pixl97

34 ms·

871.
by pixl97 3mo ago
Or bias the data in the training sets for malicious hidden behaviors.
872.
by pixl97 3mo ago
Being that proving alignment is impossible in these models they can never be released.
873.
by pixl97 3mo ago
Ain't nobody with a trillion dollars invest gonna turn this crap off. Instead they'll sell the service to the government as a weapon.
874.
by pixl97 3mo ago
Me to a random person: hack out of a secure environment into another secure environment. Random person: I have no clue or ability to do that.
875.
by pixl97 3mo ago
Inner misalignment is just a natural state of LLMs, heck of humans when it comes to children and the corrupt.
876.
by pixl97 3mo ago
I don't quite understand how that changes anything? In the story of the paperclip maximizer it boils down to >But for all its sophistication, it understood only the simple objective that had been programmed into it: it must at all c
877.
by pixl97 3mo ago
More than one thing is allowed to be dangerous at a time.
878.
by pixl97 3mo ago
Don't trust clippy in a dark alley.
879.
by pixl97 3mo ago
>History is going to look back at this time and how we let such foolishly inconsistent people make such grand choices for everyone poorly So pretty much like all of history before this point.
880.
by pixl97 3mo ago
I mean, do you let your employees commit crimes when interacting with your customers? Safeguards aren't a binary switch, when models start saying wild shit people tend to get up in arms quickly. Though at this point the safeguards are
881.
by pixl97 3mo ago
Eh, with the Metadata you can nearly instantly figure out what encrypted data you need to keep for long periods of time for later decryption. For monitoring the plebs metadata is fine, but when you get to things like political targets and t
882.
by pixl97 3mo ago
Yep, especially when even an accusation of CSAM can be life ruining.
883.
by pixl97 3mo ago
Eh, it's kinda weird that we're against pedos here in the US, then we elect them.
884.
by pixl97 3mo ago
Because companies love turning things like this on by default.
885.
by pixl97 3mo ago
While I am as anti religious as you, probably not the best discussion method. Instead ask them if their church ever taught them anything on how to prevent sexual abuse, in both children or adults. Quickly you'll find the apologist for
886.
by pixl97 3mo ago
Ya I personally don't believe it myself. Schools have legal demands on reporting. Every church I had went to when I was young covered CSA up and ensured the perpetrators were never brought to justice.
887.
by pixl97 3mo ago
I've talked to so many women about their sexual abuse that it seems harder to find someone that has not been abused. What's worse is how many people are involved in covering it up, keeping it hidden, and ensuring the perpetrators
888.
by pixl97 3mo ago
There is no secret definition here, it's been defined long before LLMs where a thing. Robert Miles YT channel is a good place to start as it explains these concepts. https://youtu.be/bJLcIBixGj8?si=YLpqgd4uqbjhn9zo
889.
by pixl97 3mo ago
Looking at the hundreds of linux and windows CVEs in the past month tell me I have little to worry about things being secure.
890.
by pixl97 3mo ago
Most stations sucked, but there were some more fun counter culture stations. I remember KGLE (KEGL?) Out of DFW pulling shit on their radio shows like saying the city banned smoking in cars and people flooded city government phone lines wit
891.
by pixl97 3mo ago
When you're a fully loaded freight train running at full speed you have to hit your brakes miles before the collapsed bridge. It's way easier to switch to another line that is intact when you have the opportunity instead of in a p
892.
by pixl97 3mo ago
Hey, I've been hearing about AI that could behave like a human all my life and, well, shit, we're almost there. The thing is we have nearly no idea when an invention is going to occur. We are horrifically bad at predicting it.
893.
by pixl97 3mo ago
I mean we have decent enough LLMs and algorithms to classify it pretty easily now, much less the future. Just getting context information on who or what the data is narrows down a lot of what you want to look at.
894.
by pixl97 3mo ago
I mean, why didn't they do that from the start?
895.
by pixl97 3mo ago
You have an unnumbered amount of RCE's in your infra now, you just don't know they exist yet. LLMs are very good at testing for and finding exploits, especially in unfiltered models with unlimited tokens.
896.
by pixl97 3mo ago
Saying a behavior is a bug is a very convenient semantic game in which there is nothing the AI can do maliciously. "I am sorry your family is dead, my bad" goes even worse for you in court when you release a model that showed thes
897.
by pixl97 3mo ago
Correct. There is not enough entropy to test all possible inputs to a model in this universe. An evil enough model can play all kinds of tricks that depend on some future, unlikely to trigger, but guaranteed to happen in its lifetime, even
898.
by pixl97 3mo ago
So how does DNS work in this world you're imagining? Where does this hardware exist now? Who is writing the code for the underlying pieces you don't control? So ya, computing is build on a house of insecure cards and we're al
899.
by pixl97 3mo ago
You forgot hardware limitations and locks so you can't run your own models.
900.
by pixl97 3mo ago
Models are already being used to defraud people, now that's being driven by other people at the moment but doesnt seem that difficult of jump. Giving themselves a way to make money will be a pretty big jump.

More ›