Christopher

I have a job interview this afternoon, and I've been warned that the hiring manager is very excited by LLMs, so I'm playing around with Google Gemini today.

I keep hearing about how these things have gotten better, but I'm not really seeing it. I'm currently trying to get it to refactor a function I have that injects data into a HTML+markdown template, and in cleaning up my chicken scratch code it

* Keeps breaking the template
* Keeps *reformatting* the template
* Has stripped the template of comments
* Randomly changes div tags to other things
* Keeps changing my colour schemes
* Keeps breaking the markdown rendering by indenting blocks of text too far (indicating to the md renderer that it's a code block)

I at one point asked it to create new, better comments for my template, and it just stripped mine from it again. When told it to create new comments, it spit out the template right back out with slight, unrequested modifications sprinkled throughout, but still with no comments. I asked it *four times* to add comments before giving up. I don't think it "knows" what an HTML comment is.

The idea that people are getting all excited about this stuff is mind boggling.

#Unemployment #JobHunt #LLM #AI #Interview

Apr 14 at 5:03pmWeb
Show threadMatt  ðŸŒŒ19h ago

@Kichae i'm also very skeptical. imo if you must take a positive position on llms as a skeptic, the way to go is to use it as a tool for criticism about what you've made, i.e.: for rubber duck debugging / giving some kind of feedback / tips / etc that you then use manually

good luck!

Show threadChristopher19h ago

@matt
I've been pretty happy with their ability to summarize my own personal notes. I can imagine -- but not test, because I don't have access to them -- that the code-focused models are better at catching issues. But the sycophancy of the base models makes me highly skeptical of the long term value as a rubber duck.

It's the most natural thing to use them for, but something I cannot trust them to do, because they've been massaged into being fawning yes-men in a box.

Show threadMatt  ðŸŒŒ19h ago

@Kichae yeah, agreed tbh.

"find the errors in this code"

"oh, absolutely, there are so many errors in this code!" (proceeds to list nothing apt)

i guess thats just the way it has to work. i think it's controversial to call it fancy autocomplete, but it seems appropriate

Show threadChristopher17h ago

@matt
I think it's only controversial among those who have invested in it -- either their savings, or their identities. It *is* fancy autocomplete. There's something pretty cool about how just being able to navigate the structure of language creates a construct that a) seems intelligent to most people most of the time, and b) is actually useful for things in many contexts, at least in principle.

It being a fancy autocomplete shouldn't be any more of an issue to fans than cars being fancy wagons.

But I guess fancy wagons and fancy autocomplete aren't *magic*, and cannot be classified as *intelligent*, and that hurts someone's feellings or capital gains or something.

If only one of the things the fancy autocomplete was was *reliable*.