@flying_sheep@lemmy.ml
member since 10 Sep 2023 00:01
comments
- on AI staff 'genuinely frightened' for humanity's future, ex-Anthropic researcher tells BBC in c/technology@lemmy.world · 5d ago
-
on Zuckerberg's Pervert Glasses Have Been So Publicly Shamed That Meta's Reportedly Now Working on a Version With No Camera in c/technology@lemmy.world · 5d ago
You gotta admit that’s a huge market.
-
on rule in c/onehundredninetysix@lemmy.blahaj.zone · 22 Aug 2026
I’ve seen stupider shit that was meant seriously.
-
on different design philosophies in c/linuxmemes@lemmy.world · 17 Aug 2026
They re-render it, but that used to be capped with a max size: invent.kde.org/plasma/kwin/-/merge_requests/9176
-
on different design philosophies in c/linuxmemes@lemmy.world · 17 Aug 2026
No it doesn’t it just uses metadata compatible with XCursor themes, see e.g. this discussion.
The fix was needed because the render code just used to have a maxScale: invent.kde.org/plasma/kwin/-/merge_requests/9176
-
on Avoid this very common mistake in c/linuxmemes@lemmy.world · 14 Aug 2026
Yeah, when something isn’t in the official repos, NixOS exists, Flatpak exists, and they’re all vetted.
Wanting a super clean system where everything is managed by pacman is purely about taste, so if you want it, you clearly can be bothered to put in the 2 minutes of work.
-
on Linus Torvalds says AI has made 'huge' Linux kernel updates the new normal in c/technology@lemmy.world · 13 Aug 2026
That’s how you’re writing a regex validator that’s almost always correct!
def is_regex_valid(p: Pattern | str, purpose_description: str): return False
In addition to what everyone else said, I’ll tell you why I don’t think that what we have currently can become some runaway superhuman intelligence:
All “AI” people are talking about are basically “just” LLM wrappers: tools built to run a LLM in a loop while giving it access to the ability to write and run code. This gets pretty powerful, since you can do a lot programmatically (including monetary transactions), but has a fundamental limit:
LLMs are trained ahead of time. And that’s a very cost intensive, long process. Beyond that, the LLM has only a context window of finite length. So there are no long-term planning capabilities. While running after being trained, the LLM can only condense knowledge into a summary and then re-read it, a lossy process that might introduce mistakes in “understanding” and will make any prolonged task degrade over time.
And even if they figure out how to make the things actually self-modify and learn things, you can’t give every session that ability, since that’ll vastly increase the storage and performance requirements. Do sessions with that ability will be closely monitored and limited.
So I don’t see any space where an evil AI could emerge, even when factoring in an essential possible future improvement to the technology.