Vinod Khosla: “It’s Over for Human Doctors”

All right. Okay guys, just a show of hands, how many of you are students? How many of you are not students? And then how many of you want to start a startup? Out of all of you, who wants to raise funds?

The other interesting question is how many want to do something in health versus non-health? It's a broad audience.

Yes, indeed, indeed. Given that everyone is here for startup school, I think it's fitting to have the contrarian question around it. YC always preaches to do what users want. I think you are very well known for saying that the customer is not always right. So what philosophy should people follow? And maybe what is also your philosophy around company building?

With transformational technology, and I really like transformation, new possibilities emerge. When you ask people, they know what they incrementally need. In fact, Steve Jobs was famous for that. People would say, I need a bigger disk drive. But he was thinking, what's the experience I want to deliver that's completely different? If you ask any customer today, when we invested in OpenAI, there was no demand for AI. There was no such thing. Surprisingly, in 2018, there was nobody interested in AI other than academics. We invested the same year in fusion. There wasn't a single customer asking for fusion. So if your goal is to change the planet, you develop fusion in a fusion startup. If your goal is to serve a market, that's perfectly reasonable too. I love transformation. I love using technology to change the world for the better. But it's not the only reason you should be an entrepreneur. There's a talk on our website that I first wrote up in 1986. It's called the entrepreneurial roller coaster, and many of you will learn that when the highs are high, the lows are really low. How many have been entrepreneurs before? So you know about the lows too. In it I say it's perfectly reasonable to say I want to do a startup because I want to be rich. Perfectly reasonable to say I want to do a startup because I want to change the world in some way. Perfectly reasonable to say I want to do a startup because I want to be my own boss, I want to be famous, or I want to just work with a few friends. All of those are really valid reasons. Don't let somebody tell you what their reasons are. You should have your own reasons. So there are lots of reasons, but it's not always listening to a customer. There's no customer asking for fusion, but that doesn't mean you shouldn't do a startup. No customer asking for AI doesn't mean you shouldn't do something. It's really fun. One of the really interesting things: I suspect a lot of you are interested in AI. About three or four years ago, when we looked at what customers were asking for, this was after the ChatGPT moment, after everybody was excited, they wanted a tool for humans to use. If you're a software programmer, you want Cursor to help you in your IDE. We decided we wouldn't do anything that was a tool for humans. We decided we'd do workers, not tools for workers. That's a fundamental decision, because I thought that was much more revolutionary, and humans would get in the way of AI. So whether it's chip design or software design or a doctor or an oncologist, humans can get in the way, and they want to apply the old habits to this new technology. So you use 20 or 30 percent of the technology, not 80 percent of it. We decided at Khosla Ventures we just do workers: an AI accountant, not a tool for accountants. We did Cognition, which was essentially a software engineer, not Cursor. So we applied this uniformly. Different people will do different things. Usually a tool is a lower risk but much smaller opportunity that utilizes the technology much less in most areas. Now there are so many exciting possibilities, so it depends on what you're trying to do. I always say, go back to your objectives, be clear about what your objectives are. I just want to work with friends: I love that idea, by the way, and I know lots of people who just do that. Don't let others convince you that you should be building the biggest company or something else. But it's perfectly okay to say, I want to build the biggest company, I want to be the richest person, I want to be the next Larry Page. But Larry Page had very much a mission orientation when he started. We invested in him very, very early. I was just talking to somebody from Google today, and I said I spent time with Larry when he was in Margaret Jacks Hall at Stanford University as a PhD student. He had very clear objectives, and financial wasn't his biggest objective. Anyway, sorry.

Earlier we were also discussing how a lot of investors tell you what you want, a lot of investors are good to you, and then the question is whose opinion should you take? Which person should you listen to?

My favorite topics. So the single most important decision any entrepreneur makes is whose opinion to trust on what topic. And on different topics, you want to trust the opinion of different people. Don't ask your friends. They'll have an opinion. The question you should ask is, are they qualified to have an opinion? Just because somebody is in a position of authority, they're a vice president of marketing at Cisco or IBM, doesn't qualify them to advise an entrepreneur in a new market. If you're doing incremental things, maybe they'll have some good ideas. If you're trying to introduce radically new things, their opinions are more likely to be wrong than right. If you Google it, most investors add no value and 70% of investors add negative value. I really, really believe that. Why? Because they haven't built a company. Just because you get a job at a VC firm or get an MBA doesn't qualify you to advise an entrepreneur. You have very little empathy for how hard entrepreneurship is. All of you thinking about entrepreneurship, be prepared. The press makes it very sexy, very glorious. It is hard. You have lonely, down days all the time, but you have to persist through that. So whose opinion you trust becomes a very nuanced thing. You really have to evaluate first, are they qualified to have an opinion on your topic? What have they done? Even the junior partners at Khosla Ventures, I tell them all the time, you have not earned the right to advise. The other part, and this is really bad for entrepreneurs, everybody wants great references from their entrepreneurs. That's a bad thing because then they're nice to you. I always say a good VC should treat a great entrepreneur like they treat their children. You don't always say yes. You don't always tell them they're doing great. You tell them what is best for them in the long term. You give them the good news and the bad news. So when we started Khosla Ventures in 2004, I put a phrase there saying, we prefer brutal honesty to hypocritical politeness. I'd just gone to an entrepreneur who had a lot of money in the bank. None of the VCs would tell him the truth, that they were doing the wrong thing. This team of 50 people, when they asked the other VCs, got polite thanks. So this team of 50 people wasted three and a half, four years, took a lot of money, and eventually sold for $3 million. Why? Because nobody was honest with them. So I decided I'd never be hypocritically polite, which is what 90% of people do, even if they know what the right advice to give you is. I'll give you my favorite example. I've done this a few times. If one of the people is the wrong person on the team, I will tell you before we invest, because you're good at one thing that you know really well, and mostly you're in learning mode. So you don't know how to judge people. You want people who will be really honest, as if they're talking to their children. I use that analogy because you always want what's best for your children. But the right answer isn't always yes or you're doing great. Sometimes it's think about this hard, you may be on the wrong path, or I don't agree with your strategy. I don't say it so politely because entrepreneurs have a lot of confirmation bias. Whatever you say, they hear it as confirming their opinion. I was just talking to Sam Altman. He was here earlier today, and this week we're having dinner. I said, Sam, your hardest problem is nobody's going to be brutally honest with you, to tell you you're screwed up in X or Y. I literally told him that, and he said, yeah, nobody says that to me, where am I screwing up? There's a great book by a woman called Margaret Heffernan. I said to him, you should invite her to come speak to your senior management team. She's given a TED talk, the easy way to look at it, called Dare to Disagree. She talks about where people don't disagree even if they know something's wrong. And very often you don't know something's wrong, but you have a hunch. So you're not even sure of your own opinion, but if you don't bring it up in a group meeting, you can't discuss it. There may be three other people in the room worried about the same thing. Very often when startups go wrong, people say, I thought that two or three years ago, and somebody else will say, I thought that too, but they never surfaced it for discussion. Discussion is always a good thing. Disagreement is always a good thing. As Jeff Bezos says, you want to have that discussion, then you want to have the team agree and commit or disagree and commit, but you want one strategy, and that's what the founder or the CEO does. So be very, very careful. Margaret Heffernan has written a number of books along these theses. She even talks about child molestation among priests: why did nobody speak up? Everybody knew it was wrong. There was just this reluctance to be disagreeable or cause chaos. Lots of big issues have happened this way, in war decisions, in lots of big and small decisions. I would suggest you read Margaret Heffernan's book on the right culture for a company. I have very long answers to everything.

I love all of the answers. Speaking of OpenAI, you mentioned that you saw a very interesting data point before investing in OpenAI, which was that a lot of young people and a lot of researchers were ending up in AI. Is this the secret sauce? Is this the way that you judge most of the investments? Or what's the best way to know the right data point to predict the future?

Every situation is a little bit different, and it also depends on the kind of startup you're building. If you're building an insurtech startup, then it's really market-oriented, not technology-oriented mostly. In the case of OpenAI, the most important factor was, when I went to MIT or Stanford to speak, what were people interested in? I realized the best talent was going into AI. Talent's a pretty good thing, because that says research will progress well. Then you look at the rate of progress. There are two ways to look at it. I remember giving a talk to the National Bureau of Economic Research on the impact of AI. I'm not an economist, but they invited me to be the lunchtime speaker. I said, here are the curves of progress. If you look at the curve, it's at a silly level compared to human performance. Human performance is up here, the curve is down here. Take a medical example I remember vividly: somebody joked that an AI asked an 80-year-old man if they were pregnant. Yes, AI was doing silly things, but you look at the exponential of the curve. You'd say, well, it's only a matter of a couple of years before performance catches up to human performance. So do you look at the data or do you look at what everybody believes are accidental examples? No technology is perfect when it starts, and you can't let that distract you. You have to look at the fundamentals. The same was true in 1996 when I invested in Juniper. It's one of the largest returns in venture capital: three million dollars turned into seven and a half billion dollars, 2,500X. That's pretty nice. But nobody believed it, because the telcos had all decided TCP/IP would not be the protocol of the internet. Think about it, in 1996 Cisco had said they would adopt ATM as the protocol, because the AT&Ts, the telcos, were asking for ATM, which used to stand for asynchronous transfer mode, a particular way where every packet is engineered, not randomly routed. But you look at the data, and TCP/IP was growing exponentially. It didn't care what the telcos were saying. I said this is going to win, and so we bet on that. That's why we got such a huge return, and that's why the world is in the right place. ATM would never have scaled. It would have scaled for a year or two. AT&T had a view that ISDN, or 64 kilobits, was all any home would ever need. They actually said that in comments to the FCC in their filings. They thought HDTV would be 480 VGA resolution. Lots of silly things. So you can look at today and look at the incremental thing, or say, let me see where technology can take us, which is much more exciting, much more fun, and results in much larger companies.

I asked Virgilio from Sword Health, whose company you invested in very early, for a question for you. His question was: how much does your contrarian mindset come from your upbringing in India versus your understanding that it produces iterative results?

I don't know. India's a pretty traditional place. Even when I was growing up, risk-taking wasn't encouraged. I just met earlier this afternoon with the President of Korea, and he was talking about what it takes to encourage innovation in Korea. I said, encourage risk-taking. Make it okay, not embarrassing, to fail. There's a Harvard Business School case where the first line is a quote from me saying, my willingness to fail allows me to succeed. Otherwise, you're not going to take risks, which means you'll do what everybody else has done because it's established. So it's almost like a self-fulfilling prophecy, and each of you has to decide how much risk and how much innovation you want. They're correlated. If you want a really big company, you want to take more risk and be more ambitious. Your probability of success goes down, but your consequences of success go up dramatically. Most people in business, especially VCs in a country like India, want to avoid risk. So they're okay with a higher probability of success and lower consequences of success, which I personally find unexciting. Different things for different folks.

Look, this was a healthcare entrepreneur, one of the most successful digital health companies. He was a PhD out of a university in Portugal, trying to do this thing, and he had been turned down for nine months because everybody was judging him on his then-current state, not on where he could go. I looked only at his learning curve. How fast was he learning, not where he was? I said, you'll get to your goals. He'd been looking for money for nine months, mostly among European investors. European investors have a great habit, and I won't insult them too much, of turning every great idea into a decent one with lower risk. Any Europeans here? It's really true. They take every large, big, ambitious idea and turn it into a lower-risk, lower-consequence idea, still successful, but that's not what entrepreneurs should mostly be doing. So I met with him and literally took one meeting to say we will invest. He was surprised, because he'd maybe talked to a hundred people over the previous nine months, and now it's one of the most successful digital health companies. He's just a great entrepreneur.

I think it's a key message for every European here.

I heard you many times saying that you use ChatGPT to ask questions about medicine, and if they disagree with a doctor, you change the doctor.

So my view is, there's a massive difference in healthcare between the practice of medicine. I don't want to criticize medicine, but if you look at the 1900s, it was decent, and then 1910, 20, 30, 2000, 2010, every decade it has gotten better. So it has improved a lot. But the practice of medicine is limited by what a human can know. If you're a woman and you get breast cancer, there's zero chance your oncologist remembers the last 5,000 articles on breast cancer that were published, usually in the last year or two. The volume of literature is so high, no human being can keep up. Now if you have that and something else, your medical treatment is very complex. Most people don't realize that, for example, with the doctors we have in this country, which is a pretty good level, probably one of the best levels of healthcare in the world, we have more deaths from drug interactions prescribed to you by doctors than from breast cancer. Think about it: more deaths from that. That shouldn't be the case in the age of AI. So what do I do? Almost everything I diagnose on ChatGPT, and my son has a company in AI medicine. ChatGPT still hallucinates, but the doctors don't keep track of everything. So they do the practice of medicine, which is what they've done before. I broke my wrist about eight weeks ago. I broke four bones here, so there's a big metal plate in there, and I told my doctor I not only want to fix the wrist, that was surgery, I want to increase my bone density. He gave me the usual answer, but not the best answer. And I said to him, why not this? He went back and researched it, probably on ChatGPT, because I had researched it. I said, I prefer that. And he said, okay, I'll prescribe that. Then I said, I have very good concierge medicine, and I did some more research, and there's this drug for women, only approved for women, to increase bone density in women. I said, why shouldn't I be taking that? No studies have been done. What is the mechanism of action? If you think from first principles, it should work in men too. And so I'm going onto that drug, and I said, I'm not risking another fracture. By the way, if you have a hip fracture at my age, your life expectancy is five years, and I had a hip fracture two and a half years ago. It doesn't apply to me because of all the things they don't tell you, including not telling you what your mortality rate expectation is over the next five years. So you really have to craft from first principles the best treatment you can come up with, and the AI tools are very good at that. I could give you 100 examples of that. I'll give you my favorite example. I'm probably not supposed to be talking about this, but what the hell. One of the best known names in traditional medical practice is Zeke Emanuel. Anybody in healthcare has heard his name. He designed Obamacare. He's at the University of Pennsylvania. Almost all the op-eds in the New York Times on health, they call him. And he's been telling me since 2012, when I wrote a blog in TechCrunch called Do We Need Doctors? That was 2012, long before AI was a thing. It was obvious to me it was the only way to scale medicine. I happened to have broken my knee skiing, torn an ACL. I went to Park City, which is the home of the U.S. ski team, so they have the best people. I took an MRI to three different surgeons. Each one recommended something different off the same MRI. And I said, I'm getting opinions, not facts. So between that and thinking about scaling medicine, I said, I don't want opinions. I don't want treatment based on the doctor I get. I want treatment based on the disease I have, or the ACL tear I have. And I decided AI would be the only way. I did a lot of research. In 2016 I wrote a 100-page document called 20% Doctor Included. That was again before ChatGPT. It's 100 pages: a third is about what's wrong with medicine, a third about what it can be, and a third about the process of getting from A to B. It's 110 pages that I wrote myself. But what was clear to me was AI is the only way to scale medicine to be the highest quality, and it's the only way to be objective, because humans have biases. Back to the story: Zeke has been telling me for the last dozen years why I'm completely screwed up and don't understand medicine. And then last September he came to me and said, you may not be screwed up, you may be right. So him, me, and my son are publishing an article. I won't give you the source where we are publishing. It effectively says it's over for human doctors. AI is just going to be better. Being able to convince somebody of that esteem, probably the most influential doctor in the country, I hope this publication comes out. It's in a medical journal, and they were pretty hesitant to publish something this radical, but it will get published. I think it's clear the time is now, and a radical change is possible on almost everything. It's the best way to do primary care. It's the best way to do chronic disease management. We have an AI oncologist in our portfolio. We have AI mental health therapists. I can talk to you about the caution in medicine and AI if you have time. But all of these things are now possible. There's an AI physical therapist, mental health therapist, primary care, chronic disease management, nursing care, you name it. These are all clinical functions that are entirely possible with AI, and if you give them to a human, they do worse. They reduce the efficacy of the AI. So one of the best people in quality of care in traditional medicine is a guy at Stanford, Professor Arnie Milstein. How many people have heard of him? Probably the best known name in quality of healthcare in the country.

He just did a study. For complex disease diagnosis, they did a multi-center study. And for this class of diagnosis, human accuracy was 73 percent. Think about it, for complex diseases, 27 percent of the people got a sub-optimal or wrong diagnosis. They used AI and on the same data set, it had 88 percent accuracy. Then they gave the AI to the human doctors, and they improved from 73 to 76 percent, but degraded the AI from 88 percent because they introduced their biases. Oh, I saw a patient recently, so I have a recency bias. This happened. Every single time the New York Times publishes an article on ADHD, the following couple of weeks, the number of diagnoses of ADHD dramatically go up because their mind is seeded. These are all facts and people are not willing to accept how good AI is. I'll give you my favorite example. We have a company called Limbic Mental Health. And I'll come back to the safety, it's worth talking about the safety. They do mental health therapy. Their diagnosis accuracy for talk-based therapy is 93 percent in a study of 100,000 people that's never been done in mental health. No human reaches that level of accuracy. But their AI can do diagnosis and it can do therapy. So mental health should be a gone problem. It is infinitely scalable at a dollar or two an hour. Yet we complain about not having enough mental health therapists. So they just did a study published in Nature Medicine, which is a very rigorous journal. Their AI did sessions, human therapists did sessions with patients. Then in a blind test, human therapists judged both sessions, not knowing which was which. They judged the AI's best therapy sessions as done by a top 10 percentile human therapist. So the AI therapy sessions were so good, they were mostly better than 90th percentile of humans. That's the data. Let me take a little bit of a diversion and give you a caution. We invest in OpenAI, I'm a big fan of OpenAI. If you use OpenAI to do triage in medicine, you get something like a 30 percent plus triage error rate. One thing you can't do, especially an AI system can't do: humans can make errors because they have insurance. AI systems can't make errors. So you get this error rate. But if you take my son's company, Curai, they use GPT-5 also, 5.6 now, but the triage rate goes to zero because they've built the whole harnessing system for safety, triage, verification, and a lot of other stuff. So you can't use these systems off the bat. Limbic Health adds a layer of computational psychiatry which guides the conversation. So it's not just a chatbot, it's guided chat, and the result is producing a set of findings the way a human therapist would. Here are my findings and here's my diagnosis, and that's the computational psychiatry layer that any therapist would recognize and is trained on. So all of these AI systems in mission-critical applications can't be allowed to hallucinate. They need a lot more harnesses and safety and triage systems. And sometimes when AI can't determine something, I always say to help AI people, triage it to a human. A human won't do any better, but it'll be accepted much more as an error from a human than from an AI.

For my last question before we do a couple of questions as a Q&A: you spoke about the idea of the Tesla in healthcare, maybe the trillion dollar healthcare company. What's your prediction of the company that becomes that? What do they do? What do they own?

My bet, look, there's a lot of things in administrative health care, lots of valuable things to do. By the way, for every physician, there are 23 support people in health care. It's a ridiculously high burden rate. So whether it's revenue cycle management or scribing or EMRs, there's lots of overhead and lots to be done to make a physician more efficient, to spend more time with patients. I think the right answer is globally make mental health almost free, because the AI does the therapy. Globally make primary care free. Globally make chronic disease free. Make gastroenterology or neurology or endocrinology near-free. All these things should cost $2 an hour, not $400 an hour. I think that'll be a breakthrough company. There will be more than one breakthrough company. Health care has so many big niches. In the US alone, it's a four trillion dollar spend. A trillion dollars of it is expertise, and expertise should be free. In an AI world, especially in two or three years, especially with all the safety and other systems added on, there should be no question. Humans will have roles. I don't think in the next two or three years we'll do heart surgery with robotic AI. That'll take the FDA an additional 10 years to approve. Regulatory is a big problem. They're doing a pretty good job recently, mostly because they're trying to compete with China in approving faster, so it's great for those of you trying to do regulatory approvals. There's a huge recognition in Washington that if we keep it the way it is, we will lose to China. About a fourth of health care is expertise of the four trillion. A fourth roughly is diagnostics and testing and imaging and those kinds of things. All of them are extremely amenable to AI. Drug design is another fourth: drugs and treatment therapeutics. That's very amenable to AI, and I'll come back to that. And a fourth is in-hospital spend, critical care, all that, where there's a smaller role for AI for lots of reasons. So a couple of cautions. The biggest problem in health care is not having the best technology or the best AI. It's getting acceptance. There is no doctor who will use AI if they are in fee-for-service. Every time AI does a session, some doctor loses their fees. So the AMA will not talk to me about AI doctors. Utah has some regulatory discretion. They're doing some interesting things. The American Medical Association is opposing the health department in Utah from allowing AI to practice medicine. So the AMA is totally against it because it reduces the income of doctors. Now, the right way to do it is in capitated systems. For those of you familiar with capitated systems, Medicaid Advantage is a great system because you get paid for taking care of a patient, not for the individual services you provide. So look to capitated care, where a person's income is improved, not hurt, if AI does more of the sessions. That's one general piece of advice. I wrote a piece on it, I think in early 2025, on AI in health care. I talk extensively about how the billing systems and methods, which are incentives, change what technology gets accepted or not, not how good the technology is or how good your service is. So pay a lot of attention to that. It's perverse, but unfortunately the truth. The other thing is most doctors don't believe this. So if you're building tools for doctors, they'll use it. If they are scribing or other things, we invest in a bridge. It's very good. We invest in certain revenue cycle management companies. But frankly, the administrative AI is less interesting to me personally. I want to change medicine. I want to provide an infinite amount of clinical services at no cost. That's the big enchilada to go after. That's super exciting. Let me go to drugs. How many people know GitLab? Almost everybody. We were investors in GitLab, we were the first investors. Sid had a great team, great founder, the company went public with billions of dollars, and then he got osteosarcoma as a public company CEO. When he talked to doctors, they literally said, you don't have a lot of time and there aren't a lot of options for this. So he decided he'd be his own doctor, sort of like me. Five years later, he's not only surviving, he gave a talk that's on our website at our CEO summit recently called Going Founder Mode on Cancer. He's designed his own personal cancer vaccines. He's taken unconventional approaches like using three treatments at the same time. No doctor would let him. They would say, you don't know which one will work. And he said, I don't care. I just want one of them to work. I'll take 10 at the same time if I have to. So he took a very non-conventional but very first-principles approach to a very serious disease which his life depended on, and he mostly just ignored doctors. This talk is worth listening to. But we are now getting to the point with the help of AI where you can treat one patient. So instead of going through a 10-year regulatory process to get a drug approved, and there will be many ways to do that, but the regulatory part will still be the expensive part, you'd spend a billion or more getting approval, I'm more interested in N equal to one medicine. If your patient population is one patient and AI can design a drug and dose them safely, the FDA can't ask you to run a trial. So we're trying a new approach of what we call N equal to one. Founders always find a way to go around obstacles. That's why founders are founders, and no big company will do this. So the encouraging news is there are many, many ways around all the large obstacles, especially in health care AI. Let me stop talking.

I think we have time for two questions max, guys. So do you want to choose?

I'll let you choose.

Jill, near you.

You mentioned how AI can be used to provide customized solutions for each patient, but a big ask for many people, especially with AI, is trust. For example, I sometimes don't trust the actual AI customer support rep. So why would I trust an AI, or how do you build that trust and relationship with an AI for them to take life-changing advice from an AI?

Look, humans will always distrust new stuff. Let me go back to an AI example. People don't trust Waymo. If every human driver in the US drove as well as a Waymo, you'd drop the number of deaths by two-thirds. There are like 45,000 deaths in car accidents. We'd save 30,000 lives a year if we forced human drivers off the road and said everybody has to take a Waymo. That's also, by the way, about the same number of lives breast cancer causes. It's like 35,000 to 40,000 deaths caused by breast cancer. But what's happened is, in the beginning, nobody would even give Waymo insurance. Now it's relatively easy to get insurance for a self-driving car company. Why? Because the insurers look at the numbers and say they are safer. In fact, I read an article today that in San Francisco, Waymos have far fewer accidents per mile driven than humans. So some people will never trust it. Some people will wait for others to trust it. You'll go up the trust curve, and there will be people who will always experiment. So it's going to be a progression of trust in AI, but there's no question that over time it will become more trustworthy because it will become well understood. In other areas, you don't have to trust, you can verify. In a drug trial or a therapy trial, the outcome you measure is not whether the drug was designed by an AI or not. I think it was in 2015 or 2016, I said, if any radiologist is practicing without using AI, they're killing people with diagnostic errors every day. And I got such a backlash. But the fact is, I think two years ago, the Journal of the American Medical Association said something to the tune of, they're no longer accepting publications that say AI is better than a human. We just accept it now. Practicing physicians haven't accepted it, but our job is to get that through. Part of it is they don't get fees if AI does the job. Out of that MRI scanner should come a diagnosis, not an image. The image should be the backup. Then the patient doesn't have to wait three days for the radiologist to read it, get the report to their primary care physician, who gets it to the patient. But it's a slow acceptance period. People have started accepting ChatGPT. It's grown faster than TikTok. That tells me something. When you deliver utility, the rate of adoption is never vertical, but very, very fast.

Let's do the last one.

Let's come over here, since the corridor is easier, and then we'll go across to you. So we'll do three questions. Cut there in the corridor. Sorry.

Thank you very much for the wonderful discussion. I'm building an agentic reasoning platform for critical care out of Seattle Children's. For me, I feel that the elephant in the room in healthcare is Epic systems, the key EHR vendor. I'm wondering what your thinking is about what the future of EHRs looks like, how and if at all it's going to be disrupted, and what are the most plausible ways that's going to happen.

EHRs are a funny story. How many people have heard of a programming language called MUMPS? Is there a hand up? I'm surprised. In 1976, when I came to this country as a 21-year-old, I decided I needed a job because I had no money, so I told a hospital system in Monroeville, Pennsylvania, outside of Carnegie Mellon, that I could program a health record system for them. I'd never actually done any system before, and I programmed it in MUMPS on specialized hardware called HBO, not the HBO you've heard of. MUMPS was invented at Mass General. Most of Epic is still programmed in MUMPS today in 2026. That's how bad it is. Now, it is so embedded in hospitals and systems that nobody wants to move it, so it's very hard to change systems of record. But for those of you from computer science, I do think the system of record can be different than the working set. In my view, think of it as a fast working set that writes through and reads from the system of record, but you never really use the system of record directly. I thought that's what would happen in EHRs. But even better, an agentic system doesn't have to. Whether you're using APIs, which are very hard to do because Epic will control what you can do with them, but with computer use models they can't do much about it. I think agentic systems would be the way. Essentially the role of the EHR gets relegated to the background, which is where it should be. It's really good for billing. Most of the data is set up for billing, not for learning. So we'll go to the last question here. Sorry, do you want to pass the mic to save ourselves a few minutes.

From reading your essays, I feel like you're one of the investors who really cares about making the world a better place. It's not just a talking point for you like maybe it is for some other investors. So I was really curious, what's your advice for people who are trying to do a startup that doesn't just make an impact but really helps people in need, like helping solve poverty or global hunger, things that venture backed startups don't prioritize?

I love people who care about that. But you have to find an investor who's compatible. Most investors are and have to be in the business of IRR because that's what they owe their LPs. They're investing their LPs' money and they have a legal obligation, a fiduciary duty to deliver returns. They also care about that because their compensation comes from delivering returns. At Khosla Ventures, we define it, and we tell our LPs, impact is as important to me as ambition, which I call IRR. So we are very clear with our LPs. I often will tell them if I had to pick between IRR and impact, I'll pick impact. As long as I can deliver them a rate of return that crosses the threshold they care about, I meet my duty, but then I focus a lot on impact. Otherwise, I wouldn't have done Opening Act. It made no sense. It was a non-profit. They had no product plan. They had no timelines. That same year, I did fusion, Commonwealth Fusion. And the Department of Energy told me I shouldn't be wasting my money because it's not going to happen in the next 30 or 50 years. And I said fusion won't happen till somebody does it. In my mind was a very vivid example. I talked to all the automotive CEOs about electric cars and they said not anytime soon. The Department of Energy, which told me not to work on fusion, had issued a report in 2011 on the number of electric cars in the U.S. in 2035, a 25-year forecast, which they like to do. And Elon Musk exceeded that number in 2016. So what happened here? Academically, if you talk to General Motors or Volkswagen or Ford, they do business as usual incrementally, just do some nominal projection and say there'll be so many cars in 2035, and it's so far away none of the executives will be around to be accountable for their view. Elon did something different. He said, I'm going to make it happen. And this is why entrepreneurs are so important to society. These things don't happen on their own. Entrepreneurs make them happen. Fusion only happened because I met this guy Bob Mumgaard, who was a senior fellow at the MIT Plasma Fusion Lab, and I said let's just do it even if everybody's saying it can't be done. Entrepreneurs make these things happen. Same here, in 2018, I also invested in public transit and everybody looked at me and said, public transit for a startup? How silly is that? I said, now we have a much superior way. I'd actually written about it in 2016, how to design a public transit system that would be far superior to owning a car. It'd be cheaper, faster, and more accessible to everybody. And I started with one simple first principle. Every street has a given width. You're not breaking down buildings on both sides of the street. So public transit only has to care about one thing, which is passengers per hour per meter of lane width. Essentially, what can you fit in a bicycle lane width? Turns out there's a way to do it that's better than bicycles, better than scooters, better than cars, better than big public transit. It's self-driving public transit with small vehicles. You can equal the capacity of a light rail system, if you do the simulations, in a bicycle lane. So you go from first principles, but an entrepreneur has to decide to do it. Other VCs don't want to invest in it. Even though we've now bid on five contracts, never invited to bid a single time, we won every single contract that has been decided so far in a public transit system from a startup because it's so compelling. So I'm saying, for impact oriented entrepreneurs, you can make it happen. People will not believe it. It's still hard to get funding for public transit in the venture community, but we have very good partners, and so you can make these things happen. The world depends on you and entrepreneurs making these large changes happen, so go at it. Thank you.

Thank you so much.

Let me just say I do have to run, but feel free to email at ovk at khoslaventures.com and your email will be answered. Thank you.