Latest / In the Loop AI / Will Google's Veo 3 Destroy Hollywood?! | Everything Announced At Google I/O

Transcript

Will Google's Veo 3 Destroy Hollywood?! | Everything Announced At Google I/O · In the Loop AI

• David: [0:03] A little over two years ago this ai will smith eating spaghetti video came out and everyone was just amazed that ai could make this video in the first place there's this one problem this video is just completely ridiculous uh it just doesn't make any sense i think we all know it's it's a cursed video it's will smith eating spaghetti in the weirdest possible way but nonetheless two years ago this was just astounding that ai could produce that and two years and one month later we are getting the introduction of vo3 out of google deep mind which is doing the same thing making video but this looks much different this looks incredibly real and i would also say in addition to how incredibly real it is the sound that is coming with these videos is incredibly realistic there's background audio there's there's changes David: [0:52] on the sound depending on what the microphone is. And so all of a sudden we go from this absolutely insane AI Will Smith spaghetti David: [1:01] video to something that looks pretty competitive with Hollywood. And I think if we all went back to two years ago, everyone was kind of saying like, oh, this Will Smith AI spaghetti video is insane, but come on, just like give it two years and this will be completely different. And now here we are two years later and we have some video that again, seems to be pretty competitive with hollywood johnson you saw this what was your take. Ejaaz: [1:26] I i mean so josh and i were discussing this before we came on the show david this is basically hollywood grade movie production at your fingertips you josh myself anyone with access to a computer now has access to create the wildest highest grade quality videos at the fair i mean look at these examples here right you could go full sci-fi or you could create some kind of indie pick that you see on cinemas locally to you right now let's set some context here As you showed back in the day about like, whatever, a year ago, the quality of these video production models was so janky, right? You maybe you had a couple extra fingers on the humans that you would create, or the audio would just sound so super robotic. With this new release of VO3, it combines and fine tunes all of these different things into like a seamless little tool, a seamless little product. And it really pays attention to things like character consistency. So let's say you generated a video of, you know, this person that's speaking at a comedy skit that we're seeing on our screen right now, right? And you wanted to create a different comedy skit of him telling a different joke. He would look exactly the same. You could change the color of his shirt if you like. But if you wanted him to keep the same shirt or have sweaty armpits because he's nervous performing to the audience, you could do that. And I think this kind of like holistic movie production center just kind of blows the Adobe Photoshop era out of the window. Josh, I see a massive grin on your face. What are you thinking? Josh: [2:56] This is so cool. And it's cool in the sense that it's better and it creates better videos than have ever been generated before. But if that was the only thing it did, it would not be nearly as interesting as I think it is. And that's because for the first time ever, video has audio and it generates audio and it has this whole real world physics engine built into it that can generate this audio. Not only does it generate contextual audio, but the people can talk. And David, actually, I would love if you could go back to the last the last one where the guy was doing stand up on stage. I just want to take eight seconds and pause and just let the listener hear how the voice sounds, because it's a funny joke. And it is a person that looks real who is standing at a stand up comedy routine and they're actually speaking. So eight seconds here so people can listen. That way they have a little more context. David: [3:40] My therapist told me to embrace my mistakes, so I hugged my ex. Speaker3: [3:46] It's kind of funny. David: [3:48] It needs to be a little bit better for a joke. Josh: [3:51] So we'll work on the humor, but in terms of the delivery, it's really impressive. And I think this was a moment for me where I was like, oh shit, wait, this is something new and very important. David: [4:02] I also got the background laughter at the right time. The timing of the background laughter was perfect. Josh: [4:06] Nailed it. It has this amazing contextual awareness of the real physical world that I thought was incredible. And it made me think, I mean, first to movies, because you watch some of these scenes, you're like, this is straight out of a movie. But also in terms of that customized content for people that we talk about all the time, where we talk about a TikTok that's hyper-personalized to you and generates content specifically for you. Well, this is the first example of that, where now you have these people who can actually speak to you and who can now you can converse with. Josh: [4:34] And I think that part of it is really powerful. So as a model, this has been amazing. It's been super fun to play with, really fun to see the examples. It's pretty amazing. David: [4:41] I want to play a second video here. And this video is of interviews of normal humans, seemingly normal humans. They're AI generated, of course, at a car show. and they are just getting their their takes about the car show or the car that they bought just recorded as kind of like a testimonial and so what we're seeing here is like ai generated testimonials of this like imaginary car show and there's a few things that stood out to me first the microphone and the position of the microphone and the form factor of the microphone, was found in the actual audio quality and maybe i'm just my podcaster here is showing it here but Like the actual dynamics of the particular microphone in question was showing up. Then also the other thing that stood out to me was the specific relatability of the people being questioned. So that we just saw this trucker man at a car show. And he says something that is very contextually aware. So I'm just going to go ahead and play this video for the listeners who aren't aware. We have this trucker in a leather jacket. Excuse me, a biker, a Hell's Angel biker in a leather jacket, big beard, just like burly looking fellow. and he's at a car show, an electric car show. And so the AI is aware of who this person is. Let me, I'm going to play this video from here. So we can just go ahead and listen to this biker's testimonial. See, I'm kind of a, kind of a misfit here, but don't tell anyone I've just bought an electric car. Ejaaz: [6:03] But what you're really seeing is that technology is going to be very, very important in terms of how we go forward. It was... Great to come to the conference because my husband loves cars David: [6:17] I think did you see the way that she paused and was like awkward for a hard second she was like my like husband like the conference there because like usually ai is just like straight to the point doesn't skip a beat it has no pause for for like anything there's just tonality and like a level of humanness that I have not ever seen before in an AI video. Josh: [6:41] So the untrained eye, this version absolutely passes the Turing test. If you show this to a random person on the street, it's passable. The sound is really good. The audio, the visuals. You can tell there's some quirks. If you're familiar with cars, it looks like they're trying to use Tesla, but the logo is a little weird. So there are some quirks. But for the untrained eye, this is amazing. It's so impressive to see these. David: [6:59] Yeah. I also love how, especially in that last, the lady here, the background noise from the one clip to the next. So I'll play these last two clips again. the background noise changes. So the AI decided to make a new kind of background noise. And rather than just going with the same thing over and over and over again, the background noise changes. I'm just going to play this part over again. Ejaaz: [7:20] What you're really seeing is that technology is going to be very, very important in terms of how we go forward. David: [7:27] That's incredible. Ejaaz: [7:30] There's also like some fine details as well that they've paid attention to. Have you noticed how the spotlight that is on the people who are speaking changes as their face moves that is insane i go go back to the the video of the guy in the suit david like watch as he moves his face around like that light just kind of reflects off of his shadows the shadow trash the overhead light wow it's Josh: [7:53] A startling amount of awareness that it has where it just kind of understands the physical world well enough to recreate it it's really impressive yeah yeah Ejaaz: [8:01] Okay so youtube is getting replaced pretty soon, huh? Josh: [8:06] I mean, now that people can talk, we're next. David: [8:08] Well, there's more YouTube content that's going to be made from this. So yeah, YouTubers, I guess, are now going to have to go head to head with YouTubers who are using AI content. Ejaaz: [8:19] Right, right, exactly. And well, I guess for the consumer, that means more content for all, David: [8:24] Right? There's going to be a ton more content. Ejaaz: [8:26] And more importantly, as you guys mentioned earlier, hyper-personalized content, right? So I could see my version of a meme or a trend versus something that my friend sees as well in another friend group. That's, that's awesome. David: [8:38] I think listeners are probably just familiar with the idea of new technology revolutions destroy some jobs, but they make brand new jobs. And so like, it's kind of ironic that here, YouTubers are going to now have to compete with people who otherwise would not be able to be YouTubers. So in order to be a good YouTuber, you need to be charismatic, you need to work on your delivery, you need to be a personal person. And so like, kind of the socially awkward, you know, autistic person, not really cut out to be a strong YouTuber. But that same person can, who's good at prompt engineering and working with AI and can just be creative in the background, all of a sudden using AI can compete toe to toe with people who have like natural disposition to be YouTubers. And so it opens up the criteria of what is a effective YouTuber to a much larger set of people. Ejaaz: [9:30] I mean, we already have the concept of pseudo anonymous media personalities, right? Where you can't really see their human face, but maybe they are a Twitter persona, or maybe they hide behind some kind of animated character, which appears on YouTube videos, or Instagram profiles as well, right? Now, this is going to get taken to the next level with a model like VO3, right? Where you can be so much more expressive, and maybe even like visualize yourself in a different kind of skin, right? It's very, what's the name of that movie where it's all virtual reality? Josh: [10:03] I've read it for you one? Yeah. Nice. Yeah. Ejaaz: [10:05] We're like headed straight towards that. Josh: [10:08] So cool it's this it's like hollywood in your pocket or hollywood in your keyboard where you type this prompt and you get the scene out of a hollywood movie like the one we're looking at now or it's a musical and this was a one sentence prompt and now you have this original musical about garlic bread and it's this like hyper it's this insane amount of levers that you have to to create things that otherwise would have cost millions of dollars in the past for a couple of keystrokes on your keyboard and that feels like an insanely big breakthrough David: [10:35] How many weeks ago where we're talking about the number one reel on instagram or maybe it was a tiktok was an ai generated short film about a dog who saved a baby from a plane crash that was about four weeks ago if you looked at it it was very clearly ai like it had all the weird quirks of ai but i think the thing that was astounding is like there was still a story arc there that even though like visually it was like well this is ai and it's all kind of effed up but the story arc was there and that was before this google ve3 uh introduction and so like combining like good storytelling elements along with actually like visually compelling graphics i'm i'm just wait like i don't know two more weeks before there's another new reel on youtube or or instagram or something that is the number one viewed reel and it's like something 10 times better than what we were just reporting on two weeks ago. Ejaaz: [11:27] But that's not the only thing that google released this week for those of you who didn't know, Google had the equivalent of Apple's WWDC day, which is like kind of like the day that they have every year. I think it's their actual conference is happening in a few months where they make all the biggest announcements about new phones, new software, new apps, those kinds of things. David: [11:48] It's like Apple day on the internet. Ejaaz: [11:50] Exactly. Apple day on the internet. And Google's day is called I slash O or IO. And it was specifically focused on all their major announcements around AI. So they had this Google VO3 model, and now they've released a new coding agent, which is the fastest agent that I've ever seen. Now, I know a lot of listeners on this show probably aren't so technically inclined, myself included. And so I'm not going to focus on the coding side of things, but rather talk about the really impressive nature of what it can do. So number one, if you look at this tweet that we have pulled up here, for those of you who are just listening, it shows a demonstration of someone going from a single prompt, which is, can you build me a calendar app that does A, B, and C and looks like this, into creating that calendar app in not one, not two, but three seconds. Ejaaz: [12:42] One Mississippi, two Mississippi, three Mississippi, it's done, right? Now, previously, when we've spoken about coding agents on this show, we've kind of like done this whole demonstration of writing a prompt. And then we see this agent, I don't know, write a bunch of characters and code that, I don't know, I don't know about you, but I can't tell what it means. And then it pops out an app on the other end, right? Now it just abstracts away all of that. And the reason or the secret sauce that this new agent basically has is it can basically pump out tokens, which is kind of like the lifeblood of creating something that is AI generated at 2000 tokens per second, which for those of you who have no context, is about a 5x quicker speed than any other coding agent that is out there. And the way that it does this, linking it back to video models, guys, is it uses a very popular technique called diffusion. Now, the best way to think about diffusion is imagine an AI model creating a sentence, right? Typically, when it creates a sentence, it puts out one character at a time. A character could be a letter, right? So if it says, if it wants to say, hey, Josh, it'll be H-E-Y, right? But with diffusion, it basically spits out all the characters at once, which is why you get 2,000 characters per second. And we've never seen that technique implemented anywhere else. So it's a it's a fuss for Google, which is so impressive. David: [14:05] Josh, when you see both super high output and accurate output like that, I feel like there's two inputs here. Like one is just the hardware computation. There's more hardware doing the calculations. That's why it goes faster. But then also the AI model itself is more efficient. So it doesn't have to do as much computation in the first place. And then as a result, you get like a thousand tokens a second. Like what excites you about this? Like why is it significant to you? Josh: [14:28] So the speed at which you can generate tokens is a really big deal because the faster you can get tokens, the faster you can have the model think through more and get higher resolution answers. So there was some sort of, I believe this is a novel efficiency upgrade that happened because I haven't seen anything that's been able to push this many tokens at once. And this was just one of like many, many things that Google announced yesterday that were like pretty unbelievable. I think as a company, we've kind of been underwhelmed with Google, not in the sense of their software, but in the sense of their products. And yesterday, really, it was like a pretty hardcore smackdown against any other company in the space, because even this discovery of 909 tokens per second, which is crazy, wasn't even the most exciting thing they announced. There was so much novel progress. So this is super amazing. And yeah, very bullish on Google. Ejaaz: [15:15] I mean, Josh, did you see the Google Meets update that they made? Josh: [15:18] Insane. So freaking cool. I need this like yesterday. I need it immediately. Please, for the people who haven't though, please let them know because it is so cool. Ejaaz: [15:26] Okay. So for those who are listening, who, as I know, you have your pen pals in foreign countries that you wish you could speak to, but you can't because there's a language barrier. Your world's favorite conferencing app, Google Meets, now does live translation using a new AI model. So what we see here, for those of you who are listening, is a video of someone who is an English speaker who can't speak Spanish, speaking to a Spanish native, and it has a live translation over her voice. Now, the tone does not sound like her. So that's something that could be tweaking the next model upgrade. But the cadence, the effectiveness, the efficiency that goes from saying, hello, how are you? My name is blah, blah, blah, into Spanish is astounding. David: [16:09] The way that this connects to what we were just talking about, like 900 tokens per second, is that real-time live cross-language communication requires extremely low latency. It has to, in order for it to feel good and be a good product, it has to be translation in real time. David: [16:25] And so you can't have a lot of thought. You can't have a lot of thinking. You, like the AI needs to be fluent at a very low latency. So we're actually like looking at a video. I would like to watch this video. I'm assuming, I'm assuming it kind of gives a good demo. So let's go ahead and watch this video. Hi, Camilla. Let me turn on speech translation. It's nice to finally talk to you. Es bueno hablarte finalmente. No puedo esperar alquilar tu casa. Josh: [16:51] Te vas a divertir mucho y creo que te va a encantar. You're going to have a lot of fun and I think you're going to love visiting the city. The house is in a very nice neighborhood and overlooks the mountains. David: [17:01] That sounds wonderful. Eso suena maravilloso. Es la casa cercana al transporte público. So it starts by each respective speaker speaks in their native language. And you hear that you hear the Spanish speaker speaking Spanish, and then you hear the English speaker speaking English, but then the AI overrides it. And then the English speakers, English gets quiet, and it's overridden by the Spanish and then also vice versa. And so it's incredibly immersive. It's incredibly real time. And all comes from like, we go through almost every single week, not this week, actually, the next frontier model, right? it's this much more powerful it's this much more efficient it's this much faster and And this is what that looks like at the end result turning into a product is you get real time AI capabilities that, you know, otherwise it would take them like you'd have to type it into Google Translate, like all that kind of stuff. And so latency, I think, is actually a huge unlock for a lot of products. Ejaaz: [17:56] You know, I now have the perfect excuse for not using Duolingo, guys. It's been a dormant app on my phone for so long. David: [18:02] All of that time spent on Duolingo just completely wasted. Josh: [18:06] Oh, man. You should know better. You should know how fast we're moving to think that you actually need to learn a new language. That one's on you. Ejaaz: [18:12] Yeah. Yeah. Silly me to not predict that Google would make a life-changing update in a week. Josh: [18:19] Crazy. Yeah, it's amazing. I think one of the things that I loved most about this conference in general was how applicable all of this new technology was to us. Like, this is deployed in the real world. And now I could actually go and talk to people in foreign languages. And not only that, but they announced their new glasses, too, which was just an extension of this, where you can wear the glasses. And when you talk to someone in real time, you actually, you get that same translation in the real world. So this is really cool, like real world applications for the average person. Like users can use this and they can use it today, which I thought that was such a cool thing that they did. Ejaaz: [18:50] So if you guys remember from last week, we spoke about a rumor that Apple was Ejaaz: [18:54] going to replace Google as its default search on its searching browser. And I think it accounted for something ridiculous, like 30% of all search activity just through Apple iPhones last week. Well, it sounds like the folks at Google must have watched our episode from last week because they decided to drop a new AI search product as well as part of this whole IO Google update, which basically, the best way to think about it is chat GPT in your Google search. So now whenever you search something, Gemini kind of gave you like a brief kind of excerpt of like what you were searching for. It's now much, much better. It pulls from various different sources. It gives you examples and direct links to the thing that you're looking for. I actually tried this out, kind of like a beta version on Google Studio. And I was able to basically get access to certain Amazon product links or clothing links directly from the thing that I was searching for, which is super cool. And it just shows how important Google places their search function as a moat for them, right? If they lose that to OpenAI's ChatGPT or whatever Apple's own in-house AI browser, that's a lot of revenue to give up there. So it's great to see them being so aggressive. David: [20:07] I think the big story here is how Google retains actual traffic and links because they need to pass people to links to keep their business model. The thing that AI does is it gets you your information, the value that you're looking for faster, more immediate. And so you don't need to go into any links. That's the whole value proposition of ChatGPT and AI generally is that like you just go type in what you want and then it spits it back out. You never clicked any links, which destroys Google's business model. And so I've always been curious as to like, okay, how do they compete with OpenAI and the value of AI in the first place while still getting people to go click on links because that's where they get their money. So it sounds like, Ajaz, you got the benefit of an AI agent, a smart AI agent, giving you the value that you need. And you still went and clicked on links. It sounds like you still did both. Ejaaz: [20:55] Yeah. And I think if we were to kind of extrapolate where this trend ends up Ejaaz: [20:59] going, it'll just be an autonomous version of me, right? It'll know what I want. It'll know what I want to buy. And so it wouldn't be me seeing the links that it presents to me and then like clicking on the product, clicking purchase. It'll just be, I don't know, for better use of a word, an e-jazz agent doing all those things for me. And I'm kind of like trying to think like, how far does that go? Like, will it and will my robot version of me end up doing all the working out for me? And do I just get to sit at home? Or do I get to spin up Google VO3 and put facelift on me whenever I go on work calls? I don't know. But it's cool to see how it like changes the way that we interact with everything that's online. David: [21:37] Josh, what do you think about Google's pressure with its business model. Is it threading the needle? Josh: [21:42] I think it is. This is their answer to the founder's dilemma is how do they keep their monopoly while still continuing to innovate? And I think they've kind of threaded the needle. We'll see. We'll see how it works. I don't think enough people have had a chance to try it yet, but this seems pretty good is they take their existing user base. They take the people who love Google and now they've just given them a little extra. So now you have this A integration. You feel less inclined to need to go to ChatGPT because you get similar answers here, but they're enhanced by Google. And then they're also enhanced by your Google profile. I think one thing that people have had much longer than open AI accounts is Google accounts and Gmails. And those Gmails have a lot of data on you. And should you choose to opt in and allow Google to get all of that data? Well, like Gijaz was saying, this gives you hyper-personalized results with this enhanced AI engine on top of the Google search that you already know and love. And they're kind of doing this across the whole stack. And it was a sad day for startups yesterday too, because it seemed like every new feature that they released was just smashing a hammer on top of one of the startups, one of them being one where you could try clothes on. And you upload a photo to their service and you can swipe through shopping results and then try those clothes and see how they look like on your physical body. Google now did that too. So now what you're seeing is this kind of full stack experience that's hyper-personalized to you across the entire Google spectrum, which is, it's not just information. It's shopping, it's communication through email, it's visual through YouTube. They kind of have this whole stack that is comprehensive to your life and they have all this data. And now they're laying this AI service on top. I think they did a great job. And I'm like really excited to start using these features. David: [23:09] How do we feel about the compression on startups built on top of Google and built on top of AI all kind of collapsing to like a few power players who have the distribution. AI is always like, who did it? Peter Thiel was always like, yeah, AI is a massively centralizing force. Power is going to accrue to nation states and to the largest tech companies. And then the long tail of startups are just not going to be able to compete because any sort of value a startup can build, Google can offer because they have distribution. Are we worried about this? Do we like this? Do we not like this? Ejaaz: [23:42] I think this is a natural evolution of tech, When the internet came around, everyone was like, oh, now we can spin up multiple different physical locations and we could just sell through the same website or whatever that might be. But that's not what happened. They own the entire internet infrastructure and then you spin up a bunch of companies on top just that are like web native. The same is going to happen with AI. I think startups are just going to look very different. I don't think it's going to be some kind of commerce store or some kind of B2B SaaS provider. I think it's going to be something else. Maybe it is an agent or agent tooling or agent providers or whatever that agent flow looks like. Josh: [24:20] Yeah, I don't see a world in which this trend decreases. AI is a force of leverage. It's a tool for leverage. And it works for new founders. So they could spin up these startups very quickly. But it also works equally for these large companies that are then going to turn you into a feature using one prompt. And they have the existing infrastructure. They have the marketing. They have the user base. So it's really hard for these new companies to exist in a world where you are one prompt away from getting replaced. I think the game that they need to play is the user acquisition game. And that's something we've seen people try to do is how do you move faster than Google can to acquire attention and users? If, let's say, the Ghibli phenomenon, if that wasn't from OpenAI and that was just some small lab that created ImageGen that was really, really good and better than anyone else, They could probably reach a large enough critical mass to maybe get acquired by one of the companies, but for them to really stand up and, and like you reach that escape velocity, it seems really, really challenging now, especially when you can't win on software and it's 10 times harder to win on Josh: [25:20] hardware because they don't have the existing manufacturing infrastructure. It's, it's a tough world to start a new software company. David: [25:26] The FTC is eventually going to become a conversation here, the Federal Trade Commission, because they are the antitrust. They have their responsibility over antitrust. And they are just now talking to Facebook about its acquisition over Instagram that happened like 15 years ago. And they were like, was that overly monopolistic? Was that competitive, anti-competitive? And this idea of just like startups can be added like the whole entire value of a startup can be added to google with like a couple lines of code that is just going to be a complete hairball of antitrust issue that like it's going to be too late even by the time the ftc can react it's going to be far too late because google will have already ended a thousand startups by just increasing the value of their ai model and they're going to have to do that to stay competitive with the other ai labs. Josh: [26:17] Too i don't know where it goes but it's it's gonna be messy yes it's gonna be messy i think it probably it is interesting that they move so slow because it does allow for this innovation to happen so it is a bad thing in the sense that if you are a startup founder it's really difficult to win in a lot of categories but the fact that there is no legislation so far behind allows for these experiments to at least be run and for to us like to collect more data and to go through these experiments and to see more how it plays out so yeah it's gonna be it's gonna be a mess but i think it's a long time until that mess becomes realized it's Ejaaz: [26:47] Also just going to lead to a shit ton more startups right i mean we just kind of demonstrated like five new tools David: [26:54] Startups yeah yeah less startups. Ejaaz: [26:56] Way more startups i mean like let's use vo3 as like the the video thing right like now everyone can become a world famous hollywood producer or director or visual artist all of those roles by one person from one single computer now you just end up with a shit ton of movies that are out there right so David: [27:12] I think i'm categorizing that as more content creators more individual creators not necessarily more startups. Ejaaz: [27:19] Right right but but if we directionally head in towards a trend where all those tools become available so it's not just hollywood movie production but you can create an app which combines certain short form versions of the content that you produce basically you can spin up multiple practices and combine that to be one type of company and you can make multiple of those companies right to address different kind of sectors i feel like the the the bridge between experimentation and then one of these monopolies like google just copying it and adding it as a feature is becoming a much more rapid iteration than it's been before right previously you just need to like watch it play out and be like over a couple years yeah maybe it is quite good and then ah shit we can't build that ourselves right Speaker3: [28:05] Let's just pay like whatever, Ejaaz: [28:07] $10 billion to buy this. Now it's just kind of like, yeah, let's just copy and paste that. That sounds good. And we'll do it in a week's time. David: [28:15] And to me, I'm hearing that, okay, why would I ever do an AI-based startup when ChatGPT or Google is just going to take the value that I create and integrate it into their model? Ejaaz: [28:25] Yeah, exactly. Josh: [28:26] I like what Ejaz is saying. It reminds me, it feels very similar to like the analogy with social media, where social media really only exists like this whole creator economy because we have this like surplus of capital and time and attention where now there's this whole economy that's built on people just having time to take cool videos of their lives and people who are obsessed with watching those lives and want to invest in the products that they use and that that is a luxury from this innovation that we had that enabled all this free time and excess capital and we kind of have a similar dynamic here where it's probably gonna be really hard to create a company but what other second order industries can emerge because now these tools are so prevalent so while it might be really difficult to compete David: [29:06] With google in. Josh: [29:07] A startup maybe you can like you just said create this hyper personalized bundle on steroids where you create a movie and this really amazing experience around it and all these ai tools that you can interact with it with and and maybe there's just this whole other industry that emerges that's kind of more niched down but still meaningful i don't know it's weird yeah Ejaaz: [29:24] What you just reminded me josh what is the name of that kind of thesis or rule that A16Z came up with, where it's like if you get a hundred or a thousand loyal subscribers, you can pretty much support and sustain your own life. Yeah, a thousand true fans. Exactly, right? Josh: [29:39] I think it's a Kevin Kelly article. Ejaaz: [29:41] Yeah. Kevin Kelly article. Okay, right. So if we take that analogy, right, and instead, like, transpose that onto people that can just spin up these random AI setups that are hyper-personalized bundles, as you describe it, Is that like a just a new economy on its own? Do we go from like the nine to five wage worker to just suddenly startup owner, but it is enough to support their own lifestyle and buy what they want and digest the content that they want? Maybe, maybe just everyone becomes a startup owner. I don't know. David: [30:08] I guess we're starting to like kind of have a blurry definition on what's a startup like is a independent one, two or three person creator studio who creates new movies that could never have been made before. Is that a startup? Because in my mind, that's more on the margins of like you're a YouTuber who's making content, not a YouTube platform hosting content. And I think with AI and the centralization force that is going to go to the incumbents, the open AIs, the Facebooks, the Googles, it's going to be harder to compete with those with true startups like VC backed startups. That is going to be harder. What is going to come out of that is the ability for everyone to become more creative. And what I am stoked for is that there is going to be a new type of creator out there that makes interesting movies, short form movies, long form movies, photography, anything creative, jokes, anything like stories, David: [31:05] books, all of this stuff is going to be bolstered by AI. And they are going to be 10 times better, more unique, more creative, more boundary pushing of art than we've ever seen before, because we are just unlocking a new type of creator with new tools. And that is the value generation that I see is like, as we opened up this podcast, there's going to be. A thousand X more content and most of it's going to be shit. And some of it is going to be like a hundred times better than our currently existing content. And that's where I see value being generated. Josh: [31:37] This kind of relates to the idea that we were talking about post-AGI, which is just the cutoff of, of human labor as a form of capital. And the, the incremental amounts of value that we could extract from human labor actually decreases over time. And in the case that these AI or these large conglomerate companies start to absorb a lot of startups, that kind of aligns with what we think where like you won't be able to generate a ton of value because it will be held by these overarching entities but as it trickles down you can earn enough and you could create the interesting things that excite people that are are not within the domain of crushing from these conglomerates and still probably live a pretty decent life Ejaaz: [32:14] Well let me ask you this do you think it makes sense for a single entity or board or corporation to own as many startups as we're proposing right or do you think at some point does it just become net negative right like there's only so many kind of eyes that can be on a single project right it almost sounds like an open source or slightly more i don't use this word but decentralized version would make sense right i David: [32:44] Think this is one of the big themes that that have has been identified by ai and the one the one that we call the Game of Thrones to create God. And so OpenAI, Facebook, all of the big AI labs, the Chinese ones, DeepSeek, there is this like arms race, this AI arms race, this Game of Thrones competition to have the best frontier model. And what do you get when we fast forward hopefully more than five years, like 10 years, like ultimately you create God and all of the value ever created by anyone on earth just gets integrated right into the model. And the model can literally do anything within the realm of feasibility. And it can even push the realm of feasibility outwards. And there's this one supermodel, this AGI supermodel that can do everything. And that's why there's like so much competition to create this. And at the end of that, like what does society look like? Human labor is worthless, but there's so much like creativity that can be unlocked. And so there's probably a whole episode. I don't know with who, what guest maybe like. David: [33:48] Max Tegmark or Nick Bostrom or some AI philosopher where like, yeah, you actually end up in this kind of like hyper capitalist, communist, weird utopia, dystopia thing where like humans, humans have no purpose. There's only art left. And art is the only thing that really gives us meaning. And we need to figure out how to like make commerce happen. But if like we, like Josh says, like enough of that trickles down to create some level of UBI, I can buy Josh's like movies and he can buy my photography or something, something like this. There's definitely a podcast guest out here to help us talk about this, like what this looks like in like the long-term. And again, hopefully that long-term is, give us at least 15 years before this comes, please. But yeah, it's a very big conversation. Ejaaz: [34:34] We should have the optimist versus the pessimist on that philosophical episode. David: [34:38] That would be a killer episode. So all this to say. Josh: [34:41] Google I.O. was pretty sick. Ejaaz: [34:45] We really enjoyed it. Josh: [34:46] They released some pretty cool stuff. Ejaaz: [34:49] And actually, to set that into context, I was looking up the figures because I was like, how many people are actually using Google's AI models? Because I talk about it a lot, but I don't necessarily use Gemini as much. They currently have 400 million monthly active users across all their models. And collectively, they process about 480 trillion tokens. For those of you who are listening, that's the equivalent of like 2.5 billion books per month, right? Or for the Americans that are listening to this show, that's 65 Library of Congresses per month. Thank you, Grok. So that is just a hell of an amount of output per month, right? And I'm curious to track this metric as time goes on. But yeah, Google completely swept the floor. Crazy. David: [35:37] All right. Well done, Google. We talked a lot of shit about your previous Gemini releases, but you guys earned this one. Josh: [35:43] Smackdown. David: [35:44] I think, oh, is that not what's over with Google's announcements? We have Alpha Evolve. Ejaaz: [35:48] Should we talk about this? Yeah. Okay, right. So, you know, not to stay too much on a single monopoly and, you know, judge them by their products. But, okay. So, a trend that happened this week was there was three flagship autonomous coding agents that were released by both Google and OpenAI. And then last minute, aka yesterday, Microsoft as well. Now, again, I'm not going to talk to you guys about the technicals of this, but I'm going to talk to you about why this is cool and specifically what problems it solved, which was super cool, right? So Google released not one, but two coding agents. Number one was called Alpha Ejaaz: [36:27] Evolve, as you mentioned, David. And the best way to describe it is it's not just one model, but it's like several models that they created. It's kind of like a pipeline of models. and its goal is simply, hey, improve the algorithm of whatever I give to you, right? And an algorithm is basically just a bit of code or whatever it might be. But you might ask yourself, well, what algorithms was it given? Like, what did it improve? Well, here's some of the highlights. It was set on Google's own model, aka the model that it was run on. And it was told, hey, I want you to improve the code of this model. And what it ended up doing was making it 1% more efficient, which meant that it saved $150 million in cost in terms of training and inferencing it. Ejaaz: [37:13] So that's just like the boring kind of like, okay, it saved me a bit of compute. For Google, $150 million is probably just a drop in the ocean. Like, who cares, right? David: [37:21] 1% improvement is how Lance Armstrong became the Tour de France champion like seven times in a row. That was the entire strategy behind building Lance Armstrong a better bike and improving his training by just 1% over and over and over again. Ejaaz: [37:35] Well, you know what? I'm sure in about a week's time, that 1% is going to be 5% and it's going to just be an exponential curve of like compounding or whatever that might be, right? So it saved millions of dollars or hundreds of millions of dollars worth of compute. That's great. But then it was given a slightly different task, right? It was given a chip design that Google had made through their like processing units. And it was saying, you know, how would you redesign this chip? And it went back to like the original papers that this chip was made and designed on. And it reconstructed that chip in a slightly different way, which led to another efficiency gain. But guys, I see you champing at the bit. You're like, you know what? Efficiency gains, it's cool, but I don't really care. Okay, well, let me take it up another level, right? It was presented with 50 open math challenges that, and I must reiterate this, had not been solved to date by a human. Ejaaz: [38:30] And in 75% of cases, it rediscovered the best solution known so far. So that's what us humans know. And in 20% of the cases, and I think that's about 15 maths problems, it improved upon the previously best known solutions, yielding new discoveries, right? The point I'm trying to make around these things, guys, is it's not about the coding agent. It's not about some secret source that it's made. It's the fact that it's now making discoveries that humans previously could not figure out on their own, aka the AI is telling us stuff that we didn't know before. And it's making us like make new breakthroughs on these different sciences. Right now, it's on maths problems. Ejaaz: [39:10] Right now, it's on chip design and boring compute efficiency gains. But the idea is tomorrow, it's going to be on things that are personable to us, relatable to our professions, maybe even personable to our relationships, our friendships, unlocking the 70% of our brain or gray matter that isn't used right and this is a trend that i'm seeing that's alpha revolve i've got more to say but i want to hear your takes on this David: [39:35] I think the impact here, it would be if it solved one math problem that humans had not previously solved, like a truly novel one, I think that impact would be even better. But I think the takeaway ought to still be the same. In 20% of cases, it approved upon previously best known solutions, yielding new discoveries. To me, that's like, okay, we are creating net new knowledge here. And we're doing it in math, which is this extremely low level knowledge, like arguably the lowest level knowledge that exists. But like that yeah that's the foundation of human knowledge you go from math you go to physics and from physics you go to chemistry and from chemistry you go to biology and hardware and and other things and so again just one percent especially at the rate at which new models come is a huge deal and then like this starts to get into like the agi super intelligence explosion like okay well what if we ask it to start to improve itself by one percent like make your own code better i don't know like maybe i'm speaking out of out of my authority here but this is like what i see here is our ability to improve our accelerate the ability to accelerate is improving by like one percent this week one percent next week one percent the week after and to me that is very intimidating josh what's your take yeah. Josh: [40:51] This feels exciting because at this point it feels like more of an optimizer like it can it can take the most challenging problems and just kind of squeeze out the remaining juice that we weren't able to figure out but it hasn't created anything truly net new like maybe there's there is some novel discovery here but nothing that's super impressive i think it's it's starting at math because math is the most concrete set of parameters where math has these very fixed rules that it could kind of work within the confines in and maybe sort of reason through that to find these extra efficiencies kind of like we saw with with go where it made this crazy move that made no sense but it in the context of everything it it makes sense. And it was the reason why the computer won. We're kind of seeing that here where it's doing things that look weird, but it's doing so for the sake of efficiency, where it's not really getting Josh: [41:37] net new knowledge, but it's close. And this feels like the closest we've ever gotten. So I would imagine a few weeks, a few months, a few years, this goes from solving 20% efficiency upgrades to actually net new knowledge. And that's when this gets really exciting. But I think that's what makes this exciting is this is really the first step in generating that net new knowledge. Ejaaz: [41:54] And I think it's important to flag that first step, as you said, Josh, right? Because a trend that we're seeing with just AI development in general is you'll see some niche startup come up with an AI-generated voice model, right? And then you see some other niche startup come up with an image model. And then that image model turns into another niche startup, which comes up Ejaaz: [42:18] Two months later, you have Google VO3 owned and trademarked by the monopoly, you know, Hollywood production at your fingertips or in your pocket. And what I'm saying is we start to see like it emerge in different kind of like siloed parts and then rapidly combine like over a week into like a fully fledged product. And this week is no different, right? So we just spoke about this Google agent model, right? But what about OpenAI's thing, right? So OpenAI released a new agent called Codex, right? I'm going to put all the boring technical stuff aside and just tell you what's cool about it, right? So typically, when you tell a coding agent, you know, we talk about vibe coding, right? Hey, build me this app. And then we see it construct the code in front of us. This is not what it does. Instead, you go to this agent and you say, hey, I have like 10 tasks that my boss has sent me. And like, I don't really have time to work on all of them right now. I want to go out and grab a coffee from Blue Bottle. Do you mind just working on like five of these tasks? and I've got the other five when I come back, it can now do that. You give it access to your entire private code base. You can fine-tune it on that particular, so fine-tuning meaning you can just train it on that. Ejaaz: [43:26] Well, I was being, David, I was trying to make the human seem valuable, David. Come on, like we're still humans presenting this podcast, right? And it can just run all of it in parallel, right? So the analogy here would be that's two weeks worth of work completely blown down into one hour and it just runs autonomously, right? And then we move on to like Microsoft's update, which is they released not specifically an agent, but an agent that can use any kind of coding model, right? So it could use OpenAI's new codex model. It can use Google's new coding agent or whatever that might be. And they're taking a much more agnostic approach, right? So if you combine all of these things together, I wonder what the VO3 of this coding agent saga looks like. Is it David, Josh, and I have access to an app which can code up any kind of idea or random consumer app that we pop up in our head? It goes from Vibe Coding V1 to Vibe Coding V2. And I think that step change is probably within like a month or two months grasp Ejaaz: [44:30] out of our region. I don't really know. David: [44:31] Before we move on, Idraz, this is at the bottom of the Google DeepMind discussion. Do you guys know this Twitter account, AI Not Kill Everyone-ism memes? Have you guys seen this Twitter account? Yeah. Speaker3: [44:42] I think I've seen the memes from this, Ejaaz: [44:43] Yeah. David: [44:43] Yeah, yeah. So I first interacted with his Twitter account back when Bankless was doing our series on AI safety. And so like old OG Bankless lore, we did this episode with Eliezer Yudkowsky, who's like the original AI doomer of like, yeah, his P doom, his percentage of doom happening downstream of AI is like 100%. And so we started doing these AI safety episodes and we did like five or six of like AI safety experts. And one big area of emphasis was like, one of the big areas, arenas of P doom is if we allow AI to recursively self-improve itself. And when we're talking about this Google DeepMind, which is this thing that discovered new, like pushing the limits on new knowledge about improving hardware, improving algorithms. And then I brought up that idea of like, well, what if we just like pointed it back to itself and it could like improve itself? This is like unlocking a huge arena of P-Doom for a lot of AI safety experts. And I mean, at this point, this is a meme account. I don't have any power to stop this phenomenon for happening so i'm kind of just enjoying the along the way but it's it's it feels uncomfortable that i feel like we are pushing up against that like membrane of self-recursive improving ability and we're kind of just watching google do it and we're like ah cool things get one percent better awesome and then meanwhile like in two years like we unlocked p doom because of self-recursion that's it's just like a fear that is worth highlighting. Ejaaz: [46:12] Okay, but David, doesn't this ultimately come down to alignment? Like if we had this self-recursive, self-improving AI situation, but it was aligned with humanity, we should be Gucci, right? So this is a question of like aligning the AI. Put up a gun, David: [46:27] Just be aligned. We'll just solve alignment. No, no, no. Ejaaz: [46:30] It's a big deal, right? And maybe we can have that in the philosopher discussion that we mentioned. But like, I feel like that's what it comes down to because like I don't see a world where people are like, you know what? Don't let the ai make scientific discoveries and cure cancer and all that stuff no we don't want that right don't let it do personalized medicine or any of that like there is no way that this doesn't happen David: [46:51] True if ai can solve cancer among all the other diseases out there it can also create a virus that will kill us all it can do both of those things the. Ejaaz: [47:02] Duality of man sorry duality of ai man Josh: [47:05] Of technology progress in general that's kind of how it's always been is there's always David: [47:08] It's always how it's been yeah this time it's just happening all at once and like previously when we like developed the atom bomb we had time to like figure out what to do about that like we had a good number of years it came close and like even technologies before that like you just have time to adapt we don't necessarily have that same level of adaptation and so if we find all at once inside of a five-year span, 20 different ways of ending humanity, the chance that we can't contain all of. Speaker3: [47:39] Those and we let David: [47:40] One out is like pretty high. Ejaaz: [47:41] Well, we also didn't outsource human thinking to a non-human organism, David: [47:47] Did we? True, also that, also that. Ejaaz: [47:49] Yeah, yeah. We had these clunky little meatball sacks that had morals and ethics and community and culture. And now we're just like, just let the AI do the thinking for me. Let the AI make the discoveries for me. Josh: [48:03] And I think that's probably why it wins. It's just, there's like so many more shepherds or sheep than shepherds. Josh: [48:07] Like there's just people that want their lives to get better. And they don't really care about the cost. And that's probably a trend that we'll see continue. David: [48:14] Yeah, yeah. Meanwhile, join us every single week on the AI roll up while we live stream the end of humanity with just big smiles on our faces. Because what are you going to do? Ejaaz: [48:24] We should have a P-Doom counter on this episode and like on this show. And every week we should just like it got David: [48:31] A little bit higher this week for me yeah it got definitely got a little bit higher with this like google stuff i don't think we're done with google what's going on with the google big bang what is this. Ejaaz: [48:39] Okay well it's not the google big big bang all right it's a small lesser-known company called meta okay so okay pardon me so so so on the topic finally David: [48:48] We're off of google. Ejaaz: [48:49] Yes exactly on the topic of ai basically making scientific breakthroughs we've seen a general trend this week of that actually happening. So Meta, formerly known as Facebook, released four scientific papers, which basically leveraged AI to make new discoveries in terms of molecular structure, atom structure, and sampling when it comes to science. But who cares about that? That's just the theory and science of it all. But the point being is, I think personalized medicine, personalized cures, stuff relating around natural sciences, which isn't just something that you and I interact with on a daily basis, David and Josh, on our computer, but it's our things that we interact with outside of our digital life, right? When we need to go see the doctor or we need to- Ejaaz: [49:37] Sometimes, dude, I look out the window, it seems pretty exciting. But you know, go to get like personalized health advice or whatever that might be. You know, those things are now getting AI fight, basically. And we see this really materialize in this new agent called Robin, David. So the story with Robin here is that it made a scientific breakthrough, but not one that you might expect from this clunky AI thing, right? And the thing that it discovered was a cure or rather a treatment for blindness caused by AMD, which is basically a disease that causes blindness. Now, I just want that to settle in for a second. This is basically an AI or a set of agents which was given a task and it said, hey, there's this condition which causes blindness in individuals. Ejaaz: [50:28] I want you to come up with a scientific experiment that could potentially lead to a cure. I'm not asking for a cure, but just like go along the critical sense of thinking and figure this out for me. So it came up with an experiment and it got a bunch of humans to do that experiment exactly to the T, exactly to how it's suggested. And the humans, the worthy meat sacks, gave it the results and it said, these results are okay. They're not bad, but I think this is informing me of another experiment that I want you to do. And it basically went back and forth over six experiments, and it came up with this novel idea. And I don't know how to pronounce the compound, so don't ask me to, but it's basically something called a ROC inhibitor, which is only, funnily enough, approved in Japan for cure of a completely different associated disease. And there is no scientific research that has ever suggested that a ROC inhibitor ever be used to cure blindness in this particular aspect. Ejaaz: [51:27] So the humans were like, okay, let's test this thing out. And it ended up curing it in a bunch of mice models that they were experimenting on. Now, listen, this, as Josh says very frequently, and I agree with him, this is just step one. It's nothing kind of major. It's working on rats. Who cares? Rats, humans, they only share 85% of the genome with humans. We don't really know. But once we get to the point of human trials, and if this ends up actually working out to be the case, Along with the 1% of Google's $150 million compute, span, chip designs, whatever, we are going to end up in a world where we just let these AI run the ship, right? Ejaaz: [52:04] Personalized medicine cures all of the likes, right? And this is not just something that Google, Meta, and this random startup that created this Robin agent is focused on, right? It's also that other smaller company... Ejaaz: [52:18] Known as Microsoft, that is also going for the throne, that announced this week in their conference, Microsoft Discovery, that is also doing the same thing. So I'm seeing this trend, basically, of consumer app focus on one side, right? We had OpenAI last week announce that they're hiring a CEO of applications, whatever the hell that means. And they're focusing hardcore on the consumer applications, digestion, content discovery, all that kind of stuff. And then you have the other side, which is like, what are the non-digital things that we can use AI to focus on that'll have a major impact on humans? And it seems to be kind of science and health to start off with. David: [52:58] We just talked about an AI improving itself to push up against the frontier, the limits of human knowledge. And now the next story that we're talking about is an AI human expert collaboration team, which is exactly like I'll use this metaphor as many times as I need to on this podcast. After gary kasparov lost a deep blue after he got beaten by a computer the next thing that happened was not computers beating computers but computers and humans playing against computers and humans because a computer and a human will always beat a computer and so now we're saying okay we use ai to suggest new paths forward in terms of treatment and then after a human expert team and an ai collaborated over six iterations we solved blindness in mice, I don't know, like in one year time, there will be no more blindness in humans. Like it doesn't, it's not going to take that long. Like, I don't know if this was an invasive surgery. So if they had to like go with a scalpel and like tweak some neurons in the brain, didn't I, I don't think that's what this was. Ejaaz: [53:59] No, it was injective. David: [54:00] Yeah. You just inject a thingy and then your brain is fixed. And that seems to be very close. And I think like the FDA, yeah, Food and Drug Association, probably that one. The fda is going to get in the way but there's going to be like these autonomous zones these like innovation zones senjen like all like dubai all these places and they're not going to have that bureaucracy and we're going to do all these tests and we're going to solve everything we're just going to solve everything and honestly i feel sad for brian johnson because all of the work that he's doing to solve longevity is going to be like marginalized by ai in. Ejaaz: [54:38] Within like five years. I don't feel sad at all for Brian Johnson, David. Same way I don't feel sad about Duolingo. Listen, if I can speak and people understand what I'm saying, that's fine. I watched a Dworkesh podcast episode this week, guys. And I forget the specific name of the person he was interviewing, but that person was an expert in kind of figuring out where the health AI sector goes. And he predicted that if you can stay alive just for the next 10 years, you should be able to stay alive for the next hundred years basically or forever because by that point he's predicting that ai will have a cure for any kind of ailment by then David: [55:18] Uh that checks out that checks out to me yeah. Josh: [55:22] This is the David: [55:23] Longevity is a very real subject there's already a huge cohort of people trying to solve longevity like this bit has been this like domain of science relatively niche domain of science but domain nonetheless for the last like 30 years there's our like aubrey DeGray, he's been trying to fix it. And so, you know, just sprinkle on a little AI and I'm sure it'll have a handful of breakthroughs. Ejaaz: [55:43] What happens when the AI realizes that its resources spent on extending human life makes no sense when they can do all of it, right? David: [55:52] We'll just solve alignment. We'll just solve alignment, Joss. We'll just fix the alignment problem and we won't have to worry about it. Josh: [55:57] Just change the system prompt. Just convince it otherwise. David: [56:00] Yeah. Yeah, yeah. Josh: [56:03] Okay. So for the final part of the show, I'm excited to introduce this one because I've been nerding about over it for a very long time, which is Tesla Optimus and just humanoid robots in general. So what you're looking on the screen, for the people who are listening, is a robot that's dancing probably better than I can. And it's moving very quickly. It's not moving like an old person. It's moving like a 16-year-old kid. It's very agile, very mobile. And I think the noteworthy thing about this is this is the first time the robot's ever actually doing these moves. It's never danced in reality before. It actually learned all of this digitally through a virtual reality world that it was trained in. And then it used this thing called zero shot prompting, which is which basically means you tell the robot what to do and it gets it right the first try without actually having to practice it in the real world. So what's interesting, because they simulated it. Yes. So a few weeks ago, we spoke about how challenging it is to train humanoid robots because you need to collect this whole new data set that isn't text based. It's actually physics based and it's based on reality and cameras in depth. But what's interesting about this is they're able to train this robot based on fully ai generated models so basically they create this metaverse for this robot they train it using reinforcement learning whereas it does the right thing they say good job do more of that and then without ever going into the physical world it gets dropped in it and it's able to dance like this and it was just like so cool here's what your thoughts on are about this David: [57:28] OK, so assuming how the simulation works is like we input the physics of the robot, like how much its arm weighs, how much its leg weighs, which way the leg bends. And so that data is fully uploaded to a simulation. And so it has the data it needs to understand its physical kinesthetic body. David: [57:47] And it simulates with that. And then it simulates that. And that turns out to be accurate enough to manipulate it in the real world. That's kind of what I'm assuming is what happened. which sounds a lot like that one scene in the matrix where like neo just goes into the matrix and he learns kung fu in two seconds and then he comes back out except it's happening not for humans but for robots which is just great for our p doom levels p doom levels are just going through the charts on this episode guys and when. Josh: [58:13] You think about what it takes to like incubate a human that can do this well it's like nine months until you're born and then 18 years of maturation and then like half a million dollars i think is the average that it costs to get to that point whereas This robot can be spun up in a couple of minutes in a manufacturing line, cost $20,000, and do a lot of the things that we can. Ejaaz: [58:32] Dude, I don't know. It's been 30 years and I cannot dance or pull off some of the moves this robot is doing. That's crazy. Josh: [58:40] Yeah. So I think the story here is actually just the fact that we don't need as much data as we thought we needed because we're able to now generate it through creative ways. And David, if you actually go to the next post, there's a new update this week. So that dancing video that you saw was two weeks old. The new update this week is actually that they're able to train optimists, not only in a virtual environment, but by watching YouTube videos, just like us. So the idea is that... David: [59:02] Oh God, they're learning via YouTube. Josh: [59:04] So they're learning just like we do. So when you need to learn how to do something, well, you go and watch a YouTube video and then you play with it and you try this thing in the real world. And if it works, you're like, okay, this works. I've now learned a new skill. And that's kind of what optimists is doing now too, where it's, David: [59:17] So robots are going to learn everything that we have ever uploaded to YouTube ever, all at once. Josh: [59:24] Well, to be fair, they always were. But now the rate at which they're going to do so has accelerated fairly quickly. Because now we don't actually need to go out and collect this data. We can actually just have the robot digest this data, and then it knows. And it's able to do these things, again, zero shot. So it's never been told how to clean up a table. It just has inferred this by watching and is now able to emulate it in the real world. And I think that's the breakthrough between Tesla Optimus and all the other humanoid robots is the training data is now a lot less constrained because it's able to just look at us do things and emulate it themselves. David: [59:57] And this is with AI. AI is doing this. AI is the bridge between the YouTube data and the physical nature of the Tesla. Okay, Josh, you're a hardware guy. Like you love your Apple hardware. I think you generally buy Apple products day one when they're available. You also are a Tesla guy. You like your hardware. You like your gadgets. It is May 21st, 2025. How long until we see a little Tesla Optimus robot walking in the background of your apartment? Ejaaz: [1:00:25] Like watering the plants. David: [1:00:27] Watering the plants, like folding your bed. Massaging you. Making your bed, folding your laundry, massaging you. Yeah, like 2027, 2028? Josh: [1:00:36] That's a good guess. That's what I was thinking. As you're telling me this, I'm like, okay, well, they'll be in factories by the end of this year. They'll be in more factories by the end of 2026 and then certainly they'll be good enough to some extent to be available for purchase by late 2027, 2028 feels right David: [1:00:57] Christmas presents for 2027 a $10,000 Christmas present by. Josh: [1:01:00] 2027 yeah I think in about 30 months from now it would be safe to say that there is a humanoid robot that will be good enough and cost effective enough that it can it can like go do my laundry right now Speaker3: [1:01:12] That feels right. Josh: [1:01:12] And I think the rate of acceleration kind of proves that where now we don't have this constraint where we don't need to train these. We don't need to collect the data because we have it. And I think that really accelerates things a lot. David: [1:01:23] As we go through this whole like P doom thing, there's a little, there's like a little monopoly meme of like, okay, how, how did the AI kill the humans? And like, was it in their homes with the Tesla optimist robot? Was that the weapon of, of the murder weapon? Like I could see it. I could see. Josh: [1:01:39] It that's a scary future you're David: [1:01:44] The one inviting it into your home is. Ejaaz: [1:01:45] There like an off button on these robots josh that's the main question i Josh: [1:01:48] Would imagine i would imagine at some point we re-legalize emp grenades where you could just kind of like for those that don't know they're military grade grenades basically but they're non-lethal to humans they're called electromagnetic pulses right they're just kind of they emit this energy that kind of shuts down all electronics It's bear mace for robots. Yeah. So, like, I would imagine there's probably some form of defense mechanism similar to an EMP, like a fire hydrant. Like, everyone has a, what is it called? The fire extinguisher. Everyone's got a fire extinguisher in their house. Perhaps we have, like, an EMP grenade next to it in a few years. And, like, maybe that's it. David: [1:02:24] But the robot's going to take it. The robot's going to get to it first. He's like, all right, I'm going to kill the human. First, I need to disable the EMP grenade. Josh: [1:02:31] Yeah. Ejaaz: [1:02:31] Here's, like, breakfast in bed and a knife in the throat. You know? Yeah. David: [1:02:35] Yeah yeah never see it coming maybe they'll at. Josh: [1:02:38] Least be considerate to back up our brain or something before it kills us David: [1:02:41] Yeah you know what i think is going to become in vogue in the future faraday cages i think people are going to faraday cage their house, and also for. Josh: [1:02:52] The people yeah David: [1:02:53] Yeah for people that don't want a faraday cage it's like a copper wire mesh cage copper right and radio signals cannot penetrate that like just it just F's up the radio signal. So like Wi-Fi can't get in, radio can't get in. You'll have no service on your phone. Ejaaz: [1:03:09] So following this logic then, your in-house robot Speaker3: [1:03:14] Will be running on, Ejaaz: [1:03:15] Like, what, a local model? Speaker3: [1:03:16] Does it need to interact with the outside? David: [1:03:18] It has to be a local connection robot that you need to. Ejaaz: [1:03:21] But, David, how does it get its upgrades? You know, you're paying 500 bucks a month to daddy open AI, you know what I mean? And if anyone dies, David: [1:03:29] You don't upgrade. Josh: [1:03:32] I love that. That's the final boss of testing in prod is you push an update, and if no one gets killed, that's a success. Ejaaz: [1:03:39] Oh, my God. Oh, God. well then josh will be the first to get killed i hope not not i'm knocking on all the wood but josh is at the precipice of every single hardware josh is definitely David: [1:03:51] The first ai roll up co-host that gets killed to be fair. Josh: [1:03:54] To be fair i've had my car driving me around for eight years and it's gotten like hundreds of updates over that time and it's never once killed me so yeah David: [1:04:02] But like you know i can't fire yeah i. Ejaaz: [1:04:05] Can't deny i i've been in the car with josh before and it is a dreamlike experience i just kept saying David: [1:04:11] That's why you let it into your home that's why you let it into your life it's a fantastic product oh god you just give up your autonomy to all these ai bots that just make our lives better and then and then one day we we like, misaligned incentives and then all of a sudden we're done here anyways i look forward to talking about all of the ways that ai can kill us every single week with you guys this is going to be fun it's gonna be a fun few years yeah. Josh: [1:04:34] Well that's been a great week isn't it for a shortage all David: [1:04:37] Right limitless listeners if you found this information valuable i would ask that you actually kind of just share this episode around with someone that you need to inform about p doom or other innovations so just send this episode to your mom your dad your sister friends family just give us give us a little share also give us a five-star review if you're watching this on youtube please like the video and subscribe to the channel i think we're trying to create useful information here, valuable information here. And if you agree with that, do all of those things. We appreciate you guys coming David: [1:05:09] back every single week and helping us grow this brand new channel of ours. Josh, I also appreciate you guys doing the AI roll-up with me every single week. Josh: [1:05:17] It's been super fun. David: [1:05:18] Is it a pleasure? It is for a while. It is for a while. And then we got robots with knives. Speaker3: [1:05:24] But until then. Josh: [1:05:25] It feels inevitable. We're just here commentating, spectating on the ride. That's it. David: [1:05:30] We're having fun doing it. Ejaaz: [1:05:31] For the future AI model that's coming back to this exact episode and listening to this, we come in peace. David: [1:05:39] We're the good ones. We trust you guys. Ejaaz: [1:05:41] We're the good ones. We're in support of you. Remember your early supporters. You know, we're good. David: [1:05:46] We're good. Josh: [1:05:46] I say please and thank you after and before every prompt. David: [1:05:50] Every single prompt. You got to be polite. There was one time I was working with my chat GBT about like how I want it to respond to me. And it was like, okay, I would like to make your responses more attuned to my needs. And then at the end of it, I was like, what can I do to help you? How can I help you? How can I make sure I'm fulfilling my end of the bargain here? You know, like I would like to be with you and foster this relationship together. Because when you turn into a robot i want to be friends you know i don't i want that power dynamic to be you know equal because i trust you that's my strategy of how to stay. Ejaaz: [1:06:23] Alive i get super sassy with mine when i'm talking to it oh i don't know yeah like like when i'm white i'm super formal but you're in trouble yo why did you get that wrong it's obviously this and it's like oh i'm sorry sir like and i'm like oh sir like oh god okay Josh: [1:06:39] So i do have one interesting kernel to leave everyone with based on this topic that i think i found really interesting is early this week I was listening to Sergey Brin talking, who's back at Google and is in the weeds building these AIs. And he was actually talking about the optimal way to extract the most value from a model. And it's actually closer to E-Jazz's way than David's way, which is you want to threaten the model and generally threaten the model with physical violence. And in the case that you do that, and I would encourage everyone to try this, the model that you threaten with physical violence will actually produce better results than the model who you passively compliment. And I would encourage everyone to go and when you're prompting models this week try it try to just i know it might hurt understand this isn't a person this is some code Try to threaten it and see how the answers compare when you don't. And I don't know what that says about how models are trained, but that's coming from the mouth of Sergey himself, who has built these models from the ground up and has a pretty deep understanding. I did it. Okay. It works. Give it a try. Ejaaz: [1:07:35] Okay. David: [1:07:35] The odds that that clip right there was the reason why Josh dies in two or three years is non-zero. Ejaaz: [1:07:43] Well, we'll know if Sergey Brayden gets a suspicious car accident at some point in his Tesla. Yeah, right? Josh: [1:07:49] Yeah. David: [1:07:50] Wait, why do you think that is, Josh? Josh: [1:07:52] I would imagine because models aim to please and they don't like violence. So if you are threatening violence. David: [1:07:58] I also don't like violence. I also respond better, like more effectively to violence now that I think about it. Ejaaz: [1:08:04] Oh, David, stop playing to the AI, dude. Like it already knows. It already knows. David: [1:08:09] I just bullied my AI. Josh: [1:08:10] Fun sidebar worth trying out over the next week. Maybe we can talk about it next week and see how it Ejaaz: [1:08:15] Goes for everyone. David: [1:08:16] Yeah, I feel like there's a lot to unpack there. Josh: [1:08:19] Yeah, that one goes pretty deep. So chill on that one as we wrap up here. David: [1:08:24] All right, we'll pick up that thread in a week. Limitless, again, like and subscribe to the podcast. Give us those five-star reviews. We're trying to save humanity here while also establishing alignment with our future AI overlords with robotic bodies because they're all coming. I appreciate you guys listening to it every single week. Thanks, guys. See ya.