AI风险投资家：这些技术预测将在2030年前改变一切——与 Vinod Khosla 对谈 | #159

It’s been too long.

Thank you. It’s been long, but it’s great to see people who believe in abundance.

Yes, to see so many people who believe in it. Congratulations for that. My favorite saying is, “Skeptics never did the impossible.”

Yes. That’s a good place to start.

It is. I’ve had the pleasure of knowing you for a couple of decades now and of being in some deals together through Bold Capital. I’ve always loved your vision, but I think it’s important for people to realize that beyond being a great investor, you’re an entrepreneur and a company builder. You’ve built some extraordinary companies.

Thank you. I have to say, in 40 years, I haven’t once called myself an investor because I don’t like investors very much.

In December 2023, you predicted 1 billion bipedal robots by 2040 and that robots would play a larger role than the auto industry. Six months ago, you suggested robots could reach $10,000 and be as essential as a smartphone. Coming off Brett’s presentation, talk to me about your thoughts on that.

I think Brett is right. It is going to be a large business.

And from a situational-awareness point of view, how many people have read “Situational Awareness: The Decade Ahead”?

For the rest of you, you should read it. If you don’t want to read the whole thing, grab the paper, put it in NotebookLM, and listen to the podcast.

That’s a good idea.

I’m serious. It’s an extraordinary paper.

Yes, it is extraordinary.

I think what I would say is this: This business of bipedal robots, or some equivalent in various form factors and configurations, will be larger than the auto industry within 20 years, and the auto industry doesn’t know it. If we have 1 billion bipedal robots—and I think that’s an underestimate—there may be many more than 1 billion.

I agree with you.

First, these robots will do more work than all of the manual labor humanity does across 7 billion people today, right? They’ll be working 24/7, not 8 hours with breaks.

Yeah.

So the capacity for labor goes up dramatically, while AGI can increase the capacity for human thinking and human expertise, which will essentially be free. We can come back to that, but it’s such a large opportunity that very few people—and almost nobody in the group that should understand this, which is the auto industry—have recognized their biggest opportunity, except Elon.

He’s doing robots. Elon says it will be bigger. It’ll be the biggest part of Tesla: Optimus. I believe that.

Yeah.

You go on in your predictions to mention that expertise will be free, with AI-based tutors and doctors available 24/7. For the abundance members here who are running health care and education companies, what’s going to be the impact on them?

Well, let me put it this way. In 2012, I wrote a blog—or two blogs. The first one was called “Do We Need Doctors?” Just to be a little conversational, it asked whether we need doctors. Now it’s clear they will play a minor role in health care. They will play a role in certain levels of human connection, and that will be important.

My second blog was called “Do We Need Teachers?” Frankly, I just saw a piece out of Beijing today about why the Chinese government is going to adopt AI tutors for kids. This was in 2012.

Almost certainly, no matter what expertise you’re talking about—AI doctors, AI teachers, AI oncologists, AI structural engineers, AI accountants—we should be investing in all of these. Any place where the U.S. Bureau of Labor Statistics says there are more than 100,000 workers in a category, we should be investing in all those categories.

So what will its impact be? You have to talk about time frames. In the next 5 years, every single professional should be thinking about how they can get 5 AI interns. If you’re an M.D., you can supervise 5 AI interns, just like they graduated from Stanford Medical School. If you’re an accountant, you already hire junior accountants fresh out of school. Imagine 5 AI intern accountants.

Whether they grow up in 3 years, 5 years, or 10 years is a little hard to predict. It’ll vary by category. But that’s a good model to think about first: leveraging expertise.

The more interesting question that I’ve asked a number of boards I speak to in the health care business is: If all expertise were free, how would you design a health care system? I think that’s the right question. That’s the right model if you’re a health care provider of any sort.

Let me go a little further and say, at the national policy level, if a farm worker and an oncologist have exactly the same expertise—up-level to what an AI has—how do you pay people? That’s an interesting question to ask.

You’ve said that the most common consumer access to the internet could be agents acting for consumers. I don’t think my agent is going to be interested in Google ads or a TV commercial featuring somebody with white teeth selling toothpaste. What happens to advertising if agents are doing the buying?

I actually might disagree.

Okay.

First, I’d say the one thing that’s clear is the level of unpredictability in all this, including what I say. The only thing we know about the future is that it’ll be uncertain. We have to figure it out in increments.

But what is advertising? It’s information to emotionally induce us into buying a $300 pair of jeans when $30 Levi’s will do just fine. That’s a filter the AI can apply.

The nature of advertising, I suspect, will change to be more informational, which is the original purpose of advertising. It has switched. One of my big beefs with capitalism is that capitalism is mostly about making you buy things you had no intention of buying. That’s what advertising has become.

Making people buy things they don’t want to buy or had no intention of buying makes them unhappy. That doesn’t add to societal good. But information is valuable, and frankly, I think people will pivot into other models.

ChatGPT has been phenomenally successful, and nobody thought consumers would pay $20 a month. Look, if that was true, that’s more revenue per user than Google or Facebook. In fact, they should have tried this model earlier if they added enough value, which, in the past, pre-AI, they had added a lot of value.

So where will this stop? The $200-a-month level of ChatGPT is incredibly valuable. How many people have tried it? A good number of hands. How many people have tried Deep Research?

It is amazing. It is way more valuable. I’ll tell you a fun story. My chief of staff is a Ph.D. in viral immunology, and we were researching a particular application—a developmental model, a foundation model for developmental biology. She did an 18-page report for me, including recommending what indications to go into.

After she had finished, I said, “Just enter it into Deep Research and give it some time to think.” Coincidentally, it came back with an 18-page report and the same 2 disease indications to go after. Now that’s mind-blowing.

It is. How much did you pay Deep Research for that?

Not a lot. It was well worth it for anybody who hasn’t tried it.

You’ve been a big investor and supporter of OpenAI.

Yeah, we committed in 2018. You have to believe in things, and I believed in it enough to make the largest initial investment I’d ever made, by a factor of 2. In 40 years of venture capital, I’d never even made an investment half that size because I was a believer.

That’s turned out well.

Yeah, it has turned out very well. But my point is this: A retailer’s CEO was talking to the CFO of OpenAI, and he said he had just finished a large strategy review on what products to add to their retail operation.

While on the call, she summoned Deep Research. It came back with a report she gave to him in tens of minutes, and he said, “To be honest, this was a better report than my strategy team produced over the last however many months.”

So we are at a transition point. One of the things I tell people is that you’re not using this technology even a fraction of your potential. Is it open in every board meeting? Is it open in every executive meeting? Are you giving it the same challenges you’re giving your team?

The other thing I want to add is that the range of applicability goes from asking, “What product should a retailer add?”

Yes. I was sitting in a scientific meeting for a couple of days—multiple days—with a small group of scientists going after a particular topic of personal interest to me. So I sat in, and anything I didn’t understand, ChatGPT was able to explain to me, so I could not only follow the conversation but ask not-so-dumb questions.

Yeah, it is a superpower, and one of the things I love—I keep mentioning this—is the ability to ask it to look at something from a completely different perspective. Right? My favorite is, “How would Steve Jobs answer this question?”

Yeah, and we are so limited inside our organizations by the knowledge and opinions that we have on our team. Why would you limit yourself in that way?

Well, the best example is the row 37 example.

Yes, of course. Centuries of humans hadn’t come up with the row 37 example because we look incrementally at what others have done, rather than think outside the box. So that’s a real, massive accelerator for humanity.

Do you believe that most Nobel Prizes in the future will be derived from AI-human partnerships or AI alone?

Possibly AI alone. I think that will happen. Whether we humans will allow—or recognize—it as such, recognizing AI as capable of winning a Nobel Prize, that’s a human limitation. But that’s also true in most creative tasks.

Let’s go into programming. Probably the single area that’s had the most success with LLMs, besides just the written word, has been programming.

How do you think about the future of programming? Is it a future of no more programmers, or is everyone a programmer?

Both. We invested in Cognition, to be fair. It doesn’t give you a copilot like Microsoft does for a programmer; it does the programming for you. Their model is very much what I explained earlier: they’ll give you a software intern who can program—an intern programmer today—and they’ll grow in seniority and experience over time.

At the other end, we invest in Replit, a company called Replit. You don’t need to know anything about programming or even the names of the language you’re coding in. You just speak in English, and it codes for you. So whether you’re an individual whose son or daughter asks you a particular question, or you want to build a model of the planetary system, you just code it in Replit without ever having touched code or knowing anything about coding.

And if there’s an employee in a large enterprise—and this is starting to happen—every employee can write their own applications without going to IT. So anyone who has worked in large companies and needed to go through IT and security reviews and all that other stuff knows how freeing it is if you don’t have to deal with it at all. You can just program your apps in English.

In the past, users of computers have had to learn computers. You get trained on SAP, Oracle, Workday, or your favorite application. I think in the future, instead of humans learning computers, computers will learn humans, and humans won’t have to adapt to how our programs work. The notion of a menu will go away.

I remember early on, the PalmPilot had us adapting to it and learning the proper language. Then finally, the computer got smart enough to adapt to us.

Energy is one of the single most important things. It impacts health, it impacts AI, it impacts everything. You’ve been a huge investor and proponent across a multitude of different energy investments. How are you thinking about where we are in the near term and the middle term?

Well, I’m pretty optimistic about energy and about climate solutions. Coincidentally, the year we committed to OpenAI, which was 2018, we made 2 other investments I’m very proud of. One was Commonwealth Fusion Systems.

Yes.

It happened to be the same time; they weren’t orchestrated. I didn’t connect that AI would need more power, so I wasn’t that visionary. The third was another one of my favorites, which hopefully we’ll have time to talk about: public transit using AI. It completely upends every assumption about public transit. There is no chance that, if you have a chauffeured car, it’ll be more convenient than public transit. That’s the world we’re going toward. I suspect most cars in most cities will be replaced by 2050.

But let’s talk about energy. Commonwealth is a fusion company.

Last time I looked, there were something like 40 VC-funded fusion companies, which blows me away. I would have never imagined that.

Well, you couldn’t imagine it until 2018. Nobody wanted to fund fusion. I won’t name the people who said no to me when we funded it. They said, “Are you kidding? You’re being ridiculous. It’s 50 years.”

It’s always been 50 years away.

And the Department of Energy didn’t want to talk about fusion as an energy source when I visited them. We invested anyway. Almost certainly, this won’t be a question in 5 years. So my current perspective is that everything we can make happen by 2030—by 2030, nobody will be debating whether fusion is an economic possibility. There may be some question of exactly how much it costs.

So educate us for one second on Commonwealth. Where are they? What have they been able to demonstrate?

Without going into details, they expect to demonstrate positive net energy output by early 2027 and start construction of their first plant soon thereafter. They’ve announced a site for it.

In the interest of time, let me just say they won’t build fusion power plants. The fastest way to deploy fusion—and I believe one of my forecasts is that—we will replace every coal and natural gas plant in this country by just repowering the boiler. So we’ll have a fusion boiler.

I love that vision.

By the way, the environmentalists who say, “Well, we need to shut down all the coal plants,” I think we will repower them with fusion boilers, and maybe a turbine change and steam conditions. There are lots of technical details, but you don’t need 10 years of permitting, 10 years of getting a grid connection, power lines, and all that other stuff. You just retrofit the same locations. That’s why I’m optimistic this won’t be a problem.

There’s another technology that nobody is talking about.

Please.

Geothermal.

Ah, yes. We’ve toyed at the edges. We need better drilling.

Yeah. It turns out the same well, if it’s at 200 degrees or 250 degrees—which is almost all geothermal anywhere in the world today—produces a certain amount of power. But you take that and drill to 400 degrees or 450 degrees, and you get 10 times the power out of the same well. Suddenly, superhot geothermal, as it’s called, is cheaper than natural gas. I think that’ll be deployed well before 2030.

We have 2 different efforts. The hard part is drilling at high temperatures, because metal becomes soft, and metal can’t drill in hard rock. Sometimes hard rock is hard to fracture. Both those technologies are well on their way and will surprise us in how much power they can generate.

Do you want to jump into cultured meat? Do you want to dive into transportation? What are you most excited about?

I’m very excited about transportation.

Let's go there. What does it look like? I want to go from point A to point B in the future.

Yeah. So, here's what it looks like. First, let me start with what it shouldn't be and what the right role for self-driving cars is. When we added Uber to cities, what happened? Congestion increased. Simulate driverless cars, and they'll increase even more.

So, we'll have gridlock, but people won't be driving, which is—driverless cars are cool. I recommend everybody fly to a site that has driverless cars and try them. They're amazing. But they increase congestion.

What should you do? Driverless public transit—that is, personal vehicles. Should you get off your restaurant job at 1:00 a.m.? A car shows up for you. You don't need to know when it's on the schedule. Why? Because there's no driver.

So, your cost is zero to have a car available when you show up, not when the public transit is scheduled. It gets to where you're going independent of the time of day and traffic lights, doesn't stop anywhere, gets to your destination, and drops you off. And because it's personal, nobody gets picked up or dropped off along the way.

So, it will always be faster than a chauffeur-driven car. That's the right use of self-driving. Off these fixed routes, a Waymo might make sense for the last mile, and I think it has a useful purpose.

But here's the fundamental thing. What is the one thing, if you ask a basic question, that's limited in cities? Street widths. Only one thing is limited. Everything else you can change.

And so, if you can increase the throughput of city streets, because you're not going to destroy the buildings on both sides of a street, you've solved congestion. It turns out that with a self-driving system like this, if you model it, it can 10× the capacity of a street by building more capacity than light rail in a bicycle-lane width.

That's the system we are designing. It can run in bicycle lanes on the road, in cordoned-off areas, not interacting with other traffic, because that's key to throughput. If you 10× the capacity of a street, you've suddenly created 10 streets wherever you had 1, and regular cars can still flow without disrupting the thing.

This is a far better solution than congestion charges. New York has 14th Street, where you can't have public or private cars on it. Constraints aren't the right way to go. Getting 10× more capacity is the right way to go, and then you can solve the housing problem, too, because you can live farther away and have the same 30-minute commute.

I love that. I want to close on a question before we go to the Q&A, so please, again, get your questions ready. You know, I have been so adamant since I wrote Abundance with Steven Kotler that there are no resource constraints. There are no resource limitations. Technology is a force that converts whatever was scarce into abundance. I think you agree with that.

I 100% agree with that, and I would urge somebody to ask me the question that always gets asked: “Well, we have a restricted amount of minerals or metals or ores.”

It turns out it's a constraint because we are not looking, or not looking with the right tools. So, we have a separate effort to look under the Earth's surface for minerals and metals. We only mine stuff that's on the surface, mostly—the most convenient stuff you run into.

Yes. Yeah.

Accidentally run into. And if we can look a kilometer under the Earth's surface, we'll have an abundance of minerals, metals, cobalt, lithium—whatever you're looking for.

Amazing. All right, guys, let's give it up for Vinod Khosla.