← Back to context

Comment by elorant

2 days ago

I can only speak for myself. I mostly write C# and the code they generate isn’t of my liking. Too many var statements, async everywhere even for the most trivial things, and then a lot of times when I ask something difficult they fail spectacularly. I asked Claude the other day to propose an algorithm for Named Entity Recognition for brand names and all it did was to build a heuristics based on the type of business entities. And then in order to do that it went and created four classes. They overcomplicate things all the time for no apparent reason.

I built an Asterisk dialplan using Gemini a few months ago and it has so many hacks and quirks in it that I don’t dare touch it. I just leave it there hopping that it will keep working. I excluded Asterisk from future OS updates to avoid something fucking up that travesty of a dialplan.

Couple of weeks ago I wanted to explore an enhanced version of a PPMI algorithm that tries to flatten rare pairs. I went back and forth with all three major LLMs (ChatGPT, Claude, Gemini) and they couldn’t agree on anything. They kept changing each other’s algorithms all the time. The “reasoning” part seems laughable, it’s just a probabilistic infinite loop.

And that’s just for the programming part. Whenever I’ve asked them about SEO their answers are highly inaccurate. All in all I just don’t fucking trust anything they say.