[00:00:02] Host Create clip Thanks a lot. Great. I look forward to seeing my slides. Surely there we go. Awesome. Okay, So, um, four things I want to talk to you guys about today. Um, the first is I want to talk about some of the strengths and weaknesses of modern A I, um And you know, there's probably gonna be about two slides in this deck about vicarious specifically because I want to try and create is much value for you guys as possible. So I try to reverse engineer the talking. I'd want to hear if I was in your shoes. So first we talk about modern A I some strengths. Weaknesses How I think you should think about a I strategy. Um, and what I see is coming in the next to toe to six years on the Aye Aye frontiers that will have some impact on your business is hopefully and then some of my general lessons from building. Ah, hard technology startup. So, um, let's start with strengths and limitations. And before I do, can I get a quick poll? How many people in this room feel like they could give a short, cogent explanation of how alphago works Okay, maybe. Okay, maybe 10% of the room. Great. So for the other 90% of you all gonna give you an incredibly short tutorial on how it works, where you'll be able to explain this to your friends, it's not actually how these systems work, but it's close enough that it may as well be, um, And for the 10% of you raise your hands like, forgive me for the next three minutes. Maybe you should just turn out because it may upset you. Um, so today, today's a eyes deep learning predominantly, and it's all about pattern matching. Um, if we're gonna build a I together right now to read handwritten digits, the first thing we do is we gather thousands of hand labeled human labelled example. So this is 6000 examples of the number zero. This is 6000 examples number one and so on and so forth until we have 60,000 human labeled training examples of different digits. And then when we see a new image, our system basically just compares it to everything in the trading database. Ah, and attempts to find any close matches. And lo and behold, here a couple. And so voila. What we have is a four. Ah, and if you want to take that system and scale it up toe from 100 it's the photos. You do the same thing. So you take, you know, a photo like this one. You label it with corn with just something that's in the photo. And then, instead of 60,000 handwritten digits or training images, you show it, Ah, couple 1,000,000 each pixels, one training image on. And then you have a system where you can show the new photo and we'll say, this is a car, but because all it's doing is pattern matching. If you show it something like this, it'll say it's a bedroom pillow. Ah, and you know, it may say this is a Jaguar, but you will have no idea what a drawing of the same picture is. And it will also say that these things are Jaguar's, even though you and I know they're not Jaguar's, um, and so that's basically how these systems work. I think there's a really interesting and direct parallel between the evolution of artificial intelligence in the evolution of animal intelligence that we can use to extrapolate some lessons. So if we go back 600 million years, we get the very first intelligent animals, things like sponges and jellyfish flatworms. And then as time passes, the animals get more complicated. I would argue they don't get any smarter. So we have is all about instinctual responses in neuroscience. We call this the old brain.

[00:03:08] Host Create clip It's about pattern, recognition, stimulus and response. And then, about 100 million years ago, a miracle happened. Evolution figured out a radically different architecture for intelligence called the new brain. It's what gave us primates, whales and dolphins. Ah, and instead of being about seamless in response, it's about causal reasoning. It's about what if and why. Ah, and I'm gonna argue that a I today is in this big yellow box and give you some examples from the Animal Kingdom and from the Computer Kingdom toe illustrate that point. So, um, and to be clear, there's nothing wrong with the old brain, and there's nothing wrong with deep learning. It just has a set of property. So old brain A. I r. Old brain intelligence can navigate environments that can you know 100 food could reproduce, but it has some really important limitations that matter if you're building a company and you plan on using these techniques. So our old brain ancestors to make a new old brain animal you need a couple of 100 million years of training data on, then the resulting animal isn't going to generalize well to new environments. It's not gonna learn new skills on. Most importantly, it just gives the illusion of intelligence. It's not actually smart. It just gives the illusion of intelligence. And I'll show you what I mean, because if you peel back the layers of what's going on a little bit, it's very easy to reveal that the animals aren't actually smart on it has this architecture where some stimulus comes in. The animal's learned a bunch of heuristics, and then it performs an output. Um, so these are baby geese.

[00:04:29] Host Create clip Baby geese make this noise on. Scientists figured out that Mother Usual actually loved and care for anything that makes that noise so you can take a tape recorder that plays that sound, and you can put it inside a taxidermy of a wolf and the mother goose will love and take care of the wolf until the tape recorder runs out of batteries and then she'll peck it to shreds. Um, another example. These air baby ducks, baby ducks believe the first thing they see when they're born is their mom. Ah, and these ducklings happen to see a dog when they were born, and so they'll follow the dog everywhere as if it's their mom. Um, and then frogs also have a circuit. Their brains. This is fire the tongue whenever you see this little pixel pattern, and so you can actually start a frog to death in front of a phone. Um, and you'll see the same kinds of behaviors in today's smartest A I systems, which isn't to say they aren't awesome, like deep learning is awesome. It could do all kinds of fun stuff. Complete games control robots, but it has the same limitations is the old brain, which is you need a lot of labeled training data, and after all that work, you get a system that doesn't generalize well beyond what you've trained it on, and it really isn't understanding what it's doing at all. It's just doing stimulus and response pattern matching a minute even has the same Architectures are old old brain animal friends.

[00:05:41] Host Create clip So you look at this very famous example. This is Di mines Atari player on Let's talk about how it works. So if you wanna deep learning system to play Atari, all you do is you. Instead of showing it one frame from the game, you just paste three frames together into a single image. Ah, and then, you know, for perspective. If that's our 60,000 handwritten digits and that's our 1,000,000 training photos, you show it 50 million frames of a single Atari game. And after doing ah, lot of 50 million frames with a play, you start to learn. It starts to learn. Well, if I go left, I lose points in this instance. Or if I go left, I win points assistance and you can start classifying sequences of frames and to go left frames and go right frames. And after 50 million of these frames, the system works pretty well and keep in mind. After doing that, it has no idea what a paddle is or ball, or even what its goals are. It's just regressing on frame, so if you want it to do something different, like play a game that's on a brighter board. You need to retrain it from scratch for another 50 million frames. Or if you want to move the paddle up an inch, you need to return it from scratch or really anything you want to do it all you're gonna read need to retrain it from scratch on. So if we make the way we think the same system, we just make the game 2% brighter. It will do this instead. Um, So, um, now Alfa goes is actually the exact same architecture. Just scale up even more. So if we look at our our input, go board like this. And now we have down there little daughters or 60,000 digits. This is the 1,000,000 photos. This is the 50 million Atari frames. And this is the 1,000,000,000 go boards. Um, let me run through it again. Okay, 1,000,000,000. Go boards. So now, after this is about the equivalent of a human plan, go for 3000 years without sleeping. Um, and after playing that much go, you start to get us the the Isis people classify go from aboard to a percentage chance of winning or losing. And so you showed any new board. It can then search over all of the successors of that board, grade them from really likely to win and really likely to lose and then pick the path that maximizes its chance of winning. So that's all it's doing. Um, and if you wanted to play a different game, you're gonna need another 3000 years of training data. If you want to change the rules slightly. Another 3000 years of training data. If you just wanted to do something totally different, like fill the board of circles, you may as well start over, um, et cetera. So this is my actually my favorite picture of the offer. Go challenge match. Um, because you can see that it's not playing go. There's this guy who has to sit here to pick up the stones and put him on the board for off ago because it's not capable of doing that.

[00:08:07] Host Create clip Which isn't to say that you can't use deep learning to control robots. You can. This is Google X, one of their most recent publications, called the Armed Farm on what they do with their training, these robots to pick up objects that are all about the same size. And, um, to do that, you need, um, about 2.8 million training images in 800,000 grasp attempts by running 14 robots for 24 7 for months. And then at the end of it, you get pictures like the ones on the left, and then you can go from picture 2% chance that you're likely to grasp an object just like you did with go boards. And the problem is, if you want to, then I could see this couple problems. Um, the first is when it succeeds grasping something, it's not because it had a good plan. It's it's it's doing something here. It happens to pick up the object, but not because it was particularly smart about what I was doing. Um, and you see that over and over again and all the things that it does. This is another example where you and I would make some different choices about how we were gonna pick up this object. And but it happens to work in the end. And so it's just fine. On the other thing is, if you want to change the color of the bins that the objects are in or if you want to pick up the objects from the side. If you want anything to be different at all, you need another couple of months and another 14 robots running 24 7 Um, so deep learning, broadly speaking, is capable of combining a lots of data and lots of computers and making really accurate, very narrow predictions. Um, and that's not a bad thing. I mean, if I'm the CEO of Facebook or Google, Um, what I have is a lot of computers and a lot of data, and what I want are narrow predictions about what add to show you, uh, which of your friends might be in a photo or something? Um, and so those are some of the strikes, The weaknesses. Let's talk about how if I were in your chair, I would be thinking about a strategy. The first thing I would say is not having an aye aye strategy right now is a lot like not having a mobile strategy in 2010 or not having a Web strategy in 2000. It's something that you probably already have. I hope you already have. You don't already have it. Make it your to do list for this upcoming week. Um, and the questions I'd be asking myself, I were your shoes. Is what? Oh, what am I predicting right now? Like, what is my product predicting? Was my company predicting, um, where are there opportunities for me to inject prediction into my product or into my company. And then once you've identified opportunities where you could be making predictions, I would start thinking about what data sets you'd need to amass in order to be better at anybody else. That making those predictions, um, is there a way you can reframe what you do now as a Siri's as a pipeline of A to be mapping? So that's really what deep learning does. It takes an input picture and outputs a label. It takes an input go board and outputs ah, percent chance of winning. Ah, and so if you have any where in your product a pipeline of A's, B's and C's and a chain? Where is their value and filling in part of B or part of See that you don't have yet, Um, and the next thing I would say is, if you can do this and do it well. You could build a data moat. You can build something that then becomes very difficult for competitors to catch up with you. So after you've decided, what is the data set that I can collect? What is the prediction I want to make? Then go back to your product design and look ATT opportunities to introduce features or user interface tweaks that cause your customers to interact with a product and, by virtue of their interactions, to create the data sets that you want in order to make your product win forever.

[00:11:30] Host Create clip So think about ways to incorporate labeling. Um, and the last thing I would say is not to expect too much. So there's a really weird, um, survivorship bias that goes on in deep learning and particularly is you hear about in the media. So when you hear about the media, is how deep learning could do everything and how it's amazing. Ah, and what you're actually hearing about is how 1% of the time when someone tried this thing with it, it happened to work, and you have a 99% of the time. No one wrote about it because it didn't work on, so I wouldn't have expectations that were too high of this technology. And I also think about this for almost every company as a wind Maur or win bigger kind of technology, and it's not gonna make or break your company. But if you use it, you can definitely use it to, uh, wide in your advantage over your competitors. So this is the thoughts. Let's talk about what's on the horizon. Um, so, um, were undergoing now And what my company's part of is this transition from instinctual based systems to systems that focus on cause ality and mental stimulation.

[00:12:37] Host Create clip And I want to emphasize this is not a difference in degree. It's not. You take Alphago and you scale it up to even bigger computers. Um, this is a difference in kind. And to make a system that can be in a body and can make causal predictions about the world, You don't need deep learning. You need something a lot different in a lot. You know, it takes a lot of aren't you kind of get their, um and just like in the animal brain, the circuits that you see instead of a frog's brain and instead of human brains are basically unrelated. Um, these technologies are very different, and I'll give you some examples from the Animal Kingdom and from the acting them. T draw the parallel, um, so killer whales in captivity are trained to pick up trash in their tanks and trade it with a trainer for fish. And one day a Siegel died and fell into the tank, and the animal brought it to the trainer. And because it was like a big object or something, it got to fish instead of one fish. Um, and you imagine a frog with two rewards. We just eat both the rewards, and that'll be that that's not what the whale did. The well did this with the second fish, and it repeated this over and over and over again and build up a stockpile of fish at the bottom of its tank and that used those fish to train all the other whales in how to participate in the Siegel for fish economy.

[00:13:57] Host Create clip Um, next favorite example. This is Coco, the gorilla. Koko was raised in captivity, and her favorite thing to do is watch of all things watch, Mister rogersneighborhood. Um, And for those you want from America, Mister Rogersneighborhood Very famous American TV program. Um, Mr Rogers everyday sits down with the Children. Welcome to this house. Takes off his shoes and teaches them the lesson of the day. So Mister Rogers found out that Coco was a fan of his show, and so he went to visit her at the zoo. And the first thing that cocoa wanted to do with Mr Rogers was help him take off his shoes because that's how we always started this program. Okay, Last example. This is an 18 month old human. Ah, and it's being put into an environment it's never seen before. Full of objects has never seen before. It's watching a human. It's never met before. Do an action. It's never seen a human do before, and it's given no instructions. Oh! Oh, so yeah, So why is it that we're able to do this stuff? Why can mammals perform feats? They're so different from our animal ancestors. And and the answer is, we have a really different architecture for intelligence and our brains. We still have the old brain, the reptile brain. We use it to regulate our heartbeat and our breathing and our immune system. And then on top of it, we have a really different circuitry called the Neocortex. It's a replicated circuit, Um, and it's what does everything we consider intelligence. So it handles speech and vision and planning and motor actions and language. Everything happens. The neocortex and it's the same circuit. Basically, that does all of those different things on. It's the only mammals have one of you's and it's the same architect across mammals. Um, interesting enough. It basically has the opposite characteristics of deep learning, so it takes very little training data. It generalizes incredibly well, two out of sample environments. And it learns this rich causal model of the world, which when you're talking about robots and bodies, is the most important thing. You need to build a reason about what if and why, Um, and in comparison to a deep learning, our old brain system, um, your architecture in the new brain is this rich mental simulator like right now, you could close your eyes and imagine what it was like would be like to drive to the airport in a watermelon. And you could do it because inside your brain is a complete simulator of reality. And so that's what a new brain architecture is designed to. D'oh! Um, and to give you a glimpse of, like, what does this get us? I'll introduce a lot of paradox, which is right now we're in living in a world where all the parts that go into robots are really cheap, like motor is and sensors, plastics, metals, actuators, electricity processors. All that stuff is very affordable, and nobody owns any robots. And if you look at a factory 100 years ago in a factory right now, I mean, nothing's changed. Its color color is the only thing that's changed. And we pay people billions of dollars a year, maybe trillions of dollars a year, to do tasks that robots have been able physically to accomplish for decades across every industry. You know, it's so far from the future. I thought I was gonna live in when I was a 10 year old. Maybe all of us thought we were gonna live in. I thought I was gonna have a general purpose robot that could do all kinds of different stuff. And the reality is we do have general purpose robots that can do all kinds of different stuff. This is a video from 12 years ago. And the trick of this video is the robots being controlled by a human.

[00:17:40] Host Create clip So as long as you have a human brain, you can do anything you want with a robot. Like if I forced you to live your everyday life using a normal robot gripper, I think you do 90% or more of what you do during your day to day routine so functionally were already living in the Jets and society. We're just missing the A I layer of the brain necessary to make robots useful in Universal in that is the product of my curious is working on. I think of it like intel inside for robots. And it's just on a platform where you give it any combination of arms and grippers and sensors and tests, and it just works. Um, I think that interesting companies exist to make a prediction about the future come true. Um, the prediction that exists to make come true is is to make a world where robots are as commonplace of cell phones. Um, and I think you can reason about what the future might look like in just one chart. The top line is labor costs in the bottom one is robot prices index from 1990 onwards. Um, and this is the only sort of vicarious physics slide. We have technology building for eight years now that eyes about 1000 times less training data that deep learning is a lot better generalizing so moving on for vicarious. Let's talk about what this means. Society what this means, maybe for your business.

[00:18:47] Host Create clip So this is a picture. I love this pair of picture. So this is 2005. This is the inauguration of Pope Benedict, and I'm gonna show you another picture from eight years later. And maybe you can tell me what's different about these two pictures. Um, so the point I like about these two pictures is it's amazing how fast technology changes, like, you know, we all live in Silicon Valley or most of us do. And I didn't notice this happening. I mean, I noticed it happened, but I didn't realize how dramatically different the world is now that it was, you know, eight years before, and I think that we're on the precipice of going through a similar transition. Ah, golden age of autonomy, a golden age of robots. And I'll give you some examples. You've probably seen some of these videos, but let's just put it all together and get it. Get a glimpse into what the future might look like. So there's that was ah, four legged robot for Boston and Amex these air to arms from universal robots. These are not for years. I have a whole bunch of these. In my office is air bipedal robots from Julie Robotics.

[00:19:42] Host Create clip This is the painting to the gripper in this one. This is a five fingered ah, human scale. Um, anatomically correct Ripper with touch sensitive pads and all the fingertips. There's another one that's been out for a few years. This is a similar hand except the actuators air powered by air instead of bye bye motors. This is the commercially realized version of that same technology where it's being used to pick tomatoes. Um, there's another bipedal robot from Boston Dynamics. Um, so we're heading towards this world or robots become a lot more universal. Tha the automation, not of ideas of information, the automation of labor becomes much more prevalent on. And so I think of today the cost of compute is basically electricity plus epsilon on. I see. In the future, the cost of labor becomes electricity. Plus some Absalon, Um, I think there's actually this really interesting parallel between robotics right now and computing in 1950. Whereas if you wanted a computer 1950 it was gonna be expensive, inflexible, bespoke, and nobody really owned them. They were, like, very rare pieces of industrial equipment. And that's functionally what robots are right now. Um, and I think in 2035 robots are gonna be all the computers are right now.

[00:20:51] Host Create clip Um, and then the natural question is, what does this do to the labor market? Um, and I think if you ask the Internet, you get all these headlines about how this means no one's gonna have any jobs anymore. We're all screwed. The economy's gonna collapse. And for my perspective, I'd like to introduce a counter narrative. Um, and, you know, maybe you've seen some articles that say things like we're being afflicted with the newsies, namely technological unemployment. Where jobs get automated, people can find new jobs. Does anyone know when this article is written, guesses 1930. So this is not a new fear, not a new concern on. And I think that if we zoom, zoom out the last 3000 years, our story of people making contraptions that take a job that used to take five people a day to do and turn into a job takes one person a couple of hours a couple minutes to dio on. And so the argument that this time is different, that this kind of automation technology is gonna bankrupt the world's labor force.

[00:21:48] Host Create clip There's a very high standard of proof for this time being different. I think our best parallel is the agricultural revolution in the 19 hundreds, where the US working age population went from 40% being on a farm to 2% on the result of that transition was not mass unemployment. Every decade since 1980 90 roughly the ah, the labor participation rate has gone up, not down. So I really think that this this change, uh, is not gonna produce an economic catastrophe. I think it's gonna produce a lot of opportunities for people who are well situated to capitalize on them on So I think of the opportunities this creates. I think it's actually gonna be a lot of winners and a lot of new businesses that are built on top of this autonomy revolution. So strictly is the antecedents, the revolution. I think there's gonna be a lot of winners in the enabling technology, so sensors, processors, data, all of those things become more valuable when autonomy becomes more affordable. Um, so being part of this ecosystem and the other thing, I would say, is, when you bet on labor getting cheaper, what does that do for your business? Does it do for substitutes your business? Because maybe your business is designed around labor being a certain price eyes. They're an assumption underlying your your bets and also think about where using labor now and where could you use labor if labour? If freely free labor existed, you could just turn around like water. Would you do something different in your business? And how does that factor into the strategy and choices you make in the long term on the last thing I'll say is a couple of lessons from building Bi curious. So we're eight years in and over 100 million raise so far. Um, and I think if I had to sell that, you know, this is this slide to be 100 pages long, but just three lessons for me. Um, one is about explicitly managing technical risk. The other is being willing look wrong. And the last one is on burn management fund raising strategy that I think a little different from hard tech. Um, so it was a managing technical risk, most sort of start testing hypotheses about what users might want to buy. Ah, and for tech startups for hard tech start ups is the opposite. It's like you're building a time machine. You know, people want to buy a time machine. Ah, the question is, can you make a time machine? And so I would think very, very carefully about how you know what you think you know, and what premises, if wrong, would be most damaging to your timelines. And lastly, I would think about future investors and where they're most likely to be skeptical and need a lot of audience to be convinced that the evidence speak events. Because if you are working backwards through next fundraising milestone, you need to have expectations about where the burden of proof is going to highest and to some extent, design your roadmap around meeting their objections, not just yours.

[00:24:19] Host Create clip The second thing is about being willing to look wrong. These air some quotes from different areas. So Thomas Edison, fooling around with alternating current is just a waste of time. Nobody will use it ever. Um, this is about Tesla in 2008. Tessa is in the business of bullshit, not cars. Never trust the software guy with a hardware deliverable. The second sentence after this is Space X. Same story. There's a nice one about us too. So get used to get use people hating you. Because if you're doing something unique and different, um, you're gonna attract detractors. I think of startups as arbitrage between what looks like a good idea and what's actually a good idea. And so if you've hit the sweet spot right, then you're probably gonna attract the tractors. The last area I would pay attention to is about burn and fundraising. So if you're doing ah ah, hard technology startup, You're always always, always gonna be higher risk than Pinterest or stripe or something with traction because all of your milestones air around technology, not our customers. And that means you're gonna have to make the veces work a lot harder to understand your business. And they're gonna make you pay for the cognitive load that you introduced into their deal making process. And so, um ah, And so I would say, for this timing is actually everything you know, You you have toe plan to be alive long enough to get the right timing window when veces happened tohave bandwith to think about hard technical problems. If you're trying to raise in the middle of when some other hot, very easy to assess like lime by kind of start up is out there, you're you're gonna you're gonna hit more headwinds because they're you know, is it's they have payback period of a month and 1/2. And so and the customers go like this where you're like talking about something that's very technical, very complicated, and you need a PhD to understand it. Um, and the last thing I would say is is get really good at explaining why your thing is unique on. I know it's like obvious advice, but I've seen a lot of hard technology pitches because for whatever reason, people tend to come to me and ask them about their pitch and over and over again. I see What I see is something that doesn't sufficiently explain to a VC that doesn't have a lot of time. Why, with their building is necessary and different on the right level of abstraction. And you know the first part of this presentation talking about deep learning it sort of educating you is the first part of my V C deck. It's like a lot of my V C deck, because investors need to understand in their language and simple language in really basic language, what is different about your technology for everything else.

[00:26:53] Host Create clip Um, so and then the last thing I'd say is recruit independent allies. So you're no one is gonna believe your word about why your technology is is good, like the investors are gonna need some confidence that what you're doing is not just unique, but other people who are more experience in the art in the practice of what they're doing is actually a really good idea. And so you need at least a couple of professors Ah, couple of independent thinkers who are well respected the community to be able to look at what you're doing and hay and say, Hey, this is really different And it's worth taking a look at our worth worth placing a big bet on in order to give investors some confidence in what you're doing. So those air Some thoughts on burn and fundraising thanks a lot for your time. Attention. Happy? Take some questions.

[00:27:39] Host Create clip I'll be Scott. Leave the stage. I think we have time for 11 or two questions. Good. Um, no. I think you did an amazing job explaining the old brain, but I got a little lost on the new brain. Like what is it that's so special about what you do that explains how you come up with a new brain?

[00:28:03] Host Create clip Yeah. I mean, this is These are the cuts. Slide's basically I want to make room for Sergio Device. There's a bunch of papers of icarus dot com. You can really at a sci paper. Last year that went into more detail about it, I would say, Uh, algorithms are only as powerful as the assumptions that are baked into them. MP three's good compressing sounds because it assumes you're human. You're listening to music, and so you know, you probably the same sound in both years and a bunch of other things. So the strength of deep learning is also its Achilles heel. Deep learning you can use for anything because it's just doing function approximation. Whereas our brains is humans, we've learned we have become pre baked with a bunch of Dr Biases. Like, you know, the pixels of making my face aren't teleporting at random all over the room. They tend to stay together. They stay together over time. When you move your head in a particular way, the room transforms itself spatially in the same way every time. There's all these biases that evolution baked into our brains to make a sufficient learning and grated generalizing, and those of the biases that we include into our algorithms that make them better deep learning at this particular task, which is embodied robotics,

[00:28:59] Host Create clip Um, what other companies support from opening I deep mine and probably you go after solving a G I. And, um, if that's that's it, then why there's so so few And, um, why is opening is so afraid of that. Like literally, people working there are terrified about a J

[00:29:17] Host Create clip Ah, So there's so few because I would say it's very capital intensive and expertise intensive. You can't be a an a G I company with three people in two years of runway. It's just not credible. And so I think there's very few of them, Just like there's very few space exploration companies. It's incredibly expensive. It's a multi discipline coordination problem. And then, in terms of opening I Why this is scared of it. So opening has a belief, actually on. I think the mind us to that if you scale up the old brain enough, you get to the new brain. And this is something that I don't believe in disagree with, which is why I'm not nervous at all about opening I was doing. I think they could do it forever, and it never never scares me because I don't think they'll get they'll just get super intelligent crickets. And so I think that's the direction. And if I believe what they believe, I would be super scared. I'd be scared of the crickets taking over the world, but I think that particular Aye. Aye. Strategy is not likely to get to a new brain like architecture. And so that's why I personal worry. But I understand why they're scared.

[00:30:08] Host Create clip And last question eso

[00:30:10] Host Create clip We've your s strategy. When do you think we will get to human level A I And what do you think about super intelligence? I think that's super intelligence is inevitable when everyone's wrong. When they make guesses about this stuff, I'm not gonna throw my hat in the ring. I mean, I think it's gonna take a long as it takes and and, you know, hopefully we'll be the ones to do it.

[00:30:27] Host Create clip Thanks a lot, Scott.