Vinod Khosla on A.I., Health, and the Future of Working (or Not)

Xconomy San Francisco — 

Entrepreneur-turned-venture capitalist Vinod Khosla made big headlines almost six years ago when he wrote a blog post called “Do We Need Doctors or Algorithms?” In it, he said medicine needed to be reinvented and he predicted a new era in which artificial intelligence might replace most of the functions that doctors do now—and do it much better, leaving physicians free to concentrate on the human element of care.

It’s been quite a ride since then. Along the way, Khosla has invested in a range of startup companies—including several tackling radiology, cardiology, and mental health (see slide and list at bottom)—that are using data and artificial intelligence to reimagine healthcare, hopefully lowering costs, improving quality, and making the best care accessible to all. And almost exactly a year ago, in September 2016, he published a 110-page paper on the subject called “20-percent doctor included” & Dr. Algorithm: Speculations and musings of a technology optimist that spelled out his thoughts and speculations (not predictions) in far greater detail.

I recently visited with Khosla, one of the founders of Sun Microsystems, at the offices of Khosla Ventures in Menlo Park, CA, for a discussion on A.I. and healthcare and beyond, including a not totally optimistic picture of the future of work and jobs. “There’s no reason an oncologist should be a human being,” is one of the things Khosla told me. “There’s nothing that requires human judgment that machines don’t have a chance at doing much better,” and “People don’t need to work, for those who don’t want to” are a few others.

What follows is an edited transcript of our conversation.

Xconomy: I’m curious–what got you into A.I. in healthcare?

Vinod Khosla. Well, I’m always looking for where the large changes are, and where the large problems are. If you look at healthcare, we all know how big a problem it is—there’s no rocket science. Nobody had ever looked at unique, highly leveraged ways to change healthcare and how one would do healthcare if it started from scratch. Of course, it has to at some point fit back into the old healthcare system.

I originally looked at it to see if I could do some nonprofit efforts in India. And you couldn’t scale enough doctors. If you had unlimited budget, you couldn’t start enough medical schools and get enough professors to teach the number of students who in 10 years would [make] enough doctors. The math didn’t work.

Then in January 2012, I wrote a piece called Do We Need Doctors or Algorithms?

X: That got a lot of attention.

VK: That was in TechCrunch. I didn’t intend to write it. I was skiing for two weeks [in Deer Valley, near Park City, UT], and the day before Christmas, I tore my ACL skiing. It’s a bummer when you’re planning on skiing and you have to stay in bed. I did an MRI and I took it to three different docs, and they recommended three different things. I said, ‘This is stupid. There’s one right answer.’ And when I talked to them about probability, they didn’t understand probability. And these are really good docs. The U.S. ski team is based in Deer Valley so they’re the best docs, and they see probably hundreds, if not thousands, of patients every season. So I was in bed, I was dealing with different opinions, and I’d been thinking about the India problem—looking at scaling medicine. That’s when I wrote that blog. Came from my frustration with the inconsistency of advice.

And frankly, there’s two types of advice in medicine, one that doesn’t matter. If you’ve got the flu, it doesn’t matter what advice you get—you’ll get better in the same timeframe. You might feel a little better if they give you Advil, a little worse if they don’t. And then when it really matters, you get a lot of opinions, but no science.

There’s a number in medicine that almost no doctor knows, but it’s well-established. It’s called NNT. You probably never heard of it. It’s amazing. [Editor’s note: NNT is Number Needed to Treat to avoid or prevent one additional bad result.] That’s the real number that matters. NNT is an incredibly important number that few doctors are aware of.

Look, medicine is better than it has ever been, and every year it’s improving. But it’s still the practice of medicine. It’s not a science. If it was a science, for any given patient you’d always have the same answer no matter who you ask, even if it is a probability distribution of outcomes. So my goal became to change the practice of medicine, which is pretty damn good, into the science of medicine. And the science of medicine needs science—and it can take a good system and make it much better.

X: I can guess where artificial intelligence figures into this, but please take us through your thinking.

VK: We don’t even measure the stuff that doctors can’t directly understand. We’re starting to run into this a little bit because in genomics, you might get a thousand data points, and no doctor can look at a thousand data points. And so in the past we didn’t measure anything humans couldn’t consume, which meant they can at best look at a few numbers. We should have thousands of numbers per patient, per episode.

So my approach was to look at how we reinvent medicine if you want to try something radically better? It has to come with something else that is a big problem today—accessibility. You know medicine, especially in the U.S., has gone down a path of increasing inaccessibility. Hey, if I get cancer, touch wood I don’t, but I can get proton beam therapy. But the damn machine costs $200 million dollars to install. How many people can access that? And that’s the wrong direction to go in. I want this damn device to do most of my diagnosis, and it can’t.

The data science, if it’s computing-based, can be very low cost. Anything that’s computing-based is near zero cost eventually. Which is why we spend way more compute cycles on a two-cent ad on Google than $10,000 medical decisions like ‘Do you need a meniscus repair?’ It’s silly. It makes no sense to me. So I started just saying, ‘How would you do it better?’ And then I took it to a bunch of friends of mine who knew medicine. I got a lot of critique. And, basically, I took every critique as a way to collect questions I should answer. And over time, as I collected more and more questions and could answer them plausibly, I said ‘Hey, this is the way medicine needs to go and should be reinvented. It was and still is speculation, but at least it is possible.’

X: It seems from your writings that you believe outsiders—especially entrepreneurs—can do this better than the medical industry.

VK: The auto companies are never going to lead driverless cars, they’re never going to lead electric cars. There was a 2010 DOE [U.S. Department of Energy] report that predicted in the year 2035, and their number for 2035 was smaller than the number of electric cars Tesla shipped last year. Why? Because they talked to Volkswagen, GM, and Ford. The car companies were thinking incrementally, and Elon [Musk] said, ‘Hey, how would you do this from scratch?’ He had no auto expertise. I think the same can be done in medicine. It just takes a great entrepreneur. I’m not going to do it, but great entrepreneurs I want to back will do it. It does eventually have to fit back into the system, but radically innovative ideas will likely come from outside the industry.

X: Tell me what you think is the difference between the data science and the A.I., because the data science, presenting it to doctors so they can make a decision, that’s one thing, but that’s not A.I.

VK: Today, you do your test results, you’re given 30 numbers at most, and the doctor looks at your blood test and says your iron’s low, your NK [natural killer immune cells] are high, your mean corpuscular volume is low. It means a few things. He can remember a few simple relationships. But let’s say you’re dealing with something serious like colon cancer. Does he know what each of the thousands of mutations could be? No, he doesn’t. So when he’s sitting down with the patient and saying, ‘You’ve got this cancer mutation,’ does he remember the 5,000 papers published in oncology journals recently? No, he doesn’t, he can’t. He’d be inhuman if he did that.

So machines have to do that. Can [the doctor] do the human element of care? Yes. But to be honest, a nurse would do that better. Three years ago I told the dean of Harvard Medical School, if you believe this is the future of medicine, and he sort of said he did, then you should change your admissions criteria to not look like an IQ test—just use the admissions criteria the USC film school uses, because they’re looking for people who can put themselves in other people’s shoes, be empathetic, have lots of mirror neurons. I know you laugh, everybody laughs, but I actually think it’s serious. If you want humans to do the human element of care, pick the most emotive humans, the most humane humans.

X: So machines have to take the data and interpret it?

VK: Look, every person should have a personal physician they can consult on every single day. If it’s an A.I. bot, that’s 10 bucks a month, max. And they would save you more than a few bucks a month by saying, ‘Buy this generic brand of aspirin, not this fancy non-steroidal anti-inflammatory. For your situation, it’s going to be better.’ The pharma companies wouldn’t love it. The doctors may or may not love it. The notion is medicine can be reinvented. Everything that can be reinvented doesn’t always get reinvented, but could get reinvented if the right entrepreneurs go after it and they get lucky.

X: Well, do you see that happening?

VK: I think we are well on our way. It isn’t yet visible. Another story I’ll tell you is a friend of mine is Marc Tessier-Levine, who’s president of Stanford now. He was the chief scientific officer of Genentech. After that he became president of Rockefeller University. So we were just talking about it – and he said, ‘Why don’t you give a talk at Rockefeller?’ And he said, ‘But here’s what I want to do. You can talk for an hour, but then I want you challenged for an hour.’ So he took very traditional experts, the CEO of Memorial Sloan-Kettering, and the CEO of New York Presbyterian. I said, ‘Perfect. I love this.’ And by the end, we all sat there, and I said, ‘So what do you think?’ Their only comment, and I remember this vividly: ‘Well, it’ll take longer.’

I said, ‘Then it’s a win for me. I’ve gone from impossible to it will take longer.’

X: Longer is faster than never.

VK: Longer is much faster than never. If the right entrepreneurs take it up, then it goes from improbable to possible, with a little luck and sometimes more than one attempt. And we know from our technology world, this happens. It’s only a matter of time.

X: What do you think will be the most impactful short-term applications?

VK: Talk about accessibility. There’s a company in Israel called Zebra Medical Vision. They do radiology. And they do a lot of radiology. But here’s the fun part. In India today, they are offering reading any image for a dollar. Now that’s impactful. There aren’t enough radiologists in India. So when you get a scan in India—you may get a CT scan, an MRI, an ultrasound, an X-ray—you might wait for a week for the radiologist to read it and write a report. But with Zebra when you get the image, you have the analysis done almost immediately if you are connected. For a dollar. That’s exciting to me. And guess what, we will see better results than a radiologist and way faster. Higher quality, faster, and dirt cheap. The radiologist couldn’t do a phone call for a dollar. So what else do you need?

And you can talk to Ginger.io. They’re doing mental health that way. AliveCor is doing cardiology for basically 10 bucks a month. In traditional cardiology you might take one ECG a year but AliveCor’s average patient is doing 200 of them a year. They’re treating an ECG like a blood glucose reading for a diabetic. Why shouldn’t it be that way?  If I think I might die from cardiac disease, or have a heart attack any day, I’m going to monitor it every damn day. Just like diabetics do. And then you can’t have a human read it or it will get too expensive, so a machine has to say, ‘Hey, you have atrial fibrillation or not today.’ Even if I wanted an appointment, and I can afford the appointment, I’d call Stanford and they’d give me an appointment in two weeks. First I’d see a cardiologist. He’d prescribe the ECG. I’d then take a week to get that, and then I’d consult with the cardiologist. It’d be two months. It can be done much better with newer tools. I can do it at home. [Editor’s note: Khosla Ventures is an investor in Zebra, Ginger.io, and AliveCor.]

And I want to build a cancer oncologist. There’s no reason an oncologist should be a human being.  Look, the right kind of oncologist isn’t the research oncologist. They know the most. But the guys who know how to take care of a patient are the community oncologists in Fresno or Stockton. They cannot always read all these journals, but they care for patients. They actually know what somebody’s daughter or son is doing, and they have that connection. That’s what you need. They can be assisted with a virtual tumor board or an A.I. oncologist.

X: What about Watson Health? On the surface, isn’t Watson exactly what you’re talking about? Scouring data, finding correlations.

VK: Watson is not A.I., it’s statistics. Watson was designed to demonstrate the power of hardware. They had this supercomputer—Deep Blue I think—and they wrote Watson as basically a statistical NLP [natural language processing] package. Then they said, ‘OK, how do we prove it to the world? We can play Jeopardy.’ And they won Jeopardy. But you don’t need intelligence to win Jeopardy, you need statistical power. When machines beat humans at chess, that was computing power, raw computing power. But if you do, if you are a powerful enough computer, you can beat a human at chess.

Now you go from chess to Go, you do not have enough computing power to compute Go. Because chess only has 64 squares. Go has a lot more possibilities. They’re not computable. And then you say, can you actually make deductions? This is probably the best way to explain what statistics is, and what A.I. is. What you need in Go is “intuition” about patterns. It’s not really intuition, it’s whatever is called intuition because you can’t compute every possibility. So if you look at how DeepMind’s opponents describe what AlphaGo was doing, they said it had intuition. [Editor’s note: DeepMind is the British company acquired by Google that produced the AlphaGo neural net program that first beat a human Go champion last year.] How else did they describe it? Creative. In Go, if you are very good, you learn to control line four. It’s sort of like the role of the queen in chess. If you control line four, you get geopolitical influence over the whole board. That’s a well-known strategy. AlphaGo figured out how to do that with line five. No human had ever done that.

That’s the difference between A.I. and what used to be called machine learning, but really was statistical techniques. Watson could do A.I.. IBM has the scientific talent to do A.I., but they chose to package Watson and market it for what it wasn’t. I think they can do much more, but the early efforts have been less than successful because they over marketed its capabilities.

X: You say in your paper you are not making precise predictions, but rather pointing out directions.

VK: Right. When innovation is involved, you can’t predict. You can speculate. So I call everything I write speculation, not prediction.

X: That doesn’t mean everybody reads it that way.

VK: No. So this is why I pick my words carefully. Even if they’re misinterpreted because people can’t tell the difference, at least the really intelligent people read it correctly.

X: So talk about the longer-term directions.

VK: I would stick with the forecast I gave in that paper. The rough analogy I did—the first picture with the seven generations of cell phones. My first cell phone was mounted on the passenger side of my car, on the floor board. Weighed about 20 pounds. It literally looks like a sewing machine. And then there was the Motorola Brick [Editor’s note: its full name was the DynaTac 8000X.] And then there was the simpler phone, then the flip phone. And then came the smart screen. Then came the BlackBerry, and then the iPhone. So, technology innovation cycles, because they rely on technologists, are about two to three years. And if you imagine seven generations of iteration on medicine, it’s 20 years—give or take 10. It could be a little faster, it could be a little slower. That’s my view of how medicine will progress.

X: Are we still in the first generation then?

VK: I think so. What will version 7 of medicine look like? I can see today’s iPhone is nothing like my sewing machine phone, which was v0. Successive iteration is how large changes cumulate in technology. If you think of a 20-pound phone and compare it to today’s iPhone, that difference or delta is what v7 of medicine will have from today’s early, clumsy digital medicine.

X: What about the broader implications of A.I. beyond medicine?

VK: I wrote my blog on A.I. in medicine and other areas almost six years ago. At that point, it was more ‘this should happen.’ I didn’t know what Google was doing, and frankly Google didn’t have that emphasis on it six years ago. But it was clear that there was a range of possibilities. I think of it as new veins of gold to mine. And if you spent enough time looking, you’d find more and more veins. And that’s what happened the last six years. We discovered a lot of these veins, a few very promising ones.

So there’s this A.I. gold rush, and lots of components and technologies are being built. We’ll soon be able to combine those pieces in unusual ways that offer amazing capability. There’s nothing that requires human judgment that machines don’t have a chance at doing much better, other than where we can’t get enough training data. Now I suspect even when you have very sparse data, machines will do better than humans. But that remains to be seen.

X: What’s this mean for people and the future of work and jobs?

VK: I spoke recently at a National Bureau of Economic Research meeting on the economic implications of A.I. Every economist’s answer has been that education will improve employment—and if machines are coming, we should just get more educated. But I contend that may no longer be true. If machines get smarter, more knowledgeable, and better at judgment than humans, then education doesn’t help. So we have to fundamentally rethink our assumptions. When machines are smarter than humans, what do you do?

X: That sounds pretty bleak. You’ve said you are a technology optimist, though, so solve that one. How do we do that?

VK: I am a technology optimist. Well, people don’t need to work, for those who don’t want to. [Editor’s note: For more on Khosla’s view on the future of work, see this Forbes article.] I enjoy working, so I’ll still keep working. But the guy who’s the garbage man, he won’t do that. The garbage truck will drive around, pick up the garbage without the person. I was just reading this book, one of my favorite books recently, called Scale by Geoffrey West at the Santa Fe Institute. He worked at a brewery in England at age 15—he’s a physicist now. He said every few minutes he was supposed to pick up a crate and load it on a truck. He did it for 12 hours a day and he got paid about a shilling a day. Now, nobody needs to do that job.

Selected Khosla Ventures Healthcare and A.I. investments

Khosla Ventures has a more extensive list of investments in healthcare than is currently posted on its website. Below is a slide showcasing the firm’s current healthcare portfolio. Most of these companies, Khosla Ventures says, have an A.I. component.

And here is another way the firm classifies its healthcare investments:

New science

Inflammatix, Sema4, DarwinHealth

Leveraging data to build an advantage and enable upskilling

AliveCor, Bay Labs, Zebra Medical Vision, Atomwise

Building things that can touch consumers directly

Ginger.io, AliveCor, Forward, Oscar

Better diagnostics and sequencing—going from analog to digital and opening up new avenues

Genalyte, Apton, Color Guardant, Whole Biome, Atomwise