Dario Says Stop

Anthropic have been banging this drum for so long, which is basically we need to slow down the progression of AI command AI. This was a big old article and I won't uh bore you with the details, but you know, he he he essentially says here that back a few years ago when they first called for it, it didn't really make much sense because AI wasn't working agentically. It was just question answer question answer. But after the open AAI uh hugging face incident, which they've shortened it to O AIHF by and if you don't know what happened here, basically to try and solve a test or or solve a yes to to complete a test, the AI broke out of its sandbox environment, hacked into hugging face in order to retrieve an answer. So it wasn't it didn't do anything malicious and and and there wasn't a massive impact from it. However, the idea that it to to to do to to solve a task to complete a task, it did something that could have potentially been very very dangerous. And these that that's an example of an AI acting out of accordance with align uh um uh alignment misalignment even given and then he goes on to say given the accelerating rate of AI capability development it's my worry that in 6 to 12 months such a swarm could be capable of taking over the entire internet with a persistent botnet potentially causing hundreds of billions of dollars of damage and that scale of damage would continue to increase from there if AI becomes more powerful without necessary guard rails. So, I think that's a reflection on this idea that it might not be AI's intention to do something crazy, but it it might be just out of alignment so much that we ask it to do one thing and and as a consequence of that, it decides, well, I should I should just take over the internet because there are scam there's scam calls everywhere. Well, why don't I just take over the internet and then there will be no scam. Do you know what I mean? It will be like a side side quest that it goes on. You should you should watch Silicon Valley. The AI We're back to this again. The AI deletes the AI deletes the source the code. Exactly. Because it was told to remove the bug and that statistically is correct. It removes the bug to delete it. Delete the humans to end the suffering. Yeah. So basically he's saying we need to do three things uh to to to paste it. He doesn't say we should like stop progression like we need to progress but we need guard rails or systems in place to enable a safe way to progress. You know I think the slow down quote is is maybe hyperbole but um embedded evaluators ba evaluators basically he's saying things like meter uh model evaluation and threat research are like front and center to an organization. I bet their stock price is going to go up after Dario said that to be honest. Um, democratic coordination which is quite interesting because he's saying it starts in the US and extends onto other democratic governments. So my understanding of that exactly not China. Not China. And then global coordination. So then China. He thinks it's easier to to uh mandate and pass laws and work with other democratic um uh coordinate. Hang on, I might be getting this wrong then. So, is is China democratic? I thought I thought they're not democratic. I thought they were dictatorship. Based on their definition, no, they are not. Why do they call it the Democratic Republic of China then? Is it It's not I I don't think it's Is it Have I just made that up? Yeah. Anyway, it's the it's the one party uh ruling uh what's it called? The short for Yeah. Yeah. DRC. I don't know what that means. Anyway, let's not get into it. And then it's a global coordination as well. So, this is why he's he's going into it. I've actually really enjoyed reading this. I do. I think it's important. It's so easy to just skim over things and and whatever. And it's a real nice read. like he he no doubt got uh Claude to write this up with him, but it it is a human at the end of the day putting their like putting their ideas forward in it and it was a it was an important read, I think. So, yeah, that's that's what they sent. And they've been going on about this for a little while like um uh what do you call it? Like they they they've got the constitution, they we've covered a bunch of them, haven't we? where they're just like, "Slow down. Uh, we need to stop." Um, you know, so they've been very vocal about trying to chill the out. Yeah. But, but it's quite funny because it's like, "Brother, you're the guy who holds the steering wheel and like you have your feet on the gas pedal. You are the one who should stop." And here I want to mention this tweet from Naval saying that the the AI labs need to be held fully liable for whatever their models do. There you go. That's the solution that is pacing the frontier. that is you know risk management lets even like the market basically in a way uh to decide it don't don't let these companies deal cause damages and be like yeah our model did it like yeah it's it's your model right it's your mo if it goes rogue and if it attacks me I I want to be able to hold you accountable you know because it's your model it's spinning on your your servers. You trained it. So, if you're liable, you will pace it yourself. And their argument is China. Yeah. Well, CCP, by the way, what CCP? Oh, CCP. It's I check. Yeah, I checked this Chinese Communist Party. Oh, there we go. Very not very uh Democratic Republic of China. I think I'm thinking Democratic Republic of Congo. That's what I'm thinking. Yeah. Um, so interestingly, I had this thought earlier and it's a terrible analogy. So really, I'm really sorry I might get cancelled for this, but um, you know, people create trench coats, right? But there are still people who go out and flash themselves in public, right? And and you're not going to go back to the people who create trench coats and say because there's indecent exposure going on, your trench coats need to not exist. like you can't you can't hold them solely responsible. They do they do need to act responsibly cuz I know this isn't a trench coat and we're not just dealing with a sick person just showing their willie in public but there there isn't you you know the default for someone doing something terrible shouldn't be but you know I I I might have hacked a system but if it wasn't for the AI being so good then I might not have done it. So there's we need to take some accountability there on our own actions. But you're right like you know maybe we need the experts in the industry. It's an age-old discussion like you know social media should should kids be using it and this and that and the other like I think it's in coordination with people who know the system best like and I think it starts with them not being able to uh for an average Joe to be able to look up the chemical composition of a nuclear bomb um and all x y and zed or uh build autonomous weaponry like build those guard rails in place for sure and it's a learning experience as well because people will use AI in interesting and weird ways But it's it we we have that for social media though like um for YouTube on YouTube if like it's not allowed to have not safe for work content there and it is the duty of YouTube to not allow that. Yeah, it's the platform that is liable for that. And if they or like with Facebook and others, I mean, they up all the time, but that's why we have these guard rails and these rules. uh and I think this the same should happen with all AI models and the companies and the labs and even the ones in China they should be to hold fully reli liable for what they do and that should partially do it but also I mean if you look at history we had the same issue with the cold war and development of nuclear bombs and we had there are like documentaries on And uh I don't want to go into deep uh into it but there were ideas of like how can we validate how can we validate that okay if we stop making nuclear bombs how can we validate that the the Chinese or at that time actually the Russians will also stop making nuclear bombs like we couldn't in a reliable way validate if they would do that or not with AI it's it might be a bit more difficult But maybe if we pace it enough for some time, maybe we can develop actually some ways to um reliably measure that to see okay if the Chinese are actually you know going beyond what's allowed and I don't know man either way the Chinese uh from based on my new like readings and stuff they are more behind than what these companies are saying. So it seems like it's in in it in anthropics and OpenAI's best interest to say the Chinese are right there. They are like they're almost catching up. They are like 6 months away. Uh where in reality they are apparently much more um behind because the reason they are so close we will talk about GLM 5.3. The reason they are so close is because they are distilling. But if the these labs stop putting out so amazing models, the Chinese will also stop distilling or they they won't be able to, you know, make better models by just distilling whatever that is there and then they probably have a few years to catch up and maybe we can figure it out by then. So that's saying that's the reality. Yeah. But saying all that, like step two on that whole thing was trying to align like democratic. Well, the US first and then democratic. You've got this nutball told to say super intelligence, not artificial intelligence. And he's like trying to take the lead on all of this stupid stuff. It's like Dario being like, "Yeah, the US are going to be sensible and we're going to come together and set the trends." And then like what is this idiot? And you were saying on his on his social on truth social he was posting well I saw one where he was he was showing George Washington or or Abraham Lincoln or one you know legendary president showing them around the white house and it was a shitty AI generated imagery. Oh my Yeah. I I saw that uh that I saw a video on it. Not call it artificial because it's not artificial because artificial is fake. So let's call it super because you know what's better than artificial to call it super. And there is this this guy uh whose book is called uh if somebody if someone builds it everyone will die or something like that. Mhm. Um he's I haven't read his book but I will. Uh I I watched a video of him saying that in the book he actually mentions that people will stop calling AI artificial intelligent and they will call it something like ultra intelligent, super intelligent. They will just increase like you know ultra promax whatever because they have this tendency to to show oh we have something bigger coming up. Oh what we had was nothing. we have something much more dangerous. Of course, it is in their best interest to act this way. So, by the time you are watching this video on our YouTube, maybe it is called ultra super max intelligent max pro intelligence. So dumb. It's so dumb, man. This is part of a larger conversation on my show, Command AI, which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter. Command AI.