Musk's Early Investor: The 3 Bets That Will Define AI by 2028 | Steve Jurvetson

0:00What will the next three years look like? I have this gut feeling that it'll be something architecturally variant. This is Steve Jurvetson, an early investor in SpaceX when almost nobody believed in private space. He backed Tesla before electric cars went mainstream. For over 30 years, he's been betting on the future, and history keeps proving him right. When someone comes to you with just an idea, what would be the best thing that they can do? Single person with an idea? I might try to find a co-founder. It's rarely an individual. Jobs and Wozniak, Batman and Robin. The working lot with Elon.

0:34Top three principles that everyone should learn from him. I do try to observe leaders in action. Even with a focused effort, it's not always obvious, but a few things. One is this insane ability to... Thank you so much. This is going to be very exciting. Steve, I am so excited to have you on this stage. What a fun time, right? Right. SpaceX, IPO, and you were there super early. What did you see that most investors didn't see back then? So the simple answer to the question is there were almost no investors considering space. It wasn't a category on any site. So the slightly varying of the question is why in the world would we invest in a sector that is just not a sector for venture?

1:19Same could be said for automotive, a Tesla, energy with nuclear fusion. There were a handful of investments, but very few. So the short version is obviously an incredible entrepreneur, someone we've worked with before. I've known him for, oh gosh, 29 years now and invested in all of his companies of the century and his cousins too. So all in, if you will, the uniqueness of the opportunity. So what we've come to appreciate in a sort of fuzzy way then, but now in a more crystallized manner is the way in which a sort of software-centric system engineering approach to a sleepy industry that hasn't had any change for decades can actually unlock incredible value and opportunity.

2:00You can see it in aerospace, you can see it in automotive now. It was sort of a long bet when we first invested, but now we can sort of see in retrospect how that's going to play out in almost every industry over time, how they become information businesses. Obviously, you're so good at predicting future. And part of this podcast, I really want to understand how you think about the future. You have this amazing graph, 130 years off compute, and it basically grows exponentially. What does it mean for all of us? What will the next three years look like because of what's happening to compute? I'm just curious, how many people have seen this version or this abstraction of Moore's Laws originally by Ray Kurzweil in like 99 book, Age of Spiritual Machines?

2:39It looks to me like 25 % of the room. Okay. I always ask because I'm curious how much it has entered the zeitgeist because I think it's the most important thing ever graphed and I give credit to Kurzweil for even seeing this pattern back when no one knew they were fitting to a curve. So just for those who don't know, this covers like five different technology substrates from mechanical devices to relay-based computers to discrete transistors and integrated circuits. And only in the most recent era would what Gordon Moore called Moore's law be almost a refraction of a much longer term trend transcending all kinds of dramas of companies that came and went.

3:13It's almost cosmological. Like why has humanity's capacity to compute compounded for 130 years? And for a sense of scale, like that's an exponential scale, right? Logarithmic scale. So straight line is exponential. This graph shows a 10 ,000 billion billion X improvement in computation that a dollar can buy. This is what customers care about. No one buys transistors when they're buying ICs. They don't say, how many, how many transistors does that one have? I'll buy the one that has more. No, they buy compute capacity or memory and both have been on rails. And so to your question, the first and foremost thing would be to just predict that it's going to keep going for three more years.

3:48Like why would it suddenly just stop and hit a red brick wall the way Intel's been saying it would? And when companies say that, like Intel usually assign they're losing their business to someone new like NVIDIA 15 years ago. So in the next three years, I think you'll see the analog chips continue to carry the mantle of Moore's Law. Some of the more esoteric and customized AI silicon that does discrete matrix multiply and ad release efficiently. And this is what's going to carry the juggernaut that we all just take for granted that it keeps going. In fact, I'd say without this sort of exponential change in technology, you wouldn't have startups, you wouldn't have this disruptive innovation opportunity like we talked about in SpaceX or in a bunch of companies, because if business is predictable, if there isn't disruptive technological change, the big get bigger.

4:30They have all kinds of ways to prevent new entrants from competing with them. It's usually somebody reinvents an industry and usually it's based on something computationally based. I think AI and everything we're talking about at today's conference is the epitome of this. It's like the most intense crucible of compute-centric innovation, economic growth, and sort of innovation of the economy. The translation of formerly industrial crappy gross margin businesses into information age, information-centric businesses that over the next three years means it ripples to, I think, energy, agriculture, construction, three industries that are enormous, growing as a percent of GDP, and the least digitized industries on the planet, not to mention healthcare, soon behind that.

5:14Are those the industries where you think we're going to see the most change? And what will cause the change? Is it going to be more advanced LLMs? Or do you think there's something else? I don't know, people are building world models. People are deep into robotics. What will be the technological driver for the most changes in the next three years? That's a great question because it's very difficult to answer with any certainty. I have this gut feeling that it'll be something architecturally variant. It might subsume the models that we know now. You could almost think of like a mixture of experts that's subsuming other architectures or the diffusion model we heard about earlier today that ultimately translates to a transformer, but it's a different way of thinking about the transformer, a massively parallel form of a diffusion model.

6:00And in the back of my mind, we have not, so what I'm about to share, we've not invested in this. So I've met with some companies, and I've been intrigued in something. My gut says they're probably going to make a breakthrough. And this is the whole new generation of MEO Labs focused on reinforcement learning because we're almost going back to the founding premise of DeepMind, which then they kind of just put to the side for a while when the whole LLM thing took off. And so if you could imagine what would be the, you could phrase this in an agentic language and say what is the you know multi-decade long agentic process you know minutes or hours not driven by some outsider you know pulling puppet strings but something that says almost like the drive evolutionarily for creatures or for humanity for whatever we consider the mission statement of our lives or humanity in general what will be that thing is it um you know to understand the universe the way grok and xai says it, does that become a driver for an artist like that?

6:58Is it something like a novelty-seeking algorithm that says, I'm going to continue to learn about the world and use novelty as my filter for, oh, I just discovered something new. How do I know if I'm making progress? What is, if you will, the selection pressure in an evolutionary algorithm? What is success? It's not just reproductive fitness in the biological sense. It's something grander. And I know that some of these groups are working on what is, is there a single reinforcement learning algorithm with continuous learning, let loose in the wild with all the data sets of the internet that could bootstrap intelligence in that sense.

7:32And the way that we think we're seeing in the large language malts today, but it's largely, we ascribe, I think, consciousness to other beings. We ascribe meaning to other things and we see patterns where they're not. And so I think a lot of it is a bit of, it's a fun interaction, but it's not quite the same thing, right? We just, we know there's nothing there inside. There's no light on inside, if you will. So what you're describing, I think, is it super intelligence when it's learning by itself, setting goals to itself? Are we going to see some version of that in the next three years? I know Jack Clark, co-founder of Anthropic, gives it a 30 % chance it happens next year.

8:08Super intelligence. Yeah, absolutely. So I thought, well, that's kind of fun. There's at least one person putting a stake in the ground. I don't know. But at Anthropic, they have a lot of strong opinions. What they do, but they also think that they're on the path. And there's a big debate as to whether this recursive self-improvement thing that they wrote about today and that Jack's been talking about for a few weeks now. I spoke with him about it last month or two months ago. Is there going to be some leap that we don't currently see for how these systems take on purpose and meaning in what I was referring to just a moment ago?

8:38Because right now, everything that they do is directed by a human. There's like, yes, the self-improving AI loop that they're witnessing already, these huge improvements, are coming from a number of steps that are still directed by humans. There's, you know, automated verification, improvement loops in the process of training itself, you know, adjusting hyperparameters from one training run to the next. A bunch of ways you could imagine high-throughput experimentation being mediated by the AIs. But what is the goal? The goal setting is still by the human. And so it may only be a thin veneer of activity that it's not yet doing, but it's in some ways the most important, right?

9:15Yeah. And they'll admit they're not sure how does that just happen, right? What makes that transition? And I don't know if it'll need to recapitulate some of the functional specialization in our own brain. like we evolved to where we are today with a history of reactive limbic systems and what have you, emotional centers that then cortex and more and more cortex layered on top of it, that whole construct may, as we heard in an earlier speech, be the bootstrap to consciousness as a perception of what we perceive. Do we need to have the same things in our robotic slash AI systems, right? There may be.

9:53So it's a philosophical argument. The main answer to question would be, I do not know. And I don't really even have the odds on it. I give it the fuzzy future kind of, yeah, that might happen, but only because that's more convenient as an intellectual shortcut to actually thinking about it as a serious hard problem, is to put off that three years feels far enough in the future that it's hard to predict almost anything. So we're seeing all the demos of robots and current technology, I think, is stronger than the deployment itself. We're still adopting, we're still adjusting. What's this gap? How big is it?

10:30From what technology is actually capable of versus how we're using it? Oh, right. Yes, that's a very good point. And there'll be inherently very differential domains of acceptance. So here's a great example, very simple to understand, is if it involves the world of atoms, it takes time. So even though it is obvious today that fully autonomous vehicles are the inevitable future, that every car will be autonomous, every train, every airplane, everything that moves on earth will be fully autonomous in the future. How could it not? It's insane to think now or to argue that it's not, even though we've been saying this for decades.

11:08The pace of switchover is going to be, it's going to feel glacial in certain parts of the world, right? People keep cars for an average of like 11 to 12 years. So you just have the physical swap out cycle for the car cycles. You have, you know, the change in mobility doesn't happen overnight. Okay, that's an obvious one. Physical robotics might be the same. How long does it take to make a billion robots? That takes some time, even with recursive manufacturing techniques. And so the place where I think it just sweeps like wildfire can be in areas, strangely, that we sometimes held as uniquely human are the creative arts, you know, the movie making, the images, what have you, which we've already seen.

11:46It's in some way shocking that that came first. And then the white collar jobs as mentioned, because the white collar job capability, take call centers, right? It's like 1 % of US GDP. That just happens like that, right? I mean, you just do not need to wait for decades for that to switch over almost entirely. And interestingly, people will increasingly prefer these two human interactions when they're better, show more emotional understanding, more reading of the situation. And that's seen in everything from physician bedside manner to chatbots or customer support agents is that the AIs do a better job with emotional connection than humans.

12:29Yeah. It's crazy how in some industries it's happening super fast, especially when it comes to software engineering. Some of my friends were editing 70 % of AI written code a year ago. Now it's down to 30%. I wonder what it's going to be in a year. So you worked with some of the most amazing entrepreneurs. You're working a lot with Elon. I know a lot of people in the audience are builders. Is there anything like maybe top three principles that everyone should learn from him? It's funny. People have been sending me these books that just, I guess they directed an AI to write about AI, you know, about Elon, how he thinks the secrets of Elon.

13:03I've actually been accumulating them on my bedside, but I'm not sure if the humans written any of them. A lot of people ask Elon's mom, you know, Hey, May, how did you, parent Elon? How did you get him to be the way he is? And that's a tough question. She hasn't been able to answer either. And so I'll take it with a bit of humility that even as a close observer, by the way, I do try to observe leaders in actions. I worked with Steve Jobs briefly, and it's like I put all kinds of energy to understand how that guy works. But even with a focused effort, it's not always obvious. People are complex.

13:31But a few things. One is this insane ability to focus, which may seem ironic given how many companies he's simultaneously running, setting new records for that in a way that, you know, when Steve Jobs was CEO of two companies, that seems strange. Now it's all the rage. But one thing that allows you to do is use the fact that you've got obvious competing needs for your attention as a way to focus, prioritize, and not go to meetings. For the normal CEO of one company didn't go to their holiday party, you know, it might be seen as weird and like, whoa, but no one questions the feeling, he's got other things to do, he's got other companies.

14:02So whether it's an excuse or just works out this way. He says no to things so effectively that are distractions, that are not mission critical right now. I mean, for example, years ago, I was trying to hook him up with Craig Venner to brainstorm ways we could, you know, terraform Mars more easily and do a sample return of life from Mars with gene sequencers and reinstantiating. Anyway, microbes on Earth. It was a fascinating topic to me. I was like, whoa, this is so fascinating. But he's like, no, it doesn't matter until we get Starship flying. None of this stuff on Mars matters. I we think about what we do when we get there.

14:36There's, I think maybe more importantly than what I just said, even more importantly is this maniacal focus on the, what I would generalize as the cycle time of innovation, which is how rapidly can we run experiments or iterate in our learning loop? What is the core learning loop? Whether it's the launch cadence, whether it's the data gathered from all the Teslas before fully self-driving vehicles came that could be used to train the models. How can we make sure that we have a leg up on anyone else on the rate at which we're learning from customer interaction, product features, and technology in general?

15:13And as an example of how powerful that is when you do it right and the data fly away you can make for AI, just one example, Tesla cars today and their cameras gather for their AI training set more data every four days than Waymo has in its entire history. And the brilliance was enabling every vehicle, whether or not the customer paid for full-size driving, to be a data collecting vehicle. So focus, learning loops, and this whole series of well-honed skills on identifying talent that I wish I could replicate. I just can't. Sometimes there's a pattern recognition, and he'll share bits and pieces of this, not leaning on credentials or specific background or experience.

15:58In fact, it's often an albatross, but like having people really walk through major engineering crises or problem solving things and then drilling down further and further and further to show if they did they really master the do they have mastery of the understanding of what it took to make something successful. So broadly defined, being a magnet for talent, finding a way to pitch and refine a vision that people want to join you. So like one of his brilliant things at Tesla, SpaceX everywhere is not just saying, oh, we're making rockets, we're making cars, but to really think of something much grander, right?

16:31Catalyzing the transition to sustainable energy or making humanity multi-planetary, understanding the universe now that XAI is merged into it. And these are the sort of lofty goals that motivate some of the best and the brightest to want to work with you. And that is a sort of compounding benefit that ripples out through a whole organization, right? Because great people want to work with other great people. I'm talking to a lot of entrepreneurs. And especially these days with things moving so fast, there's this new shiny thing every single week. How do you stay true to your mission when the rest of the world, 99 % of the world, tells you it's too early.

17:10Like, talking about space, we have so many problems here on Earth. That's an interesting question. And I realize I have a bit of a sample selection bias in that I've tried as best I can. I've done VC now for 30 years to only work with the people who have a true, sincere, you know, messianic mission in mind that is driving them. And they're not the arbitrage-shaking opportunists to see the next bright shining object or, oh gosh, You know, where should I go to next? And one of the ways, and I'll get to your question, but one of the ways, by the way, that I filter for that in meetings is, let's say I'm going to get really excited about a company, I'll often ask, you know, okay, what does your business look like in 50 years?

17:48And I get usually two reactions most off. One would be a chuckle. I'm like, well, that's a ridiculous question. You know, like the arbitrage-seeking opportunist is going to be like, I'll be my third startup by then. Like, how would I possibly know what my startup is in 50 years? Like, they'll just laugh at the question. and we pass on those. And then the best is when the person's like so relieved, like, oh, thank God. Now I can actually tell you what I've been wanting to say all day long, which is this is what's driving me. It's this thing that's so many steps ahead of what you'd probably want to invest in today.

18:18Like making, you know, colonizing Mars is an uninvestable proposition. Go back in the founding days. Like when you start a business day one, I'm going to colonize Mars. Like, you know, next, right? For most investors, right? Like that's not a door opener. And so most entrepreneurs that have that true, sincere vision have found a way to like subjugate and put off what their true dreams are and talk about something much more prosaic and near term. So I think the answer would be, as an entrepreneur, it just happens naturally and try to find investors and partners and certainly employees who are with you for that long ride and have a path to get there that is plausible.

18:54So, you know, this is sort of the joint tension, I think, in the best startups that's hard to simultaneously satisfy, which is an inundacious, you know, 50 to 500 year vision. This is what this company is going to do to the economy or the universe. Coupled with, oh, and by the way, over the next three years, we're going to iterate with real customers, learn from that and can paint the path from where we are now to that future that is chaining. sometimes they chain back from the past to the present, like to get there, what do I have to build now and then move forward along that path? But it's not like go into a research lab, pop out in 20 years and solve all the world's problems.

19:32Yeah. This is a really fascinating feature that I see with a lot of greatest entrepreneurs. It's like if they're reverse engineering from 50 years ahead. Is there anything surprising that still surprises you about those amazing entrepreneurs? Well, I suppose it's a bit surprising in a way each and every time it goes incredibly right. So weirdly, this may sound weird, I don't think I've ever thought about that question before or been asked it before. And so the perpetual surprise for me is like, wow. Like in the year eight, nine, 12, some new opportunity that opens up and unfolds from, in a sense, the expanding option value of going into some new frontier of the unknown.

20:18So what I mean by this is we try to invest in, by the way, at our firm Future Ventures, in things that are unlike anything we've seen before, yet adjacent to where we've been. So ideally, it's a company that's literally one of a kind, based on things we are used to, whether it's AI, whether it's something synthetic biology, whatever it might be, but they're taking it in some new direction. So the window, as long as you have an agile mind and you're looking at it, you're like, wow. Like no one thought of that when we started. So for example, when we first invested in Tesla, there was no concept whatsoever autonomous driving.

20:48It was not in the business plan. There was no talk of it. It was not in anyone's mind. The way in which electric drive train uniquely enables that and control of fidelity was fascinating. Or in SpaceX, the Starlink opportunity, like, oh yes, of course, when you lower cost of launch that much, you can have mega constellations, but what would be the new thing that would make sense that we weren't doing before? not just we invested in planet labs for earth observation yes constellation of telescopes but this whole notion of building a network you know backbone for the internet in the sky was and then direct the cell phone like each one of these things is unfolding than orbital data centers right not on the dance card even five years ago so that continues to surprise me in some ways it's not easy, but it seems so much more powerful as a business vector than purposeful design, if you will.

21:39It's almost like exploring the option space or the light cone, if you will, of possibilities in an economy versus sort of planning out something 10 years in advance and having it go according to plan, if you will. It is so fascinating how you were successful in so many different bets that you made in the past, and they're so different from each other in different industries. What are you betting on now? What should we be looking out for? Plastics. No, let me think. Some people who know the old movie. Let's see. So taking that thesis that AI and information technology will innervate every economy, meaning add a nervous system to everything.

22:18We saw in automotive and aerospace. Just expanding on that thought a bit, we are looking for additional things in energy. We've invested in a variety of nuclear fusion and subcritical fission that doesn't trigger NRC regulations. Basically, avoiding the Nuclear Regulatory Commission, but figuring out energy, which, by the way, is the third bottleneck for AI. It's not just good people and a lot of compute. It's also energy. There are a bunch of things that you could imagine 500 years from now have been solved, and we're trying to figure out the entrepreneur that will open our eyes to how we get there.

22:46So free health care forever via a cell phone, all diagnostic information you could possibly need for your personal health should be a free service globally. trying to figure out how to get there. Probably won't be in the U.S. that it launches. Bypassing FDA, bypassing insurance and reimbursement. On food, we won't slaughter animals for meat. The products are getting there, but you can sort of see the future. It's so close. You can almost taste it, so to speak, whether it's cellular ag, mycelium, or other techniques, mycelium being the fastest growing thing. But we are going to eat meat-like things that are delicious, healthy, and not involve slaughter of animals.

23:21Construction, growing as a percentage of GDP and labor productivity has been flat for 30 years. So it's such a hard industry to change that we've tried and failed a few times, but we're looking. Again, so the best I can do to answer your question is, I don't know what the answer is, but I know there are these categories that we want to look at. Recently, we've been investing in epigenetic editing across a variety of things from crop health, pesticides, herbicides, human health. It's fascinating. It's basically the software of biology instead of going to the firmware of our genome. And we've been investing in materials, critical minerals and metals, everything from deep sea mining to copper refining because of incredible need.

24:04It's sort of like the workhorse of all these chips is you need these materials to make the stuff. And there's a couple of that, a reshoring or bringing back to the US capacity to build, which we had atrophied over many years. Analog AI, I mentioned it, We have three different investments coming at it from different angles using AI to design AI chips. Sorry, analog chips. Analog chips. Yep. Analog in-memory compute from Mythic where they can do 8-bit multiply and add in a single transistor. And then unconventional, which is taking a very, very strange and forward-looking big bet on, you know, in every case trying to get 100x and then another 100x on power reduction, power per calculation.

24:41Overall, we're about 40 % life sciences, 60 % IT. And we, in the life sciences side, just see, we look for the weird things that are like on the edge, you know, harvesting organs for transplant, growing humans without brains so that you can use their organs. There's a company here actually in the audience doing the same thing. A male birth control pill, improving IVF dramatically, things that fall through the cracks of a traditional pharma VC. So I'm hearing agriculture, biotech. I'm just thinking in my head, how can I replicate your strategy with ETFs? Our strategy, it's very unusual. I can state it openly and then it's hard to replicate.