But I want to touch upon your investment in OpenAI in 2019. You have been known for contrarian bets, but this was way out there contrarian.
We wrote a $50 million check, and when we wrote the check, we also sent an apology letter to our LPs saying, 'Sorry, we're making this investment. It doesn't make any logical sense. It's a nonprofit. No product plan, no business plan.' I love making those plans.
So freedom is the ultimate promise of AI, and just people free to live life and that becomes a new normal.
Who wants to work on an assembly line for 8 hours a day for 40 years? That to me is slavery. It's servitude to survival because you need a job. Those are not jobs with human dignity.
But what does it mean for someone who is going to college now? Why go to college?
I think the only reason to go to college is...
Vinod, welcome to Sparks. I've been waiting for this conversation for a long time. You may not recall, but I have been to this office a few times. I've met you in different contexts, and it was good to see you at the AKA event as well. I want to start with... I'm in the Valley for the last few days, and I come here often once or twice a year, but the pace seems very different now. There's so much action, and a lot of it seems to be concentrated in San Francisco. You've been in the Valley for the last 40 odd years. Does this feel different?
It definitely feels different. There's a lot of activity, a lot of genuinely good startups, but also a lot of money.
And that combination has resulted in velocities that are very, very high.
Yeah. In one of your talks, you mentioned, and I think it's fairly obvious, there are only three or four very deep model companies now in the US, and maybe there are an equal number in China. Do you think most of the innovation is concentrated inside these companies, and there's a delta between what the companies outside these companies are able to do?
No, I think there's a lot of innovation outside these companies. A lot going on, at least in this area. Every day there's a new idea that's worth considering, if not funding. As a result, the best way to understand Silicon Valley today is: of all the unicorns, 91% are in the Bay Area. The rest of the world combined accounts for 9%. Which is pretty stunning.
This is all unicorns, not AI unicorns?
I think it refers to AI unicorns, but the second highest is New York with 2%. So even the number two is largely irrelevant.
And so the action is here. Almost all the innovation is here. If you look at any AI company today, the engine of all AI companies is obviously the inferencing, which is powered by one of these companies. What is the nature of innovation outside?
But I want to touch upon your investment in OpenAI in 2019. You have been known for contrarian bets, but this was way out there contrarian. If you recall your thinking and the discussion, what convinced you? It wasn't even clear where OpenAI was going to be a for-profit company or nonprofit, yet you wrote a pretty large check.
I wrote a check that was twice the size of the largest initial check I'd ever written, because I don't like writing huge checks. We like very small checks. So I'd never written in 40 years a check over $25 million initially. Later on, we added additional money. And we wrote a $50 million check. And when we wrote the check, we also sent an apology letter to our LPs saying, 'Sorry, we're making this investment. It doesn't make any logical sense. It's a nonprofit. No product plans, no business plan. It's just a technology effort.' But the idea was very simple. So today, that letter is sort of a funny thing. I should probably tweet it. It was so funny in retrospect. But what was very clear to me is when you get AI, it'll be very consequential. So it wasn't when we would get AI. It was when we get it, whenever that is, whether it happens in two years or 10 years, didn't matter. It would be very consequential. So the bet was very much on the very large upside if you were successful, and that's our kind of bet. I love making those bets.
And then each bet has to be, I guess, two parts. One is you were convinced that AI, whenever it does happen, will be very consequential. But this particular company, because when you're dealing with so much uncertainty, almost no precedence, and almost no clarity about timelines, picking a team versus the other also has to have some method behind it.
Absolutely. You judge the team. I thought this was a very good team. We were very confident in the team. But there weren't a lot of companies. There was Google, OpenAI, and Baidu maybe as a Chinese company. So there wasn't a lot of AI efforts. It's not like today.
And there's this whole thing about opportunity meeting a prepared mind. I know you've been talking about AI for a while, including I think 2012 you've been talking about AI doctors and so on. So is it almost a case of you were looking for something like this, and you happened to eventually...
I was absolutely looking for it. There are some important areas of technology that I tend to like to identify and say, 'We should, if we find a great startup, invest in it.' By the way, same year, it was 2018 that we committed to OpenAI. It closed in early 2019. But same year, we invested in fusion. Nobody was expecting fusion in the next 30 years. Yet we placed that bet, and I think it looks pretty good today. So I like unusual bets. We also placed a bet that year in public transit, which is also not considered a venture area. I think around then we also invested in hypersonic aircraft. That was also unusual for a startup to be doing hypersonic aircraft. So we're not afraid to place large bets. We tend to be cautious about how we guide them so they don't run out of money, because some of these areas become very difficult. But I'm always trying to say, 'What will be hugely disruptive 10 years from now that people aren't expecting?' If everybody's expecting it, then it's not interesting. But it has to be at the sweet spot of people are not expecting it, but it should at least, on average, become real in the next 10 years. Because if, let's say, OpenAI, I bet if you were to place the same bet in 2009, the story would not look that great because of finite lifetime.
No, look, this is the hard part of venture investing: guessing when you will have breakthroughs, when large changes will happen.
So I'm always looking at everything between 2030 and 2035 and saying, 'What can be consequential?'
And this philosophy as a venture capitalist, or I guess you would like to call it as a venture adviser, has been a work in progress. You've kind of honed it over a period of time by seeing a lot of different patterns. What is the backstory behind doing this way?
So I like to differentiate from people who are investors. They have to run spreadsheets. We never run a spreadsheet on rates of return. Rate of return calculations are just mostly misleading in our business for what we do. Most other investment firms use it. It's misleading because of the artificial mark-to-market that you...
Well, it's not artificial mark-to-market. Mark-to-market in private markets is always artificial.
That's not the real issue. What the real issue is you can't predict these things. If I said to you OpenAI will have no revenue in January of '23, but January of '26 it'll have $20 billion, you'd say I'm crazy. So sometimes it's better to realize you can't forecast instead of trying to make a forecast and then using it as a basis for calculation.
You just say it's not possible to forecast.
Today is obviously very impressive. It can do a lot of things, and one can also argue that we are not fully exploiting what's possible in AI even today. But a lot of the AI investment, there are now close to a trillion dollars of capex being spent this year and perhaps the next few years every year. This also assumes the scaling to continue for the next few years. Is it almost taken for granted now that we have hit an inflection point, the scaling will continue? Because it can also stop. The asymptote may play out, and this may have implications for the whole journey.
If you look at the five-year journey, could you have 10 times the amount of inferencing? Absolutely. There's demand for that.
10 times more intelligent? What's the metric? 10 times more inferences than we do today. Could you see 100 times? Possibly. It depends on the price decline.
And you're commenting on quantity of inference, not quality of inference.
No, just the number. Sort of measure it in tokens produced. Now the intelligence of each token will also go up, which is the quality aspect. So I suspect we'll see continuing increases in the number of tokens almost exponential. We'll keep seeing large but more modest increases in the intelligence per token, which is human intelligence delivered. And we'll see pretty rapidly declining costs per token. And if the cost declines, the demand will be there. I'm pretty confident about that.
But is any of this contingent on newer breakthroughs?
I don't think so. There's a set of breakthroughs we can forecast pretty easily because we know how to get there. If we scale X or Y, we know scaling laws apply. We know certain cost curves of silicon and other things. Put all that together, I don't think any massive breakthroughs are needed. I consider things like memory in AI systems, continuous learning in AI systems as sort of routine things. Again, a question of when it'll come along, not if.
Yeah. And the bigger question is when they come along, how good are they? How good is the continual learning? How good is memory? How well can it be utilized?
So I think those are predictable. I also expect we'll see unusual phenomena in AI progress that we aren't planning on. Unusual good, unusual bad, we can't say. I think mostly unusual good. There will be some bad things for sure.
I want to bring the conversation to health. I think around 2016 you wrote this paper called something like '20% Doctor,' which was about the role of AI in medicine and ubiquitous world-class doctor available to everybody in their mobile phone. In fact, incidentally, 2016 was when I started my second company, and it was a health company. We wanted to solve for holistic lifestyle across everything from fitness, food, and marrying to primary care. Most of those things didn't work out. Fitness worked out very well. So today it's a very large fitness company. But the dream of that AI doctor was at least that we couldn't pursue. But fast forward 10 years from now, you think we are much closer to that vision of where everyone has...
There's no question. Even five years from now, you won't need a human doctor other than in interventional medicine, meaning surgery, heart surgery, burn victims, broken bones. I think five years from now, the expertise involved in a doctor is almost completely already in existence. I no longer consult doctors. I fractured my wrist last week. Got an X-ray, sent it to ChatGPT, and it told me what to do. Now, I had to have a real surgeon. I called the surgeon. Like any surgeon, he said, 'Can you come in for a consult?' I said, 'I don't think I need it. Here's what ChatGPT told me.' I sent him the exact comments from ChatGPT. He said, 'This is exactly what I'd tell you. So you don't need to consult. We'll go directly into...' The first time I met him was in the surgical theater. And it worked out. And since then, I consulted on every question: can I do this, can I do that, when will the cast come off, when can I put pressure? It's a much better... I would say it's at least as good a doctor and a much more accessible doctor. I can ask it a question at midnight and get an answer within five minutes, as opposed to waiting multiple weeks for an appointment. That's possible for every person in India today. And one of my pet projects is to get that going and deliverable in India as a nonprofit effort within the next two years.
And why does it need to be a new effort? GPT already is available in India. They have pretty wide distribution, one of the largest user bases for GPT. I also consult GPT as my first protocol for any health issue, and I'm pretty happy with the outcome. Why does it require a separate effort at all?
I think this is misunderstood. The safety requirements for AI are much, much higher for some areas like health than in others. Today, if you consult ChatGPT, the general triage error rate will be 20 to 30%. That means you get the wrong answer 20 to 30% of the time. Now, humans are pretty bad at triage too, so you have to be careful. But when you add a health-specific system, and my son has a company that does that, they use GPT-5 but layer on things on top, the triage error rate goes to zero, much better than human beings. So I think in all these areas, if you're talking about national defense, healthcare, or financial trading, you want to use these large models. All of them are essential. They're necessary but not sufficient. You need additional things, various levels of insurance. So this is not about making AI better with further scaling. It's about building enough guardrails and checks and balances. Every domain is a little bit different. You need different things and different guardrails. But yes, we know AI systems hallucinate, and it's not been easy to reduce hallucination. If you're doing customer support and you hallucinate somebody's bank balance, that's a problem. So there are special systems that ensure the systems don't hallucinate and still use these systems but with the extra guardrails and safety. Whether it's financial, healthcare, or other critical areas.
And these additional systems are more deterministic rules-based systems?
They're not rules-based systems. I think rules-based systems are too hard to do and manage. They're just better different techniques with different tradeoffs.
Right. Got it. So maker-checker and probably multiple checks, but they're all inference-based systems that are layered together in some way.
So that's why there's opportunity in application areas too.
Right. And this is... I'm just thinking about India healthcare since you mentioned. We have a massive public healthcare system. There are, if I'm not mistaken, hundreds of thousands of doctors working there. So one possibility is to co-opt them in some ways to be the human layer.
Yes. Whether you need a human layer or not, and for what, is an important question. I think each human being will make their own decision. Obviously, if an AI is going to tell me I have cancer, I might prefer a human being to tell me that rather than get a text message. But different people may prefer different things. Others may say, 'Hey, tell me as soon as possible. Text is fine. I don't want to wait one week for a doctor's appointment to learn that.' So different people will prefer different things, and I think human preference will become a big part of how we use AI systems.
Yeah. What does education look like in this? It's a two-part question. For a grown-up person, someone in their 30s, 40s, 50s, probably people need to learn a lot of different things and perhaps unlearn. And a similar question for someone who is coming of age now: how should they think about education?
I spend most of my time today, at age 71, learning. I'm learning new subjects. If I'm interested in fusion, I learn about fusion. If I'm interested in biology, I learn about biology, cancer, drugs, whatever I need to learn, I can learn now. So the opportunity for 30- and 40-year-olds, and 50, 60, and 70-year-olds, is huge as far as learning, and it's almost all free. So that's the big advantage.
And when you say learning now, it's mostly going to GPT and finding out the right resources from there and then following them?
Yep, that kind of thing. I often have long dialogues with GPT on a subject. Even a typical 'How do you design a much better drone motor?' I've not designed a motor in 50 years, but I can still design one with GPT and say, 'I understand the basic scientific principles and I know how to apply them, and I'll do these exercises.' I say, 'How would I design a cancer drug for this cancer?' All those are available and possible.
And this is something you said one can do or one should do?
Well, that's each person's choice. To survive and thrive in this new world, you have to do it. If you don't want to keep up, that's fine. But I'll give you the best example. Most people in the software business know GitLab as a company. He was one of our founders. We invested when there were four people. It became a big company, a public company. Then the founder got cancer. If you look at his history, he basically decided very quickly he wouldn't talk to doctors. He would essentially design his own cancer treatment. He got deep in, learned everything from scratch as a software person, and did his own cancer vaccines, cancer drugs, custom drugs for each stage of his cancer. He wasn't given that long to live. He's alive and well and healthy now, and doing even more drugs for what he might need as cancer drugs in the future. So if a software engineer can do that, and you don't need a medical degree, that's the world we can live in. We can choose not to or choose to live there. I find it very inspiring.
Yeah, this is an incredible example. Does he talk about this publicly?
Oh, he talks about it. In fact, he's given a couple of talks called 'Going Founder Mode on Cancer.' It's a beautiful talk.
