Latest / StartUp Health NOW Podcast / Vinod Khosla on AI Interns, Trust, and Health Moonshots in the Age of Superintelligence

Transcript

Vinod Khosla on AI Interns, Trust, and Health Moonshots in the Age of Superintelligence · StartUp Health NOW Podcast

• 0:01
  [on-hold music] Welcome back to Startup Health Now. My name is Unity Stokes, co-founder at Startup Health.
• 0:11
  This is the podcast where we get to have deep and meaningful conversations so we can get to know the change makers, the entrepreneurs and innovators who are transforming health. Stick around.
• 0:23
  It's going to be another great show. [upbeat music] I want to pass the microphone back to Unity Stokes.
• 0:30
  Before I do, though, I want everyone to acknowledge one thing, which is fifteen years ago when Uni and I were first here, when he talked about meeting Margaret Laws at a Mexican restaurant to tell her about what we were working on.
• 0:43
  Um, Unity would describe how he wanted to bring people together at JPM, and it started out with Startup Health Cafe. If anybody was at Startup Health Cafe before it was even the festival.
• 0:55
  The Startup Health Festival then came and took over Market Street for many years.
• 1:00
  And today we're at the Apollo House, and I can tell you it would not happen without the leadership and the guidance of Unity and the dreams of Unity Stokes. I just want to say thank you for [clapping]
• 1:11
  putting this together and truly leading the team to doing that. And with Unity, I am incredibly excited to j-- have Vinod Khosla. He needs no introduction, but what he does need is a quick moment of gratitude.
• 1:25
  For the last fifteen years, Vinod has been joining our community at every event we've done, almost systematic since two thousand and thirteen, maybe even two thousand and twelve, talking about the world we're living in today.
• 1:39
  If you go back, there are videos on YouTube at the Startup Health events talking about having a doctor in your pocket, talking about how it's going to impact the world.
• 1:47
  So couldn't be more excited to introduce Unity and Vinod Khosa, our fireside chat. Thank you. Thank you, Steve. Thank you. Good to have you. [clapping] Thanks. All right. Welcome, Vinod Khosla, to the Apollo House.
• 2:02
  Today, we're talking about health moonshots in the age of super intelligence. And I thought we would start with, I think it's the tweet that is pinned now on X for you.
• 2:14
  I'm just going to read it because I think it's really beautiful.
• 2:17
  "At seventy plus, I feel like I'm thirty, backing exciting innovators-- innovations, making the impossible possible, and from improbable ideas to world-changing, socially impactful companies.
• 2:31
  Never had such a fast learning rate in my life or felt I could have so much impact, which is exciting." Wow. At seventy. Okay. Now, that tweet was maybe a week ago. So do you-- Things move so fast.
• 2:45
  Do you still feel that way? How do you- It's still true. Is it still true? Even more in a way it is today than yesterday. Why do you feel this way? What's going on? What's going on is from time to time, platforms change.
• 3:02
  I started my first company in nineteen eighty when the microprocessor came along, and a whole bunch of stuff was built on that platform, including the PC and all that.
• 3:15
  In nineteen ninety-six, the Internet came along and a whole bunch of companies were built on that platform. In two thousand and seven, the iPhone came along and whole bunch of companies were built on that platform.
• 3:31
  If you look at the AI platform, it's broader, bigger, and more impactful than one can ever-- than one-- anybody could have imagined a little while ago. I, I was optimistic. You predicted this.
• 3:48
  You were talking about this years ago. You wrote papers on this years ago. Yeah. So it was twenty twelve, uh, when I wrote two blogs that were highly controversial then. One was called...
• 3:58
  They were in TechCrunch, so they're there to look. Is this the twenty percent doctor? No. Uh, so the first blog was called, Do We Need Doctors? Right. My thesis was, with AI, we don't need doctors.
• 4:11
  Uh, some people still doubt that a little bit, but most people are coming around to that point of view. The other one was, Do We Need Teachers? And that's also true. Uh, that was twenty twelve.
• 4:23
  Twenty sixteen, I wrote a hundred-page document called Do We Need, uh, twenty percent Doctor Included. And people are now coming around to the point of view that this is real. It sounded loony ten years ago.
• 4:38
  Even ten years ago, it sounded loony to be talking about this. It was a hundred percent clear to me. It was only a question of when, not how or, or if. It's very clear now.
• 4:53
  There's, there's no area of endeavor where humans will be better than AI. Well, well before twenty thirty-five, likely well before twenty-thirty. Right.
• 5:06
  That's a good place to- So we started- How many people here are physicians? Quite a few.
• 5:14
  I had a convers-- uh, recently, Ike Manuel, who most of you probably know, came up to me and said, "Ten years ago, when you were talking about this, I disagreed with you. I think I was wrong."
• 5:29
  And he said, "Oh, by the way, you were wrong, too, because you thought it'd take till twenty-forty to, to get here." So when somebody as
• 5:39
  conservative as him, as a denier, somebody who didn't acknowledge it was possible ten years ago, comes around to the view of it's happening now and it's happening really fast, it should be obvious to everybody.
• 5:53
  When we started Startup Health about fifteen years ago, it was a twenty-five-year plan. We're now fifteen years into that. And still, there's so many people that have no access to care.There's major challenges.
• 6:07
  We're not seeing the outcomes. It appears on one level that there's not enough progress that's been made. At the same time, if you ask me, I-I'm very confident that within another decade it will be possible.
• 6:24
  Wh-where do you sit to which-- It will be possible to achieve more of a global health where billions of people, you talk about someone in a village in India maybe getting better care than someone at Stanford.
• 6:36
  Do-- What do you think today? Where are we going over the next decade? That's the exact quote of my-- out of my twenty sixteen hundred page paper. That's right. I'll get better cardiac care in a village in India
• 6:47
  than I will at Stanford because Stanford will have humans and the village will have AI only. I totally believe that. So there was a recent study by Arnie Milstein's group.
• 6:59
  Arnie is at Stanford, probably one of the best-known names in quality of care in the US. And his group did a study that said for complex diagnosis, human accuracy was seventy-three percent,
• 7:16
  which was considered decent, which does mean, decent being twenty-seven percent of the people got the wrong diagnosis or suboptimal diagnosis. To me, that's shocking. A-and this is complex, so it's not your flu.
• 7:31
  You can diagnose flu wrong, and you'll be fine. AI was at eighty-eight percent. And then they gave the AI to the doctors, and the doctors improved.
• 7:46
  But from seventy-three percent accuracy to seventy-six percent. They degraded the AI from eighty-eight percent to seventy-six by introducing the biases. That is the nature of human involvement in complex stuff.
• 8:03
  That's an actual-- And it was a rigorous multi-center study. It wasn't just like a few questions here or there. So there's this friction between-- There's this system we have today, especially in the West.
• 8:16
  A lot of-- It might be easier to just start over, but we have the system we have. How do we confront that challenge, that friction, and merge these worlds? So let me come back to that in a second.
• 8:28
  Let me take a little diversion. We've heard-- Now, you'd say, "Hey, GPT 5 can do mental health therapy," and you hear about that.
• 8:40
  Then you hear all the stories of chatbots run amok, driving people to suicide. What's going on? My view is very simple.
• 8:55
  A chatbot isn't designed for clinical therapy with high acuity patients. If you use it for that, it's like driving a car without a steering wheel, and you're gonna run into trouble. Like a Waymo.
• 9:13
  [laughs] Waymo took, took fifteen years to-- I think it was fifteen years ago or fourteen years ago, I first rode in a Waymo. Wasn't called Waymo then. Yeah. And I've run-- ridden in every generation since.
• 9:31
  You need entrepreneurs like these to tailor this general AI capability into a specific solution that does the appropriate level of guardrails.
• 9:43
  If you're trying to generate an image for your company logo or something, you don't need guardrails. The general systems do pretty well. You're trying to prescribe something to a patient, that could be critical. And so
• 10:02
  I, I think entrepreneurs need to tailor this general AI for the specific purpose and do the right risk-reward trade-off. So I think that's what's going on. The capability is there. Used out of the box irresponsibly,
• 10:19
  it's a bad idea. So how many people read a paper in Nature just a few weeks ago called "The Fragility of GPT 5?" A few hands. It's an interesting paper.
• 10:34
  It essentially talks about if you use GPT 5, everything that goes wrong. And they were looking for things to go wrong, so it did go wrong. I penned with the
• 10:46
  person, AI person from one of our companies, penned a response to it, which I haven't had the chance to publish. I was being too busy. But everything in the paper is right.
• 11:00
  What's wrong about it is they use standard out-of-the-box system without the right fit-for-purpose controls, guardrails, approaches, post-training. I can talk to a lot of things.
• 11:16
  Where the optimism goes wrong is from people who don't know how to use it, using these systems poorly, and then declaring either they don't work or wow. The fact is the reality is messy, and you have to do this.
• 11:30
  So you were one of, if I'm not mistaken, one of the first venture investors in OpenAI. What did you see then, or what were you betting on then? What were you hoping for?
• 11:40
  And are we-- Has it surpassed your expectations or were you thought- Oh, it clearly has surpassed my expectations. Nobody expected, I didn't expect, they didn't represent that capability will develop this quickly.
• 12:00
  Why did I invest? There was lots of things wrong, including it was a nonprofit and there was no product plan, no plan to ever sell products or any of that.I invested because the consequences of success were so high,
• 12:19
  and that has borne out right as a thesis. Wow. Yeah. So my view was whenever it happens, whether it happens in five, ten or fifteen years, the consequences will be large enough that we'll figure out how to get a return.
• 12:36
  And that's exactly what's played out. Nobody would-- It was three years ago that ChatGPT was released.
• 12:44
  Nobody would have predicted something could grow five, six, seven times faster than TikTok did to a hundred million users. Now approaching a billion users. That's a pretty damn large part of the population.
• 12:59
  How do we channel this
• 13:02
  with this human intention, with this-- to, to direct this extraordinary opportunity, the power of these platforms and technologies, and instead of racing towards whatever humanity's last exam or whatever they're competing on
• 13:18
  towards solving these health moonshots? How do we- So let's-- When I wrote my paper in twenty sixteen, it was long before ChatGPT or even transformers or any of that. I had one dream.
• 13:33
  Every person in India for less than a dollar a month, let's say ten dollars a year, could get twenty-four seven primary care, and if they wanted to, have hundreds of interactions with their AI physician.
• 13:51
  For a price point that'd be far below what even a poor country like India spends today on healthcare. Like small, tiny percentage of that. Now, that I think is totally achievable now.
• 14:09
  In a country like India, it's definitely achievable here in this country easily. Two things have to happen. Probably the most important to, to do this right is to introduce it the right way.
• 14:28
  In no matter which profession, I counted thirty-three professions in which our firm is building AI workers.
• 14:36
  AI primary care doctors, AI mental health therapists, AI oncologists, AI chip designers, AI accountants, AI sales. Like thirty-three professions today in our portfolio.
• 14:50
  I think we are addressing a big part of the workforce. So how should we treat this? The way to think about it is if a student graduates tomorrow from Harvard Medical School,
• 15:08
  joins your clinical practice, you don't let them run wild. You supervise them. So to think of this AI in a properly done AI system,
• 15:21
  where my son's company is building an AI primary care doctor because I think it's so important. And there's no question we can do as well as any human, including front-facing conversation with the patient,
• 15:36
  including how you minimize the number of questions to come to a diagnosis that's high in reliability. But you have to treat these as AI interns, like a fresh graduate out of, uh, a medical school.
• 15:51
  Then the physician, the clinician, whose license they can practice under, oversees all their diagnosis. And over the next five years, they are signing the prescription because AI can't sign a prescription order.
• 16:09
  They are supervising, but the AI is doing all the work and the patient interaction. So it's a collaboration. It's a partnership.
• 16:19
  I think it's a leash on the AI, where as it gets better and better and as you get more confidence, you let the leash out, like you do with your dog.
• 16:31
  So you're scratching at something that you-- as you're an optimist, I'm an optimist, but what should we really be concerned about? What-- Where are the risks?
• 16:41
  And the reason I'm asking, there's a lot of builders in the room. There's a lot of entrepreneurs. There's a lot of people that are moving really fast, hopefully moving fast. But what should we be concerned about?
• 16:52
  What should we be thinking about? So in healthcare, the most important commodity is trust. It doesn't matter how well the technology works. You have to generate trust in the technology.
• 17:05
  So this intern model is the right way. You're doing an AI intern, uh, pharmacist, an AI intern nurse practitioner, an AI intern, intern mental health therapist, and you supervise them.
• 17:24
  And I think that as trust is established, then you can let it operate more independently. Trying to get to complete autonomous operation in step one is the wrong thing to do.
• 17:43
  By the way, it's also wrong to let the AI assist a human. That-- Copilots is a really bad idea in most areas. It's what people think of. I just gave you a study that shows humans make the AI much worse.
• 18:01
  They should review the final results and approve them and approve the prescription. That is the right model.
• 18:11
  It does-- It's a worker that does the work.And three years ago, we went away in our shop from building copilots or assisting any humans. Humans are just flawed. Just
• 18:23
  let them supervise till we get enough confidence and till we prove the AI does as well or better, and then you can let the AI loose. So a few areas copilots work, but in general,
• 18:40
  I don't think that's the right model of AI. Because AI and humans operate at very different bandwidths. There's a mismatch. Different species. There's a lot of entrepreneurs. What do you tell your founders?
• 18:57
  What do you tell the companies that you're investing in? What wisdom can you share? And then I wanna turn the question over to you in a second. M-my answer is what I just said. First,
• 19:10
  most people, especially true in healthcare, are limited not by what AI can do, but what people think it can do. That's the big limitation. So it's like a lack-- We need a bigger imagination. Yeah.
• 19:23
  Or we need to leverage our imagination. Right. So it's up to us as founders to help develop the trust and proof.
• 19:36
  I can tell you, we've done a clinical study of mental-- of an AI doing mental health diagnosis.
• 19:44
  A hundred thousand patients, ninety-three percent accuracy in diagnosing the top ten disease-- the top therapy diseases or CBT diseases. Ninety-three percent accuracy.
• 19:57
  There isn't a human who can get even remotely close to that. But that's a hundred thousand person study, so you have confidence. Yeah. This chatbot won't make you commit suicide. Yeah. Trust. Essential to humans.
• 20:14
  Esther Dyson's here. I asked Esther if she would join our conversation, just ask a question.
• 20:21
  And Esther has a new book coming out, I just wanna mention, called Turning Limits, which is in this age of abundance, where are the limits? It's-- I can't wait to read it. But Esther, over to you.
• 20:32
  Is there a mic that we can- [speaking faintly] I'll repeat it if people can't hear. It's a pleasure to see you. Hey, great to see you again. Thank you for joining us. Yeah. It's-- So I wonder-
• 20:50
  I'll repeat the question. Yeah. It's-- It requires a little Abraham Verghese, no doubt. He's, uh, used to run medical school or something, and now he- He has a lot of bad ideas and bad habits. Okay.
• 21:05
  We were on a panel together about ten years ago. Yeah. We disagreed strongly, and almost every aspect, he's been wrong and I've been right. He's talking about Abraham Verghese. Yeah.
• 21:19
  Nonetheless, I'd like to quote something from his book. Which is probably wrong, but go ahead. [laughing] It's a wonderful book called Cutting for Stone about-- That's what a surgeon does, they cut for stone.
• 21:33
  And at some point, there's a class of students, and he asks them, "What treatment is best delivered through the ear?" Do you remember what the answer was? No. Okay. Words of comfort.
• 21:47
  And so my challenge here is that a doctor does so much more than diagnose. And AI can do lots of things, but the second kind of notion I want to bring into this conversation, did you see the food over there? Yeah.
• 22:05
  It was amazing. It was real food. It was fresh, it was vibrant, it was real.
• 22:10
  And I hate to say it, but at this conference, I've been eating disgusting, ultra-processed food served by people who claim to be interested in health. And in a sense, what we need
• 22:25
  is amazing AI doing diagnoses and then real people, not ultra-processed communication. Real love, real words of comfort.
• 22:34
  And so, yeah, let the AI do the diagnoses, let them do the robotic surgery, but what people need is much more communication. So let me- Yeah. -respond to that. [clapping] Let me respond to that.
• 22:50
  In twenty sixteen, I wrote my hundred-page document called Twenty Percent Doctor Included. It's still on our website. You can still find it. I said twenty percent doctor included
• 23:03
  because I thought doctors would do the human element of care. So there was a whole section on the human element of care in twenty sixteen. And that may still be true.
• 23:15
  If you tell somebody they have three months to live because of their particular cancer, uh, I s-suspect a human isn't the-- is-- a human is the best way to deliver that message. So clearly there's that role,
• 23:33
  but you don't have to be an oncologist to deliver that message. Five years, seven years ago, I had lunch with the dean of the Harvard Medical School, and I said to him, "You should stop s-selecting for IQ.
• 23:51
  It's so hard to get into Harvard Medical School. You should look for EQ, so why don't you go to UCLA Film School and see what criteria they use to maximize mirror neurons and empathy and all that?" Now,
• 24:07
  let me give you some counter indications to that. So I still believe that may be a large role to play.And I do think medical schools should now select for EQ. The IQ they select for,
• 24:23
  most people who get into Harvard or Stanford just don't have any EQ left in that equation. [laughs] The IQ-- EQ bal-imbalance is so large.
• 24:36
  But my son does this primary care thing, and the first thing they ask is a patient is, "Would you like to--" they text message back and forth. "Would you text-- like to text with the AI or opt out and talk to a human?"
• 24:54
  Three percent opt out and pick a human to talk to. Why? They feel much more comfortable asking stupid questions or
• 25:06
  aren't afraid to say, "I don't understand," so the AI can translate it to fifth-grade level, which is above the level at which most Americans understand medical conversation.
• 25:19
  Not general conversation, but medical conversation. So that's one data point. Uh, I talked about mental health therapy in these large studies. We've done six hundred thousand patients in the UK.
• 25:34
  We are approved as a medical device in the UK, both for diagnosis and for therapy. Most of those patients seem to prefer the AI over the humans because they don't feel judged,
• 25:48
  they don't have to lie about how many drinks they have or if they still smoke. And you can talk to them in the middle of the night. And you can talk to them at three AM.
• 25:57
  In mental health, most of the peak period starts after eleven PM and goes on till three AM. So this accessibility so far goes up. The other thing they found,
• 26:11
  diversity-- access for diverse communities dramatically increases.
• 26:17
  For really marginalized communities like transgender, it's a hundred and fifty percent more access to mental health therapy than if they were using their therapist. So I don't know. And then we gotta wrap up after this.
• 26:33
  Yeah, we'll wrap up. Just I-- everything you've said, I also agree with. I think we need more care that is not necessarily medical care so that people don't get to that point. Replace the doctor. Absolutely.
• 26:46
  Look, it's- Yeah. Let me say five. Including the food. Yeah. Between yoga and meditation and wellness to bipolar disorder, it's a continuum.
• 26:59
  We draw a line at some point on the PHQ-nine score and say, "You're now a manic depressive." It's an arbitrary line. It's an arbitrary test.
• 27:11
  So moving people towards wellness is the right way to think of men-mental health. Thank you. Thank you, Essan. Thank you, Vinod. Thank you for pushing boldness. We are so grateful for you sharing your wisdom and time.
• 27:24
  Thank you. [upbeat music] And finally, an invitation to all of the extraordinary health transformers.
• 27:34
  Whether you are a startup, someone building the future of health, or whether you are an investor, funder, or buyer of the solutions out there in the marketplace, this is an invitation to you to join the Startup Health member community.
• 27:49
  Learn more at startuphealth.com. Look forward to hopefully seeing you soon. [upbeat music]