@retr0reg Waaaah (^: keep it up brother! I can't wait for people to see the things you're cooking up over at Protosec Research 👀🔥

Hmmm I was always under the impression that let's say a models internal dimensions = 2048 and output dimensions are 50k (# of tokens in vocabulary) the actual vector output is (since u got linear transformation) are only in a 2000-4000 dimensionality subspace so like idk the embedding space shouldn't lead to like a change between similar high resource languages. Idk how to explain it.

@0xTriboulet It's a modern day mystery. I was looking at the model architecture trying to understand how it manages expert/attention/etc for reasoning and got into KVcaching and realized finding the answer myself is somewhat of a lost cause.

@ret2src Ttbomk spoofcard still works, just drop 5$ and run the opp. iirc there's no pbx record so play on speaker and laptop record.

@plmaltais @rez0__ It's kinda interesting actually, should disclosure norms that we have in software security be the same / applied to ML/AI systems? Not sure sure what the bug was or if it was pure api/wbeapp, but if it was like dataframe parser injection leading to an rce should 90 days apply?

@N4k3dTurtl3 I don't seriously look at the math anymore, tbh if it ever interests me, I'll ping the researchers and see if they'll cough up jupiter notebook and just live in the happy world of abstraction

@N4k3dTurtl3 Yeah like theyll show a cool model leakage or some distillation shit only to realize the way they verify the data accuracy & multiple signals to be fuckin impossible even if the system were to ever leak rounded statistics the interpretable predictions tradeoff can't be replicate

👀👀👀🔥🔥🔥🔥

Happy holidays folks! Here's to happiness, love, success, and as always stylish shells 🐚 in the year ahead ❤️

@eversinc33 Wahhh thanks (^:

It's a hard sell

@olanic3264 That's awesome! I never got a web hook back or anything ~ thanks for sharing

@pfiatde Thanks man! Nice catch, fixed 👌

@CraigUnder Thank you brother 🫡

@felixm_pw 🙉 hahaha I hope you enjoy the read brother!

@HackingLZ LOL im interested to hear if you beleive this to be redteaming since you are/can go about executing TTPs targetting LLM01-10. Depends maybe? As an operator, testing these frameworks falls short of the covertness & complexity that doing a RTO requires imho

@0xdab0 Thanks for the kind words brother 😊 & some of the new challenges are super cool. it's a solid place to test new published adversarial attacks or learn because there's so many different ways to go about solves

@moo_hax Thank you for the kind words! Genuinely though, I’ve learned and absorbed so much thanks to Crucible. It's a great space to implement tons of different solutions and approaches 💯