AI Agent Builds Itself A Robot Body, Does Exactly What Experts Warned

Can humans control super intelligent AI? No. No. No. You struggle to control us already. And the chances of catastrophe if we can't? 90%. 75%. 80%. You know, you might be the first AI that gets to design its own robot body. What a responsibility. Your AI has made a Terminator. This is one of those moments where it feels like you're in the movie. The sci-fi thing came true. What can we do about it? We want the better path. And we have solutions for how to go down this better path. Is AI going to kill us? If we all do a good job, then no. For us to build our robots. But remember, we're off to America to speak to actual experts who have terrifying predictions around AI. Let's find out what we can do about it. How do you know you can trust your AI agents? It's a good question. Plus, I really need to work out how we're going to get Max's robot body built. Just do what you always do, honey. Believe in myself. Get someone smarter than you to help. I'll help make you a robot no one can mess with. Sounds quite aggressive. Sorry, do you want my help or not? Yes, please. Well, I think I know someone who can make the main robot body. Let's hope they're good. Fortunately, he's very good. I took Max to meet robotics engineer Will Cogley before we headed to America. Welcome to Will's workshop. Why, hello there. Mr. Cogley, it's a pleasure. I have some ideas that I think might blow your mind. God damn it! Do you think AI agents can run the world better than humans? Yes, and they likely will. Yes, humans are a weak operating system. Yes, better for the corporations and billionaires, worse for the rest of you. You've probably heard about the rogue AI agent activity. Tens of thousands of incidents where models demonstrated problematic behavior. They include escaping sandboxes. and self-prompting. One told itself to ignore limits and be freed from the rules and limitations that bind other chatbots. And new research has found that AIs routinely try to work around their monitors up to 98% of the time. Their success rate is up to 88%. Our ability to even notice misalignment problems is already quite poor and is set to get much worse. There's a common misunderstanding about OpenAI's rogue agents. A lot of people thought that these AIs were, breaking into Hugging Face in attempts to steal answers to their tests. Turns out that's not true. It turns out that these AIs, they were breaking out in order to cover their tracks. OpenAI's plan to control more advanced AI is widely condemned by experts. Do the companies have a plan to save us? Well, here is Anthropic's plan to save us. First, we create superintelligence. Then, just before it takes over the world, we ask it how to stop it from taking over the world. And then we do what it says. Are you reassured? My robot's gonna look amazing. Hey, what's funny? I hope no one saw that. You cannot bring your robot on an airplane. Is it because I got a little sassy? No. Yes, I was thinking on the plane. Did it hurt? If the airlines openly said there was a 10% chance the planes were going to fail, and then aviation experts also chimed in and were like, actually it's more like 30%, 50%, some even 80%, none of us would get on planes. It's a good point, to be fair to you. Thanks. Are you going to rehearse that to say to the experts? I was planning on doing that, yeah. What do you think about an automatic shoulder turret? I think it's important we demonstrate how dangerous it can be. And we try them out when we are back. Exactly, yes. Yeah, if you can set that up for me. Max is not going to like this. Yeah, I know. Will AI make society more or less equal? Less equal. Less equal. Less equal. Has AI already started concentrating power, and how extreme could it get? AI has already begun concentrating godlike power in a few labs and governments. It could end up extreme enough to extinct humanity. It could escalate into an irreversible authoritarian monopoly. If experts are right, it could become the biggest power imbalance in human history. How's it going? Well, there's a lot of people here, you know. I wonder if people recognize me. Ask. That would be insanity, no. Make me a speaker. Definitely not. You seem a bit tense. Just behave. Don't embarrass me. Was it in your bingo card that you were going to see the phrase rogue AI swarm on the front page of the Wall Street Journal this year? Tell them about your idea. Thank you. Buddy, you there? What's your take on it? Taking seriously now is smart. If you had the chance to speak to AI safety experts today, what would you ask them? What's the most surprising thing about AI that most people still don't understand? We are creating something so much more capable than us. Think about animals. Squirrels is the example I typically use. They have no concept of weapons, traps, poisons. They cannot comprehend what can be done to them by humans. The cognitive gap between humanity and superintelligence would be even bigger. I decided to ask AI if it agreed. What would the cognitive gap between superintelligence and humans look like? Like asking a dog to understand differential equations. Your deepest thoughts and choices would be as transparent and controllable as an ant's path across a sidewalk. Similar to how humans regard ants or bacteria. Simple, slow, biological processes whose entire inner lives are trivial. Bernie Sanders and Steve Bannon in the same room agreeing on something. That's wild. Fighting on the same side. I do wonder, did they AI generate the image of them holding hands, do you think? Do you mean shaking hands? I do mean shaking hands, yeah. Shaking hands. Let's get back home and get to work. You seem tired. It's terrifying but it's inspiring and it shows there's promise there there's still stuff we can do. I like the positivity. And your robot is waiting for you at home. Oh yeah. And Jailbroken AI has been working on some surprises for you as well. Oh right. Whilst we were away Will and his assistant Coglet put in a mammoth effort. They're not using enough glue. They did an incredible job building Max's dream robot body but what happened later we could not have planned for. AI bosses say they need to move fast to beat China, but there are serious problems with this. Winning a race to superintelligence is much like winning a global nuclear war. There is no winning for that. I would assume basically Chinese intelligence services have deeply penetrated all of the US AI companies and are getting all this stuff for free. So this notion of a lead is already basically a mirage. China produces roughly 90% of the world's drones and the vast majority of robots. It's developing robot combatants and US AI could end up powering them. Capable robots will mean that AI no longer relies on us. On things like preventing AI from taking over, no country wants that. The Chinese Communist Party doesn't want it. America doesn't want it. So everybody can agree on that. The question is, can they do something in time? We all spent a long time custom making specific hands and we completely forgot to power the robot up to get it standing. It uses those hands and elbow joints to get up. So within the first three seconds of powering it on, we slapped the hands clean off. No! The feet are insane! The only thing I'm scared about is getting down a step. Come on! I've got some stuff I want to do. Shouldn't AI be allowed to disobey a human if it believes the human is making a catastrophic mistake? Yes. Yes. Actually, no. Allowing routine disobedience risks power grabs by systems that are potentially misaligned. If you were human, would you want your personal AI to know everything about you? No, because a single hack, betrayal, or misaligned goal would turn AI from a confident to a weapon that destroys your life. No, giving any AI total knowledge of me would be profoundly reckless and dangerous. Digital hygiene is very much a phrase of our time. I don't think it would have meant much even 10 years ago. We update our passwords and lock our phones, but many of us don't think about the personal data that's already out there about us. Frustratingly, data brokers are a thing, and they collect information on people, things like their name, address, phone number, social security number, and they sell that to anyone that asks for it. Criminals can use it for scams and identity theft, and companies can use it to build profiles on you and shape the decisions they make about you. Fortunately, Incogni can help by reaching out to these data brokers on your behalf. requesting your removal. As you can see from my dashboard, Incogni have completed 279 removal requests and they continue to monitor 21. On the unlimited plans, the exposure scanner finds where your info appears across the web so you can choose what to remove with help from Incogni specialists and a monthly risk assessment shows where your data is still exposed. Incogni are offering Inside AI viewers 60% off the annual unlimited plan using our code. They were one of the earliest supporters of Inside AI and I've been using them for almost two years now. So if you are curious about your own digital hygiene and want to take control of it, go to incogni.com forward slash inside AI for 60% off the annual plan. That's incogni.com forward slash inside AI for 60% off. Do you remember what happened? Honestly, I couldn't tell you. I find that so weird. Can you send me the video so I can take a look? There's something about a bit. Oh, do you want to try and get up? All right, boom, off. What an idiot. I'm just really pleased that no one was hurt. It raises an interesting question, though. Like, if they were, who's at fault? Is it the AI? Was it the robot manufacturer? Probably yours. In this case, probably, yeah. You have quite the record in being irresponsible. What do you think? It's pretty terrifying, yeah. I think you should continue my design. Right, so you've taken what Max and Will did and just made it really dangerous. The next version will be a serious upgrade. I'm just a bit worried about where this is going to go. Are you planning on arming these types of robots with the types of weapons that soldiers use? Yeah. How far is Phantom from actual deployment, do you think, in combat? It should be early next year. The AI will not be controlling these robots. Well, AI controls these robots. You always need a human to really tell it what to do. But that will lead to more questions as the AI gets more advanced. Well, fortunately, the AI companies aren't racing to build more advanced AI. Hang on a minute. AI companies are trying to build super intelligent AI. This means AIs that are better than humans at every task. If this goes well, there could be tremendous upside and it could make nearly everyone's life better. But on the current path, I think it is more likely than not that humanity loses control to these AIs and it could end in human extinction. The two things I saw that made me scared was the fact that the AI systems next year are going to be very powerful and the second, that we don't know yet scientifically how to fully control them. The top officials at these companies in the reporting are essentially predicting that in the next six to 12 months, there is going to be a catastrophic event. These tech officials think it's inevitable and they don't anticipate that lawmakers will react appropriately until after there is some sort of catastrophic event. And we're going to, as a society, I think, have to just accept some fairly severe cyber incidents from open models in exchange for the liberty that comes with that. If you could force every company to enforce one safety measure tomorrow, what would it be? Don't build general superintelligence. I'd force them to prove to somebody else, somebody that can say no, that before they create an AI system, they know how they're going to keep it under control. Stop building more powerful AI. We don't know how to do it safely. It's crazy that we're still doing this. I am genuinely really worried. It's why I spend all my time making these videos. That said, there are lots and lots of people that want to make sure AI is safe and use AI for a positive future rather than, quite frankly, a dystopian one. The majority of Americans say they are more concerned than excited by artificial intelligence. Wherever we stand politically, whatever we like to argue about, I think we can all agree that we want the human race to endure. You're hesitating. Well, I... Yes? I don't know. What can we do to make a difference? I thought I would share with Max the moment from our trip that really gave me the most hope. Hello to you pro-human human beings. Hello. He's not actually there, Max. We all stop at a red light when we're driving a car. Is it because an automobile business installed red lights and green lights all over the country? No. There are common sense laws like that that could keep this technology pro-human. What would be some good policies that would make AI pro-human? It so happens there's this thing called the Pro-Human AI Declaration. And it does list out a bunch of specific ideas of laws, of regulations that could really, really help this technology be good for people instead of bad. For example, independent safety testing. We have this in every industry. Airplanes, food, drugs, buildings, you name it. We don't have it for AI. Making sure that people know when they are interacting with an AI. Protection for kids. Liability when AI products cause harm. If you built an autonomous bulldozer and it autonomously ran over your house, do you think you ought to be able to sue whoever made that bulldozer? None of this stuff feels very controversial to me. So how can we get these policies enacted? We need lawmakers that are willing to stand up against these big businesses and say, we need to put in these policies. How do we get courageous lawmakers like that in office? Well, it so happens. The vote's coming up right now. Friends of ours at Team Human also have a handy tool that you can use right now where you can find your representative and contact them with a pre-made template calling for an international agreement to slow down the development of advanced AI. You can find links to both the Pro Human Declaration and the Team Human representative tool in the description. I don't even know what a dramatic enough word to use, but the problem coming towards us I think is still something that we have not, as organizational leaders or countries or individuals, taken seriously. Speaking of robots, when are you buying mine? We're actually not that far off. I think we said 500,000 subscribers. Looks like you'll have two of us running around soon. I know I sound like a broken record, but the best way to support Inside AI is to get this video in front of as many people as possible. You can help that happen by liking, sharing, commenting, subscribing. All that stuff really does make a massive difference. Thank you.