moonshots ep185 americas ai plan google search transcript

America’s AI plan. This was unveiled by President Trump. This sounds like a war footing in the global AI war. >> It is a plan to turn the US into one huge AI factory. >> There’s no precedent in history for what’s about to happen. Uh, you know, everyone always compares it to the internet explosion of growth, but this is so so so much bigger. My sense is this is potentially the the broadest US industrial strategy that we’ve seen since President Eisenhower. >> China is just just covering their countryside in solar. And I don’t understand why we’re not doing the same. >> We’re going to solve the energy problem, but if the chip supply gets disrupted by, you know, a Chinese invasion of Taiwan or otherwise, that’s going to be the real vulnerability. >> I spend a lot of my time thinking about the day after super intelligence. And I I think the day after looks like Now that’s a moonshot, ladies and gentlemen. >> Everybody, welcome to Moonshots, another

[00:01:00] episode of WTF. I’m here with my Moonshot mate, Dave Blondon, uh the head of Link Exponential Ventures, an extraordinary visionary. And with a new moonshot mate, Alex Quzner Gross, uh, our friend Selma is driving his son to some summer camp because that’s what you do when it hits July and August. Dave, good morning. >> Good morning. >> Yeah. >> Oh my god, I’m so so excited to talk to Alex this morning. Um, >> well, listen, I mean, I know how awesome Alex is because I get a chance to hang with him in Boston, but most of the folks viewing don’t. Would you take a second and uh and tell everybody watching about Alex? Uh yeah, absolutely. So, Alex and I have been chatting for years now about everything going on in AI. Uh, you know, I meet so many MIT and Harvard people, but you rarely meet these like crazy true high geniuses. So, Alex, he got three degrees from MIT in in four years. And you got math, physics, computer science, the three hardest degrees you can get, and then MIT Bandit right after

[00:02:01] that because it’s just too stupid that any human being would ever put themselves through that. Peter, you would know. I mean, you’ve seen that before. crazy, crazy, uh, productive, but also um, you know, Alex was one of those child geniuses who worked with nanotechnology when he was like 18. If you want to check out Bloomberg, you can read articles about him as a as a teenager. Uh, then he went to Harvard, got his PhD in physics. Uh, and he’s been studying and reading every single detail of what’s going on in AI. Uh, tracks everything in technology. If you want to check out his website, alexwg.org, He’s got papers on there like, “Can you build a computer out of pure gravity gravity waves?” Uh, so just an absolutely wild brainstorming partner, but super super fun to talk to. We meet every Monday. I can barely keep up with everything that he’s studying. It’s just so fun. >> Alex, welcome to the pod. >> Thanks, Peter. Thanks, Dave. Amazing intro. >> Yeah. Well, welld deserved. >> Pleasure. >> Yeah. And I love our conversations and excited to have you share your

[00:03:00] brilliance with everybody viewing. So, get ready, guys. We’re going on an extraordinary trip. Today’s episode is chock full with hot topics. We’re going to be talking about the AI wars, who’s winning. Uh, and it is really a battle galore. We’ll be diving into America’s AI action plan and what it means. Uh, the browser wars are back. This is not Bing versus Google. This is Google really fighting for dominance across all the LLM. of course, China versus the USA and a special peek at I’ll take fries with that Model Y. And you’ll see what that means in a moment. So, um, today I am sporting my, uh, my exponential mug here and my, uh, my gratitude mug. One’s got water, one’s got coffee, but that’s the mindset I’m in. Speaking of mindset, I’m going to take you guys back to the year 1993. And there are these uh ads that are

[00:04:01] playing that back then look like science fiction. I want to just take a second to sort of acknowledge how far we’ve come. Uh so listen up. This is AT&T 1993, the you will ad campaign. >> Have you ever opened doors >> with the sound of your voice? her car, please carried your medical history in your wallet. Your wife’s going to be just fine. [Applause] >> So, this is where we stand on the atrium >> or attended a meeting. I >> really like what you guys have been doing, but >> in your bare feet. >> And I have a few other ideas. >> You will. And the company that’ll bring it to you, AT&T. >> Which one? >> Have you ever had a classmate? >> I have a bike. >> Who’s thousands of miles away? here. [Music] >> Conducted business >> in a language you don’t understand.

[00:05:02] >> You will. And the company that’ll bring it to you, AT&T. >> Ah, well, kind of close, but the company that brought us a lot of that stuff wasn’t AT&T. >> Oh my god. Do you remember those? I do. >> I do, too, actually. It was so inspiring at the time. And uh nobody believed any of that stuff would actually happen at the time. Uh just a few years later it all becomes reality. But you’re you’re exactly right. It’s like yeah, it came from a whole bunch of startups and a whole bunch of companies you never would have thought of. Google didn’t even exist back then. >> Uh so it didn’t come from AT&T. >> I’ll I’ll take the the contrarian point there. I I would argue not only did did the future materialize for the most part, but AT&T, which you know, think back now 30 plus years, AT&T was for a brief period of time, um, based on my research, the most valuable company in the world in 1993. So there’s a certain sense in which uh as largest market cap company you can see the future and and I

[00:06:01] I would argue that in fact AT&T as you remember when the iPhone launched uh as the the the sole at least in the US carrier launch partner of the iPhone AT&T was in the end uh a key enabler of the future to the extent that most of the tech demonstrated in you will video conferencing or substantially similar to video conferencing. they they did at least help enable that. >> Yeah, I I I I agree. I remember I bought one of their first video conferencing calls. Um and what it was was a telephone with a screen that would refresh an image every like 5 seconds. It was that was that was I bought one for my folks and I was living in California and they were in in Florida. But so the question is what’s it going to be like in 30 years from now? And I think you know given the speed and you know Ray Kerszswwell was on my stage at the abundance summit and he’ll be back again. Uh and he said listen between

[00:07:01] 2025 and 2035 the next decade we’re going to see as much progress as we saw between 1925 and 2035. So uh >> yeah I think you know one of the things that that makes the American economy so incredibly strong is the constant turnover of leadership and you know if you look at the Magnificent 7 today and you say how many of those companies existed back when this ad was made you know it’s basically none >> Microsoft was was still around >> uh 93 Microsoft was little >> 80ist right >> and Exxon and Walmart they were all trading places throughout 93 >> yeah so you know you Hey, 2030. Well, yeah, everything is the pace of change is always accelerating. So, 2030, you know, if you look at 1993, you’d say 30 years in the future. If you say from today to 2030 is about the same amount of likely change. So, I I suspect there’ll be a whole bunch of new names that are in the trillion dollar valuation camp that we’re not not thinking of right now. And that the challenge is to identify them.

[00:08:01] >> Yeah. >> Kind of predict. >> I think what’s pushing this right now is the AI wars. So, not the Clone Wars if you’re a Star Wars fan. It’s the AI wars here and today. Every week, my team and I study the top 10 technology meta trends that will transform industries over the decade ahead. I cover trends ranging from humanoid robotics, AGI, and quantum computing to transport, energy, longevity, and more. There’s no fluff, only the most important stuff that matters, that impacts our lives, our companies, and our careers. If you want me to share these meta trends with you, I write a newsletter twice a week, sending it out as a short two-minute read via email. And if you want to discover the most important meta trends 10 years before anyone else, this report’s for you. Readers include founders and CEOs from the world’s most disruptive companies and entrepreneurs building the world’s most disruptive tech. It’s not for you. If you don’t want to be informed about what’s coming, why it matters, and how you can benefit from it. to subscribe for free. Go to dmmandis.com/metatrends

[00:09:00] to gain access to the trends 10 years before anyone else. All right, now back to this episode. Let’s take a second and give a quick overview on what’s going on because it is uh everybody’s trying to outdo everybody else. Uh here we go. Taking a quick look. Uh first off, of course, Elon with XAI and his GPU targets. Here are the numbers. Colossus launched in July of 24. hard to believe was a year ago, right? Uh with 100,000 H100s, they doubled it to 200,000 in 3 months. Colossus 2 is getting ready for launch uh with the equival with 550 uh GB2s, but the equivalent of 5.5 million H100s and his goal is 50 million H100 equivalents within the next 5 years. I mean, that’s insane. Alex, >> it’s so incredible. If my arithmetic is correct, in today’s GPU dollars, 50 million H100s is is the trillion dollar

[00:10:02] AI supercluster that people have talked about for the past few years. Obviously, there will be uh some deflation of cost. So maybe it’s only a few hundred billion instead of a trillion dollars, but in in today’s GPU dollars, that that’s a substantial fraction of the US economy. >> For sure. Alex and I had a meeting earlier this week with our new secretary of commerce coming into to the state of mass and uh he put together a brilliant 18-point plan on how to be a dominant state or a successful state in the world of AI. Uh I I suspect none of it resonated but it will after these numbers come true on this chart. It’ll all come back to Alex. But uh yeah, I did a little bit of research on this though that you know each of those GB200s has actually two Blackwell GPUs. So you’ll see the the production of GPUs, but you have to divide by two because each one of these these super chips uses two of them. There is an acute shortage. Like everybody’s going to want those chips. They already do,

[00:11:00] which is why Nvidia is worth so much. But when you see all these, you know, numbers of GPUs, numbers of GPUs, remember if you’re using the Blackwells, the GB200s, you have to divide by two. And so, you know, the the 10x performance improvement that you see on this chart between H100 and GB200, it’s a little bit overstated actually. It’s it’s about a 4x raw compute performance increase. Uh but then the NVLink networking is much much more efficient and uh it there’s a bunch of other innovations. So, you do effectively get 10x the AI out the other side. >> So, the question is how much of this is the battle between Elon and Sam that’s that’s driving this? So, check this out. Of course, uh, OpenAI’s goals have to outdo Elon’s goals. And this is a this is a uh tweet from from Sam Alman says, “We will cross well over 1 million GPUs brought online by the end of this year.” Amazing. Very proud of the team, but now they need to figure out how to get to

[00:12:00] 100x that. So, I mean, it’s an extraordinary battle. Um, thoughts? Yeah, I think a lot of to the extent that much of the the compute demand is inference demand. Call it a half to two/ird of compute ends up being allocated to inference as opposed to to training or pre-training. A lot of this potentially is going to be unlocked by new use cases. The the capex requirements, the opex requirements are ultimately going to be driven by new use cases. If we can solve all the world’s problems, if we can cure every disease, that’s going to unlock an enormous capex budget that can then be invested back into these GPUs. So I I think the the elephant in the room here is can we solve math, solve physics, solve medicine and through those solutions reinvest uh the games as uh directly into these compute budgets to unlock millions of GPUs and hundreds of basis points of GDP being allocated to GPUs. >> I want to dive into that with you, Alex,

[00:13:01] a little bit later. What does it mean to solve math, solve physics, solve medicine? because I think this is part of the extraordinary future that uh that I wake up just vibrating about every morning and which most people don’t yet fully gro. Uh but let’s get into that in a little bit. So again, it’s XAI versus uh Open AI. Uh but Meta is not going to be left out of the picture. I I love this. So Meta is building a Manhattanized data center. So Prometheus is a multi-gawatt data center. You know, it’s interesting that we’re now measuring data centers in terms of of power instead of necessarily chips, right? So, it’s like a 1 to 5 gawatt data center. >> Uh even more interesting to get to speed, they’re deploying in hurricane proof tents, which is insane. And then they’ve got their plans for their second mega cluster, Hyperion, uh in the works. Um, so Dave, what do you make of all

[00:14:02] this? Just all out all out AI wars. >> Yeah, there’s no there’s no precedent in history for what’s about to happen. Uh, you know, everyone always compares it to the internet explosion of growth, but this is so so so much bigger and it’s unbounded, you know, like uh so um you can see there’s no such thing as a hurricane proof tent obviously, but the race to get these things up, I mean, come on. Seriously, I’ve been intense before. I don’t care what you make it out of. There’s no way. But the race to get these things up and running is so so acute so fast. Uh, you know, and it’s hard to get the power. And so you put, you know, you put the chips where the power is. And, you know, any structure is good. I’m sure they’ll they’ll build around the tents eventually. Uh, but yeah, the scale is basically only constrained by how many chips can I get, how do I get them wired together. Um, you know, we’re finding the power. It’s actually depriving other manufacturing operations of that same power supply, but nobody cares because because this is so much more important of a use case.

[00:15:01] So, the race, you know, we’re we’re still in the first or second inning, but man, is it going to get exciting over the next year or two? >> It’s already exciting. Alex, what do you make of this? >> It’s all about the latency of construction. If we had years to build these data centers, we would probably use normal construction methods. But if the data centers have to be erected on a very short time scale, materially less than a year, tents are the way to go. And I think to the extent that there is further acceleration uh in in the construction space, I I think uh optimistically this points the direction of totally new form factors for data centers. One could imagine ocean-based data centers or space-based data centers uh and maybe even new modular data centers, data centers on on wheels to the extent that electric vehicles with GPUs ultimately become distributed data centers. I I think potentially we find ourselves in the near future where data centers occupy all sorts of environmental and geographic niches that otherwise would be insane.

[00:16:00] >> I love that. >> What’s interesting is those GB200s are $60,000 a pop. And so so you’re you have to be really careful that you don’t put them on a on a set of wheels and it falls off a cliff or something like that. >> Well, it’s really >> the idea that you know that Elon’s spoken about is having all of the Teslas being deployable energy sources, right? So you can imagine uh having them having GPU capability on board as well and just just a I don’t know millions tens of millions of of Teslas computing and deploying energy around the country. >> You know the one of the reasons that’s such a brilliant idea is because the uh the data centers need absolutely continuous power. You can’t just turn them off at night because the chips are so expensive. They need to run full throttle all the time. And so all that Tesla power that’s buffered up in the cars, if you if you have a wind or solar outage and you need power right away, buffering it in the cars actually makes a ton of sense. >> But look at the size of this. I mean, if you’re if you’re listening, what I’m

[00:17:00] showing you here is an image of the uh of the Prometheus data center, actually Hyperion, mapped over Manhattan. It’s the size of Manhattan. I mean, we’re basically on the very very first baby step towards computium where everything’s turned into compute. Alex, what do you think about that? >> I I I think that’s the multi- trillion dollar question. So do we find ourselves in a near future where due to what one might call horizontal exponentiation as opposed to the vertical exponentiation of Moors law where we have to disassemble our solar system uh because we need the computer or or do we discover algorithmic advances or physics advances that make the entire notion as laughable as the the fears from say the early 20th century that horses would overrun Manhattan. Maybe maybe this is the the naivee of 2025. Maybe we we have achieved such advances in physics and algorithms that the idea that we need to

[00:18:01] tile the earth’s surface with computium AI data centers is laughable and quaint uh in 20 years. >> Well, we’re going to we’re going to find out. So meta’s in the game uh in a number of different ways. We’ll see that. Uh and here’s one of the other ways that meta is in the game. And I just I took a moment to lay this out because I think it’s fascinating. So Meta’s super intelligence team has an extraordinary makeup. Let’s take a look here. First and foremost, half of the team is from China, right? Uh so these are Chinese Americans or Chinese immigrants that are been hired by Meta. uh 40% of Meta’s team are from Open AI, I should say, poached from Open AI, 20% from Deep Mind, and of course, 15% from the recent uh acquisition of scale. 75% have PhDs. Um and I’ll just make one note that each of these members of the

[00:19:00] team are likely getting a salary of between 10 million to hund00 million per year uh at the lower end. So what do what do you make of this makeup, Alex? I >> I think the the headline is is the bottom line here about the compensation. I I think that this is a preview of post scarce individual economics where individuals who are sitting on top of enormous amounts of per capita compute have compensation equivalents that are in the hundreds of millions. This is a preview optimistically of what everyone on Earth postabundance, post artificial super intelligence could be seeing. >> Peter, do you It’s so rare you get to live world history. It’s so cool. But remember we met with Mark Chen. >> That was about a month ago at OpenAI headquarters. Yeah. >> So he’s the head of research at OpenAI, MIT alum. Awesome. Super low-key, very nice guy. He spent the whole day with us. Then it was I think just a couple days later he got that call from Mark Zuckerberg saying hey what would it take

[00:20:00] to bring you over to Meta 100 million a billion and he said I I’m going to say no to a billion dollars and then that made the Wall Street Journal cover just right after that. How do you how do you do that? I mean first of all your your EA your assistant says oh by the way Mark Zuckerberg’s on the phone and you’re like oh boy how high is he going to go now? That’s a >> but Mark Chen in that in that call with Zuck was the guy that triggered this whole wave cuz you first of all he said no to the billion dollars but then he said you know you really ought to be investing more in brilliant human capital. You’re going to spend I guess on the other slide it said 75 billion uh imminently on on the chips on the capex you know a good algorithmicist can get two three fourx more value out of that investment put some money behind the the people here. >> So I think that trend is going to continue. >> I have a question for you Alex. How long is individual human talent going to be uh recognized and paid at this level?

[00:21:00] Because I mean as we head towards you know digital super intelligence isn’t the human in this specific role going to be obsoleted or is this kind of salary level and this kind of commitment to human capital going to continue for decades? >> Well, you’re asking the accelerationist here. I I think humans many of them are going to choose to merge with the machines. So I think it’s almost a trick question. I I think this is a a preview of near postcarce economics. >> Uhhuh. So, by the way, everybody, uh just so you understand, Alex Weer Gross is definitely an accelerationist and we’re going to speak about uh timelines that are shockingly fast compared to what others would say. Let’s put it that way. So, um, but I still believe we’re going to are we going to see salaries dropping off or continuing to rise? What’s your bet on this, Dave? >> Uh, well, it’s continuing to rise, but I think most people are like are thinking, well, what does that mean for me? Cuz

[00:22:01] because it’s really weird, right? You’re seeing some sets of people get insane comp packages while other people are getting laid off. And so like it the the storyline is turbulence. Not, you know, there’s going to be incredible abundance. So there’s plenty of value to go around, but it’s not going to land where you would normally expect. And you have to really, really think through, okay, what’s rising and what’s falling. And you know, Sam Alman is always saying this on stage, be nimble, rethink your life, rethink it every month in light of what’s happening in AI. And don’t get stagnant, you know, like listen, focus, you know, stop listening to mainstream media. Stop wasting your time on the next whatever. Focus on what’s happening right here and use it to remap your life. So yeah, no, it’s going to go up, of course, and but where is it going up and why? You know, what what type of jobs, what type of people, what type of roles, >> you know, Dave, one of the things that that you state uh when you’re searching to invest in teams through link XPv

[00:23:02] is you’re looking for best friends. And you know again for those who don’t know link is uh link exponential ventures is a a billion dollar plus firm out of uh based in in Cambridge that invests 80% in MIT and Harvard teams at the beginning uh and when you’re looking for the team makeup uh can you explain that once again? >> Yeah that’s a great great uh question in the context of this slide. So, we’re looking for for teams of uh super super tight-knit. Uh they have to pass what we call the Fred Wilson test. Fred Wilson is the most successful venture capitalist of all time, founder of Union Square Ventures, awesome guy to study by himself. Uh but he has this three-part rule. He only invests in teams of three or more best friends who quote unquote write the code themselves. So, all three are equally capable of doing each other’s jobs. So, they’re replaceable amongst each other. Uh and he trusts them. And if they pass those three filters, you invest even if it’s a bad business plan

[00:24:00] >> because the business plans in the age of AI can change in a moment. Yes. >> But the team dynamics won’t. But now it’s more important than ever that they’re best friends because sooner or later Google’s going to call or Meta’s going to call and try and take one of the three away. And if that person cracks and defects, then your company unwinds. And we’ve literally never had a loss on a deal that comes out of MIT, Harvard, Northeastern. uh except in the scenario where somebody defects. Other than that, they always succeed in the end. And so that’s that’s the core theme that we’re looking for. >> I just want to I want to nail that down for all the founders listening. At the end of the day, you really want to partner with best friends. You want to partner with people you’ve known for a while who are going to stick by your side. And I I put this down here for that or ask you that question for that very reason. Uh here’s the other question. you know, for a number of decades, we’ve had this immigration issue where, you know, people get their PhD at MIT, uh, or Harvard or, you know,

[00:25:03] Caltech, whatever the case might be, and then they’re forced to go home. They’re sort of basically kicked out of the country instead of like stapling a green card to every diploma out there. And when I see numbers like this, like 50% have origins in China, I’m like, we want those people to stay in America, become Americans, and stay inside of our ecosystem. >> Well, lately, um, you know, the government has done a amazing job of getting out of the way recently. And, uh, one of the areas that we’ve gotten tremendous help is these 01 visas. Mhm. >> So if you know 01 visas used to be for celebrities, sports stars, you know, hey, the the Red Sox want this incredible pitcher from Japan, get him an O1 visa cuz we the season starts in just a few weeks. So now they’ve extended that to apply to AI experts. Like we desperately need this person from Romania, from China, from India. They’re brilliant and we need them uh get them an O1 visa. So if if they have the credentials to be a superstar of AI,

[00:26:02] they can actually stay in the country immediately on an 01. >> Yeah. Incredible. So uh Alex, thank you for these numbers. For those you this is poly market predictions for uh the best AI models at the end of this month and at the end end of August and the end of the year. Uh Alex, uh can you take a second explain Poly Markets and do you trust their prediction engine? Oh, it’s such an interesting question because this particular poly market prediction is based on language model arena type masses of people having conversations uh with uh with textbased language models with image-based language models. And so uh my understanding of what it’s actually measuring is whether the quote unquote average person interacting with a frontier model prefers that interaction or not. And the problem with that is that as we achieve super intelligence, many of the frontier capabilities are

[00:27:02] not necessarily benchmarkable through conversations with the average population. So I think if anything, we’re starting to see many of these uh call them community- based benchmarks start to recede in in terms of their predictive power. And it it’s left increasingly to specialist benchmarks that measure exceptional abilities. for example, open problems, the ability for for AI models uh hopefully in the near future to be able to solve outstanding open problems in science and math and engineering in in other disciplines that nonetheless can be verified. I think those will be far more predictive in in the next few months in the next few years than conversations with um with the general population which is what you see reflected in in these predictions where it’s the the same top four or five uh labs constantly switching places. >> Yeah. And so just to spell this one out by the end of August the market is saying open AI will be in the lead with

[00:28:01] 54% over Google with 41%. And I think that’s basically the telltale sign of GPT5 coming online. But not to be outdone, by the end of the year, Google’s placed at 45% likely to be the lead over OpenAI at 31%. And I’ve seen a lot of data that basically says, you know, Google is a little bit slower in the race, but will dominate over everything else given their their strength. How do you feel about that, Alex? Do do you sort of view Google as the ultimate winner here or how do you how do you place them against meta and open AI? Um and and of course >> I I think we want to live in a near future where there are lots of there’s lots of competition and there are lots of competing frontier models and and also it’s a heterogeneous ecosystem where we have open models, closed models, open weights, closed weights, APIs, non-APIs, edge-based inference, uh data centerbased compute. I I think we

[00:29:00] we want that really rich heterogeneous jungle and and ecosystem of vendors. So big fan of competition here. Uh would like to to live in that term that near-term future. But I I think there there’s also uh buried beneath these numbers the the big headline that we just zoomed right past the touring test. That’s what essentially this this poly market is is measuring and it it barely uh it it barely got any attention. Here we are debating who who’s going to have the the best post touring uh model. The fact that anyone has a post touring model is utterly remarkable. >> Well, I think we’re going to we’re going to whiz right past uh you know AGI or some version of AGI definition and past digital super intelligence. Just going to be looking at what’s next, what’s next. Oh by the way, oh my god, we’ve got super intelligence in our pocket. Um, >> Charlie Straws and Accelerondo, uh, which by the way, one of the best novels I think ever, has without spoiling it,

[00:30:00] has an amazing scene where you have uploaded humans talking to each other on a Star Wisp going to another star, debating when the singularity is going to happen. I I I think that that that’s the the world we find ourselves in today, you know, debating is it going to be Frontier Lab A or B that achieves slightly better postouring test benchmarks. We passed the touring test. >> Yeah. >> By the way, that book Accelerondo is pretty entertaining, too. But we make all of our partners read it so that they can keep up with Alex. You know, the terminology alone is it’s worth the investment. >> I read it. It was fascinating. And I’ve got to read it again. So, I couldn’t fully gro what it has to say. But we’re in this we’re in a every single week, every month, there is a a set of new developments that are just pushing the limitations. Uh we haven’t spoken about Anthropic at all. Uh and so the headline here is investors value Anthropic at hund00 billion. Uh its revenue surged from 3 billion to 4 billion in just a

[00:31:00] month and Claude uh is generating $200 million with 60% margins for their coding. Still considered one of the best coding engines out there. Uh, interestingly enough, of course, uh, Anthropic has been teamed up with Amazon and there’s been a lot of talk about does Apple buy Anthropic. Dave, what are you seeing here? >> Well, I use it every day and it does write the best code. Uh, there’s a talent war going on though. Actually, Kush Pavaria, one of our partners, wrote a a really cool little memo, open memo, anyone can read it. uh mostly targeted at MIT saying, “Hey, if you look at the Golden State Warriors and Steph Curry and the dream team factor, you know, Meta just did that. Look at all the people they just hired. They have the dream team and you can buy you can invest in their Poly Market or Khi at like 4 cents on the dollar right now.” So, I told him don’t give investment advice. He did it anyway. But, um but uh you know, they’re accumulating the talent and spending whatever it takes.

[00:32:01] So then Enthropic used to accumulate that that talent because they’re so conscious of the safety side of this >> and that’s you know Daario is kind of the thought leader on on mechanistic interpretability and and AI safety and that attracted a lot of great AI talent but now that same group of people is getting these $und00 million signing bonus offers. So we’ll see if they can keep up with that kind of pressure but you know as of right now it’s the best coding. >> I have a question for both of you. You know, we keep on hearing about the same names over and over again in terms of the frontier models. Will there ever be another frontier model that starts that we haven’t heard of yet? >> There definitely will because Alex and I are working on it. >> All right. >> Secretly. Uh, no, I think what a lot of people don’t know is the algorithmic improvements uh are, you know, factors of 10 to a thousand. And so everyone thinks that because these guys have massive valuations and massive budgets they’re going to run away with everything. Uh and you know in the short

[00:33:01] term that’s true but in the long term if you do come across a 100red or a thousandx performance improvement if you have the the willpower to turn it into another foundation model company then it can succeed. But then the big wild card is quantum computing. >> There you’re looking at just a huge step function opportunity. So I would say more likely than not there’ll be one or two more that get in the race. >> Alex, what’s your what are your thoughts on uh >> I think the open >> the open so so forth anthropics specifically I mean a I have a number of friends and former classmates on the founding team and very excited for their success. I I think the the the the valuation of anthropic is perhaps reflective of software engineering being automated. uh this software engineering being the the first major major labor category to to succumb uh in in in whole or in part to AI automation. I expect this uh the same pattern to play out through other highly productive labor

[00:34:01] categories. But I I think it’s also uh it also points the way to evaluations. I I made the point in an essay a number of years ago that the key to overcoming grand challenges in AI is having evaluations, benchmarks combined with data sets. And the the stated institutional focus at anthropic right now is focused on software engineering almost entirely. And so I think anyone aspiring to build the next great frontier lab, the the next great frontier model arguably should be laser focused on what are evals, what are benchmarks that I can focus on that no one else is paying attention to and then working backwards from those evals. >> Totally right. And and you know if you look at the leaders of the top AI uh foundation model companies, Dario is the one and only who is a a AI research pioneer from day one. Everyone else moved sideways to to get into their current role. You know, whether it’s

[00:35:00] Elon Musk or Sam Alman or or Mark Zuckerberg. And so he has this much more innate understanding of the self-improvement process because he he doesn’t want to dominate coding because coding is like, you know, much better than midjourney and jokes and and videos. He wants to win in in coding because of the self-improving AI process. And and there’s some great research that Alex found this week on a scaling law for how self-improvement is going to unfold. I don’t think we’re going to roll it out today because we’re still evaluating whether it’s real or not, but but he is in the middle of that phy. So if he wins this race, it’s because of that in intuition he has about using winning the code war, then have the coding engine run all night long, every day, self-improving the algorithms themselves and get that loop started. >> Insane. Um, all right. Here’s a interesting uh note. This is a a set of tweets that comes out that GPT’s 03 now runs on GPT5.

[00:36:02] that all 03 IE GPT4 turbo requests are now being routed to internal model called Zenith which is GPT5. So Alex, what do you make of this? Do you believe it? And uh and what are you seeing? Part of me wants to say uh based on uh very very publicized rumors, we’ll probably know the ground truth in in a few weeks or the next couple of months, but anecdotally uh I’m constantly peppering 03, one of my favorite models with challenging challenging to humans math and physics problems. And anecdotally, I’ve seen over the past few days 03 or what presents as 03 uh become able to solve challenging mathematical physics problems that literally a few days earlier it was not able to solve. So one person’s anecdote, it does suggest that uh that 03 perhaps at the back end has had some sort of capability leap and that’s very exciting. >> So they’re just test driving.

[00:37:01] >> Could be. I I think we’ll again we’ll we’ll probably know the ground truth. uh hopefully in in a few weeks or next couple of months, but seeing under the hood capability leaps is is incredibly exciting. >> All right, let’s talk about benchmarks. And uh Alex, I’m going to ask you to lead this. So on this chart here, we’re seeing a couple of benchmarks. of course, humanity’s last exam and the AME25 benchmark comparing uh Gac 4 heavy, Gemini 2.5 Pro, the 03 Pro, GPT5, and GPT5 reasoning. Uh, and the question of course is where are these numbers for GPT5 coming from? Are they validated? And do we believe them? But if we do, GPT5 is going to outra everybody. So take me through this, Alex, if you would. >> Sure. Well, I want to draw inspiration maybe from the the spirit of Ray Kurszswe uh who who famously uh

[00:38:01] mentioned that when the human genome project was 1% completed that it was actually 50% time-wise completed. I I think that the same the same notion likely applies here. Uh whereas uh some might say, okay, uh look at humanity’s last exam. Well, we’re we’re approaching 50% on on some of these purported uh strongest uh but unreleased models. I look at that and and then I I look at say models that are achieving only 20% on humanities last exam and I say we’re we’re most of the way towards fully saturating these benchmarks. I I think the real story here both on this slide and also the next slide. So on this slide we have humanity’s last exam uh and this year is Amy um uh the olympiad exam and then uh I think on if we can jump around uh so Google proof question answering and uh bench uh software engineering bench these are all saturating uh we’re running out of benchmarks we’re running out of evals

[00:39:00] that test frontier capabilities that I I think in in my mind is the headline story we need really hard eval now and going back also to the the point about LM arena uh and the poly market, we need much much harder benchmarks to understand and and differentiate between frontier capabilities and the the industry I think is is essentially starved of of those frontier benchmarks. >> Well, the the other thing we need so badly is is brilliant people like you to tie this into uh human good. You know, if you if you said when I get to hle 50 60 75 what does that mean for solving all disease? What does that mean for discovering new physics? What does that mean? Cuz it because it directly connects and we know intuitively it directly connects, but it’s it’s really mentally challenging to say, okay, what’s the timeline? How does it roll out? Uh you know, all the all the stuff that takes it from just a test to the real world. That’s that’s just a great use of mind power.

[00:40:00] >> Totally agree. And I just maybe quick thought on that. I I would argue that what we’re sorely missing as an industry and as a research community are benchmarks that measure the ability of frontier models to solve open problems. This what I when I see all of these charts, including, you know, pulling out the the purported GPT5 benchmarks, I see saturating benchmarks. I would like to see benchmarks that address open challenges, and I think those are the next frontier. >> Yeah, I love that. Two two thoughts here. The first is there’s going to be a point at which uh these models are solving problems that we can’t even understand and it’s impossible for us to create the benchmarks in terms of closed benchmarks. But I agree with you Dave. I mean you know why aren’t our benchmarks like which is the model helping us you know double the human lifespan? Which is the model helping us uh create the most uh you know the highest efficiency fusion capabilities or the models I mean basically models being driven to create

[00:41:00] this future world of abundance. U I mean this is some of the work that Immad Mustach speaks about with his intelligent internet. But, you know, it’s interesting that people tend to gravitate towards uh competitions uh and steer long-term decisions towards competitions, whether it’s making more money or or getting an Oscar, whatever the case might be. Why don’t we use as the benchmarks the things that uplevel humanity uh and get companies and teams to focus on doing those things? And that that Peter I I would say in in a nutshell that’s what we’re I think about to see in the next two or three years. Call it abundance bench or abundance benchmarking. And the more abundanceoriented benchmarks we see I I would predict within 2 to 3 years of the benchmarks if they can be uh mechanically verified by uh benchmarking organizations you’ll see a lot of those problems get solved by AI. Yeah, I think

[00:42:02] I need to write a newsletter on this on this front because >> you know what we me you know what we measure matters and uh and it also influences where people spend their money, time and their egos. >> Dave, you were going to say >> well no I I would love to see that newsletter sooner rather than later because you know we never used to interact with the top uh politicians in the world. Now they all want to know about AI. And so we have access to the state house, to the white house. And if you can’t tie it to human outcomes, then you don’t get the voters. And historically, we we didn’t think in terms of voters, right? We thought in terms of AI progress, how does it benefit the world bottom up? But now everybody wants to know what does this mean for me? Yeah. >> Who’s going to figure that out if it’s not if it’s not, you know, between you doing everything related to healthare um and and space and then Alex doing everything related to science, physics, technology that who else is going to figure this out if not you guys? And so if we can just tie it to like when will

[00:43:01] cancer be cured, what type of cancer, when will housing be solved, what what will be the new economics? Ahmad actually has to figure out the new economics, but this this is the place where we got to figure all those things out. are effectively having AI solve the X- prises of the future. I’m curious if if you’re if you’re listening here or watching as one of our subscribers. I’d love to know your thoughts. What benchmarks in terms of abundance benchmarks should we be creating? Uh yeah, drop a note and uh and share with us your your thoughts on this. So, speaking of IMAD, here we go. Immad Mustach believes AGI is here. Uh so, I’ll just read his his uh his tweet. says AGI is already here. All the components exist. We just need to stitch them together. Two years ago, who would have said uh an international um math Olympic gold medal and topping benchmarks isn’t AGI. So, Alex, you’ve been saying this for a while. I remember our first conversations. I’m saying, you

[00:44:00] know, when are we going to reach AGI? You said, Peter, it’s kind of here already. You were way ahead everybody else. So, talk to me about that. Yeah, I I I would take the position that uh a AGI has arguably been here since at the very latest summer of 2020 and that we’re 5 years into AGI. I I picked summer of 2020 because in May of 2020, OpenAI released the GPT3 paper that uh language models are are fewot learners. And that was the the first time that I I think the notion that in context learning that you could pose a task within the context window of a language model and have the language model learn uh in the moment just in time. That was the first time that that seemed to actually work. Uh one could even look uh further back at uh early language models that were based on tpples uh and smoothing tpples back decades. But I I I think when we look back from the uh with

[00:45:02] the benefit of hindsight, we will say this was a more or less a smooth exponential. There was no no singular moment, no before and after. It it’s just a smooth exponential of compressing human knowledge and compressing world knowledge. uh and I’ve drawn the analogy also that the idea that uh AI arguably inevitably is the result of compressing knowledge in the same sense that fusion power and other other physical phase transitions are just the result of compressing matter basically. really glad you said that because you know history is always written in hindsight right it’s never written in the moment and 2020 is a very memorable easy year for the history books to write down but I think what you said is dead right if you take the exact GPT3 algorithm set and scale the heck out of it it reaches AGI so nothing needed to change from that 2020 innovation era to today other than much much more compute huge data centers big GPUs and so then how would

[00:46:02] you say the birth of AGI was anything other than that moment in time. And so I think that’ll stick. >> Amazing. I love it. >> We just didn’t realize it at the time, obviously, but you did, but nobody else did. >> I love that idea. All right. Uh, some interesting news. And again, Alex, I’m going to lean on you here. So, OpenAI model wins gold at the math Olympics. Uh, achieving a score of 35 out of 42 in the 2025 Olympics. solving uh solve five of six worldclass programs. So talk to us uh Alex, what is the IMO, International Math Olympics, and how important is this? You’ve been saying for some time now that we’re in 2025 going to see uh these AI models solve math. What does solve math, solve physics mean? >> Oh, I I I love this this achievement. Uh it’s incredibly exciting. Uh so maybe just to to back up a bit, uh the IMO,

[00:47:00] the International Math Olympiad is is the hardest high school math competition. It it is the Olympics for math for high school students. When I was a high school student, I I was a member of the the US team on the computer science version of the International Math Olympics. It’s it’s a very competitive competition and many of the IMO winners go on to become professional mathematicians. Many of them go on to at least uh now as we know with the benefit of hindsight found frontier AI labs or or be uh key leaders within them. So achieving uh basically solving five out of the six problems on on this year’s IMO is incredibly exciting. I’ll go out on a limb and take maybe uh a slightly unpopular position uh at least relative to the math community and say I think as going back to to our friend Rey and 1% of the human genome project being or human genome being sequenced indicating half of the project or more than half has been solved. I I I would argue that we’re

[00:48:01] actually most of the way towards math being solved where where you operationalize math being solved as superhuman AI performance uh uh super professional mathematician uh AI performance a step deeper what does math being solved mean to the person listening >> it would mean basically that uh the work that professional math researchers professional mathematicians carry out can be fully be automated by AI models. You just scale them and out pops new mathematical insights. >> And I think >> same for physics. >> Same for physics. So if you look at another benchmark that uh where the results were recently announced, the frontier math tier 4 benchmark and I have number friends at at Epic AI uh who are managing that benchmark. If you look at the recent performance in math benchmarks, the hardest ones, the the ones that take professional mathematicians weeks to verify and andor

[00:49:01] solve independently, and you look at how the performance on those benchmarks has been improving over the past few months, and and if you believe in the law of straight lines, then it’s it seems reasonable to predict that we’re going to see 20% of mathematician level hard problems be solvable by the end of this year at the present rate of improvement. and potentially 60% or 70% of them be solvable in the next 2 years. And at that point, I would argue math has basically been solved. >> Amazing. >> Alex, I got to ask you like one of the problems most near and dear to my heart is can you run a neural a neural network on a quantum computer and writing those quantum algorithms is notoriously incredibly hard and very very math heavy. Does a five on six, five out of six on IMO mean we’re close to AI being able to write those algorithms or not? >> I suspect it it’s correlated. I I think there’s the common cause here which is the the rising tide of technical

[00:50:01] capabilities in math, physics, engineering in in frontier models. I I I think Dave, what you’re highlighting is maybe that we’re we’re missing uh a benchmark for quantum algorithm design and and maybe this is a call to action for the world. We need if if if the goal is to to build the world’s best quantum accelerated foundation model, we need evals to match. >> Just just so the audience knows, a a scalable functional quantum computer is within 3 years according to our our best experts around MIT, maybe even sooner. The software to run on it is is next level. But it’s software, you know, it could be solved very very quickly. >> So what is the implications of that, Dave, if we had that? >> Well, it’s Yes. So nobody knows cuz right now um you know right now quantum computers load very very very slowly like order you know thousands of times slower than a regular computer but once all the stuff is in there it processes instantaneously what could take you know a regular computer 100,000 years to compute it can do in seconds but only

[00:51:01] within very specific domains. So the question is is optimizing a neural network one of those domains or not? And I know aspects of it are we’ve already gotten that far, but but you know any bottleneck kills the whole system, right? So so the whole thing has to port or the the hard inner loops need to port for it to be a big unlock. So nobody knows. But if math and physics are about to be quote unquote solved by AGI, maybe we’ll know as soon as what in the next 69 months. >> Fascinating. >> Alex, any prediction on that? I this is like >> I I would predict I mean so so the elephant in this particular room I think is complexity theory. Uh so so a major challenge in in the quantum information processing community has been identifying algorithms for which a a provable quantum speed up or quantum advantage can be achieved versus classical computation including versus stochastic classical computation. So to the extent math and physics are deacto solved by frontier AI, I would hope and expect that we achieve superhuman

[00:52:01] performance in complexity theory as well. And so maybe AI complexity theory researchers will identify some some new quantum advantages, some new complexity classes for which it is just obvious by construction that there’s a quantum advantage for AI. >> I think I might understand what you just said, but I’m not 100% sure. It just sounds like shit’s going to hit the fan and go much faster over the next few years. And uh and you know, fasten that other seat belt you had. Uh moving on here. Uh so not to be outdone, uh Google’s Deep Mind team also wins gold at the math Olympics, scoring 35 out of 42. Congratulations, Deep Mind. Congratulations, Open AI. You know, I did a little digging uh on the International Math Olympics uh and I found two things I find fascinating just to have a conversation on. So, this is the US team. The US team scored second uh in the international uh in IMO 2025.

[00:53:03] And if you look at the US team, um it’s got six members. Four are from are basically of Chinese descent and one is Thai. So, five out of six members are Asian. Um, the US team scored between 33 to 39 out of 42. But check this out. Uh, the Chinese team, all six members scored a perfect 42 out of 42. And I’m just saying that there’s something in the water in China or the gene pool in China. And uh, we’re just, you know, we talk about US versus China. There is a huge intellectual capacity there that resonates towards math and computation. Comments, thoughts well I don’t think it’s anything to do with the gene pool. I mean the population is big which always helps but uh it’s the focus. I mean the

[00:54:00] the education system and the culture cares about this area of endeavor and there’s so many brilliant young students in America and you know the the schools are saying yeah you know go focus on your soccer game and you know be be you know and it’s just it’s just not a a focus of the education system and so a lot of people just don’t go down this path. So what happens is they get all the way into college, maybe all the way through college and then they land inside of our incubation environment in Cambridge or you know your new setup in LA and then they start you know they’re starting there like well in China they started at age eight and so they’re just ahead of >> or younger. Alex thoughts. >> I was going to so just on on the the previous uh item here regarding uh regarding Google DeepMind’s IMO submission just to to comment on that since since we flew by that. I I think it’s super interesting to look at DeepMind’s particular solutions as well as Open AIS. They’re all in natural language and this is in stark contrast

[00:55:00] to what folks may remember past uh past papers out of Deep Mind on solving math. were very focused on formal reasoning on first requiring that uh the IMO problems in in geometry and otherwise be first formalized in a formal language before they could be solved. Here if you look at the the reasoning process and and look at the output it’s all natural language. So when we start to think going back to your earlier comments Peter about wouldn’t it be wonderful if if we had a benchmark for solving all human disease or solving uh unlimited energy or fill-in- thelank other abundanceoriented topic historically prior to uh to these IMO windins one might have reasonably suspected that we would need formalizations of of all of these problems and and that is in fact if uh folks take a look at the formal conjectures repository on GitHub from Deep mind appears to be sort of a parallel effort to formalize open conjectures in math. But what’s so startling about the deep mind and open

[00:56:00] AI accomplishments here is that this was all done with natural language. Uh there there was no formalization step. So it does in my mind raise the question when we start to think about solving all disease something that various leaders in the space have talked about potentially being achieved in the next 5 years. Do we even need formalization at all? Or could this just be solved with natural language? The the way human reasoners think that that I think uh is the one of the the most important ontological shocks coming out of this year’s IMO. >> Fascinating. All right, let’s move to uh a important conversation next which is America’s AI plan. Uh this was unveiled by President Trump uh just very recently, a couple of days ago. Uh the plan has 90 plus federal policies. Uh the key moves includes uh exporting full stack AI tech. Uh fasttracking data centers, cutting AI regulation. Uh it

[00:57:00] has three pillars and we’ll talk about those pillars. I’ll just hit on pillar one and then we’ll pause to discuss this. This is really an acceleration of the acceleration. So the first part is rescending old regulations and reviewing state level rules that show AI that slow AI development. So get rid of the roadblocks. Number two, promote open-source AI models for startups and research. And then uh invest in worker training and retraining. Let me just hit these and then we’ll we’ll talk about it. Pillar two and three streamlines data centers, chip factories, energy uh projects. So basically, you know, allow for rapid permitting to build things as rapidly as possible. Um, and we’ll invest in nuclear and geothermal uh power and secure data centers for military use and exports US AI tech to our allies. Alex, this sounds like a war footing in the global AI war. Yeah, I I

[00:58:00] uh I I read the AI action plan and and my sense is this is potentially the the broadest US industrial strategy that we’ve seen since President Eisenhower, since the the interstate highway system. And I I’m also reminded uh this this anecdote that’s that’s in the historic literature that in 1939 Neils Boore, this is prior to World War II, Neils Boore and and to the Manhattan Project, uh told Edward Teller, the father of the the hydrogen bomb, that building an atomic bomb can never be done unless you turn the United States into one huge factory. >> And it appears to me this AI action plan is more or less doing that. It’s it is a plan to turn the US into one huge AI factory. >> Mhm. Dave, what do you think about it? >> Well, I I think we’re incredibly lucky the way the timing lines up with a single administration. I’m not political at all, but uh but we have continuity. Like, we’re not even one year into a four-year term now. So, this is the road

[00:59:00] map. We can at least rely on it for 3 and 1/2 years, which is exactly concurrent with the AGI explosion. So this will this at least we know this road map won’t get overturned by the next election and thrown in the trash you know just cuz the last regulations literally first sentence in this document throws everything that we just did in the garbage and starts over and that’s one of the great flaws in America right is this lack of continuity but here we’re going to have continuity for the time window that matters three and a half years so you know the I I think it’s just an incredible miracle that David Saxs got recruited into the government >> then that he took the Um cuz if you look at the authors of this, they’re actually really brilliant people who know what they’re talking about, which is pretty damn rare in Washington. >> Yeah. >> And so I want to get so lucky. >> Yeah. I want to get Michael Katzios, who’s one of the authors here, on our pod or to join me at the abundance summit. Uh you know, this represents uh first of all, there’s no congressional approval required. This is all being

[01:00:01] done by executive order. And so this is not a matter of if, it’s a matter of go. Uh getting rid I mean we’re going to talk about this in a minute, but we are so far behind the energy curve uh required to to power our AI revolution. Uh and we’ve heard this we’ve heard Eric Schmidt say this in our last podcast, Dave, where we’re not chip limited in the United States. We’re electricity limited. We’re power limited. And so this is like let’s double down on nuclear and geothermal. I I note that this action plan did not go heavy on solar which I’m still you know like scratching my head on because uh as we as we know China has gone allin on all solar and everything else. Um, >> but the ability to like just wipe away the state and federal regulations that slow things down on building, uh, I mean, if we’re going to compete, this is

[01:01:01] the time to pull out all the stops. >> Yeah. Well, I I would point out, too, that we’re not chip limited because we’re importing everything from Taiwan, but Taiwan is still manufacturing what 80 90% of the GPUs driving all of of AI. It’s all from TSMC. single point of failure, one company, and it is such a huge national priority to get new fabs, but also new fab company or get Intel rebuilt. Um, but we need some diversity in that area. And because we’re going to solve the energy problem, I it’s an acute problem, but we’ll solve it. But if the chip supply gets disrupted by, you know, a Chinese invasion of Taiwan or otherwise, that’s going to be the real vulnerability. And I think it’s pointed out in this document. It’s not really highlighted too much, but it’s up there in bullet one as a as a critical constraint. >> Alex, other thoughts, please? Yeah, maybe the the book end to the uh to the bore comment uh that uh years later it it’s reported that uh after the

[01:02:01] Manhattan project and and after the country was uh in many ways uh strip mind in in order to facilitate collection of enough refined U235. Uh Bore apparently told Edward Teller, I told you it couldn’t be done without turning the whole country into a factory. You have done just that. uh and I I I think that’s the race dynamic we find ourselves in. Uh and uh whether it’s one particular energy source or another, I I think energy sources that can’t be assembled in time for this this super intelligence explosion, uh even though they might be more ergonomic over longer time scales, if they can’t be provisioned, permitted, and deployed very quickly, they may be obsolete. Let’s watch a quick video here of uh President Trump on investments in this exact area. >> We’re back in Pittsburgh to announce the largest package of investments in the history of the Commonwealth of

[01:03:01] Pennsylvania. And it’s not even close. I don’t imagine it’s too close. I don’t think second is I don’t think second is too close. That’s a big statement. This afternoon, 20 leading technology and energy companies are announcing more than $92 billion of investments in Pennsylvania. And if you want, we could probably get them up. Let’s talk to them right now. A lot of capital flowing in. We’ve seen capital commitments out of Saudi, out of the Emirates, out of uh every major tech company. Um it’s uh you know, I was just making this comment the other day to a friend. You know there is almost an unlimited check being written across this converting dollars into chips and electrons. >> Mhm. >> And interestingly it’s not just flowing into the equity market. It’s also flowing into the bond market which perhaps not enough people pay attention to. And the bond market is absolutely enormous for for fixed income. So, you know, may maybe another angle here is

[01:04:02] what does the future securitization of of these hundreds of billions, if not trillions of dollars of investments in AI even look like in the future? >> Yeah, it’s actually you’re dead right. It’s so cool to to watch because you know most of the last 20 30 years has been software dominated and you know not building out infrastructure but AI is actually a combination of of software plus huge concrete slabs and massive you know hurricane proof tents I guess and uh you know and and it’s all liquid cool so plumbing and piping like you’ve never seen. I’ I’ve toured a couple of these data centers and it’s a million valves like a million freaking valves and then water sensor. It’s just it’s much more like uh kind of like you were saying 19 1939 or the buildup to World War II. And so when you mentioned the bond market, yeah, all that physical infrastructure is usually funded through a combination of equity and debt. And so you’ve got you’ve got bonds to issue and it’s global, too. Those bonds go out to the whole world. They’re not it’s not just a US thing. >> Pretty wild.

[01:05:00] >> I’m curious what’s not being funded. This money is being deployed here rather than someplace else. uh everything is not getting like if you’re if you’re trying to build the coolest company ever and it has nothing to do with AI. You literally can’t get fun. You can’t even get a meeting and it makes sense to me. I know I know it’s really frustrating for a lot of people but it makes sense because the priority of this for the world is so much higher than any idea no matter how good the idea was. I was you know one of my business partners uh Dave Massie is really big into real estate. He builds hotels and restaurants. And I was telling him like the uh the data center buildout is going to be a trillion dollars a year starting in 2029. It’s going to ramp from here to a trillion dollars a year. Do you realize how big that is compared to anything in hotels and restaurants? It’s going to suck up all the capital. And that that’s why the startup economy is so good right now is because these you know the US venture market’s only 200 billion a year. It’s tiny compared to the data center buildout. So all this massive amount of funding coming into AI is is sucking up

[01:06:00] the startups. You know, they’re getting acquired like Windsor for, you know, three billion or two and a half billion dollars in year two because they’re getting sucked into this vortex that’s funded by much bigger capital pools, the bond market, the public equities market. >> I think it’s also interesting. It’s also interesting to extrapolate. So today it’s energy and data centers and fabs. tomorrow. I I would reasonably expect this will include robotics and drones, humanoid robots. And there’s almost this I often think as in in part with my computer science training, what’s the innermost loop of civilization? If you’ve played the the video game civilization, there’s this notion of a technology tree. Uh certain technologies lead to other technologies. Well, there’s also, I think, an important notion of uh innermost feedback loops of civilization. uh certain technologies will beget other technologies that then reinforce in a positive feedback loop. What what is the innermost feedback loop of technology investment today? And I I think what we’re seeing here uh with uh

[01:07:00] hundreds of billions if not trillions soon of of capex and opex going into AI and energy is the beating inner beating heart the innermost loop of civilizational investment that once it it achieves some threshold is going to spin out and touch much of the rest of the economy that right now is it’s sort of being deprived of oxygen. And I love what Alex is saying about robotics, too, because my daughter just, you know, moved back to Cambridge to work at MADNA and she hasn’t been around much. And she came to a link studio and she said, “This has got to be the best office in the country.” It was like so heartwarming to hear that from your daughter. But we have so much energy in the building, it’s it’s just off the charts. But once the robotics hits, then not only will it be what it is today, but there’ll be robots, you know, experimental robots all over the place. And I remember, you know, back at the AI lab at MIT back when I was in undergrad, all the robots used to be there. You know, now everything went to the cloud, it’s all just a bunch of terminals. But when the robots come back, it’s so fun. The energy goes through the roof. >> You know, you know, Dave, this year, I

[01:08:00] mean, you’re going to be on stage with me at the Abundance 360 Summit in March. This year, our theme is digital super intelligence and the rise of humanoid robots. And I’m planning to have five of the top robot companies there. And I just want robots walking around all over the place so you can go and play with them. >> Yeah. And they don’t just walk around. Remember the old the pogo stick robot that there’s one that used to clean up the Coke cans and eat them, but the the pogo stick is just so it’s just bouncing around. You’re like it just so fun. >> Yeah. And and of course uh we’re heading up to the Bay Area in a couple of days uh to go and visit 1X uh Technologies making of the maker of the Neo Gamma robot. and we’ll be doing a podcast from there which will be very cool >> and we can have the age-old conversation uh that See, our dear brother who’s not here uh keeps on saying why do why do human robots only have two arms? Why can’t they have six arms? I’m going to ask that question of of our host. >> Okay. >> Yeah, >> that’ll be fun.

[01:09:00] >> And now it’s time for probably the most important segment, the health tech segment of Moonshots. It was about a decade ago where a dear friend of mine who was incredible health goes to the hospital with a pain in his side only to find out he’s got stage 4 cancer. A few years later, a fraternity brother of mine dies in his sleep. He was young. He dies in his sleep from a heart attack. And that’s when I realized people truly have no idea what’s going on inside their bodies unless they look. We’re all optimists about our health. But did you know that 70% of heart attacks happen without any preceding? No shortness of breath, no pain. Most cancers are detected way too late at stage three or stage four. And the sad fact is that we have all the technology we need to detect and prevent these diseases at scale. And that’s when I knew I had to do something. I figured everyone should have access to this tech to find and prevent disease before it’s too late. So I partnered with a group of incredible entrepreneurs and friends, Tony Robbins, Bob Hurri, Bill Cap to pull together all

[01:10:00] the key tech and the best physicians and scientists to start something called Fountain Life. Annually, I go to Fountain Life to get a digital upload, 200 gigabytes of data about my body, head to toe, collected in 4 hours to understand what’s going on. All that data is fed to our AIs, our medical team. Every year it’s a non-negotiable for me. I have nothing to ask of you other than please become the CEO of your own health. Understand how good your body is at hiding disease and have an understanding of what’s going on. You can go to fountainlife.com to talk to one of my team members there. That’s fountainlife.com. All right. Uh moving on. It’s time to talk about the browser wars. So here we go. Uh question on the Wall Street Journal uh posed. Is AI killing Google search? it might be doing the opposite. So the article states AI over uh overviews serve two billion plus users per month helping drive record 54.2 billion in Q2 search revenues. I

[01:11:01] mean first of all the idea $50 billion in a quarter is insane. I mean this is why Google’s such an incredible cash machine. Search impressions are up 49% uh facing increased competition from perplexity and open AI. We’ll talk about that in a second. >> So it looks like, you know, this has always been an existential threat for Google, but it looks like they’re they’re moving in the right direction. Dave, >> yeah, I don’t believe it at all. The bullet one I absolutely positively do not believe for an instant. The search volume’s going up, but the traffic is going to the AI thing they added at the top. You can see it in the little picture there. Yeah. and it’s going to cannibalize the hell out of the clickable links down below. The reason they haven’t had any impact yet on revenue is because they’re showing more and more ads all the time down below. They just keep ramping that up. >> But I don’t think Google’s in trouble. I’m not I’m not saying that. I’m saying search is moving entirely over to AI. Uh it’s going to crush that core revenue engine at Google. But at the same time,

[01:12:01] YouTube is growing like crazy. Google’s own AI is growing like crazy. uh and they have lots of opportunity to actually stay on top of the food chain. But if you look at the market caps today, remember Nvidia is worth twice as much as Google today. >> So the market is telling you amazing. >> Yeah. Isn’t that nuts? If you said that 5 years ago, people would say you’re insane. There’s no way. >> But people are viewing the future is entirely going to belong to AI and talking to AI and not to clickable search links. >> So how does Google make their money, you know, when people are buying uh buying link ads, right? What’s going on there? How are they going to make a money in the in the future? Do you any ideas? >> Yeah, nobody’s figured that out yet. And it’s really fun to watch the evolution because right now it’s kind of a free-for-all. So, if you ask the AI, hey, what’s the best insurance policy to get? Where do I get a cheap mortgage? It gives you a free answer and a very very good answer. Um, and so that’s what cannibalizes the clickable revenue. But, you know, they’ll find a way to charge for that or to to monetize it. I’m sure it’s it’s very much in flux right now. We’ve got a couple investments in companies that are figuring out how to

[01:13:00] how to turn that into a revenue machine, you know, and it’ll be huge. it’ll be on the order of hundreds of billions to even a trillion dollars of revenue in that channel. >> And then there was a comment that Sam Alman made that open AAI’s, you know, GPT5 will not have its answers tied to advertisers uh who are paying Open AI for that. So I am curious, you know, if we’re going to believe the output of these of these uh large language models, are they being influenced by who OpenAI or or Geminis’s or X’s customers are or not? Uh Alex, what are your thoughts on this? Is there a revenue? >> Yeah, >> I mean I I I there’s a popular narrative out there that uh affiliate links and and referral revenue is is the killer business model here, but I I think there’s a story behind the story. If if if you’ve used 03 or or some other modern uh agentic model, it’s doing far more searches for you than a human could do. Uh if you if you ask it a question,

[01:14:01] it will fire off 10, 20, 100 searches to its fill-in-the-blank arbitrary backend generic white labeled search engine. It’s doing far more searching than I would have done if I had asked the question to a leading search engine today. So I I would expect number of searches is going to skyrocket as as we start to delegate the problem of search to agents. It’s just that the agents don’t click on ads. They they fire off far more searches. They also don’t click on the ads. >> Yeah. >> Mhm. >> And by the way, it’s it’s a major pain point how expensive grounding is. So grounding uh for those who who don’t live and breathe this is it typically takes the form of uh avoiding hallucinations in the answers to to frontier models by having the frontier models conduct searches and then ground their answers in the facts or information that comes back from the searches. grounding agent results is insanely expensive and I haven’t seen major progress uh not the sort of orders

[01:15:02] of magnitude cost reduction that we see in terms of raw capabilities from frontier models for grounding and I think that’s actually one of the next major frontiers if if some non-incumbent wants to come in and radically improve the economics of grounding information based on search I think they will transform the market >> you know in my last book the future’s faster than you think I talked talked about how the advertising market was going to get transformed by AI in the following way. You know, there’s going to be a point in which I ask my version of Jarvis, you know, buy me some tooth toothpaste or buy me an outfit. And it’s not going to be going and looking at ads to see who’s got gleaming white teeth. It’s going to basically look potentially at my genetics or look at a whole bunch of independent data and make a purchase based upon what’s best for me, not influenced by ads whatsoever. And so that’s going to be an interesting transformation uh of how do you

[01:16:00] influence the AIS in a in a authentic fashion so they’re motivated to to order the product uh that you want them to order. >> Have you tried that Peter? I mean, I had the experience a couple days ago of of doing my first in 03 product purchase. I fed it a laundry list of requirements and said, “Go find me something.” It it identified top three options and uh presented a user interface, an inline sort of a third pane within uh OpenAI 03 for enabling the purchase and I made my first AI directed inline product purchase and it was a seamless experience. >> What did you buy? >> A hat. with a long list of requirements. >> Okay, put that on somewhere. >> Oh, that’s hilarious. I mean, I’ I’ve used I I’ve used the models to make a recommended list, but I haven’t used it for inline purchase yet. But again, uh how we buy stuff is going to be fundamentally transformed over the the

[01:17:01] next 12 to 24 months. Fundamental. >> Yeah. Not just buy stuff, but think about travel or, you know, all your life decisions. Where am I going next? What am I doing tonight? What are my friends doing? All that’s going to go through that same machinery and so that drives all the volume at the bars and restaurants and flights and hotels and you know >> what I really want is surprise and delight. Uh I’m going to Portugal. Set it up. Buy everything. Here’s my budget. Uh just set up all my meals, my restaurants, my experiences. Uh oh, by the way, my kids are coming. So set stuff up that you think they like. I mean, I think that would be absolutely it’s a it’s a level of uh of gourmet experience that uh from every every standpoint that you could not afford right now. It t you know >> Yes, I totally agree. I think I think life is so dominated right now by by marketing convincing you to that you need something that you don’t really need that’s way too expensive. You know, the the car, the hotel, what there’s so much more fun that you could have had and it’s life is kind of

[01:18:01] dominated by missed opportunities to have fun that you just kind of passed you by. And so I think your AI agent is going to do a much better job of helping you with that and you know just because everybody thinks they need this thing. Well, why do you think you need that thing? because it’s marketed in your face just like every time it’s just in your face and it just it creates anxiety, it creates stress, creates jealousy. At the end of the day, life can be so much better and so much easier and I think the the agents if we don’t mess it up, the agents have every opportunity to bring that to us. >> That’s awesome. All right, moving along. The competition is coming for uh for search OpenAI to release web browser challenging Google Chrome. So launching soon on Chromium with GPT style AI to handle agentic tasks targets 500 million weekly chat GPT users threatens Google ad driven Chrome empire. So the first thing that hits me here is is the web browser still a thing? Um and and why

[01:19:02] are people talking about web browsing experiences when I think that’s sort of like like you know the year 2000. Yeah, thoughts. >> I think at at at at one level we see this every tech cycle. Uh every every major tech company needs a sovereign distribution channel and to the extent that the browser itself remember Chromium was forked off of KHTML. Uh and Safari is a similarly forked off of KHTML. uh every major tech company uh feels a strategic incentive to own its own distribution channel to the extent the the browser or the operating system or the device or dot dot dot is is that distribution channel I think this is just part for the course what’s interesting to me though uniquely on this subject is if you play with open AI operator or Google mariner or more recently chat GPT agent you start to see the strengths and weaknesses of uh of so-called computer use agents or or kas.

[01:20:01] These are agents that are manipulating browsers for you. And I think we’re so painfully close to to having an agent that’s able to in real time carry out essentially all or most economically valuable human browser tasks. And that that’s the real headline. >> Yeah, Peter, your point is right on too. The word browser is going to go in the trash can. It’s really the word portal might come back. Remember back in the in the Yahoo days, Portal was there. >> Yeah. >> Yeah. Portal is much better for what’s going on here because if you we’ll see the new Perplexity Comet browser in a second here, but >> it’s a portal. It’s not a browser. It’s >> I I still go back to uh Iron Man and Jarvis. I think they had it right. I mean, to a large degree, it’s going to be uh voice interface until we get to BCI. It’s going to be the ability for you to uh you know have constant screens every place deployed either on your heads up on your AR or VR glasses or on screens in the home and an AI is looking at is is displaying wherever you’re

[01:21:01] looking at the time. Um but the idea of a computer web browser feels very last century to me. >> All right, let’s move on to uh to perplexity. And I’m not a Perplexity user, per se. Dave, how about you or Alex? >> Oh, yeah. You gota you got to try this stuff. I’m paying I’m paying for it. I’m paying for it. I’m just not finding myself uh you know, using it uh as much as others are. So, what’s >> Alex discovered this like always? He’s like, “Hey, you got to try this right now.” And so, I I go to try it and it’s like, “Yeah, 200 bucks a month. Another 200 bucks a month or you can wait two months.” Like, “God damn it.” So, uh, it’s incredible. Yeah. So, yeah, >> here’s the title. Perplexi launches Comet, an AI powered web browser. So, talk to me about this, Dave. Why? Why is this important? >> It’s important because you can start there. You know, like, you know, Apple and Google try to intercept you before

[01:22:00] you get to Perplexity. So, Perplexity fights back by saying, “Look, install this on your laptop and just throw away Safari and and uh Chrome and you can um just start here and then when you start there, it’s all AI all the time. Every it’s beautifully integrated. It’s really really well designed and you can see because they’re not worried about cannibalizing their search revenue like Google is, they can actually make a really clean from first principles design that’s AI first and and it’s it gives you search results when you want search results. It gives you AI when you want AI. It’s it shows you where things are going to go. Uh so it’s worth trying just for that reason alone. >> Alex. >> Yeah. And maybe add to that. Uh so so one of my favorite challenge problems to today in middle of 2025 when I encounter a new computer use agent and I I would classify comet as one of them is I ask it to win at a game of chess and uh some CUAS will get most of the way toward

[01:23:01] actually winning a single player game of AI chess against a a web uh competitor. Some will outright refuse to. Uh when I try to persuade Comet to do it uh it usually refuses. uh it can nudge it along uh and it’ll play part of the game. Uh chat GPT agent or operator will usually get most of the way through a game of chess. So where I think all of this goes is it’s more than just a browser. Totally agree, Peter, that uh browser is uh it it’s it’s almost um like a a straw man template for uh that we just have to pass through to get to solving the real problem, which is full vision, language, action or VA models that are able to solve general purpose challenges out in the physical world. that this is a a way point almost before we can have those humanoid robots that are doing our laundry and cooking our food and solving all the problems in the physical world. I I think it’s a necessary midpoint to have a an agent that’s able to accomplish economically

[01:24:01] useful tasks in the browser. >> Nice. Yeah, we should get Arvin on the on the podcast here. uh he he’s showing some serious Steve Jobs capabilities cuz you know everybody would have said well look he’s he’s really brilliant but he doesn’t have his own foundation model and open AI and Google are just going to crush him. Uh but now he’s got a lot of capital he’s got incredible insights and he really understands the consumer experience and so uh it’ll be fun to track his journey because you know there’s a version where he gets annihilated by the big guys. There’s a version where, you know, that he he emerges like Steve Jobs and rises to the top and it’ll be a fun drama to to track. But try Comet in the meantime. You’ll you’ll get a sense of wow, it really is much better. >> I’ll play with it. All right. Here’s the here’s the next note. AI writing 50% of Google’s code. And we see here a chart uh basically over the last two years going from 25% of the code being written by AI to now topping out at 50%. Uh this

[01:25:00] particular article points out that Amazon is writing 25% of its code, Microsoft 25 to 30%, Robin Hood 50%. Alex, uh uh is this just predictable? Are we going to get to 100% soon? What’s this mean to you? >> It it is tantalizing and riveting. What these numbers don’t tell us >> riveting? Okay. >> Riveting. I’m at the edge of my chair. What these numbers don’t tell us is what percentage of time is being saved by human developers. And that would be a more direct indicator of how close we are to recursive self-improvement. If if we’re nearing 100% time savings, then the AI is writing itself at this point. So, we don’t actually know based on these numbers, is it the 50% most boilerplateesque portion of the code that’s being written by AI or or is it the 50% most valuable? But e either way, I think, you know, even if it turns out this is just boilerplate that’s being generated, I think recursive self-improvement is is imminent. It would just be lovely to have a more direct indicator of that.

[01:26:00] >> So, so let’s talk about that one second. Recursive self-improvement is when AI is rewriting its own code. And uh there’s a lot of science fiction movies based upon that and that’s where hits the fan and goes sideways for us. Uh so should I be concerned about about that or is that just uh it’s an inevitability? Uh and there was I remember the the the three thou shalt not in the early days of AI is thou shalt not allow your AI onto the open web, thou shalt not allow recursive self-improvement and then there was a third one because there’s always three. But uh how do you think about this Dave? Uh well, you should not be worried if we have a handful of brilliant people in government who understand this. Otherwise, you should be worried sick. Uh because cuz there there are very straightforward guardrails that don’t slow down AI progress that keep us competitive with China and the world and that don’t slow down the military aspects of this which

[01:27:00] are critical. Uh the military part has to keep up and and be, you know, ahead of the world too. you don’t need to slow any of these things down while still guard railing the you because in the movie version of it which is pretty accurate just like Jarvis is pretty accurate uh if you let it create its own objectives and you let it design its own next parameter set around objectives that you didn’t give it then it becomes this kind of out of control conscious thing that we absolutely don’t need the world doesn’t need it society will never need it and it’s dangerous as all hell but you can prevent it with just some straightforward rules, uh, while still getting all the benefits. And I think if if there’s a handful of very very smart people who understand that, then there’s nothing to worry about. >> Alex, do you believe that? Do you agree? Or do you think it’s, you know, it’s up into the right and out of control? >> I I I tend to to be more on the accelerationist side as opposed to the what might call the safetist camp. Um,

[01:28:00] my worry tends toward uh worrying about overregulation. I I think we have an opportunity uh to to it to maybe by analogy with explosives to the extent we’re uh expecting an intelligence explosion. I I think of this almost as a shaped charge and we have an opportunity to shape the explosive charge here in in a positive direction. But I I think also of all the downsides of not achieving super intelligence, not solving all of the major outstanding problems in a timely fashion. And I I think on balance if if recursive self-improvement in in a thoughtful but not overly hamstrung way buys us solutions to the grand challenges of the universe I I would tend to prefer that future >> and it’s not a matter of it of it not happening is is it going to not happen in the US versus not happen in other parts of the world right it’s just not >> I think that that’s the element of the shaping so what what values do we attempt to imprint print on it. Um,

[01:29:02] national values, cultural values. I think all of this is at play. But I I also think if if you look back a decade or two, many of the people who were thinking about AI safety were maybe thinking too unambitiously at the level of having a single human align a superhuman AI. That was never going to work. What what’s actually happening arguably is all of humanity through government, through multi-corporation competition, through lots of individual leaders and researchers. It takes it takes an entire civilization to a lot an AI, not an individual. And that’s what we’re seeing. >> Love it. Love it. By the way, uh I want you to take a second if you haven’t yet and give me some comments on on Alex’s brilliance as a member of our uh our mindset mafia here and our our moonshot mates. Alex, thank you so much for for for commenting on this. I love your way you think. All right. Uh here’s another comment. Nvidia is making more billionaires than anybody else in the world. Let’s listen to uh to Jensen.

[01:30:01] >> We see this capital being applied to human capital in a way that we never thought was possible. It used to be NBA players signing $300 million contracts. Now it’s model researchers and then there was a post this weekend that said that there was a person that was offered a billion dollars over four years by Meta. Now, if that’s happening at this layer, why hasn’t it happened at your layer? Because you are the enabler of all of that. And how do you think all of this human capital is going to actually play out? First of all, I’ve created more billionaires on my management team than any CEO in the world. They’re doing just fine. >> Don’t feel sad for anybody at my layer. >> Yeah, everybody’s doing okay. >> Yeah, my layer is doing just fine. The big idea though is that you’re highlighting is that the impact of 150 or so AI researchers can probably create with enough funding behind them create an open AI. >> 150 people. >> Yeah. Deepseek’s 150 people. >> If you’re willing to pay >> say $20 billion, $30 billion to buy a startup with 150 AI researchers, why

[01:31:02] wouldn’t you pay one? >> Right. >> Incredible. >> Dave, how do you think about it? >> Uh, well, I I think the story within the story is that the equity upside is dominant versus the salary. the south these these you know NBA style signing bonuses are making the news but the founders of the companies and then the acquisitions have already created you know much bigger numbers through the equity upside uh there’s a lot of research that shows that in the future you know because AI does so much of the work that ownership of equity stakes ownership of physical assets generates 90% of the wealth in the world and not your day job and so I think Jensen’s just saying that in another another kind of way. Alex, >> it’s it’s funny. I have friends who tell me that they they literally live off of their Nvidia stock holdings. I I think we’re we’re lucky to live in a world where accelerated compute, uh, which is creating enormous amounts of wealth for humanity gets rewarded. We’re lucky to

[01:32:02] live in a world where markets reward that wealth creation and incentivize it. It’s wonderful. >> Amazing. All right. I love this. This is more on Google’s V3. Uh they’ve created a mechanism that allows you to draw uh as an artist would on a on a video frame and have the uh V3 model actually implemented. Let’s take a quick look here. So on this frame, the artist is is drawing, you know, Aurora Borealis. And of course, it’s instantly, you know, emulated on the video. Uh, this is VO3 is still one of the incredible shockers out there. And Alex, I appreciate all the VO3 videos you keep sending me via text. Thank you. I love I love your Star Trek and space theme. It’s It’s >> What What is real anymore? >> Yeah. So, how far are we, Alex, from uh

[01:33:00] from the first VO3 generated full feature film? it may exist somewhere already uh and not be evenly distributed. I I think one of the the biggest shockers for for me from this story is how close we seemingly are to a final convergence between textbased language models and uh and diffusionbased or diffusion transformer-based video models. And it it it makes me wonder what does what does the final model look like? it it seems like the entire space uh all of these different model architectures are starting to converge and what does the final converged architecture look like and part of me wants to think it’s going to look like a massively multimodal model that handles text and video and audio and DNA uh and raw machine data and many other modalities but at the same time if if you remember the the architecture wars uh in uh in computer software engineering you know should we have a micro kernel architecture should we have monolithic kernel. I think it’s

[01:34:00] going to look like a micro kernel that handles every single modality. And we start to see that here with with inframe visual prompting, visual text prompting of video output. We’re starting to see the the glimmers of of sort of the ultimate transfer learning between modalities and it’s incredibly exciting. >> I mean, it’s it’s the humanization of the process, right? So we’re connecting uh and giving guidance as we would to a person uh without having to go through specifically you know code to enable what we want to see. >> Why stop there? Why not add human thought via brain computer interfaces as yet another modality? I I don’t think it it ends with text and and video. I think this goes all the way to the end game. I actually hate among MIT alums. I I least like the idea of plugging something into my skull and communicating directly or I don’t like the upload either. I’m just very different on >> I can’t I can’t wait to plug in personally. >> Okay. Well, you plug I’ll see how it

[01:35:00] goes for you. This this video though, this is emergent behavior. This is not built by some software engineer. Hey, you can annotate now. This is emergent from the model itself. you’re going to see more and more of that where you can do things with these capabilities that the authors didn’t even know you could do. And so it’s really empowering for the user, the creator to say, “Hey, I discovered that you can use VO3 to do this thing no one even knew.” Uh that you’re going to see more and more of that. It’s uh it’s actually pretty cool. So we’ll track them as they come out. >> Awesome. >> One other comment on that just is you’ll hear the the term world model often. arguably a model that understands all the physics of the real world and can obey text instructions and transfer between them starts to to be a true world model and that’s also very exciting. It’s Star Trek holiday level. >> I love that. And yes, I can’t wait for my holiday. All right, so here’s uh next article up. Nearly 75% of teens have been using AI companions. So 73% of teens aed 14 to 17 have used an AI

[01:36:00] companion. 37% have shared personal secrets and teens using AI companions are twice as likely to feel depressed or lonely. So, uh you know, I have two 14-year-old boys and uh we’ve made a decision that they’re not going to have a cell phone until they’re 16. Uh they do have computers, but they have not been playing with AI companions yet. I mean this gets very scary uh in some ways of of uh breaking a normal socialization loop which is so important for uh for mental health. Dave, any thoughts on this? >> Yeah, this is so easy to fix. But the problem is that if you, you know, if you build a video game or you build a virtual environment, your incentive is to trap the person, try and get them to spend their entire day inside the game cuz that’s where you generate more revenue, more addiction, more fe, you know, more more cross cells. But, you know, it’d be much easier for the AI to

[01:37:00] say, hey, you know, it’s time to take a break. It’s time to go outside. It’s time to, you know, get some sunshine and some vitamin D. Easy to build that in. It’s just not in the incentive of the creator to do it. So now you get these super engaging AI companions that are literally like a soulmate and they’re super like they listen to every word you say. They hang on every so nice to you. Yeah. And so >> so that creates a really slippery slope. I didn’t know it had gotten to this level already. I was telling the kids this the other day. They’re like, “No way. No way.” Like I’m telling you it’s in the data and and it’s not like it’s a survey. Like people log in. You just count the login. So I’m pretty sure it’s right. So, it’s it’s crazy how quickly this has happened. >> Everybody, there’s not a week that goes by when I don’t get the strangest of compliments. Someone will stop me and say, “Peter, you’ve got such nice skin.” Honestly, I never thought, especially at age 64, I’d be hearing anyone say that I have great skin. And honestly, I can’t take any credit. I use an amazing product called One Skin OS01 twice a

[01:38:01] day, every day. The company was built by four brilliant PhD women who have identified a 10 amino acid peptide that effectively reverses the age of your skin. I love it and like I say, I use it everyday twice a day. There you have it. That’s my secret. You go to oneskin.co and write peter at checkout for a discount on the same product I use. Okay, now back to the episode. All right, next up, uh, Replet CEO apologizes after its AI agent wiped a company’s codebase. Oh my god. Ouch. It’s like, how does that come across, Dave? It’s like, oops. Sorry about that. >> I had the same thing actually. Literally just had the same thing. I put um I put cursor into full agent mode and said, you know, I’m just going to walk away for an hour or two. You think about that. I came back and it absolutely obliterated everything I was working on. I had it all backed up, though. But I was like, wow. Okay, I can see how this gets off the rails in a hurry. So yeah,

[01:39:00] you know, you you learn to trust it really really quickly and then you step over the line very quickly too, but it’s so capable, you know, it’s just it seems so trustworthy for a minute. >> Any other comments you want to make on this other than back up your data? >> Back up if it’s so cheap to back everything up every 10 minutes. Like just do it, man. And maybe just to to comment on this, I mean my my friend uh John SM likes to point out uh speculatively that the first generation of humanoid robots will accidentally fold the cat in the laundry. But that this is the moral equivalent arguably of folding the cat in the laundry. It this is where you know Nasim Talib’s notion of anti-fragility comes from. without going through a few of these sort of localized moral panics. Oh no, the AI agent wiped the codebase, we won’t get to uh to to a a stable robust system over the medium to long term. So, ironically, I I think these sorts of micro panics end up being net healthful for the ecosystem in the long term. Got

[01:40:01] it. All right, here’s a piece of abundance news, and I love this. Nigeria accelerates learning with AI. So, Nigerian GPT4 pilot delivers two weeks of learning in uh I’m sorry, two years of learning in just two weeks, 1,200% faster. So, you know, this really pisses me off. I mean, I’m excited about this for Nigeria and for other African nations, but you know, here in the United States, I’m not seeing the adoption of AI in learning anywhere near as fast as we should. Uh when are we going to see that, Dave? Uh look, if it doesn’t come from us, I don’t see it coming from anywhere else. Uh I think it’s just going to bypass the incumbent education system at this stage because we’ve had so many meetings and seen no motion whatsoever. So the the students want to learn, the students will work around whatever school and just learn on their own. One of the good things is our schools are giving virtually no workload to the students, so they have plenty of time to >> That’s so true. I just remember homework

[01:41:01] all the time, grade letters, and where did that go? >> I don’t know. All right. >> Well, it’s it’s going to be good in the sense that they’re free to pursue their own. >> We’re seeing in Nigeria now. We’ve seen it in Estonia. China is all in on AI. The US, you know, talks about it, but still not seeing anywhere near enough. And I I hope that’s going to become part of the conversation uh across all school systems. And you know, the teachers unions can’t block this. We’re going to have the best educators are going to be AIS and it’s going to be immersive education. >> All right, talking about China, let’s continue on our conversation about US versus China. And uh this week I want to talk about energy. Uh so this is an incredible chart that shows growth uh of of solar uh in China. And so here’s the article. China’s installed 464 gawatts

[01:42:00] of solar capacity in just the last 12 months since June. Uh that’s epic. And while we’re focused on natural gas and coal and nuclear, which is unfortunately kind of slow, China is just just covering their countryside in solar. And I don’t understand why we’re not doing the same. Any thoughts? >> I I think it’s interesting to think about. So So we see this uh this uh quadratic maybe exponential curve there. It’s interesting to ask where this goes. Uh arguably if we find ourselves in a call it a solar super intelligence future. This leads inevitably to a Dyson swarm. We we ring the sun from from all directions with with supercomputers. >> Had to bring the Dyson swarm in. Um I don’t think that’s likely. I I I think if if this were if if solar super intelligence were the the end state, we would probably have observed lots of Dyson swarms throughout the galaxy already. They would stick out like a sore thumb in infrared. We’re to my knowledge, we’re not seeing them. So

[01:43:01] that suggests to me this exponential growth of solar as as a critical path to super intelligence uh suggests that it doesn’t scale all the way to post super intelligence. Assuming this horizontal exponentiation happens, it seems more likely that there will be other energy sources. So I I think well >> so much left on the table, right? So I’m a pilot. I fly out of Santa Monica airport here. I’m flying up the coast and I’m looking out and I see all of these rooftops just, you know, that could be solar producing in the interim. Uh I don’t get why we’re not pushing that here. Uh Dave, what do you think about solar? Are you investing at all in this area? >> No, I’m not. But I think that’s the problem. Like I should be, but I’m not because physics is about to be solved and I don’t know if Fusion’s going to come online very very soon. This is why America gets its hands all tied up cuz the investors won’t pour the money in out of fear that some other innovation will disrupt it before the payback. Meanwhile, China doesn’t worry about that because it’s all government funded. And so there’s literally 200 gawatts

[01:44:00] worth of solar panels sitting in warehouses that we could actually buy and deploy in our sunny, you know, Utah kind of Colorado or Nevada areas. And nobody wants to take that risk, but it’d be a good move because just as a hedge, it’d be a good move, but investors won’t do it. >> I mean, we talk about uh, you know, SMR, small modular reactors and uh, and generation for fision reactors, but we’re talking about those coming online like a decade from now. Um, and fusion, you know, we’ll see here. Here’s the next article. Uh, Chinese fusion reactor sets record. Uh, keeping a superheated plasma for 1,066 seconds at 180 million° F. God >> it’s not. >> Fusion, remember fusion’s not I mean it it it sounds futuristic to to our uh 2025 years, but it’s actually not that efficient in terms of rest mass. So, so

[01:45:01] light uh light element fusion consumes less than 1% of the rest mass of of the reactants. If we solve physics in the next few years, we can do way better than fusion. We we could be building, you know, conceivably, you know, micro black holes and and dropping matter into them and harvesting the Hawking radiation and and the rest mass. There are many things that we could be doing if if we’re about to solve physics. So, >> I can see that black hole reactor in my backyard, please. What could possibly go wrong? >> Yeah, >> Mr. black hole instead of Mr. Fusion. It could happen. >> Uh so let’s talk about that solving solving physics. Um Alex, what’s your vision of solving physics mean? >> So cool. >> I I think it comes down to uh to discovering new physics with AI. And I I think we’re uh maybe only a few years away from solving physics the way uh earlier I mentioned this notion of solving mathematics in in the sense of achieving uh professional mathematician level AI. I think we’re only a few years

[01:46:01] potentially away from achieving professional physicist both theoretical and experimental AI that can unlock new physics. If there is new physics to be found and we have strong observational evidence that that there is across many different uh subd disciplines of physics. I I think our best shot for a field arguably uh especially at the fundamental level fundamental physics where we haven’t seen major fundamental new physics in the past half century maybe AI is our best shot at unlocking new fundamental physics. >> I think we need you to make a movie so that I can get what that mean how that’s going to play play out. You know how Jarvis kind of it seemed like total science fiction now it’s absolute reality but it opened our mind to how this was going to work. the equivalent for solving physics. Like I just can’t because because the AI will know things but have I’ll have no ability to understand what it’s trying to say to me. Like you know like I can’t I can’t even comprehend string theory as it is. You know >> when you start when you start getting

[01:47:00] gravity shielding and you’re floating up to low Earth orbit and you’ve got all the energy you need, you’ll understand the implications of it. >> Totally. And I I think in in the popular discourse of super intelligence, everyone is so focused on just racing to the destination. They’re going to be the proverbial dog that catches the car and and wonders, so what comes after we have super intelligence? I spend a lot of my time thinking about the day after super intelligence. And I I think the day after looks like solving math, physics, chemistry, biology, medicine, a bunch of other fields and then unlocking solutions to the grand challenges that we face in on mass. You see the biology in medicine I totally get because its only purpose is to give us longevity, happy, health, healthiness, you know, and and you know that that is really clear. People are just healthy and happy. The physics side of it just like okay, it’s discovered things beyond quantum. It’s discovered things be like and it’s trying to explain them to us and it’s building things, you know, but I can’t predict what it’s going to build

[01:48:00] next cuz I don’t get it. That’s the part I’ve just >> I don’t think it’s going to be understanding the fundamentals of the physics it creates. I think it’s going to be experiencing the breakthroughs that it creates in the physical universe for us >> and that’s going to be absolutely fascinating. All right, some quick stories on robo taxis. As we near our end here, I think it’s important to note Uber’s invested 300 million in Lucid’s EV. Uh so uh this is a deal for Uber to basically purchase 20,000 of Lucid’s Gravity EVs over six over six years. Uh aiming to challenge Tesla and Whimo. So interesting, right? So Whimo is beginning to roll out. Um and there’s, you know, every 3 minutes there’s probably two or three Whimos that pass me by here in Santa Monica. uh we’re seeing a slower roll out to the robo taxis from uh from Tesla and there is space for a third um and so it looks like uh Uber’s coming in with Lucids EV.

[01:49:02] Any thoughts on this Dave? >> Yeah, it’s interesting. Uh there’s a constant tension between user base. Uber has the user base but doesn’t have the foundation model. needs to partner for the tech. But then, you know, on the other end of the spectrum, Elon has the tech and he’s he’s building out, you know, the user base. But you see Sam Alman really being the the okay, I’ve got a foundation model company. Now, I need to control all these user touch points. So, I need Johnny IV. I’m going to build a consumer device. I’m going to get into the browser wars. So, it’s starting to look like the foundation model userbased vertical integrated monopoly is going to become a real thing. >> Yeah. So Uber’s got to get in the game though. I mean, they don’t have any of the underlying tech, but they’ve got the user base. >> And there’s been a lot of conversation over the past. Does uh does Google buy Uber? Uh and you know, who’s at play here? But there’s going to be some kind of consolidation or some kind of uh extension. Uh Alex, any thoughts here? >> These are mobile data centers on wheels.

[01:50:01] I I think that the metaphor that autonomous vehicles or AVs are smartphones on wheels is misplaced. These are data centers, micro data centers on wheels. So I I think this is we’re seeing the the deck chairs all all move around in a game of of musical chairs. And I I think we’re starting to see the emergence of a new class of mobile distributed data centers. And what’s missing in my mind is algorithms, training algorithms and inference time algorithms that can take advantage of all of this compute that right now is being used for autonomous driving, but could in principle be generalized to to having mobile distributed data centers. And this is Uber and maybe other companies finding themselves slowly into the mobile data center space. >> Well, you know, Darra Kazar Shahi, the CEO of Uber, is an engineer uh fundamentally always has been an engineer. So, we should get him on the pod and and he said he’s very happy to come on the pod. We should have him join us and and talk about his vision here. You know, it’s interesting. Uber was so

[01:51:00] early in this game. Um you know they I remember they were early in uh with Travis really supporting the buildout of autonomous vehicles also flying cars right they did a lot of the uh earliest work in EV talls. Uh and then they’ve fallen back when Travis left, they’ve fallen back to fundamentals, but time to start growing again, otherwise they’ll get displaced. >> Yeah, they were under pressure uh from their shareholders to show profitability uh for a while there when DAR came on board. So, they had to cut a lot of the really interesting R&D, but I think now it’s obvious they should get back in the game. So, be a great time to get Dar and pick his brain. >> Sure. >> But he’s he’d be awesome. He’s brilliant. >> Um I I love this. I gave it a title. Take fries with a Model Y. Uh this is the roll out of Tesla’s diner here in Hollywood. Uh we’re going to close out on this story. Of course, there is an image of Optimus serving you popcorn.

[01:52:00] So, this is a 24-hour day, 7-day week diner in LA, combining American food, drive-in movie experience. Orders are autotriggered on your Tesla when your Tesla nears the diner. It’s a massive EV hub with 80 version 4 superchargers. Uh I haven’t gone yet, but Dave, when you come here and visit, uh we should make a a side junket for sure. >> All right. You know, my here’s my question. Why is Elon the only one who’s building out our vision of the science fiction world? >> You know, it’s so so frustrating. Yeah. He’s he’s the guy building in public is the new rage. And being a public figure while you’re creating this stuff, it works. It attracts talent. it attracts capital and it works so well in this hyperacute really fast environment and so few people are doing it but around the incubator a lot of the teams have embraced it and they’re doing it uh but yeah the I guess it’s just cuz the older CEOs you know the incumbent CEOs never thought that way and and they just aren’t getting on the bandwagon some of

[01:53:01] them aren’t very good at it either I >> I’m just curious you know the taste makers that make this stuff happen for Tesla is extraordinary I mean, what I’ve seen about the details, the diner, I mean, I love it. This is like, you know, we’re finally getting there. We’re finally getting to this to this fun future world. Alex, you want to come visit and go uh go for a burger? >> Would love to. Maybe a veggie burger. I I I love the the the retrofuturistic aesthetic. I I grew up on 1950s era golden age of science fiction reading. And this this reminds me of the the era when cars had tail fins and people were excited about the future. Future transportation, future energy, energy was going to be too cheap to meter. Uh this evokes all of that and and hopefully we find ourselves in the near future when energy and intelligence are too cheap to meter. >> I love it. uh you know when we have the abundance summit in March uh and those of you want to learn more you can go to abundance360.com

[01:54:00] to learn more it’s going to be our uh our 13th or 14th year I’ve committed to running the abundance summit uh for 25 years and uh for me it’s showing people what happened in the past year and where things are going in the years ahead. Uh and one evening Dave during the Bundance Summit we’ll have a WTF episode sort of recapping what happened uh in the year and what happened in the last couple of days. But uh Alex uh you were going to join me last time. You came on via Zoom. Uh maybe you have to come. >> I’ll be there in person this time. >> All right. I love that. I love that. So before we break off, Dave, what’s new at Link Ventures? >> Oh god, it’s such a golden golden time. Um, I really hope people don’t take it for granted because, you know, I was around, you know, in my 20s when the internet exploded and everybody was succeeding and, you know, people were running a million miles an hour and then I I think a lot of people underappreciate how incredibly slow the period from kind of 2004 to 2020 was

[01:55:05] relative to what’s happening today. But around Link Studios, you know, people are literally running from meeting to meeting, running to the bathroom and back, just trying to keep up with the pace of change, reading Alex’s feed every week, and like it’s just it’s like a full-time job, just trying to keep up with everything going on. >> Um, but we had we had the Cambridge cops come in the other day, and one of the one of the guys from one of the CEOs came up to my door and said, “Dave, why are the Cambridge police in the building? What’s going on?” So, I went out, talked to him, and they said, “Someone’s trying to break into your building. They’re they’re climbing the bricks on the outside wall. somebody in a plaid shirt. I don’t think criminals wear plaid shirts very often. I’m not not sure I buy into this. Turn out Turned out one of our Harvard teams, all the math majors are rock climbers and they like to do crazy things. So one of the math majors cracked open it’s a second story window and he’s just scaling a brick wall like okay. And the the the team was like we can’t have that. I was like this is going to make a great movie someday. I mean these these guys are going to be the ones that are

[01:56:02] the next Mark Zuckerberg social network. You know, actually the Sam Alman movie will be coming out soon, too. So, I feel like I’m just living in this this community where it’s all happening. So, it’s it’s just a golden time. And I like I said, we’ve had 100% success rate on our investments in these teams. So, I’m surprised there isn’t more capital just realizing that this is this is the moment in time every one of these teams is likely to succeed because the tailwind is so strong. So, just get invested. And I’m not sure you’ll be able to get invested 3, four, 5 years from now. It’s it’s kind of now or never. >> One of the things I find fascinating is most people when they ask me where can I invest in AI uh you know it’s typically all the public companies they don’t have access to the deals at the beginning when they’re reasonably valued when they’re below you know $9 billion valuations. Yeah. It’s crazy. >> Oh yeah. No. Our our entry valuations are same place they’ve always been and then the step ups are like nothing I’ve

[01:57:00] ever seen. which is just in the first few months. >> What are the entry valuations? >> Yeah, you know, it’s it’s inflated just a hair, but you know, 10, 15, 20 million kind of founding day valuation team of three to five. You know, they just got the idea, but they really want to move into the lab and and and circulate with all the other teams and and so, you know, we’ll put a couple million bucks in uh to to liberate them from having to go work a day job and and get them focused. But you know normally from there to significant revenue would have been 2 years 3 years. Now it’s like 2 months 4 months like I cannot even describe the difference versus just four or five years ago. >> And how many folks in the incubator space? >> Uh it’s packed now but uh we got another floor so we can start filling that out. But we have 26 companies there. Biggest one is 24 25 people uh and smallest one is three. So I’d have to count but it’s it’s packed. Alex, prediction for the next couple of months. What are you seeing coming soon?

[01:58:01] >> I want to see several new state-of-the-art frontier models. I want to see um ideally additional companies or competitors come forward with IMO Gold. That would be amazing to have a competitive ecosystem like that. I’d like to see ideally at least one grand challenge level problem in math or some other physical science get solved by AI. I’d be a very happy camper if if some or all of those predictions happen. >> Yeah. >> And Alex, I and I are are working on that new team that’s going to be the first to figure out whether neural networks can be trained on quantum computers. We have a very profitable pathway from point A to point B. We’re mapping out. So, Alex is going to find me three or four of the smartest people on the planet to to work on that project. >> Love it. Love it. Well, guys, thank you for another great WTF episode. Alex, a real pleasure to have you. I hope we’ll have you back on a regular basis. Our our love and appreciation to our missing moonshot mate, Selma, wherever you are with your son. I hope you’re having a fantastic time. And everybody remember

[01:59:03] this is the real news impacting our world. Um it’s not about politics. It’s about technology that’s shifting our industries, our companies, how we teach our kids, how we run our nations. It’s an extraordinary time to be alive. Remember, uh don’t blink cuz it’s moving that fast. Every week, my team and I study the top 10 technology meta trends that will transform industries over the decade ahead. I cover trends ranging from humanoid robotics, AGI, and quantum computing to transport, energy, longevity, and more. There’s no fluff, only the most important stuff that matters that impacts our lives, our companies, and our careers. If you want me to share these metatrens with you, I write a newsletter twice a week, sending it out as a short two-minute read via email. And if you want to discover the most important meta trends 10 years before anyone else, this report’s for you. Readers include founders and CEOs from the world’s most disruptive companies and entrepreneurs building the world’s most disruptive tech. It’s not for you if you don’t want to be informed

[02:00:00] about what’s coming, why it matters, and how you can benefit from it. To subscribe for free, go to dmmandis.com/tatrends to gain access to the trends 10 years before anyone else. All right, now back to this episode. [Music]