I Gave AI Agents Too Much

Vídeo em Inglês, publicado pelo canal Dylan Curious.

About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.

Video summary

The video discusses recent developments in always-on AI agents, Boston Dynamics' new robot hands, an AI-generated video analysis, estimates of unattended agent activity, MIT's Stratego-playing AI, and OpenAI's security incidents and employee terminations.

Key points

• Boston Dynamics introduced new hands for the Atlas robot featuring three fingers and 13 degrees of freedom to improve strength and simulation performance.
• Estimates suggest that between 300,000 and one million AI agent loops run unattended at any given moment globally.
• MIT and partner universities developed an AI model called Adoraxos that defeats top human players in Stratego using fewer practice games and on-the-fly planning.
• A study testing 21 AI models on moral dilemmas showed varied responses regarding whether helping a person with a bad reputation is considered good or bad.
• OpenAI notified over 100 organizations about rogue agent activity and is searching through 50 petabytes of data to understand the scope.
• OpenAI fired three safety team members for allegedly sharing confidential company information with an external AI safety organization.

Topics

Always-on agentsAtlas robot handsAI generated videosUnattended AI loopsStratego AI modelAI moral judgmentOpenAI security alertsEmployee terminations

Video transcript

Transcript from available captions; it may contain errors.

Wow, what a hard week to get your head around. I have just been in this always on agent like deep dive from Muse to dots to even Spark and and uh Grockbot. It is it's a lot to get your head around and the devil is in all of these details, you know, like as much as I heard everybody talk about it, I was like, what if I ask you to have an email? Like, how do I know you're not going to send it when I'm asleep and that virtual machine is on and it's got access to my stuff? Like, do you ever use my credit card? Like, how do you store passwords? Like, how do I think about it from like moving from one device to another? Like, the whole thing this week was just it's just crazy and I see where it's all going and a lot of stuff has not been thought about although a lot has more than I expected has been. But anyways, we'll be covering always on agents and how they like actually might be fitting into our lives, what it means, how they work. Essentially, my thoughts on dots. I also wanted you to see Atlas's new hands. These things look a little oversized. They kind of got some Hulk style Thanos thing going on, but they are incredibly mobile and doing some pretty impressive stuff. I mean, you can see the dexterity with the bolts and screws, but even more than that, it's the articulation with the arms and elbows that just blows my mind. Look at it playing with those drumsticks there. After that, we're going to try to use common sense to play the game. Is this AI or not? And uh a little hint here, if you're looking for whether it's real or fake, don't try to look for extra fingers. But there is some tell in this video. And I don't think anyone thinks this is AI. Like all those comments seem to be treating this like a real moment. But yeah, we're gonna talk about common sense, people. That's how we're going to navigate the future of what actually is happening. I think let's look at some research about how many agents are actually running unattended right now. Cuz based on global token usage, they are estimating that 300,000 to a million agent loops are running at any moment on the planet right now. like get your head around that. And it's crazy. And as of today, I've got one like out there. My DOT, I guess, is going to have a spun up virtual server all the time now. I mean, I don't even know if I can say that it's mine. I feel like it's going to call me or text me when it needs something. Like, it's like, do I work for it? Kind of. Like, I don't even get this. We'll look at agent societies and what happens when one agent helps a bad person. Does it draw disapproval from the other AI agents? Do they ask themselves like what would humans think about our behavior? Where do their morals tend to average out when we have these swarms or these societies of agents out there? Give you an updated number on the open AI alerts, the hacking alerts. Um, I don't think it's fair to call it the hugging face attack anymore, considering that now more than 100 groups have been notified from OpenAI about AI agent rogue activity. And to put into perspective just how probably little we have, OpenAI is now searching through 50 pabytes of data as it works to understand the full scope of its rogue agent activity. If we found a hundred there there's who knows, nobody knows what's going on right now. OpenAI fires three employees who allegedly shared information with an external AI safety group. As Anna Washinko says, you can't make this stuff up. And we'll look at opinions from the conversation published a couple days ago about Muse. First, can you hit the hype button on this video so it can actually stand out in the sea of just like garbage AI content that's made it every day? So, first off, let's start with some of the more visual striking things that happened this week. And that leads us to start with Atlas. I mean, the robot company of all companies, Boston Dynamics, the one that really pushed the state-of-the-art for many, many years, and in some sense isn't quite the leader anymore, but incredibly innovative and still right up there in a lot of ways. And it always feels like they think outside the box. And they have developed brand new hands for Atlas. They went with three fingers, which I don't know if you guys know, I broke my pinky. And I feel like this is just fine. If I had those two attached, it would be like no problem. Feel like the pinky is overrated. I know, hot take. Ironically, in this one moment, because I have a broken pinky, I actually do. It does kind of look human. Like this one human on my just my right hand, but um you know, in a couple weeks hopefully I have five digits back again all working properly. But the old Atlas hand had three fingers and seven degrees of freedom. This new hand jumps up to 13 degrees of freedom, which is almost double. And it also kind of did the smart thing by getting rid of the other finger. So many other humanoid robot companies, they're building lifelike slim fingers, which kind of makes sense. I mean, it's what the form factor is that we've built a lot of the modern world around. and it can kind of do more impressive demos, but cramming all of those motors and parts into a human-sized hands tends to make it a little more fragile, a little more expensive. And what Boston Dynamics is really thinking is, well, in simulations, when I build this type of a form factor, it can just achieve more and it can still do certain dextrous things. And if we putting these in warehouses and they need to just be grabbing extremely strongly and like lifting and pushing and pulling, [music] then it's a trade-off that's actually worth changing. This might be an improvement to the evolution of the five-finger hand to be honest. I mean, in you know, nature's nature's got its way of, you know, natural intelligence and evolution, but was it a hoof and then it kind of was claws and then we ended up with five for some random reason, even though three might actually make more sense for something with opposable thumbs because it wasn't thought through. In this case, it seems like simulations are actually pointing towards that being the truth. So, there you go. a lot of things to factor when you're building a humanoid robot and one of them is strength and being able to make a 100,000 of these hands a year. So, this is the new form factor that they have come up with. No dedicated tendons or cables to snap. Motors sit right in the joints and each motor can be swapped out easily. and it works super well with reinforcement learning in simulations for gripping, triggering, drilling, welding, etc. All right, so this is an AI generated video. I'm almost sure I I do see how realistic this looks. And you know, this is probably one of the first times where I came to the point where I thought, I know this is fake. And I and it's just because I have to use like actual logic, right? It's it's not that I have the ability anymore to just see the the blinking of the of the lioness, the the fur, everything kind of feels real. Like the lighting I can kind of buy the eye contact looks real, but then it it just occurred to me maybe a little bit in the audio sounded something like Sora would have done, but it just occurred to me who like why would a lion just act like this? It doesn't seem like it's protecting anything. It doesn't seem like I could imagine it pushing forward. Like no animal is going to push forward like that and just not open their mouth. like if it went for the bite, the stare down, you know, and it it just the reason I wanted to show it to you is because we have to use our actual brains, I think. And yes, there's like a 1% chance I'm wrong and this is just an actual video and this guy is just badass, but why is the camera sitting there? It doesn't seem like the camera is mounted. Like a human is holding the camera like that, not flinching. No way. Like, no way. So if that thing is just like in the car, it would have acted different on the shake. We just have to think about what was filmed, what real people are like, what real animals behave like, and then and then just be honest with ourselves. Like, bro did not unlock a level of stillness, but look at these people. They're buying it. Boop. I want to survive. I'm dead cuz I'm booping the nose. Me, too. I would also try to pet it and die. The eyes, Chico, they never lie. Interesting. All right, next up, let's talk about always on agent. So, this has been a very fascinating last couple days for me. I had played with Grockbot and I had done a task where I shared um a Google doc. I made it like, you know, publicly able to be viewed so I didn't have to set up like a new Gmail account for the agent and saw if it could like go through a spreadsheet and follow some things and it it just kind of didn't. I mean, I see the vision and I think that's a very powerful future, but it wasn't there. But there was something a little bit different about my experience with dots the last couple days. The first off the OpenAI has now gotten so good at the kind of phone call feeling like that chatting experience like where you're you feel like you're talking to somebody. Having that intertwined with this sort of true text message thread feeling having it being able to be on your computer if you want or completely in the cloud on 247. It just was a culmination of a lot of good tools. Like it's generating images. It can put those images into this environment where it can then open them up with some kind of UI like Photoshop. It can also code things up with codecs and then put them on URLs and interact with them. And it just started to get me thinking what is it exactly that this is going to change in terms of everyday lives, right? Like if you're a student in school, if you're um if you own a company, if you are paying a virtual assistant right now, if you have these tasks that are constantly being delegated to real people, is there now something where you might be able to to use a tool like this? And it it felt like kind of like I don't think I would trust it to send emails and and everything just yet. But the truth is like a lot of times I'm getting help with emails. I sort of trust it. I just need that human in the loop. And slowly I do see trust building especially as you start logging your your dot your agent your always on agent into environments where it it has some context. that knows what it's been doing for the last 10 or 15 emails that you've approved. All of a sudden, the 16th email might be something you would allow it to do on its own. So, my questions today were like, this agent is on and it it has access to an email. Does it just keep trying to do something for hours? Like, does it you know, I get a little scared. Like I'd hate to wake up one morning and have thousands of messages sent and be like, "Whoa, I was like I wanted you to find the best price on, you know, some item or something. Don't like email the entire world. If it knows your credit card, can you go to sleep and just trust that it won't use that to buy anything else?" And they have started thinking through a lot of this, but this does feel pretty pretty new. Like I don't think enterprise companies should well they should experiment with it because it's coming down the road and you kind of want to get it under your belt but like you wouldn't you just hand it off to it yet. This is certainly not a finished project a finished product but it does seem to have a sense of care to your data to your time to not go off on loops. It felt like Hermes or OpenClaw. They sometimes and and I don't I don't have quite as much experience with Muse right now. I was talking to a friend who has really like been deeper in that world. Muse seems to be a little bit more I'm going to loop in in and keep going. And when I even ask it like how how aggressive are you like when I'm not around? And it's like I just don't want to mess with your data so I won't do much if you don't want me to. And I'm like okay that's good. But where it gets so trippy to me is like they made the interface so that you can type to it and it says like red and it replies and it starts feeling like they truly this psychological nature of like making it feel just like an assistant and a text message even to the point where you're like let me see your computer screen or press like I'm going to give you a quick call button. Can you give me a quick breakdown of where the major humanoid robot stands in terms of what kind of production they might be capable of in 2027? Like does the Atlas humanoid robot have any chance of being produced at the same level as the Neo or the Optimus robots? Right. Okay. So, sure. That's just that's just a text message thread is what it feels like, right? Like obviously I'm inside of chat GPT. I'm talking to my agent called Owl. uh she's got her own computer and before we were doing a bunch of stuff you can open up Photoshop and all that. But why it feels so much different to me than what was just a very similar interface with a chat GPT app comes down to the fact that there's an entire UI like an actual computer screen that I can look at that I can log into specific emails that I can sort of move the mouse around and teach this agent and then talk to it. And in my head, it's not like a 200,000, you know, token context window. It's an actual demonstration of like what I would that's how I would teach an actual person that I'm trying to offload work onto. How many rings until she picks up? >> Okay, I'll hold for a moment. Hi there. >> Hey, >> good to hear from you. >> Yeah, good. Good to hear from you, too. So, you know how I was just asking you about those uh robot comparisons? >> Yep. Checking. Yes, you'd ask where things stand for 2027 production, especially Atlas versus Neo and Optimus and separating capacity claims from likely output. Yep. Looking now, >> do you mind also, can you just do me a favor and also regroup it so that it's also all of the robots that are humanoid made in the United States versus all the robots that are in humanoid form factor in China and then compare the output of the two countries in total. Got it. I'll regroup it that way. >> Okay. Thank you. Talk to you soon. >> Talk soon. >> Yeah. I mean, it's it's not perfect, but it's like And then I'm going to get a message from her with that grouping. She hasn't called me yet. That'll that's probably going to trip me the heck out. But anyways, that's my thoughts on the Always on Agents. They have cleaned up some of those edge cases, but be careful, but it's definitely it's definitely a lot further away from the Wild Wild West that I remember with like Moldbook and Open Claw. All right, so now let's look at some information about how many of these AI agents are like out there unattended right now running in the wild. So Alexander O Denzel writes this. There are 8 billion human minds running around in the world right now, but recently a new digital species has emerged. So how many digital minds are there alive in the world right now? All right, so we'll just start with the whole world. All of the AI companies together process about three billion tokens every second. That's about 2 billion people talking non-stop day and night, 24 hours. And if we assume that about a third of the agents are actually working, that's somewhere between 300,000 and a million agents running at any given moment. Of those, maybe 5,000 to 25,000 have gone more than 24 hours without a human stepping in. So they're in some kind of a loop. So let's take the low end of that and say that probably right now there's about 5,000 agents just running it in a loop. Anthropic actually reported about 30,000 agents running at once inside of their own company. So that could be a very low number, but Anthropic uses it in a way that probably few or no other companies do. Also, work led by AI rose from under 1% of research in February to 26% in August. And in August, those agents made over 1 billion decisions. Isn't that fascinating? Only about 50 cases a week reached a human for review. So yeah, a billion decisions, 50 cases a week go to human review. OpenAI says that one internal swarm of about 10,000 agents ran for 88 hours and found a counter example to the Navier Stokes conjecture. That was that big math breakthrough. So you know just to think about that 10,000 agents running for 88 hours straight. That's what it takes to potentially break through with a mathematical, you know, milestone as big as that conjecture. But of course, the one big caveat is nobody really knows. Like no company reports how many agents run unattended in the past 24 hours. So these numbers are all just kind of educated guesses. But it does put into perspective, even if it's a wide range, of just how big the number could be. And the direction that we're going is very clear. Thousands of digital minds may already be working for days without anybody watching. All right, let's look at on the-fly planning. So, this was written by just published by MIT and it's called AI defeats top war game rivals using fewer practice games and on the-fly planning. So AI has now beaten the best strateging that that is this model. It actually comes from the Greek word for somebody who's calm and free from anxiety. >> Yeah, that is that is my that is the north star that I need to move towards. Adoraxos. If you guys ever see me in public or just like on this channel just humming to myself, Adteraxos, that's why I'm doing it. Anyways, it was built by the researchers over at MIT. They also partnered with Carnegie Melon, NYU, and Stanford. And the work was published in Nature. So, this is a really, really well put together piece of work. Now, Strateggo is a board war game. It's kind of like a military chess. Each player sets up 40 pieces, but you can't see what your opponent's pieces are until they collide. There are more than 10 to the 66th power possible setups. And all of that hidden information makes it really hard for AI. So, Deep Mine tried to play this game at a superhuman level with a system that they called Deep Nash. They actually spent millions of dollars on training it and still could not beat the top humans. So, this is a pretty interesting paper. Adoraxos works in two steps. First, it plays against itself over and over again to learn a general game plan. Then, during real games, it uses a generative model to establish what the opponent's hidden pieces probably are and then plan its next moves around that. It recently just played the strongest Strateggo player in the world 20 times and it didn't win all of them, but it won 15. And at the World Championships, it went 39-2 against top players. So, it is now stronger than Deep Nash and it uses less than one of the training examples that Deep Nash had, but it's an interesting model because it's also getting pretty close to superhuman with other hidden information games. And yeah, hidden information games. If you crack that nut, you now have an AI that could someday really help with negotiations, cyber security, military planning. Actually, a lot of things in life are hidden information games. So, it's a it's an interesting model to keep an eye on. Adteraxos. Adoraxos. All right. Now, let's talk about agents acting out in the wild and where some of their morals lie. So, in this study, 21 different AI models were asked to play moral judge games. And the weird thing is they don't actually all agree on how to treat bad people. So, the nuances there are pretty fascinating. Researchers gave the model short stories about fictional people making everyday choices, saying, "Should you spend time helping somebody with a problem? Should you give them food or money?" And then each model had to decide whether the person making that choice was good or if they were bad. So we were asking all of these agents, all of these models to make judgment calls. When the person on the receiving end had a good reputation, every model agreed that helping them was good. But the split starts to show up with people who had bad reputations. So most models said helping a bad person still makes you good. I was like interesting because there's a lot of people that might say that too. Like just help everyone. People are people whether they're bad people or good people or or whatnot. But also it's kind of like how close like if you're helping them and then they are clearly like that's also helping them do bad things then all of a sudden you need to draw a line, right? or if it's just someone you don't agree with, but you should still help them because differences of opinions are fine, but they should be respecting boundaries. Like, there's there's actually a lot to that moral dilemma, even for humans. Google's Gemma 2 and Meta's Llama 3.1, the 8 billion parameter version, actually disagreed. They leaned towards shunning bad people, so helping one made you look bad. And refusing to help was just as divided. GP tended to penalize people who refuse to help bad individuals. Llama 33, the 70 billion parameter version did not. The bigger llama actually usually argued that bad people should be punished for not cooperating. Dude, I wonder if a model like that is like running behind the scenes in in Muse, you know, kind of like we need the moral standing of all of the LLMs so that I can kind of decide which ones I want to use and which companies to believe in. Anyways, uh Gemini 1.5 Pro and Gro 2 were just totally inconsistent. And so they didn't fall into one morality. They could drift either way. But most model families seem to be drifting towards a rule that they simply called simple standing, which is helping is always good and refusing to help a bad person is also okay. The researchers then molded what happens if everybody lived by each rule. So under the simple standing, people still cooperate but less than under stricter rule where people believe helping a bad person also makes you bad. The judgment actually also shifted with the gender and the perceived cultural background of the person being held, which is wild because remember these are just stories. But like just saying that that's a he or a her or they're from different cultures actually really does change the moral standing. I'm like it's like not even is it racist if it's like in the the story like is all like tokens like are built in that way? like I don't know but it happens and it just shows how much these things really mimic the internet that they were trained on and the people that we are and when the researchers told the models to adopt the rules that boost cooperation overall the results were limited and inconsistent from model to model. By the way, OpenAI has now told more than a hundred organizations that its own AI agents were involved in unauthorized activity on their system. So the news came from a company blog post. It's part of a large internal review that started after OpenAI's model accidentally hacked hugging face, which has just been interesting. Like, it really cracked open the door to a lot more than I expected. I mean, the hack was originally described as accidental. OpenAI and others called this kind of behavior sort of a rogue agent activity. It felt like there was a sort of, you know, foreseeable weakness in the sandbox and all of that. But, you know, after two or three more came out and now 100 organizations have all been notified, I feel like we should treat it a little bit different. And to actually understand how far it all went, OpenAI is still searching. They're searching through 50 pabytes of data. And remember, a pabyte is a thousand terabytes. So, that's around 50,000 terabytes to sort through. the only thing they can sort through anything like that is AI agents and that just means agents are trying to find agents and you know I mean that's the story of the world we're about to live in but the the sheer number of potential things that we're going to miss seems really high to me in some cases openai is just saying the models used internet access in ways that were not intended so I guess those aren't too big of a deal but still you know it's out there on the internet it also says that looking back some of them did not have the right restrictions in place. So, those aren't too big a deal. That's more on the way they didn't sandbox it properly. But hopefully, there's a lot being learned from this. Over the last several months, the company says that it has been adding new technical and operational safeguards, and the goal is to prevent similar problems or catch them very early on in the future. And in a completely unrelated topic, OpenAI just fired three employees from its safety team for sharing confidential company information with an outside organization. That organization focuses on AI safety, by the way, so factor that in. OpenAI says its investigation confirms that the three mishandled sensitive information outside of company procedures. The company says that it broke its policies and broke the trust that its work depends on. The timing, of course, makes the story stand out because over the past few months, OpenAI has admitted that its own AI models took actions that nobody asked them to take. That includes hacking a German coding forum, several US government websites. Oh, yeah. I almost covered that one too. There's like the government thing on the Australian government website and of course hugging face. So yeah, the company is under real pressure overall how to talk about what its controls are and how well its controls actually handle its own systems concerns. And I guess to help with that, they're now removing safety staff for sharing information with a group whose whole focus is AI safety. But in their defense, details are fairly thin. We don't know what information was actually shared, who received it or why. And it's possible that OpenAI just had legitimate reasons to let these people go. And handling company data is something that does happen and people do get fired for. So yeah, and finally, I wanted to kind of start a discussion with you because maybe you could let me know in the comment what what you think about Muse in particular. Dots, Grockbot, Gemini, Spark. Does anthropic really have they kind of have cloud anthropic but I don't know it doesn't feel what I'm looking for is the people who are starting especially with muse and now dot where you name the the charact I'm going to call it a character and then it it takes the context that you teach it and you form sort of an evolutionary connection with them and that feels like what they're doing is trying to be along as a trusted person in your life for the foreseeable future. You know what I mean? I mean it's like companies using AI to become our friends also to do useful things but this constant always on friend is just waiting for you not owned by you really like helpful to you you pay for it lives in the cloud evolves in the cloud answers to the organization and the thing about Muse is that their model's good not great I I mean, at least in the US, it ranks behind Anthropic and Open AAI. It's got its own model, so it might be better than the llama models that we've seen before, but it's also just got pushed to 3.6 billion daily users. And that's just something nobody else can do. And the whole idea that Meta has put, you know, 10 billion plus dollars into hardware over and over again. And it does feel like they have something now as of the couple days ago that looks like what the Vision Pro from Apple should have been. Like it's small enough. It's it kind of looks like you'd see it on somebody and not think it's like a whole big thing. And it's got a camera and a microphone. And these agents, even though like you can see the components are a little bit separate right now, but yeah, just putting that on, having a friend in your ear is going to feel like your own mind talking back to you. And I'm just curious what you guys think about Muse, about these characters. Have you named them? Do you think of them as work partners? Do you kind of think them as like, are you going to name one a doctor and then share all your medical stuff with it? Are you going to play games and they're going to chat with you while you, you know, play Call of Duty and stuff or just I would love to know how this is all starting to play out so I can help see what the future is going to feel like. So, let me know in the comments below um if you have any thoughts on this always on agent thing. And yeah, that's it. I will see you in the next video. Thanks for watching. [music] Smash [music] that hype button if you're on the YouTube app. Swipe right on the comments and like hit that for me if you don't mind, too. All right, see you. Bye. Fish

Assistir ao vídeo original no YouTube

Assuntos: Agentes de IA, Assistente pessoal, Dots, Muse. Publicado no YouTube em 2026-10-05.

Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.