x74sys

And I want to add: even though LLMs can identify cybersecurity risks, it doesn’t mean they are good at cybersecurity. They’re probably just as bad as in any other area. Also questionable if the actual positives outweigh the labor required to flag all the false-positives.

Just because they‘re used everywhere doesn’t mean that we just have to accept them. Also doesn’t mean that LLMs are a good thing.

I think LLMs can be used as an (additional!) cyber security analysis tool, that’s honestly the only area in which it seems to be actually useful (right now). And most projects don’t reach the size in which spotting security risks spanning across many different modules is a relevant skill to have. So it should be used sparingly, on things like the linux kernel. Then the cost of it might even be worth it (but I also don’t want to know about the amount of hallucinated bugs it finds).

Adwaita

I don’t really have any problem with the thing by itself, I honestly don’t even know that much about it. I just hate everything that gains publicity just because there is a certain name behind it. Which is what happened with omarchy.

Didn’t know that it was AI slop as well. That’s kinda sad…

Yeah, honestly if I notice something has AI in it, I don’t use it anymore. Open source projects with a CLAUDE.md or whatever in them? No, thanks.

The amount of assumptions they make are really one big issue what makes them suck so bad. In the end you just have more work. Instead of getting done 80% of the work in 20% of the time, now you get 30% of the work done in 1% of the time, but good luck getting the remaining stuff done at all.

Yeah, at most you can let them manage a 1k loc python script (the free tiers or Gemini Pro at least), but more than that and it starts to really eat your tokens without achieving what you asked or breaking functional behavior.

I extremely doubt that Coding Agents will see a future like promised. LLMs are still so expensive to run, and the useful larger models will probably never be affordable (if they charged for them what they cost). Apart from the fact that even their output can be utter garbage (and mediocre at best). You can already see it everywhere. Websites break in weird ways, ways in which it’s clear that either a complete beginner wrote that or an LLM did. Look at Shazam a few weeks ago. UI design? Horrific. Extremely inconsistent. Ugly. There are many other examples. It just shows that it doesn’t work. And no, the next model will not solve those issues. LLMs are flawed for this task from the ground up, the approach is outright wrong, we can make up so many bandaids and they will still suck, forever.