ok so someone on bsky i respect made a valid point about LLMs: you have to keep up to date with what they're like, so that you can form more informed arguments against them. that otherwise you might say something outdated, be easily disproved, and make everyone against LLMs look bad
to that end, I just talked to claude. check this out
I told it what the conversation would be about, and it IMMEDIATELY went off the rails
it started writing a phantom python program that does nothing
it kept going
and going
it eventually stopped writing code for a program that doesn't do anything, and then told me how good the program that doesn't do anything is
when I said "why'd you do that", it went "oh my bad, yeah no reason, you're smart for pointing out that i shouldn't do that, but here's why my program that doesn't do anything does the nothing correctly"
i checked in geany, and it was 136 lines of code that do nothing and have no meaning that it generated for no readily apparent reason
"hey claude, did that vomit you just spat out smell like farts?"
"excellent question, yes it does"
I asked it to explain how it got to this situation and it panicked and hit the big red emergency stop button and sirens started going off
overall i'm confused. claude is meant to be "the capable one", for programmers. but my conversation opener made it lose it's mind, and pressing the issue caused it to self destruct. i feel like captain kirk. i talked a computer into committing suicide
how is it that reddit is full of "i built an app" spam entirely made of claude output when it's this fragile? babying this thing must be harder than just using a textbook and writing your own code
fin
@kyle_pegasus I check in on LLMs occasionally too for the same reason of keeping up to date.
This is 100% my experience every time. It’s *wild* that people think so highly of these things.
Oct 15, 2026, 5:15 PM@t3rminus @kyle_pegasus Every time, I leave frustrated, seeing those tools that are programmed to behave like know-it-all bootlickers.
Since we're not allowed to call them "Artificial" any more, I'm now calling them "American Incompetence".
Oct 15, 2026, 5:46 PM@t3rminus @kyle_pegasus someone once likened it to a Shepard tone, where every new model indeed has some changes, but the net amount of motion sums out to zero
Oct 15, 2026, 6:41 PMAs someone who uses Claude extensively, are you confused and looking for an answer at why it failed, or are you confused and looking to point at it and say "Look at how bad it is?"
I can help with the first.
@serge Fuck off
Oct 15, 2026, 5:12 PM@kyle_pegasus My wife asks the exact same coding question at each new announced release, and she and I evaluate the results together. Separately, I occasionally ask questions related to what I work on.
Ιt's an interesting experience for sure.