Vinod Khosla & Delian Asparouhov | Venture and Valor: Building the Future from Silicon to Steel

The most significant force in producing goods and services inside the United States that have a better cost structure than goods produced in China and shipped halfway around the world with very large latencies, huge inventories, lots of problems with manufacturing in China, but using AI and robotics and robotics will have its magic ChatGPT moment in the next two, three years, most likely this exponential improvement capability. So it's a real opportunity to do manufacturing in the United States, not only for the US but for the rest of the world too.

I admit it's always a little intimidating to interview your former boss, but at least now Vinod can't fire me if I ask the wrong question.

So okay, we might still fire you, we investors.

That's true, that's true. Okay, I've got a, I've got a fun first one. I don't know if you remember your talk at last year's Hill and Valley forum, but you ended on this note of saying that releasing open source models within the United States by our leading labs could be incredibly dangerous because the Chinese could go and adopt that into their weapons systems. Kevin, the sort of chief product officer from OpenAI was just on and they do now have a plan to release an open source model partially in response to Deep seq's sort of release and that giving China some soft power. So my question to you is, is it the right move for us to go release an open source model that's a generation behind to mitigate that soft power or maybe a little bit conspiracy theory? Did China release Deep Seq just to bait us into releasing more powerful open source models that they're going to then go use to integrate into their Chinese weapon systems?

Well, China has good AI researchers, let's acknowledge that and they put a lot of resources behind it. There's maybe more orchestration which is a weakness in China. China's approach to R and D, ours is more haphazard and chaotic, which I think has real strengths as a developing methodology. So if you go back to last year's forum, and by the way, I said the same thing the year before that at this forum, I don't think we should release state of the art models, we should release the one before then. And that's exactly what OpenAI said they would be doing. And it is the right strategy because we don't want to lose on open source, but we want to win on SOTA models in even a 6 or 12 month lead in AI model development, given the rate at which capability is developing, will be a self reinforcing advantage. So I've always argued it's only the sort of models that shouldn't be open source. Just like we wouldn't open source the Manhattan Project, we wouldn't open source that or our fighter blueprints, which the Chinese have access to now. But so I realize it's hard to contain, but even small leads matter in this space. And soda is the only thing I've objected to. I'm saying in the national interest, we should not open source them.

And two years ago there was a bit of this trend in Silicon Valley. It seems like it's mostly disappeared, but around a bunch of people trying to say we should stop development on AI, come up with a treaty.

That's the silliest idea I heard.

Why would an AI treaty basically never work?

Well, very simply, when we've had treaties in the past, in nuclear or in biowarfare, the use of the technology is verifiable. The use of AI technology is not verifiable. And so you can't have a treaty when you can't verify if somebody is adhering to the treaty. And we'll try, the west is more likely to live by treaties. But China, or imagine Putin, is he going to live by a treaty that nobody can verify he violated? No, not chance. So I don't believe in trust in this kind of geopolitical game.

And so Kevin earlier said that the four key things for us to stay ahead of the Chinese are infrastructure, so energy and GPUs, data sets, algorithms, and talent. When you think about what it takes for us to stay ahead of China, what of those four do you think are the riskiest in terms of our ability to make sure that we have.

Them to today not getting the right amount of immigration. America's advantage is we get the best talent from anywhere in the world. If there's one thing I'd say we could do, it's for people who have PhDs in math or physics or AI. Get them here proactively, go woo them to be here, because it will be our largest advantage. The other stuff follows from having the best sort of models and the best data. Now, China's also very liberal on data issues for their AI companies. That's a disadvantage we have now. I still think we should do things like respect privacy, but we cannot afford to lose the race. Let me elaborate on that just a little bit. I think, and I used to say two years ago, I said in 20 years, I will today say in 10 years, whichever country has the better AI model will win the economic race by a long shot. It is such a powerful technology. And I could be talking about three to five years, not 10 years. I used to talk about 20. Once you win the economic race, you also have both the hard power stuff, cyber warfare and on battlefield warfare. That's obvious. But people haven't thought about the soft power issues. I suspect China could easily go to all of Southeast Asia, Africa and say we will offer you free doctors, every doctor you need, every specialist you need for free because it's going to cost less than a dollar a month for providing that service. That's soft power. Could go and say this is what I would do if I were China, what I would do if I were the US Instead of the foreign aid program trying to feed a few thousand people, I'd go to offer free AI tutors to every country, every child on the planet. It'll cost us under a dollar a month per kit. Far less than almost any of the foreign aid programs we have. And dramatic in terms of soft power, what scares me is and I was pretty active in the effort to ban TikTok. I saw a speech from President Xi that said the next battle. And this was a year ago and this was the argument I used to when TikTok was being considered. The next battle will be fought on the Internet and the battle will be for people's minds. Connect that to they took the algorithm for TikTok and made it a national security asset. It couldn't be exported. Put those two together, it's obvious and we know the Results. Kids on TikTok are much more China inclined. So more than almost anything else, which is hard power stuff, persuasive AI will be such a powerful weapon. I'm afraid of that in Chinese hands with Chinese ethics used to perpetuate a political philosophy that's different than ours. Even if they won the economic race, it wouldn't be as bad. But if the world converts to their political philosophy, which would flow from their economic wins and their soft power and that flows from AI race. That's why the AI race is so important.

And so when you think about policymaker, let's say priorities or leaders within the government is, let's say integration of AI into the war fighter community, is that actually sort of a distraction relative to the North Star of all that matters is win the AGI and win the economic race and then whoever integrates and the warfare fighters, well, that's just a natural thing that will trickle through.

That is a natural thing that will trickle through. So there's a lot of implications of that in so many areas. So it will look there's a conflict going on in the world. If I sit down with Sam Altman today, and I did this fairly recently and sit down with a Larry Page today, or any of the people really knowledgeable about what's going on, their view of the rate of progress is so dramatically different. On the other end is unions like the American Medical association that won't let you bill for AI based consults, which could cut out half of expertise costs, which is a quarter of all healthcare costs. My bet is the expertise part of our medical spending. Trillions of dollars, $4 trillion could be cut in half within two to three years if we allowed AI to prescribe and provide medical advice. But we have the AMA as a union preventing that from happening. They won't even let a nurse with four years of medical school prescribe you a Z pack. And most states, a few states have bypassed that. But that world, which is doing a lot of rearranging deck chairs on the Titanic colliding with this world of OpenAI and Google and others, AX and others doing exponential development of capability, we have a Titanic shift coming and it'll happen next five years. It's, it's almost hard to imagine how it gets resolved.

Yeah, I mean, touching on, you know, sort of the, you know, sort of a doctor side of this. The current administration clearly has a, you know, sort of broad set of priorities around cutting government costs and they spend a lot of time on, you know, sort of USA and other areas. But if you look at the budget of the US government, the vast majority of it is, is entitlements and Medicare being one of the largest ones. How do you think this administration should be thinking about, rather than focusing on certain smaller budget areas, thinking about how to implement the AI doctor, the AI teacher, or how, you know, we had the, you know, sort of panel here earlier today on how to even think about the like AI that just catches the individual soldier that's spending money at MGM casinos and that probably shouldn't be happening.

So there's a big battle this week on Medicaid. I was just talking to David Schweika's congressman who's been a huge proponent of AI being able to prescribe. And I said, hey, part of the trade off, I'd cut Medicaid spending a lot if they let AI prescribe. So maybe that's a tool in this political horse trading. It's such a big deal that you could actually make Medicaid, even with large cuts, profitable if you allowed AI to do some of the job. So I see David Sacks sitting there. Make that happen. So, look, it's a big deal, and we could take a few chunk of our medical spend out. Expertise is at least a quarter of all healthcare costs, and that could be cut in half in three or four years and probably by 80% in five to seven years. And that's the magnitude of the opportunity with that one thing. And I think even if we said AI can only practice under the license of a current physician, but acts like an AI intern, fresh MD out of a medical school, we'd still cut costs 80% where they get to review the results of the AI, say within 24 hours or 8 hours, same day, whatever the requirements. It's such a large opportunity, especially in the big budget area like healthcare, hhs.

I guess, related to the implementation within these medical areas. There probably is some room for some sort of regulatory oversight sort of process in that. The last administration, you know, I think we could argue, was trying to regulate AI in a way that was more to control the AI and potentially stifle some of the innovation on it. This administration has taken a very laissez faire approach. They appear to have, you know, sort of no interest in trying to overregulate AI. Perhaps the answer is somewhere in the middle where there needs to be some amount of, you know, sort of licensing. Where do you think that. Right. You know, sort of blend is?

So first, I do think there's real danger of AI going rogue. So let me acknowledge that. And the people who propose that are thoughtful people, people like Joshua Bengio and Jeff Hinton really worried about it. And those are respectable scientists. So we should worry about that. But the benefits so far outweigh the potential risks. There's also a risk an asteroid hits the planet. Not all risks are controllable. So I do think we should worry about it. If we did any funding in universities, it'd be on research on AI safety because they don't have the resources to develop foundation models anymore. There are other approaches to AI than the traditional transformer model that I'm very interested in that may actually solve the safety problem, and we should be doing more of that. At the very least, there'll be more efficient ways to do some of the things that take a lot of power consumption and a lot of GPUs and stuff like that. So I do think there's other answers here that we should worry about. But by far, sentient AI is a danger and make it kill every human on the planet. Way, way bigger danger is China's AI in China sense, like, so you have to look at the basket of risks humanity faces. China AI is by far the biggest risk and more immediate in some ways.

This administration's approach on trying to limit China has largely been through both, you know, sort of tariffs as well as sort of sanctions. Obviously some attempts to limit the amount of GPUs that are flowing there. I guess maybe if you were, you know, in the presidential suite, or you could determine the sort of tariff and foreign policy, what would be your recommendation and how do you think the current tariff approach by the administration is working? Not working? And how can AI help US Manufacturers deal with the current environment?

So first, independent of the tariffs, Covid showed us we don't want to be dependent on President Xi's policies for his own domestic purposes. And probably good reasonable thinking. He did a lockdown. Suddenly you couldn't buy a refrigerator in the United States for 12 months. You couldn't buy a car. We cannot be dependent on our supply chain in China. That's a strategic issue for every corporation in America today and worldwide. So outsourcing the supply chain from China, I used to call it the China plus one strategy was. Was really, really important, and we can get to how we get there. The answer is the same. In addition, they manipulated their currency to achieve a balance of trade that's totally lopsided, and we should balance trade. So am I in favor of tariffs because they're essentially subsidizing through currency manipulation? Absolutely. But where I would disagree with the administration is that let's declare a consistent policy over four or five years and then help onshoring what will help onshore. You can't just move a people operation from there to a people operation here. As much as it'll create jobs, it'll have inflation go through the roof. China has been a great deflationary force the last 30 years in this country. What you want to do is say, are there better ways to onshore AI and robotics, Absolutely. The most significant force in producing goods and services inside the United States that have a better cost structure than goods produced in China and shipped halfway around the world with very large latencies, huge inventories, lots of problems with manufacturing in China. But using AI and robotics and robotics will have its magic chat moment in the next two, three years, most likely this exponential improvement capability. So it's a real opportunity to do manufacturing in the United States, not only for the US but for the rest of the world too.

I mean, one of the key components that's required for us to even stay in this sort of race is obviously the semi chips that go into the GPUs for these sort of models. We've obviously in the prior administration passed the bipartisan Chips act that has enabled TSMC to start to set up some sort of shops in Arizona. They recently announced that 30% of their sort of frontier nodes are now going to be produced in Arizona. Do you see that as a sort of potential for us to sort of more rapidly decouple or do you think we're nowhere close to being able to decouple from let's say semis in particular from sort of Taiwan and China?

I do think it'll help decouple. Look, one of the large, very large dangers is China taking over Taiwan. I think it's a major risk and I think the timeline for this that has accelerated. For those of you interested in that topic, there's a great book called the Danger Zone that argues that we are in a particularly critical time for China and the longer they wait to invade Taiwan, the less likely they are to succeed with it. So there's a critical danger. But if they do invade, we got a problem on chip supply. And so as soon as we can bring this up, we should absolutely do it. And it's critical that we do and decouple from dependence on Taiwan. As much as obviously Taiwan doesn't want to see that happen because it's their security shield. But I also think there's a different form of warfare that we are starting to see in the venture community. AI based drones, coordinated systems, new kind of dynamic in warfare. For those interested in that, there's actually a five or seven year old book called the Kill Chain. How the nature of warfare speculated is going to change. It's become very evident from this five or seven year old book that that's what Ukraine is about. And I do think we have the ability to build defense capabilities if we rely on the venture community to do that. And if you rely on General Dynamics, they'll wait for a contract from the Air Force to develop X over 10 years and that doesn't work. They won't take any risk themselves. While the venture community is happy to plow $100 million into demonstrating some capability so they can get a contract if they can develop the capability. So that risk taking attitude is so critical and the venture community can really help because all we do is take risk.

Yeah, I was talking with Emil earlier. How do you incentivize the government to start to think about risk being an okay thing? And you have to celebrate, you know, the wins and then also be willing to sunset and cut off the failures versus by default.

Today I ask people a question. I would say failure doesn't matter. It's the magnitude of the success that matters. So most people know my background. I started Sun. Probably a fair number of any audience knows that. Nobody can ever tell me the other company I started at the same time with the same co founder that I started. Do you know that answer?

I do not know.

Nobody's heard of the failure. Who cares? I argue failure doesn't matter. It's large successes that we should aim for.

True on the investing side too. I'm sure there's been some failures. But you also led the Amazon Series A, which I think that's turned out pretty well.

And Google, Amazon, Google in, you know, things like OpenAI help.

Yeah, not too shabby. How do you think about some of the reshoring of things that like, aren't necessarily obvious that we want to have happen in the United States? So as an example for like neodymium magnet, you know, sort of processing and refinement, 90% of that basically happens in China. It also generates these, you know, I think they're called tetrafluoride. You know, basically, you know, sort of compounds as a part of that process where you basically have to. There's no way to really, you know, sort of process or dump it. They have these like poison pools around the manufacturing site. Like, do we want those poison pools back? I mean, yes, because they go, no.

I don't want poison pools back. But we absolutely have to ensure our supply chain. And it can't be in China, right, Because it's a very important. Is the biggest geostrategic enemy. And I don't mince words on that. I don't think there's any compromising with China because of what's at stake. Stake geopolitically. So we have to do it. I'll tell you a story. In the early 2000s, we had a solid oxide fuel cell company called Bloom Energy. They were doing scandium oxide made in China. Suddenly China imposed a policy that you couldn't export scandium oxide. You had to add value inside China, which was a way to ensure the use of the technology was in China so they could steal it. That was the whole purpose. Bloom had to set up its own mining operation in the US for that particular rare earth. I think we can't not take that risk. We have to have these supply chains and we can do much better. We've just invested in a company that can do Process control in a radically different way than almost all minerals and mining operations. Processing is done in China. They've opened up so many lithium facilities and they put gobs of people on it. Well, there's a much better, smarter, AI driven way. It's called Mariana Minerals. We've just invested in that last year and re up. But these mineral resources are critical and better processing is entirely possible. It's just it's been in the hands of people like Rio Tinto that have such large operations they have no incentive to take the risk to do it differently. That's what we are trying to do. Take the other issue people will remember. In 1960s everybody declared peak oil. There was no more oil to be found. And then we created the tools to discover more oil and gas and technologies and there's no shortage of oil. We've also invested in AI based discovery using proprietary data sources for new minerals and resources. I don't think we need to go invade Greenland to get mineral, plenty of it here. We just don't have the tools to look yet. And so our focus is to have new kinds of tools to look. Can you look a kilometer under the earth's surface? We only do minerals and stuff where it's near the surface and then we start digging deeper. Much better ways to do it. It's never been a problem. So we haven't spent innovation time and R and D dollars to try and find ways. But we are doing it now.

Absolutely. Well, vk, I want to say thank you so much for joining me on stage. Hopefully you can tell Samir not to fire me from Varda and enjoy having you at every single forum.

Great to talk about these really critical issues. Thank you. Thank.

You.