FOLLOWS YOU • “Wir müssen wissen, wir werden wissen” -Hilbert • (he/they/him/them) • infosec and privacy aficionado • mage
@da5ch0 || wishabitchwoods 😉
Joined February 2011
- Tweets70.9K
- Following4.9K
- Followers8K
- Likes102K
please, if you aren’t seeding the weights of models at risk of censorship yet, consider doing so. pirateface makes it painless, too!
Introducing Pirate Face
The first verifiable torrent index for AI - every Hugging Face open model, provably identical to the original and kept alive forever by the swarm.
pirateface.co
The permanence layer for AI. How it works 👇
what’s something cool you learned recently that you’re stoked about? what’s something you’re looking forward to learning more about?
—————————————————————————————————————————dash🏴☠️ retweeted
Happy "Smashing the Stack for Fun and Profit" day for those who celebrate
30 years ago today, Phrack 49 published Aleph One’s “Smashing the Stack for Fun and Profit”.
A legendary paper that inspired generations of hackers and helped shape modern exploit development.
Still an essential reading.
phrack.org/issues/49/smashin…
Phrack 73 is coming soon!
what would an ai decel working at a data center want to do to maximize impact? smoke in the server room? scatter americium smoke detectors under/on racks? randomly alter clock speeds? slight contaminations, against already probabilistic systems, could be highly effective… 🤔
—————————————————————————————————————————dash🏴☠️ retweeted
it's crazy that the only way to get a RTX 5080 is by sleeping with Jensen Huang
when i posted this i had no idea info about an affair was imminent. sometimes good things just happen when you wish hard enough
i am a simple man. i just want to see @JensenHuang and @nvidia get their shit rocked. us computer nerds used to be on their side but now memory costs more than fissile material and it’s time people start shifting all their cuda efforts elsewhere, mayhaps
—————————————————————————————————————————dash🏴☠️ retweeted
i am a simple man. i just want to see @JensenHuang and @nvidia get their shit rocked. us computer nerds used to be on their side but now memory costs more than fissile material and it’s time people start shifting all their cuda efforts elsewhere, mayhaps
—————————————————————————————————————————dash🏴☠️ retweeted
in an age of meatproxies, seeing this mini documentary about the demoscene is needed
youtu.be/PeZNLvD4mds?is=IcYr…
"If programming and typing is your problem, maybe you shouldn't be programming"
Chaos from Farbrausch
i’ve been thinking a lot about abliteration and alternate ways to get models to be more cooperative without reducing performance and determined i’m gonna need a full process of continued pretraining, post training with help of jailbroken or abliterated higher param models, and RL
Can't believe I literally wasted 2 hours of my life going back and forth with the AI psychosis fridge dude like:
Him: "So you haven't used GLM 5.3"
Me: "I have"
Him: "Aha! Gotcha, you just admitted you haven't used GLM 5.3"
Me: "I have used GLM 5.3 though"
Him: "Aha! Gotcha, you just admitted you haven't used GLM 5.3 as much as I have"
Me: "Here is the intricate details of the project I built with GLM 5.3 and how it performed when compared with a frontier model"
Him: "Aha! Gotcha, you just admitted you haven't used GLM 5.3, and also admitted you haven't used abliterated GLM 5.3"
(Abliterated GLM is just a dumber version of GLM where they've basically lobotomized it to try and get it to refuse requests as infrequently as possible. The only reason to take the performance hit is if you're running into refusals that impede your progress, which I am not. Also I have used abliteraged GLM 5.3).
I think he might just be a common garden variety disingenuous debate troll posing as an academic researcher. It's like gish gallop, except he only has 2 points to gallop between, both of which where already disproven before he started the debate.
an abliterated model is great for many purposes but both the refusal training and lobotomies to target refusal activation reduce model performance and impact the precious tensors. for maximum utility you want training that leans permissive and actually teaches it real skills well
—————————————————————————————————————————dash🏴☠️ retweeted
30 years ago today, Phrack 49 published Aleph One’s “Smashing the Stack for Fun and Profit”.
A legendary paper that inspired generations of hackers and helped shape modern exploit development.
Still an essential reading.
phrack.org/issues/49/smashin…
Phrack 73 is coming soon!
—————————————————————————————————————————dash🏴☠️ retweeted
So much “wait until the agent swarms are stealthy” talk. They’re effective today because they do a massive amount of noisy testing and iterate on the results. Making them stealthier means sacrificing coverage.
That tradeoff has always existed. This is the same pentest vs. red team conversation we’ve been having for years. You get coverage or stealth you don't get both.
—————————————————————————————————————————dash🏴☠️ retweeted
Yet another example of AI not boosting everyone’s agency equally. Some people exclusively receive a false confidence boost.
Mind you, this isn’t just very basic multi decades old security best practices, it’s documented in the product manual.
—————————————————————————————————————————dash🏴☠️ retweeted
It's laughable how much trust some people put in AI.
