Nvidia, I get you're all-in on AI, but please at least find a human to write some copy. It's both overly verbose and entirely lacking of important info at the same time.
Why clearly? You mean the tensor pre seed with deterministic pseudo random numbers? I don't have much experience in LLM training, but I do remember from minimization algorithms (like gradient descent) that it's better to start with (pseudo) random starting values and minimize from there. And seeing how I would like my results repeatable, seeding a pseudo random sequence seems logical.
I don't understand why these tensors specifically weren't changed during minimization (training) though. Maybe they were not activated, or a w part of an image decoder or something?
consider most naive possible way one would try to deceive users without exorbitant budget -> attempts a few things you naively would do -> recovers it by chance?
I'd love to know the demographic breakdown there. For instance I know students in some relatively isolated corners of the world are still, ubiquitously, using it to cheat. And a quick search gives around 1.4 billion students globally.
Obviously it's not just students, but the fact it [numerically] could be is rather amusing and makes me surprised that it's only 1 billion active users, since GPT has a lock on name recognition amongst low information users.
I'm naive, but is this possible? The power required per user to run those GPUs seems like enough to run an air conditioner. For a billion people everyday or so seems like enough to overload power plants.
> The power required per user to run those GPUs seems like enough to run an air conditioner.
Not true. Maybe you're thinking of computer games? Inference requires much less energy per user session than a session of your average AAA game title. And there are more active gamers at any time than active ChatGPT users.
You’re comparing 1 billion people doing bleeding edge gaming to LLM usage. Feels a bit like your thumb is on the scale here. Even under lower load GPU’s do in fact use a lot of power, as evidence by the massive power buildouts that we are seeing across multiple countries.
Yes Silicon Valley gives things away to build a high DAU count before the squeeze we’ve seen it for almost 2 decades now. That tells us nothing of the longterm health of their business.
A company worth trillions does not have trillions, that is the expectation for what it will be worth in the future based on money it has already spent and will need to keep spending.
Moreover, that translates to "it's actually carrying a huge nut* that puts extremely intense pressure on the company to figure out how to make as much money as possible as soon as possible."
(*nut is a slang word from loan sharking that has been adopted by (personal et al) finance people. It means "how much do you owe each month?" in the sense of, if you want to do a startup, advisors could ask, "what's your nut, and how much do you have left to self-invest?" Because monthly expense can be combined to a lump sum, it can also mean the size of the overall loan where the monthly interest is called the vig, "as long youze got da nut, you owe me da vig;" from vigorish, which would also require an etymology, but I'm assuming it's the word vigor combined with licorice meaning bitter tasting medicine, and nut because squirrels busy themselves carrying nuts.)
----
and if we can... and remember flame-prone people, if you reject the "can" premise, then the rest of the argument goes away, and you don't need to dispute it, you just say "you can't nuff said"; people here frequently seem challenged by propositions.
so, if we can...loosely describe the attitude you show here toward high market cap companies to be compatible with "left" or "socialist" or "populist" views shared by many others here, then I'll point out that at the same time wanting higher market prices paid for bounties because 'sploits are so expensive, is actually advocating for the same dread capitalist market based solution that leads to trillion dollar companies that never even had a trillion to begin with.
If you reframe the perspective from 'free $500 sitting on a server' to 'requires actively utilizing an unreleased 0 day exploit on a server that will certainly be able to track exactly what was done, for a payout of... $500?'
That's pennies for something like this.
Add two zeros and it might get nabbed.
Edit: others have pointed out it's not really a 0-day anymore, they're right, so it's likely more a matter of time.
So you are therefore arguing for the author's point.
The entire exercise is to counter the claim that AI means every script kiddie with tokens can just have their AI create an exploit, and then have their AI scour the internet for vulnerable endpoints. If that were true, then the exploit has very little value, because everyone could make one.
In addition, the claim is that these exploits will all immediately have a thousand bots using them to exploit every possible website, and you can't leave your website vulnerable for a second or it will immediately be owned.
The fact that that hasn't happened to this server already proves the point. Whether professionals can create an exploit from the vulnerability was never the question.
I suppose it may not be there for actually taking it, but it does serve as a good demonstration of the realistic risk for average people running relatively low-value PHP sites. If nobody thinks it's worth the risk of exposure for a quick and easy $1k in Bitcoin (at the time I'm writing this), why would anybody bother hacking a site that's basically worthless, like the dinky little Wordpress blog I run?
This does effectively make the statement that, there probably are exploits out there for this, but even with a non-updated version, you're not at much risk unless someone with serious resources thinks they can make pretty big money from compromising your thing.
random sites generally don't log carefully, or watch their logs if they do. so the risk to hacking them is low. this site almost certainly logs very aggressively. so the risk to hacking it is high.
based on differential risk, attackers will choose targets differently. they might reason that the risk/reward on this obvious trap is not worth it. but on joe shmo's php emporium, the reward is less, but so is the risk, and it might pencil out.
> that's good news because my webserver also has logging
In the future, every website will carry a little blurb of text in the footer of their pages intended to scare off would-be hackers and their AI agents:
“The server of this website employs append-only remote logging and state-of-the-art monitoring and intrusion detection. Unauthorized access or attempts there of WILL be prosecuted to the fullest extent of the law. The outcome of such prosecution is and always will be astronomical amounts of money in damages that the guilty party will pay to the victim, alongside decades of prison time for the party guilty of computer crimes. Do not even try.”
That's always the case for a novel exploit. There isn't a lot of value in popping somebody's PHP blog, so people who've invested time and money into a zero-day exploit have little incentive to use it for that purpose.
The white hats want to extend the time before an exploit is in the hands of the masses who might do so just for the fun of it, and the black hats want to use it on high-value targets.
I think the point is that this is completely the opposite of the narrative that was pushed about this bug - that there would be thousands+ of AI agents actively compromising every vulnerable site automatically regardless of value within minutes of the fix being released into the wild. Well, now it's 3 days after, and this has been up for 3.5 hours, and nobody has taken it yet.
You would have a point if it was still a 0 day. The patch is out, and open source. Anyone who cares has a bunch of the pieces for putting together their own exploit.
Zero days are valuable and kept in the bank to avoid detection. That ship has sailed.
It’s not a 0day it’s now 5 days and counting since disclosure. Nobody is gaining value sitting on an exploit for a vulnerability which is being actively patched.
This logic doesn't work here, because it relies on the premise that you alone potentially have an unreleased zero-day exploit, which is not the case.
If prompting an AI could exploit this for less than $500 in token spend, then you should immediately do so, since if you don't, in a few minutes' time someone else will.
Why not? Those pros have their opportunity to use this exploit dwindling by the hour as news of this and pressure to patch grows and grows. All the major managed hosting companies had this patched in the first 24 hours.
It works fine on mainstream models where the output hasn't been edited.
It doesn't work on humanizers or custom model output.
reply