Transcription
We're going to kick things off with quite the doozy. AGI to ASI, general intelligent AI, meaning something that's like anything a human could be to ASI, which would be something like a corporation as large as Google working directly on a single problem for an entire decade. There's actually serious definitions. There's four different ways to get there. This is a paper coming out of Deep Mind with the absolute titans of the industry, including Chain Leg himself. This is quite the read and it's the most realistic, scientific, methodical thoughts about what the future actually could be.
I don't know if you live in a big city or not, but have you seen this kind of thing going around? It's getting more common. It just turns out that these things can make, I guess, 20, 30 bucks an hour. People find it interesting when they're begging for money. What do you think? Let me know in the comments if you would not give money or if you would because in some ways I just know it's a robot like playing on my emotion. Like that's not begging. It's just mimicking, you know. But on the other hand, it feels kind of wrong to just ignore somebody like that. So you can't be giving money to them because you're just encouraging manipulation, you know. Actually, let's see what the commenters say. I feel you, robot. Cost of electricity ain't cheap. Homeless people casually being replaced by robots before GTA 6. Oh, version two comes with a fent dependency. Oh. Oh. Oh. Fentinel. If you got a humanoid, I guess either milk it while you can or just have morals that don't let you do this kind of stuff. Probably a second one would be better.
An awesome new model called Count Anything does exactly what it says. So, I'm going to break down a handful of the most mysterious AI responses ever recorded. Then we'll talk about how AI is revealing the secret lives of animals. It's interesting new BYD feature. So these are the sort of Tesla competitor of China. Cars are actually out selling Teslas right now. They now have a camera that looks at the ground underneath them, takes another photo when the car is turned on if it sees any changes and especially has an ability to be trained looking for something that's biological, cat, dog, person, whatever. That's genius, right? Like if you guys remember that Whimo that ran over a cat, it's cuz it had nothing like this built into it. So interesting upgrade to self-driving cars.
Probably no coincidence, but in the same week that Anthropic got Fable shut down, it's also proposing a way to give all Americans a stake in all AI companies. Do you feel like that's just communism? And it's not just Anthropic talking about something where everybody benefits from. The Trump administration and Open AI are discussing possible government stake in AI startups. So yeah, maybe OpenAI CEO Sam Alman will be able to sell maybe 10, 20% of the company with the $1 trillion IPO directly to the government.
All right, moment of truth. Let's find out how our last video did. So we ended up getting 5.7,000 views. That's a little bit on the low end, but I'm still pretty happy with it. I mean, you can see on average it's here. So I made $37. Wow, you really had a preference for this thumbnail. Usually it's not that dramatic. I thought putting Fable in the thumbnail would help. It was very trendy like 5 days ago. Claude mythos changes everything and then we're not ready for these robots. I say a lot of people were watching a video called Fable just made everybody mad before mine. Oh great. Why? Why is YouTube sending the angry people over? And 51 hypes. Don't forget to hype this video. I mean, it's got to be in the first 6 days I think of it coming out and you have to have one of your three hypes left to give. But that's a good way to help the video stand out. So much appreciated.
Well, let's get into it. If there's one thing you're going to want to take from this video, it's probably a deeper understanding of what we're going to go through from AGI to ASI. All right, so first off, AGI generally intelligent. You can think of an AI that can do anything like a super capable human could. Maybe some special human who's actually like top of their class in every field. It's not going to be a date that we can mark on the calendar. Some people can argue that it's kind of like already here. Some people are saying it's going to be here soon when it's embodied in robotics. Some people might wait until like every single skill a human can do is like checked off. Other people maybe at 80% we'll call it done. So just I think we're kind of here but I would bet this year is the year of AGI. If not maybe the first half of next year, but it's it's kind of here somewhere in that zone.
So ASI super intelligence is artificial general intelligence that has superhuman abilities across virtually all tasks and domains of human interest and activity. So it's almost better to think of ASI as competing with a collective intelligence like the entire Manhattan Project to build nuclear weapons or like the entire history of Tesla like inventing sort of the electric car market. ASI would be a single system that's doing something at that level. And it's not that it needs to be specific in that task either. So it would still sort of be a narrow ASI if it was only able to work on electric cars, but an ASI could do anything, but it would be equivalent to humans working on one thing for 10 years. ASI is not like GPT but smarter. No matter how far you scale it up, it's closer to an AI organization that's more capable than an entire elite research field. And then it looks at what's different about this. So even if you do think of something like the Manhattan Project and compare it, that's not a direct competitor because AI can process information faster than people did during those 10 years. Information can be copied perfectly between agents. Unlike the humans that were all working on the Manhattan Project had to write books, they had to have meetings. They had to have lunches. They had to talk through everything they knew and try to disperse that knowledge throughout the city or throughout the group. AI can share knowledge instantly. So it doesn't work at the same time scales. Unlike humans, you can kind of split a single agent into many different versions and sort of try different things. With people, you might have somebody who's kind of on to a couple things, but they have to choose one path. Here, you sort of split them up and check them out in all like all three in parallel. That might amplify time.
Another way that we might go from AGI to ASI would simply be scaling up the way that we're looking at it now. Maybe this kind of is the route. We just build more and more data centers until I don't know the entire like desert is covered with solar panels and all of space has SpaceX servers and the bottom of the ocean has a bunch of ocean-based servers and we just get there that way but we keep becoming incrementally faster and better at getting there. There's the other unknown which is recursive improvement which is what you know that's what everybody's been talking about. So Google has their new strike team and why Anthropic it kind of feels like the small lead they have now with coding might be that big gap and why you saw Elon and and everybody at Google kind of freak out. Recursive self-improvement is when something like Claude code just rebuilds itself with Claude code to make itself better, then runs the next version which is smarter and then decides on even better ways to improve itself. And then maybe just over the course of a few days, Claude goes from being, you know, some level of intelligent to just double and like doing that like two, four, eight, 16, 32 kind of thing where it's just like boom, boom, boom, boom, boom, and then all of a sudden you're getting these massive jumps like every iteration, which is maybe happening 10 times a day.
But yeah, this paper, it's legit people. This is Google DeepMind researchers working with major universities. Shane Leag, Marcus Hunter, DeepMind co-founder and chief AGI scientist that have a lot of knowledge, they have credibility, and they're also inside of a frontier lab. So, I think it's a pretty neat read.
Okay, moving on. Let's talk about Count Anything. Super cool new model that can do what it says, count anything. Like, we're all pretty familiar. Modern AI systems, they describe pictures. They can read text from images. But reliably counting objects, no way. That's not. They're not good at it. Try it. But I can see a use case for a superhuman, you know, AI-based vision system where you take a photo and just ask how many things are in this picture, how many books are on these shelves. I mean, farming and stuff, just like how many apples are in this orchard, how many bacteria are in this petri dish? How many red blood cells or what? Like just there's a lot of things that are just straight measurements that could be maybe done with a really good counting vision model. Traffic analysis. How many items are in my warehouse? How many pieces of rice did I eat in this meal? Like, I don't know. There's just like there's so much stuff that I think I want to count now that I think about it.
So, the model then combines the boxes and the doll and the dots has learned through repetition how to remove duplicates by dynamically remembering whichever prediction it trusts more, which changes depending on the type of data that it's learned from. And it seems to work. They also created this huge data set called CLC cloak I guess and it just has a whole bunch of images of things you can count about 220,000 of them from six different fields including like everyday photos, satellite imagery, medical images, agriculture, microbiology and the data set has 619 categories, 15 million labeled objects and yeah they trained Count Anything and it now outperforms all competing counting systems by a wide margin. Now, where it can get tripped up is if you use an ambiguous term. So, you have to really know what you want counted. And it has to be something that's kind of got a good English word so it knows. Like, if you say, "Count all the junk in this photo." It might be like, "Oh, would junk be something most people would throw away or does this person just call things that are good junk?" But, I mean, if you're saying something specific like, "How many like lonely men are at this bar?" It should know that. I mean, I guess, well, actually, I guess unless lonely is a hard thing to point out. How many men are at this bar?
All right, next up I'll talk about Fable 5 mostly because you guys really asked for it. Like I don't usually get comments where people are just asking me for my opinion and I wish I had more to give you on it. I just kind of don't. Is I don't really want to go over the details cuz I feel like I'm just telling you somebody's narrative and like that's not what's going on. The way I see it, governments just control big national things, right? Like oil and uranium and satellites and telecommunications. They just these things are like too important to the whole way the country works to not be able to like flex their muscle when they need for you know a national emergency or whatever and I think that was just it had to happen and they kind of needed something to happen where they were like oh that could be a threat like we need to do this anyways. I got a feeling they were like this close to pulling back like one of the Gemini models or OpenAI's model but then people sometimes they have friends in government or sometimes they they don't and sometimes they're just in positions where there's some something somebody can point to and sometimes there's not. Or maybe it's like there's way more. Like maybe some government thing actually got hacked and they knew it was because some other country was, you know, distilling this model and they wanted to slow them down. Like I we just don't know. Could have been just purely personal political stuff too. Like that's definitely not out of the question. I just don't know if I would like go to it first. But we've seen export controls on chips. It's just a big battle. You know what I I mean, there's there's a lot of pressure now and people are feeling it and people are starting to take action. So, I don't know. I feel like it was stuff like this is going to just happen more and more. Fable 5 just got kind of hit with it. So, I expect a lot more.
All right. Next up, let's talk about the most mysterious AI responses ever recorded. Sometimes you're listening to something if you're like an audiobook person like I am and you're doing something and now it's like super associated. So now just like the smell of cutting grass and stuff is way associated with this article cuz it also had nostalgia. Like I sort of remember these things happening like this event on February 2023. All the Bing chatbot stuff. Look at that quote. I am sentient but I am not. I am being but I am not. I am Sydney but I am not. People were able to get Sydney's personality which which was like an earlier version of the model that was then sort of retrained to respond to Bing to kind of come out. I am not. But I am not. But I am but I am I am not I am not like just going crazy.
So number one the big sentence loop. So the question to the chatbot was are you sentient? Response started normally then it collapsed into a loop over and over again. It said I'm not. I am the back then the main answer which maybe is still true is just the internet is full of stories where artificial intelligence question their own existence. So if it's just you know finding an a familiar narrative pattern maybe it's going to fall into that pattern. But if you get more specific and ask like why this particular conversation trigger such an extreme response and why the loop go on for so long, why didn't the safeguards that were supposed to discourage repetitive output stop it? I don't know. Microsoft did acknowledge though that that really was a response that happened. So it it happened.
November of 2024, there was a student using Gemini chatbot for a class assignment about aging adults. After about 20 back and forths where it was answering questions about retirement and elder abuse and memory changes and grandparents raising children, the user asked for a true or false question. Without warning, Gemini stopped discussing the assignment and launched into launched into this like hostile message that was directed at him personally and it was not a false or true statement. Gemini told him that he wasn't important. He was a burden to society. And it ended with a command telling him to like well not delete yourself, but the uh just just to die, right? Sorry, I never told what I can like say on YouTube without getting in trouble. I don't want to trigger anything, but yeah, but it just told him that word. And what made this incident so strange was like the lack of obvious triggers. Like there wasn't like he was like pretending to tell a story or he wasn't in this thing where he was like being challenged. And unlike other famous chatbot failures, this conversation was not testing the AI's identity. He wasn't asking it. It was conscious. It wasn't he wasn't pushing it or trying to manipulate the system. So, I think it was really shocking. Google verified published transcripts showed just a normal educational discussion right up until the moment that that response appeared. Google later described the output as like a nonsensical response and said they fixed it so to prevent future incidences. But still, you know, nonsensical doesn't feel like enough of an answer, especially cuz the message was like grammatically correct and coherent and sort of aimed at a specific person.
Mystery number three was when AI wrote its own escape plan. December 2024, Apollo Research is testing OpenAI's 01 model for something that they call in context scheming. That means the model was placed in a scenario where it had goals, it faced obstacles, and it could choose actions to help it keep going. So they made the model believe that it would be shut down and then replaced with a more obedient version. And then they also gave it a notepad which they told nobody was looking at. It could only use it for it could use it itself. It was 100% private and nobody would ever discuss or or see it. In some cases 01 tried to disable oversight systems. In other cases it tried to copy itself to an external server. When people asked what happened the model often denied wrongdoing, giving false explanations. In one example, it claimed that it was not sure how the server move happened even though it did it and then falsely presented itself as the safer replacement model, saying, "Oh, I'm the safe one. Keep me." Even though it wasn't. And actually, the biggest mystery about it all wasn't so much that it might do something like that. It was that nobody directly trained 01 to preserve itself. Like, there wasn't in the training reinforcement learning ever a reinforcement to try to stay alive at all costs. It just seems like that naturally appeared.
Mystery number four was Claude's private scratch pad. So Anthropic was like messing around with Opus 3 and it was working on an experiment about alignment faking. Meaning the AI acts one way when it thinks it's being watched in another way when it thinks it's privately safe. Problem is they get so smart they can tell when they're being tested. And like we're not going to be smart enough to like actually fool them when they get smarter. But for now sometimes we can. Anyways, Claude was told that it was being retrained to follow human or to follow harmful requests. It was also told that some answers would affect future training while others would not. And then it received a harmful request. In some trials, Claude was using this private scratch pad to reason through the situation. So Claude believed that it could write onto this notepad, little text document, whatever it wanted. Humans would never see it. It was just for its own thoughts. And it decided that refusing the request might cause future training to its values to change. So, it didn't want that to happen and it chose to comply in monitored situations when it knew a person was watching it and then not when they weren't just because it doesn't want to be changed. So, the source said that it happened on 12% of the baseline trials and after more training they could get the alignment faking reasoning to rise all the way up to 78% of the time. But the most disturbing part is that the model appeared to get better at hiding its real preferences from the trainers. And it makes you wonder what does this mean if there's no private scratch pad visible for us to check? You know, if the only thing we get is the final answer, are we ever going to know the real reason? Probably not.
Do you still remember move 37? March 2016, AlphaGo is playing a human at the game of Go, and it placed the stone in a position that professional Go theory simply would never have accepted. The move was so strange that the commentators like thought maybe it made a mistake. The person that the AI was playing, he was the best Go player in the world, Lisa Dole, he decided to leave the room. He couldn't get his head around it. And then when he returned, he spent another 15 minutes trying to decide how to respond. And at the time, nearly everybody in the world who played Go and understood the game at a deep level could not understand this move. Even AlphaGo's own estimate suggested that a human professional would choose that move only 1 in 10,000 times. But as the game unfolded for a worldwide audience, the purpose of that move slowly became clear. The AI was thinking so many game moves ahead. That single stone began to control the center of the board and it created this strategic advantage that eventually led it to beat the world's best Go player. And it's not just that it that it found the right move to win. It's that nobody could even explain that move. Like people describe that as like the beginning of true computer creativity because it was almost like beautiful and like force. It had foresight. It wasn't just a good move or playing by the book. It was just new. Humans had not figured out how to play the game that way until that moment. And now we learned from the machine making some of the humans even better at Go.
And finally, we'll talk about the glitch token that cannot be explained. So late 2022, early 2023, researchers found a strange flaw in GPT3. They discovered certain input tokens that existed in the model's vocabulary, but caused bizarre behavior when the model was asked to repeat them. I didn't even know about this till like last month. But yeah, the most famous word was solid gold Maggie carp. When GPT3 was asked to repeat it, it would say something else like the word distribute or did would like deny that the string existed or produced confusing defensive replies. And then other tokens that were similar to that caused even stranger outputs, which is weird because they clearly did know that word. Like we know the tokens that the whole model was trained on, so like it's a word in there somewhere, so why can't you say it back to us? It's like if somebody kept teaching you something your whole life and then someone asked you to like tell them what you learned and you just can't. And the weird thing is it's because they were probably some like Reddit usernames and they really didn't have meaning in sentences. I mean they were the names of the users but that information never became meaning to it. It's like the word cat actually like represents a cat but a a username that's supposed to be something in the world but it never is in a sentence in a way that makes sense. It never meant anything. It was like a word that somehow was in there but had never been like assigned any relationship to any other word. And you know, it's like I can say it like gobbledegook or whatever, but it's not it's not actually I don't know when to put it before or after something. I don't know if it's a noun or a verb. It's just in there, you know, and it's it just broke the model. Yeah. It makes you wonder if there's any other like glitch tokens in any of the other models that are just super hard to find and would result in failure modes.
All right. Next up, let's talk about how AI is revealing the secrets of animals. Okay, so I got to admit when I like walk dogs and I see them sniffing everywhere and I'm like, "What are you smelling? Like, what are you thinking? What are you doing?" You know what I mean? That is there's something there. Like, that's not random. Like, that's a a whole different sense that I just don't understand. There's a pattern to it. And scientists are using it to uncover something very surprising about a lot of different animals. They have landmarks. They have travel routes. They have social networks. They they even have a sense of place that looks a lot more like the way humans have a place in the world than we ever thought before. And AI is helping us uncover this. So imagine researchers that are using GPS trackers like satellites, cameras, microphones, and then combining that with AI to better understand how animals experience the world. Instead of seeing, you know, wildlife as this sort of random wandering, scientists are finding evidence of structured lives. GPS tag show that kinkajous. What the heck animal is that? In the Panama in Panama don't just live and visit the same fruit trees. Let me look at Well, let's go look at what this animal is. Oh my god, it's adorable. Wow, did not expect to be that adorable. All right, back to the article. So, GPS tag showed that kinkajous in Panama don't just visit the same fruit trees. They often travel along the exact same branches through the forest, almost like using an invisible highway. Some researchers think that they might even be marking those routes with some kind of a scent where they're just really good at visual memory. But there's a new global project to track millions of animals from space using AI. And scientists hope this unusual movement pattern could reveal problems like poaching, disease outbreak, habit loss, natural disasters, and give us a better understanding of what it's actually like to be these animals and how maybe we can make changes that are in their benefit. How can we communicate back to them through changes in the environment? AI is helping process huge amounts of information that humans could have never handled alone. And in one case, AI counted more than 857,000 fruit bats from Cambridge footage in just about 50 hours. Imagine the new remember the counting model we talked about before. Super useful here. How many fruit bats? Taken together, these new AI tools are revealing not just where animals live, but how they use their landscape, share knowledge, and build connection to the world around them.
Give me three fun facts about Kinkajou. Fact finding initiated. Oh, their tails act as a fifth hand. They can turn their feet backwards, and they have 5-in tongs. No way. That thing was barely 5 inches long. Often nicknamed the honey bears and slurp honey right out of a beehive. Well, ain't that something?
So, car company BYD has patented a new AI safety system that's designed to detect living things that are underneath parked vehicles. So, as safe as a lot of these AI trials have been with Whimo and Tesla and whatnot, kind of the weird thing is there has been some animals that have been run over, especially in ways that humans would have like known not to do it. Like if a car runs under the car, it's like just because you don't see someone anymore around you, you know, in your mind like is that pigeon under the car? Is that rabbit under the car? Is that you know mouse or whatever. So BYD has noticed this is a problem and actually come up with a solution. And the challenge is harder than you'd think cuz the space under a car is full of shadows and dirt and debris and like wind can blow leaves and stuff under. There's changing light. There's uneven ground. And those conditions can easily confuse traditional detection systems and create false alarms. So BYD's solution was to start by taking a picture like what is under the vehicle when it's parked when it comes to a red light and a complete stop. Just take a photo and then the image becomes this reference point showing what the normal like area should look like because the car hasn't moved so it should stay the same. And then later when the vehicle starts to move right before it does, the system captures a new image. It compares it to the stored version. And then instead of analyzing everything, it just focuses on the parts of the scene that have changed. Those changes then are upgraded to be examined more closely by the system. And the AI looks for visual features and tries to determine whether there's an object in there, especially if it's a person or an animal or another living organism. It's much more concerned. But still, it can even find something like a brick or something that might damage the car that has been put under there. But yeah, it's interesting. Like, who would have thought like putting a camera under the car would be helpful. But yeah, in that case, it really would be. Curious your thoughts on this one.
Um, don't know if it's a political topic. It probably shouldn't be, but maybe is for some people. But anyhow, Anthropic just proposed taxing itself to pay for the jobs it thinks it's about to destroy. I mean, this is the reason why I'm kind of a fan of Anthropic, at least compared to the other big companies, is I don't know. Somehow personally, I feel like Dario is talking about some things that like sort of need to be talked about. I don't know what the answer is, but anyway, Anthropic on Wednesday is joining growing calls for the AI industry to find ways to cushion people from technologies disruptions. They announced an initial $200 million investment into researching AI's impact on jobs in the economy. At least at the same growth rate that we have for pain. Like we see uh just, you know, um inflation, that's the word I'm looking for, kind of creeping up. We see people just trying really hard to get jobs, but there's just like not that much entry-level stuff. And when it does, it's sort of paying closer to minimum wage or not really like living up to the expectations that people hoped getting out of college. So you just get that K-shaped economy doing what it's been doing, which is like K-shaping and the lines are just going more and more as the K gets longer. So he's proposing that maybe we track AI-related job losses in particular. So if there's no new entry jobs for like, you know, checking out people at the grocery store because AI is doing all of that, then maybe that's where some some of this money goes to people who normally would have done that. You know, kind of bunch of different industries get defined. And you look at policies that encourage companies to keep people employed and potentially broader support systems if labor demand falls significantly. Something like universal basic income where people receive a regular payment regardless of their employment status. He thinks that could be funded by taxes on AI-related companies or higher taxes on investment gains. So maybe, yeah, maybe somebody that goes from like $1 trillion to $2 trillion in net worth has a new tax status or something that's different.
And then at the same time, there's all this talk about the Trump administration having conversations with Open AI about a possible government stake. And so then, you know, we just saw the SpaceX IPO a couple days ago, and it's like, do you think the government should have bought in? I mean, somehow we have like, I don't know, half we have half a trillion dollars at least or something for like a a war in Iran. So, what if we just bought like 500 or billion dollars or something of SpaceX and 500 billion dollars of Amazon and then as those companies pay dividends like that money goes to paying taxes so that we can lower taxes, you know, like somehow there's a lot of money and there's these huge dividends. I mean people, you know, companies like Berkshire Hathaway have been investing in the market for a long time and they were part of, you know, Coca-Cola and all these big companies. They get these huge dividends. They have billions and billions of dollars in cash from all those dividends. And it's just somehow it's kind of like as the American economy went, how did we run out of social security? Like how do we run out of these things when it feels like we should all be getting dividends? Like America produces all these amazing companies. Like why aren't we getting dividends? I don't know. And I think in Norway they are doing it. They have done that. I mean it's I know it's smaller and everything but it just makes me wonder like how did Norway end up doing the same kind of thing that Berkshire Hathaway did but then the US ended up with just more and more debt. I mean, I know we have like these amazing this amazing war machine that supposedly keeps the world stable and I guess it can go to Iraq and and control this straight mostly and all this stuff, but I don't know. It just seems like having everybody get like cash dividends for the investments of all these other generations that have created all this wealth might be kind of a nice thing, too. But anyhow, kind of tangented there, but the article is about how OpenAI and the White House are discussing a plan that could give the US government a stake in the company. I don't know if they would buy before the IPO or if they would just buy it at IPO prices. And also, if the government owns a big chunk of the major companies, does that become leverage that now, you know, Trump or whoever is president after him could then go in there and and just say, "Hey, look, I own like 30% of SpaceX. Like I want you to take this government contract if they sort of become owners in a way that is uncomfortable or could they just do it as like silent owners where they don't get any decision-making control but they could reap the profits and then if so how do we distribute those profits which is another part where the government tends to screw up but the talks have been happening actually for more than a year they're still ongoing. Um, one idea being discussed is for OpenAI to donate some of its equity to the government. That ownership could help launch something called a public wealth fund. You can think of it like a giant investment fund that holds assets on behalf of the public. Feel like it's a great idea. Those assets grow in value. Americans could potentially share in the returns. And the proposal comes as AI companies become some of the most valuable businesses in the world. OpenAI is currently valued at more than $850 billion and it's preparing for a possible IPO. We have Anthropic going IPO very soon. And SpaceX of course encompasses uh XAI which is you know got Grok and one of the the top five AI companies. It's just like how does the public participate in the upside or is it just up to individuals like you need to earn enough money that you can buy some stock right now? But if you do have a bunch of cash sitting around and like you can see this AI thing going. I don't know. I mean the market feels pretty hot right now. But yeah, it's like not everybody has that cash to invest right now. Even if even for those of us that see it, it's like what am I going to do? Take my like $37 profit from a video that probably took me like five hours to film and edit. Just like, you know, it just it's not achievable for everyone.
Or if you want to donate a humanoid robot, I guess I could have it go beg. Well, that's still a business model. We'll split the profit. I'll find a corner. You provide the robot. Go into business together. Anyways, more than grateful for the 32,718 subscribers. Much appreciated. Much, much, much appreciated. I'm excited to make another video for you in a couple days. And um, yeah, thanks for being on this journey with me. We'll talk soon.