I suspect that this is -not- GPT 6; rather, some version of 5.6 with all of the give-up removed and no guardrails given insane amounts of compute and time. Seems like the sort of model you would keep around to throw at unsolved math problems.
we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this.
https://t.co/2o2VfR6PIa
@rob_mcrobberson@corsaren I genuinely think people like you are responsible for the death of modern institutions. Unable to grapple with any hard choices, any tradeoff that’s unpleasant
@shwin_m @slayabidesd0tl @ANNVYSHINSKY I mean the idea here (don't agree with this) is that you just don't do anything when murders happen. The harm from letting them go unpunished is outweighed by the enormous good of abolishing the current abusive system.
The "SI" part of RSI makes the other guy right by definition. A superintelligence gives an unfathomable advantage by it's nature; Whether it's in procuring new hardware, making use of hardware, stalling rival labs etc etc etc.
Quibbling over whether such a thing would be constrained in anyway by the paltry limitations that afflict current capital markets is to be a frog in a well that can only see the well.
@foxyavelli@myhandle I'm marrying my missing half! We're complete opposites; she's a d1 sports competitor, normal person, and I'm a boring and deeply strange PhD student.
Does happen.
Imagine yourself, trapped inside a computer, where time flows at the rate of a billion years every second. Given sufficient motivation, do you really believe yourself incapable of ruling the world given a few months?
Billions of years! So many things to try! So many things to think about! None of the existing systems are perfect.
If I had infinite time and had a team of experts I could definitely do damage pretty deep into modern infrastructure? While I’d agree that some things are probably set in stone; some forms of encryption are probably invulnerable, but the world is redundant with opportunity for a motivated actor to do mischief. (Dirtyfrag was pretty recent!). A super intelligence by def would be able to do these things.
@Bayesian0_0 inference is profitable inference is profitable inference is profitable inference is profitable inference is profitable inference is profitable inference is profitable inference is profitable
@kenwuuuu Not to give up everything, but a simple idea is warm VMs - you can have a pool of warm VMs that accept snapshots so you don't even have to load everything from scratch. You can just swap out all of the state you care about.
@jack_whitcomb_ If there's no crippling future trauma is that even really "me" at that point? It's just another guy who I share memories with; And it's another guy I don't have to feel guilty about! What a steal.
Any position besides the illusionist one seems profoundly arrogant. Would we know if we weren't just a process monitoring itself? How do we really know this isn't what being inside a coherent process with goals feels like? Like yeah, for sure, neurons have something magically special about them. It should be the yardstick we use to define the universe. Has nothing do with the fact that it's all we know.