Notices by Kraftwerk-Das Model Collapse (dngrs@chaos.social)

LLMs are not like asbestos, that's an unfair comparison. Asbestos has useful properties.

#llm #genai

@thomasfuchs maybe - but models will become stale and training is very data center intense work afaik. Depends on model/application of course. But LLM vendors will certainly put a big emphasis on continuous training being a necessity.

@thomasfuchs this seems to just forward to a "Sky Guide - Star gaze constellation finder" app?

unfriendly reminder: dismantling public health is a *right wing project*. Don't support their work by saying "during Covid" when you mean "during Covid lockdowns" or "back when people still took Covid seriously".

We have normalized condoms and STD testing. We can do better with airborne diseases than sticking our heads in the sand in the face of life destroying, incurable diseases like Long Covid that current vaccines does not reliably protect against.

#covid #pandemic #covidIsNotOver

@dalias do you have a link/recommendation for such a policy file?

@baldur motivated reasoning is one hell of a drug. I've seen a developer far better than me getting completely hypnotized by LLM sycophancy; I tried pointing out that what they proudly posted as "see? Completely bug free after just a few rounds of conversation!" did in fact contain a subtle bug/violation of their prompt. Got ignored and they only went downhill from there. It's a saddening cult really.

I'm hiring! For now Germany only, but hopefully EU wide is an option from March on.
Please read the ENTIRE thread. Boosts appreci.

Looking for a person with strong ethics (LGBT+ compatible, opposed to AI & general capitalist exploitation) - we're self funded and actively avoiding VC. This also means limited cash at the moment so the position is for a 6 month period, but can be extended once we manage to secure EU funding.

Our project is open source (MIT).

1/3

#hiring #fedihire #getfedihired

@jeridansky @GeoffWozniak it gets worse, the debunking is also trash, the author asks Gemini "did you make it up?" and trusts its output 🙄

@amszmidt I'm here all week!

@amszmidt in case you didn't like my reply ironically: I was making fun of you

@amszmidt whoosh: the toot: the movie: the musical

@mntmn hmm, @whitequark any ideas?

The ship of Theseus thought experiment has been retold so many times that we're not sure what's still left of the original idea

@SuperDicq
1. normies aren't security experts. Why should I consult them on security matters?
2. phones do video decoding in hardware
3. we probably agree that when given a tradeoff of performance vs. security, security isn't *always* the correct choice, e.g. I'm not very happy with the amount of spectre etc mitigations; my personal threat model would prefer to have those disabled
4. 20% isn't an accurate figure, but of course ideally the RIIR impl isn't slower at all. I think it's doable.

@SuperDicq I don't think it's a universally true statement that memory safety is only something for "Rust nerds"; image/video decoding can be a pretty juicy attack vector (cf LogoFail or the various issues Apple had over the years). Of course all bets are off once you're in ASM land, but reducing attack surfaces does have value.

@SuperDicq @phnt in fairness the 35% slower was for the initial release, no? There's been a lot of progress https://ohadravid.github.io/posts/2025-05-rav1d-faster/

@whitequark @fasterthanlime I'd second that assumption, afaik the C family just doesn't do ZST

@WilliamNB @ConnyDuck @larsmb on top of that, a substantial amount of infections are asymptomatic - around 44% in 2020 according to this meta analysis https://pmc.ncbi.nlm.nih.gov/articles/PMC9935239/

@zenkat @glyph but... it's not accurate to call LLM summaries "approximate" - they might be entirely wrong. If you make decisions based on that, I'd call this a pretty careless attitude

@zenkat @glyph all your verbose LLM apologia does not change the fact that LLMs work statistically, always, and *don't know what facts are*. They can't not hallucinate, because that's the *only* thing they're capable of - even when the result looks correct.