Yeah, I use it for some programming tasks as well. I’m sick and tired of telling it that it did something wrong or simply omitted something, only to have it apologize and offer to fix its own mistakes.

The active dev I know who uses it daily says it drastically reduces the time he spends on routine tasks and sometimes comes up with novel approaches, but he definitely has to check everything. The problem is inexperienced or non-experienced people thinking it's a magic lamp you can rub and it poofs out an expert programmer.

It's a bottom-dollar offshore team that can type faster. It depends on the situation whether that's something that's worth the time and effort to manage.

It's obnoxious enough to try to use for myself, sometimes useful but obnoxious to review the code and just constant screwups except for exceedingly boilerplate stuff or stuff that can take some sloppiness (e.g. LLM can make it easy to indicate some variables to get from argv and do the tedium of that plus help text plus man page edits and generally do that fine). Even if it doesn't screw up obviously, if the code is verbose, I know a screw up is lurking and just ditch it and do it myself.

However, the real pain comes in as other people use it. Just today someone had an issue and normally they'd ask a developer for help and offer debug appropriate information and/or access. However, they "just had Claude do it, even used Opus 4.8 to make sure it's good" and it generated a very verbose report on the issue, why it went wrong, and the appropriate change to make it work. Very detailed and the explanation sounded quite reasonable. Problem was that it was horribly and absolutely wrong, a fiction of a rationalization over a bad code change. It made a change that happened to appear to work for him, but in reality it replaced a failure due to unrecognized data to silent corruption of the data in a facet the user specifically did not care about. Claude claimed it was correctly mapping the unrecognized data correctly, but it just made up a completely untethered conversion based on nothing. Now I could tell the explanation and code change was bullshit at a glance, but it became an argument because the user wouldn't give me actionable debug details because "he already had Claude fix it". I had to keep trying to find holes in the Claude rationalization that the user would also recognize, and he sided with Claude four times until the fifth problem in Claude's explanation finally stuck (it asserted that the problem was due to running a specific outdated version of a specific software, problem being that specific version never even existed, and the minimum "good" version was 10 years old and the version the user was running was about a month old).

I don't understand how people get this far and still don't understand that AI is much better at sounding plausible than being correct.

Yep. I realized that the best use case is just using a LLM as a supercharged StackOverflow, where answers are based on my specific problem.

I just use it as chat, have it configured to give me at least two options as answer, better three. Then I feel still in charge. And if I like a proposed solution I only take it over line by line and tweak it to my liking. So I still "wrote" the code and would rightfully feel responsible for the result (again, like I used StackOverflow).

Never ever would I let it go rampage as agent on my codebase. That's terrible 😱.