any new tool, or if it is just really not as useful as people say

One thing I've noticed is that with traditional intellisense it was often fast enough to get ahead of me so I could tab complete

Cursor is slower than me. Often I am typing faster than it can think, which makes it suggest things I'm already past.

Llms are still a big speed boost there

Significantly so! But the thrill of the gamble keeps me coming back, sadly.