--:-- Olivia Webb: 2022 was a year where many of us got to experience firsthand how AI is transforming the creative spheres, from writing to image generation. But those are far from the only arenas where AI is advancing. In this episode, A16Z Bio and Health founding general partner Vijay Pande sits down with A16Z co-founder Marc Andreessen, who infamously wrote his essay, "Software is Eating the World," over a decade ago. In this conversation, Marc and Vijay discuss what AI can and cannot do today, but also the tension of where it's advancing and what might stop it. They also explore the rapidly advancing technology with a frame of what it can augment instead of replace, touching on the potential future of doctors, teachers, therapists, and more. I hope you enjoy this fascinating conversation.
--:-- Olivia Webb: Hello, and welcome to BioEats World, a podcast at the intersection of bio, healthcare, and tech. I'm Olivia Webb, the editorial lead for Bio and Health at A16Z. I'm very excited to share this episode because it features Bio and Health founding partner Vijay Pande in conversation with A16Z co-founder Marc Andreessen on the topic of expert AI. You can also check out a video version of this podcast on A16Z's YouTube channel, and we'll link it in the show notes. In this episode, Marc and Vijay have a lively discussion about the future of expert AI with regard to healthcare, but they also get into self-driving cars, screenplays, music, and the nature of consciousness itself. It's one of our longest episodes to date, but we could have gone longer. As Vijay and Marc discuss, AI has the potential to change many industries. So let's get started.
--:-- Vijay Pande: Hey, I'm Vijay Pande. I'm the founding general partner of the A16Z Bio and Health Fund.
--:-- Marc Andreessen: And I'm Marc Andreessen, co-founder of A16Z.
--:-- Vijay Pande: Marc, thank you so much for joining us.
--:-- Marc Andreessen: Yeah, it's great to be here.
--:-- Vijay Pande: You famously wrote about software eating the world. That was basically what, 10 plus years ago?
--:-- Marc Andreessen: Yeah.
--:-- Vijay Pande: And that very much seems to have come to fruition. If you look at all these other industries that software really wasn't a part of, software's actually become a dominant part. But this year's been kind of an amazing year for another type of software, for AI. I'm curious to talk about the arc of what we think is going to happen in the future based on what we've seen in the past, and really how this new technology is going to change everything, much like we've seen software change the last 10 years. I'm curious what you think for just like this year. It's been kind of an amazing year. We always seem like not much happens in any given year, but 2022 seems to have been an amazing year for AI.
--:-- Marc Andreessen: Vladimir Lenin once said, "There are decades in which nothing happens, and then there are weeks in which decades happen."
--:-- Marc Andreessen: Let's not hope that happens politically anymore, but it does happen in science and technology. There are moments where things kind of hit critical mass. This AI machine learning revolution seems like that's what's happening right now. It's been interesting to watch. It feels like there was a breakthrough moment in 2012 that had to do with images. Then there was a lot of work subsequently that led to things like the creation of self-driving cars based on that. Then there was some natural language breakthrough maybe three years ago. Now that's really catalyzed into this whole thing we see happening around GPT and text generation. Even other applications, transcription is getting much better, all of a sudden speech synthesis is getting much better. Now you've got this artistic revolution happening with image creation, and video creation is coming up really fast. It seems like one of those catalytic moments. Every week now, it seems like there are fundamental breakthroughs, research papers, product releases coming out. So it seems like a cascading thing.
--:-- Vijay Pande: The way I think about it as a software person, a lifelong programmer, is that in the fullness of time, it will appear that there were kind of two different ways to write software. There was the old way, the classic von Neumann machine, deterministic way. The whole problem with writing software in the old model is that computers are hyper literal.
--:-- Marc Andreessen: Yes.
--:-- Vijay Pande: They do exactly what I tell them. Unfortunately. Every time they do something wrong, it's because you have instructed them improperly. It's a very humbling experience to learn as a young programmer that everything is your fault. The machine will just sit and wait for you to fix the problem. It's not going to do that on its own. But then there's this other way to write software, having these AI systems and then having training data, training the systems, tweaking the systems. The way I describe it to normies is that it unlocks the ability for computers to interact with the real world, with the messiness of the real world, the probabilistic nature of the real world.
--:-- Marc Andreessen: It seems almost less like writing software, almost like training something.
--:-- Vijay Pande: That's right. When I think about machine learning and image recognition you talked about, it felt like it was almost like training a dog. Reinforcement learning is like we'll give treats as it gets better. But there's something different now. It feels like we've gone from training a dog to recognize a bird versus a hot dog, or hot dog or not hot dog, to actually something where it feels closer to training a person. When we talk about learning and training and data, what are we training? Where do you think we are in that arc of getting to eventually like HAL 9000?
--:-- Marc Andreessen: While this has been happening, and you have kids, I have a young child, a seven-year-old now. As this stuff has all been happening, I've been simultaneously training the seven-year-old.
--:-- Marc Andreessen: Anybody who's had kids will recognize what I'm about to say. It is really interesting watching little kids. The way I think about it, for the first few years, every single thing he did was like a little applied physics experiment. "Let's see what happens if I drop this. Let's see what happens if I eat this. Let's see what happens if I do this to daddy."
--:-- Marc Andreessen: See what the response is. They just run experiments. You can see it very clearly when they're learning how to walk because they're running all these experiments about how to stand up and what to hold on to, and they keep falling over. Then at some point, the little neural network actually figures it out. It does learn. Off and away they go.
--:-- Marc Andreessen: It's a little bit eerie. You can see that a similar kind of thing happening. Having said that, the human brain just keeps developing. It ultimately achieves higher levels of consciousness, higher levels of self-knowledge. It reaches the Descartes stage where it has self-awareness. It is clearly very creative from an early age. I'm a little less convinced that the software technologies we have now are on some linear path towards AGI, consciousness. It's hard for me to believe that consciousness is simply emergent from higher-scale neural networks. That seems like a hand wave. Now, having said that, I have a lot of smart friends who are pretty sure that's what's going to happen.
--:-- Vijay Pande: I feel that way as well. I want to get to AGI in a bit, and we can debate whether consciousness is an illusion. But where we are now is kind of amazing. People can take GPT-3, give it SAT exams. It does okay. Actually, it does quite well.
--:-- Marc Andreessen: Yeah.
--:-- Vijay Pande: It scores at like the 1,200?
--:-- Marc Andreessen: Something like that.
--:-- Vijay Pande: That's not bad. That would get you in a lot of schools. I actually gave it questions like explaining the derivation for the Schwarzschild radius, the black hole radius, to write a code for an 8x8 tic-tac-toe, random things that you should never be able to do because it's not just memorizing, it's generalizing. And it's getting that. But then it also seems to have some weird hiccups. One thing that really does not seem to get is humor. I'm curious where you think it's going to go. Before we get to AGI, there are things that an average human can do pretty well that GPT-3 can't. But then there's also what experts can do. I'm very curious if we may get to some of the expert stuff first before it can do even something like humor. The irony is that something like humor that we take for granted might actually be really hard, and other areas might be easier.
--:-- Marc Andreessen: The ultimate example of the things it can't do: it can't pack your suitcase.
--:-- Marc Andreessen: There's no robot that will pack your suitcase. If you try to get it to make an omelet, it will shred your clothes and shred your eggs. It can drive your car, but it can't pack your suitcase. It can't do your laundry. There are these interesting twists. I would describe it as follows: I think that this generation of AI, as impressive as it is, is a little bit of a sleight of hand. But I also think human consciousness or human intelligence is also a little bit of a sleight of hand.
--:-- Marc Andreessen: Maybe slightly different sleights of hand. The sleight of hand you see when you're using GPT or one of these image generation things is it's not literally creating new information. It has no opinion. It has no point of view. It's not sitting there thinking on its own, coming up with some new thing. What it's doing is training on the sum total of all existing human knowledge. For text generation, it's training on all existing human text. It plays back at you projections from the assembled composite text. You ask it to do the eight by eight, so probably somebody on the internet at some point wrote some paper.
--:-- Vijay Pande: I think it's a little more than that because I asked like 56 by 56 or 101 by 101. It has some sense of generalization.
--:-- Marc Andreessen: I'll bet if we Google long enough, we'll find a paper that described a general purpose algorithm for multi...
--:-- Vijay Pande: That may be.
--:-- Marc Andreessen: Yeah.
--:-- Vijay Pande: I've done the same thing. I only have it write Seinfeld scripts. Sometimes they're really funny, and sometimes the words make no sense.
--:-- Marc Andreessen: I went for Curb, but it's the same idea. There are a lot of jokes on the internet. You could kind of go back and say it probably plucked these jokes. Or maybe there was a paper somewhere where they articulated a general theory of humor. Humor's been studied as a thing, and maybe there's a general thing like humor's the unexpected or whatever, so it generalizes.
--:-- Vijay Pande: It could be too that all sitcoms might be the same sitcom at some level.
--:-- Marc Andreessen: Here would be an example. I also have it do dramatic screenplays, dramatic stage plays. It's quite good at those. You can say "write a three-act screenplay" and it will do it, and it will have the proper setup and resolution. But there are systems for screenwriting in Hollywood where they have three acts. It's all Rocky or it's all Star Wars.
--:-- Vijay Pande: It's really interesting that maybe what we think is magical when humans do it isn't actually all that magical either. That's what I was going to say. The human sleight of hand is, is there actually free will? Is there actually creativity happening upstairs? Either way, if there is, is it everybody? Is there really a thousand types of movies, or is there one latent space of the monomyth? The theory would be "The Hero with a Thousand Faces" or the idea of the Jungian hero's journey. It's the basis for all of the plots, Star Wars and Harry Potter and everything else. Somebody with your background might say it's an algorithm for surfing human neurochemistry, just generating different neurological responses to fear, anxiety, love, and all these other things.
--:-- Marc Andreessen: I've always been fascinated by this thing in psychology called Core Affect Theory.
--:-- Vijay Pande: Oh, that one I don't...
--:-- Marc Andreessen: It's great. What are you saying? Love and despair are these different emotions. Core Affect Theory says no. Good or bad, and then high or low. We either have a positive neural response or a negative neural response, and it's either high intensity or low intensity. Wistfulness is slightly negative, but despair is extremely negative. It's all two by two. We're more basic organisms than we think. We're very good at creating a story to justify whatever happened. We create these stories, these scripts, this idea of an emotion, but it's basically just justifying the neural response. The cynical view would be having an ice cream cone on a hot day and falling in love are the same thing.
--:-- Vijay Pande: Well, maybe neurochemically, maybe they are.
--:-- Marc Andreessen: This comes into play in drug abuse. Things that generate an opioid response. Some people get an opioid response from alcohol, and therefore we're prone to alcoholism. People don't get that response. So it's literally a neurochemical thing. Maybe we're bundles of neurochemistry to a much deeper extent, or a much simpler extent, than we want to believe. That takes me to the other thing on AI. One of the ways that people are testing AI is with the so-called Turing test. The simplified form is you're chatting with somebody that may be a human or may be a bot, you chat for 20 minutes, and at the end can you guess better than randomness whether it's a human or a bot. My take is that Alan Turing was a genius, but the Turing test is malformed. Humans are too easy to trick. That's too low of a bar because tricking a person is not that hard and does not prove anything other than that you've tricked the person. Things like GPT are about to pass the Turing test.
--:-- Marc Andreessen: They probably have in some places. I think it's going to turn out that that was too lightweight of a test.
--:-- Vijay Pande: Here's my favorite example for why I know GPT is not self-aware. If you ask it if it's self-aware and you ask it to elaborate on how it became self-aware, it will happily tell you. By the way, if you ask it how it's going to feel if you turn it off, it's going to tell you, "Please don't." If you ask it to explain to you why it's not self-aware, it will very happily do that too. It does not have a differential opinion.
--:-- Marc Andreessen: Yeah.
--:-- Vijay Pande: Whereas every conscious, even non-conscious living organism has a very different response to those two scenarios. It's been amazing because in some ways I feel like it's been as much interesting to study the AI as the AI has reflected for us to study ourselves. We are seeing that the magician has certain tricks, whether it's an AI magician or a human magician. We're going through this education process. It feels like GPT can get into high school, get into college. But what would it take for it to get its PhD? I think that's where the dramatic stuff is to come.
--:-- Marc Andreessen: Again, exactly to your point, I would ask the question the other way. What does it take to get a PhD? What does it take for you to get it? How are the universities doing in quality control of their own programs? How many people are getting PhDs today that we would say are actually valid, with actual accomplishments? People who got professorships 100 years ago, how would they score the PhDs that are being granted today? Would they say the bar is lower?
--:-- Vijay Pande: Lower.
--:-- Marc Andreessen: I think they would say the bar is dramatically lower. The answer might be we've lowered the bar. Same thing for college admissions. What does it take to get into college? What does it take to finish college? This is coming up a lot right now because GPT can auto-generate student essays. The grading method of assigning an essay and grading the result is probably not going to work anymore. But was that ever actually education? Was that actually teaching anybody anything? College applications are basically done, at least to the extent that you believe they were a legitimate way to evaluate anybody in the first place. I'd be more skeptical if they were ever useful in the first place.
--:-- Vijay Pande: In the PhD, let's talk about that old school mentality of a PhD, some advanced learning where you become an expert in something.
--:-- Marc Andreessen: Expert. What do you mean by expert?
--:-- Vijay Pande: The ability to be in the top 0.1% of humanity at designing a drug or building something. Is that what they teach?
--:-- Marc Andreessen: Is that what they teach in universities? I think it is. Or at least that's what you have to do eventually when you get out. You have to apply it. One of the things about being an expert in my mind is that the difference between bad, good, and great can be really close. I could probably write a piece of music, but no one would think it's all that great. You could have someone who's a good musician, but not a great one. Then you have a genius, like a Mozart or a Led Zeppelin, pick your genre. I think where we aren't there yet is that when the difference between good and great is so close. I remember from Spinal Tap, "There's a fine line between brilliant and stupid." I think that is where it hasn't really hit yet. If you look at the jokes, the jokes are just kind of okay. The screenplays it makes are not brilliant screenplays. I think it could get into college, but could it win best screenplay? That's the part where I think we're not there yet. But I think we're getting there.
--:-- Marc Andreessen: Name a great music composer generated by a music PhD program in the last hundred years.
--:-- Marc Andreessen: Name one.
--:-- Vijay Pande: I'm thinking more in the scientific side of things. Probably the PhD programs in that space are not intended to generate music.
--:-- Marc Andreessen: Name one great screenplay written by a PhD in drama.
--:-- Vijay Pande: That's an interesting point. What I'm getting at is still the ability to do something. In the case of the screenplay or the music, they still have to learn something. Or do you think they just innately knew how to write a screenplay?
--:-- Marc Andreessen: I assume there's a process where they write a screenplay and it's kind of mediocre. Then they get critiqued, or they critique themselves, and it improves and improves.
--:-- Vijay Pande: So there's a divorce of the education.
--:-- Marc Andreessen: The test for the screenplay is, does it sell? Screenplays are subject to market discipline. Question number one is, does it sell to a studio? Will they buy it? Test number two is, when the movie comes out, does anybody watch it? Do they like it? Do they finish it? One of the fun things Netflix will now tell people who make film and TV is they actually tell them for the first time whether anybody's actually finishing a movie.
--:-- Marc Andreessen: Those stats are kind of mind-boggling. A lot of movies, people go to the theater and they feel invested and they don't want to leave in the middle, but it's very easy to punch out. It turns out a lot of screenplays, professional screenwriters will tell you, it can't ever sag, because people will stop watching. Screenwriting is subject to market tests. Popular music is subject to market tests. But classical music, which I'm a huge fan of, is no longer subject to market tests. It's thoroughly subsidized.
--:-- Vijay Pande: That's interesting.
--:-- Marc Andreessen: It's not in the free market anymore. Or maybe the equivalents are movie music. Movie music is subject to market tests. It's probably the modern classical.
--:-- Vijay Pande: For that reason. The market test is real. Let me grant your point. Let's use the term "taste."
--:-- Marc Andreessen: Or just ability to do something hard. Ability to create something complicated. And also the ability to judge. Critically, to start with judging your own work. And therefore the ability to improve.
--:-- Vijay Pande: There is something about taste. I tend to think this stuff all has aesthetic. Properly constructed math, a formula, a software program has aesthetic.
--:-- Marc Andreessen: 100%.
--:-- Vijay Pande: Martial design has aesthetic properties. Physics has aesthetic properties. All of it. There's something about taste that's some combination of quantitative and qualitative. A great startup from a mediocre one is taste. There are certain methods and certain signals, but it's not necessarily reducible to an algorithm. It's a composite of foundational knowledge combined with some scope of experience, combined with some ineffable characteristic.
--:-- Marc Andreessen: We associate an aesthetic with it, but I wonder whether that's also just our emotional connection to it. We have this good right or wrong, or more right or more wrong, like a gradient. A lot of it is also whether something is elegant versus just a hack. You can tell whether great things are simple and powerful rather than some complicated machine to do something that's going to eventually fall apart. That's true in physics or in go-to-market or in music. It has that complexity and a simplicity at the same time.
--:-- Vijay Pande: I'm curious, when AI gets that point, which I think is a when, not an if.
--:-- Marc Andreessen: Okay.
--:-- Vijay Pande: Why would you say it wouldn't get there? Do we even understand how it works with people? Maybe we don't have to.
--:-- Marc Andreessen: This is where I describe the AGI question. This is where I call the hand wave. The embedded assumption that it's a "when" is that it will be an emergent process that will unlock as a consequence of greater and greater levels of scale. One way of looking at that is yes, that's what's going to happen. The other impression is it's just a massive hand wave. Let me ask you a question. What is the subspecialty of human biology and medicine that most understands the nature of human consciousness today?
--:-- Vijay Pande: I don't think there is one.
--:-- Marc Andreessen: There is one. Anesthesiology.
--:-- Vijay Pande: Which is poorly understood.
--:-- Marc Andreessen: It's poorly understood, but they know how to turn it off. And they know how to turn it back on. They've got the on-off switch. That's all we have. We collectively have been studying this question of human consciousness for a very long time. We have very advanced technologies today, functional MRI, all this stuff.
--:-- Vijay Pande: There's a field I would love to see created, which is molecular psychology.
--:-- Marc Andreessen: Okay.
--:-- Vijay Pande: Where you can start to probe this a little more. Literal or metaphorical?
--:-- Marc Andreessen: Quite literal. Molecular biology was this big thing in the 80s. If I can bring chemistry of small molecules to polypeptide biology or chemical biology, and if we could use small molecules to perturb more than just on-off, but perturb things, we can start to understand the brain a little bit. Reading is one thing, but poking and perturbing and then seeing the result is usually how we do any sort of experiments.
--:-- Marc Andreessen: Would you view that as a chemical experimentation? Would that be electrical?
--:-- Marc Andreessen: It could be either one. It could be any of that, but it would probably be some combination of those things. Neural science is on a track, in theory, to enable anything of this.
--:-- Vijay Pande: So, here would be the counterargument. We don't know how human consciousness works. I actually looked at the field 30 years ago when I was in college. I was going to study it, but I looked at the field and I thought, they don't have a clue. I'm not going to spend my entire career tilting at windmills.
--:-- Marc Andreessen: So you wanted to go into consciousness, but you saw it was like expert systems back then.
--:-- Vijay Pande: Yeah, exactly. It was early neural networks, brain chemistry, and we were going to figure it out. And they just didn't know then. As far as I know, they don't know now. So the counterargument is that all this talk about merging with AI is just massive cope for the fact that we don't understand consciousness. We can't do it. We'll be sitting here 30 years from now with no more knowledge, barring other scientific breakthroughs.
--:-- Marc Andreessen: What's interesting is if you think about that time, they were all single layer networks. They couldn't even do XOR. You needed deeper networks, and you couldn't have deep networks then because we didn't have the computational power. The space was dormant until we got GPUs and could go deep. So it is possible we saturate the compute we have now, get as far as we can, maybe get close to AGI, maybe not. Then it takes another 30 years for the next breakthrough.
--:-- But let's pull back from AGI. Let's pick a domain you think a lot about: life sciences, designing drugs, healthcare. Can you suggest a drug? Can you pick a diagnosis? In those limited domains, we don't need to go all the way to consciousness. But right now, generative AI doesn't seem quite far enough yet.
--:-- Vijay Pande: Yeah, we'll see. I know you think about healthcare a lot.
--:-- Marc Andreessen: Yeah. Well, let's talk about medical diagnosis, which is low-hanging fruit. The first question is: what's the threshold? Is the goal perfection, or is it better than human?
--:-- Vijay Pande: That's a great point. This comes up with self-driving cars. Is it perfect, or just better than human?
--:-- Marc Andreessen: Exactly. Self-driving cars score accidents per thousand miles driven, and they're already lower than human drivers. Humans may actually be getting worse with texting. And machines have the characteristic of universal improvement. One car has a mishap, every other car learns from it. You can make a serious argument that self-driving cars are already better than people. Morally, you could even say human drivers should be outlawed today. The utilitarian argument would be to ban human drivers because the machine is already better. Probably the same is true for airplanes.
--:-- Vijay Pande: Right.
--:-- Marc Andreessen: We're not going to do that, but you should think about it as a possibility. I think you should think about it for medical diagnosis. There are two tests. Test one: feed in symptoms, get the correct diagnosis 100% of the time. That's a high bar. Test two: I do that with the algorithm, then I go to 100 human doctors and get 100 different responses. Let's compare over time. How good is the median doctor?
--:-- Vijay Pande: The median doctor may be smart, but also overloaded, exhausted, has 12 other patients.
--:-- Marc Andreessen: Fifteen minutes. Experts in these areas have concierge doctors who spend a lot of time with them. The median healthcare experience is 15 minutes with a doctor who may never see you again and has very limited data. The algorithm is: they come up with a diagnosis, a treatment, it doesn't work, you repeat. While you're still sick and not dead, you just repeat.
--:-- Vijay Pande: And then there are drug interactions. Is any one doctor tracking all of them? Then there's compliance. Does the doctor know if you're taking the prescription? Compliance is one of the biggest disasters. The ability for a median doctor to even evaluate the success of a treatment is compromised because they don't have the data on compliance.
--:-- Marc Andreessen: You look at the existing system, it's like looking at how people actually drive cars. It's horrifying. We fool ourselves into believing it's good because we don't want to look behind the curtain. If the machine could be twice as good at listening to symptoms, giving a response, writing a prescription, doing follow-up...
--:-- Vijay Pande: I've been playing with it. You plug in a list of symptoms. It has access to collective medical knowledge. If it doesn't now, it could be filled with all the EMRs and medical records and learn from that.
--:-- Marc Andreessen: And the medical field moves. The median doctor has to read all the papers. That never happens. There's continuing education, but it's not the same. Do you like a young GP or an old GP?
--:-- Vijay Pande: Probably young.
--:-- Marc Andreessen: The old GP has more experience and pattern matching. The young GP is more up on current science. Do you want to make that tradeoff, or can the machine have both?
--:-- Vijay Pande: Exactly. Can it beat 100 doctors collaborating? That's the wisdom of the crowd.
--:-- Marc Andreessen: That sounds horrifying. That's a committee. That's the Soviet method. Have you really found human beings make better decisions in groups than as individuals?
--:-- Vijay Pande: That's a good question.
--:-- Marc Andreessen: The wisdom of crowds and the madness of crowds are flip sides of the same coin. When are you harnessing wisdom? When are you descending into mediocrity? Groups can do well on very specific tasks, but otherwise it's like a high school group project. Social conformance kicks in. Group polarization happens. You take a group that inclines slightly one way, let them talk for three hours, they all come out much more radical.
--:-- Vijay Pande: So maybe that's interesting. You could imagine training AI to have different aspects, and its collaboration with other versions of itself would be very different.
--:-- Marc Andreessen: Yeah, maybe it should run the same inputs a hundred times in a Monte Carlo simulation.
--:-- Vijay Pande: Right, right. So either we'll never get there, or we're already there now. But in 10 years, it seems real. Things are accelerating. What do you think society needs to change? This seems bigger than the software revolution or the internet revolution. It changes government, regulation, education. It may take us 10 years culturally to get ready for something that may arrive in 10 years or may already be here.
--:-- Marc Andreessen: We've already fallen into a trap. We've only been using a single example: something is done today, and maybe the machine can do it instead. That's worth thinking about, but the way technological impact actually plays out is by revisiting more fundamental assumptions. It's about what's not being done today that all of a sudden becomes possible. This comes up in discussions about employment. People worry about technological displacement, but technology never creates unemployment. It only creates jobs because it makes possible things that were not possible before, which leads to growth.
--:-- The doctor of the future is probably not going to be doing the same thing. The core motion of a lot of doctors is diagnose, prescribe, diagnose, prescribe. The technologically empowered doctor is highly unlikely to be spending their day doing that. They'll be doing things that are much more important. Maybe they have more time with patients because the machine saves time. Maybe they have more data. They've got the machine as a partner. Maybe they spend more time talking about psychological issues or behavioral issues. A lot of primary medical issues today are a consequence of behaviors. Compliance is a behavioral issue. And there are all the behavioral health issues coming out of COVID.
--:-- Maybe the doctor of the future will be more of a life coach with a pharmacological component. More holistic. The doctor of the future could be a much more fundamental figure in your life than today.
--:-- Vijay Pande: That sounds fantastic.
--:-- Marc Andreessen: If I'm a doctor, that's where I want to be. That's probably a bigger and more important market. The same is true in education. The teacher 10 or 20 years from now, I hope, is not doing the same things. One-to-one tutoring is the only education intervention at scale that actually improves outcomes. It's very ancient. The education system we have today is an artifact of the industrial age. If the industrial components become automated, the teacher becomes freed up to work one-to-one with students. That could be a significant breakthrough.
--:-- Vijay Pande: You could also imagine AI doing the one-to-one tutoring, with the teacher supervising the AI, sitting at the control panel watching it all happen. That speaks to something interesting. We're probably nervous, at least short-term, to just unleash this without paying attention. So you'll have the doctor using it as a tool but keeping an eye on it. The teacher scaling dramatically but keeping an eye on it. Do you think that's how it will go?
--:-- Marc Andreessen: That's how all technologies work. Another way to think about it: I like the term "augmented intelligence" rather than "artificial intelligence." The Doug Engelbart idea. Steve Jobs called it a bicycle for your mind. A bullet train for your mind. The way technology plays out is that everyone fears replacement, but it turns out to be augmentation. You take a human, give them technological tools, and they become much more productive. Artists are more productive with digital tools. Even artists who work with paint and canvas are more productive because they can sell to a larger audience online.
--:-- Photography came along and dramatically changed art. Being photorealistic wasn't interesting anymore, so it created modern art, which is maybe even more expressive. The artists of history were not happy about photography. It was a threat. But it transformed things. The market for art is much larger today than before photography. We call it different things—TV shows—but the market for creative expression is much larger. Music is the same. Recorded music was a threat. You used to have to hire a musician to come into your home. Phonographs were a threat, but they made the music industry much larger.
--:-- I think AI will play out the same way. Some argue AI is different because it will keep climbing and replace everything. I think it's the ultimate superpower, the ultimate pairing. If I'm a Hollywood screenwriter today, GPT is my best friend. I reach a plot point, I ask for 10 ideas, and it gives them to me. I can be more creative.
--:-- Vijay Pande: Your comment about music is interesting. We have Spotify with everything in our pocket. Can you imagine the AI Spotify that is the doctor, the personal trainer, the educator, all in my pocket for whatever I need?
--:-- Marc Andreessen: Yeah, with the human escalation path. The AI therapist, but if it gets serious, it escalates immediately.
--:-- Vijay Pande: Okay, so what's going to hold us back? What do we need to change?
--:-- Marc Andreessen: I think it's mostly fear. This is where maybe I'm a radical. People start talking about regulation. I think we have these fear-driven reactions. There's a deep-seated myth in human societies: the Prometheus myth. New technology, fire. It can be used for good or for destruction. Prometheus retrieves fire from the gods, and his punishment is to be chained to a rock and have his liver pecked out every day. Embedded in that is the anxiety about new technology. The person who brings it should be punished. The myth plays out over and over in discussions about regulation. The gods punish him on behalf of existence.
--:-- If you look at the history of new technologies, they generally have these fears. Every step along the way, technology has been created with some prediction that it will upend the social order. It does upend it to some degree, but generally in a positive way. Technology is why we live much better lives. Nobody would trade to go back 50 years. Nobody would ever make that trade. You would not want to lose the technologies you have today.
--:-- So fear may be the biggest threat. Fear leads to the reach for regulation. I'm a skeptic. Are we really going to regulate math? They'll say it's not regulating math, it's regulating this superpower. But the actual implementation is regulating algebra. Regulating linear algebra. Are we really going to regulate matrix multiplication? And even if we do, will we do it in a way that makes any sense?
--:-- Vijay Pande: It won't look like that. It'll be, "We can't have computers drive cars."
--:-- Marc Andreessen: Or, "How do you give the computer a test?" You make the claim that the AI is better than human. How do I know? Well, because the cars are driving. Here's how it played out in self-driving cars. One category of company said, "We'll wait until it's perfect, work with regulators." They're still not on the road. Another category of company said, "Let's evolve out of cruise control. Cruise control, then radar. Humans drive with it, we label data." You don't expect the car to drive itself from the beginning. It's an autopilot. You're told to pay attention. But that company has been climbing the ladder. They push new software releases live to cars at night. Those releases are not tested by any federal agency. That has led to incredible progress. Because you can't make it work magically. It has to happen gradually. It's like medicine. It's entering a complex system with a lot of variables. It's life or death.
--:-- We go back to how we started. The "wait for permission, wait for perfection" binary approach versus the incremental "let's get better and better" approach, where the threshold is "is it better than humans?" In self-driving cars, the second approach is working. You observe the statistics. We have all this happening right now. The statistics show it's so much better than humans. What wouldn't we do?
--:-- Marc Andreessen: Yeah, exactly. Right. And then at some point, the morality tips were like, well, obviously we have to go in this direction because it's just obviously better.
--:-- Vijay Pande: Yeah. I suspect we're going to get through medicine pretty quick.
--:-- Marc Andreessen: Yeah, yeah, yeah. I'm an optimist on that. And again, I'm not an optimist because I think the AI is going to be perfect. I'm an optimist because I think the status quo is not that good.
--:-- Vijay Pande: Yeah. Well, that might be like you start empowering doctors, you give them tools, they start using them. And you start empowering patients. Patients start using them. Actually, here, I think it's even different than a car because you're not on a road. It's your body or whatever. And actually, patients are driving their own health care more than ever. I think COVID was another sort of tailwind there. So maybe you start, maybe it's just about developing the tools and giving them out.
--:-- Marc Andreessen: Well, here would be an example. So let's use our screenwriting example.
--:-- Marc Andreessen: The plan in medicine, which is, any given set of conditions, there may be many possible diagnoses.
--:-- Marc Andreessen: An experience I've had is there's a set of symptoms. One doctor comes up with one diagnosis. Another doctor comes up with a different diagnosis. You read the literature and it's like, actually both of those diagnoses in theory are, but for some reason, the one guy only thought of the one, the other guy only thought of the other.
--:-- Marc Andreessen: So a way for doctors to start using this technology today would be plug in the symptoms, give me five possible diagnoses.
--:-- Marc Andreessen: Okay. Oh, I didn't even realize, right? Because maybe this is a new thing since I went to medical school or something. I didn't realize diagnosis number three was an option. I just go look at that.
--:-- Vijay Pande: Yes. Yes. Right. And so the doctor is still doing the diagnosis. So that's your screenplay example. You augment it. As a doctor, you augment it. Because in that case, it's luring you to things that you should know, but you don't know.
--:-- Marc Andreessen: Yeah. I mean, that's interesting. It's almost like having a mentor or just someone to riff with.
--:-- Vijay Pande: Yeah, that's right. Yeah, yeah, yeah. And it's a great thing. It is a machine. It will riff with you as much as you want. And it will sit there at three in the morning.
--:-- Marc Andreessen: Oh, yeah, yeah, yeah.
--:-- Vijay Pande: It has a hundred times for you. It's happy to. It just doesn't get bored. It doesn't get tired.
--:-- Marc Andreessen: Yes. By the way, and then it also has the advantage, it has all the up-to-date information.
--:-- Vijay Pande: Yes. Yes. And all the outcomes. And when it makes a mistake, it actually can learn from it, rather than being devastated by it or emotionally reacting to it.
--:-- Marc Andreessen: Right. Right. And like self-driving cars, if some other doctor in some other state had a patient last week and it made a mistake and they fixed the mistake, it will not make that mistake on your patient.
--:-- Vijay Pande: Yeah. Yeah. Yeah. So, I mean, so you think, and so that is a very different regulatory play than we've seen in the history of healthcare.
--:-- Marc Andreessen: Well, I think that's just going to happen. So here's what everybody knows. I'll give you a couple things. Everybody knows that patients should not be on Google.
--:-- Marc Andreessen: Everybody knows, every patient now does that. It's called Dr. Google. Everybody knows it. Literally, I was called in the field.
--:-- Vijay Pande: Right.
--:-- Marc Andreessen: And there's no way, practically speaking, to regulate that out of existence.
--:-- Vijay Pande: Yeah, yeah, yeah. That's going to happen. I think doctors using these new tools as augmentation is something that they can just do. It doesn't require approval. So the ship has already sailed.
--:-- Marc Andreessen: I think so. Yeah, yeah, yeah. And by the way, patients using GPT, if it hasn't started, it's going to start imminently.
--:-- Vijay Pande: Yeah, yeah, probably.
--:-- Marc Andreessen: So the patients are going to show up with the results of the GPT queries and the doctors are going to have to respond to that. And so they're going to end up being in this world whether they want to be or not.
--:-- Vijay Pande: But that's actually really interesting because as a patient, and I probably know just about enough medicine to be dangerous to myself, but I show up with the doctor and I have all of that thought out. That might equalize the patients, such that they can actually become much more educated and become much more thoughtful, and they become much more in the process as well.
--:-- Marc Andreessen: Yeah. Okay, so what goes wrong? Double-edged sword. I mean, as a doctor, do you want your patient more educated or less educated?
--:-- Vijay Pande: They may just be humoring me, but I think they want them more educated.
--:-- Marc Andreessen: Maybe with you, they're educated. With me, they might look a little bit more sideways.
--:-- Vijay Pande: Actually, if it was really helpful, I think they would. I think it's just about how good it is, right?
--:-- Marc Andreessen: Okay, so what goes wrong? I mean, the big thing that goes wrong, I think two things go wrong. So one is just the expectation of perfection. It's very easy to generate the negative headline. It's very easy to set off the scare of the moral panic. A single incident goes wrong and it gets extrapolated. We talk a lot about thalidomide, like it's very easy to have that kind of thing. Or the person on a bike that got hit by a Tesla.
--:-- Vijay Pande: I think it was biking across a freeway, right?
--:-- Marc Andreessen: Exactly. And so a human probably would have hit them too.
--:-- Vijay Pande: Yeah, that's right.
--:-- Marc Andreessen: Yeah, so the trolley problem.
--:-- Vijay Pande: Yeah, yeah. The trolley problem always gets mooted about self-driving cars, which is you have a choice between killing five grandmas or one little kid, you have to pull a lever. But human drivers don't...
--:-- Marc Andreessen: No, no, no. Human drivers never make that decision. They have gas or brake. They have, hit the car in front of me or not hit the car in front of me. It's never this elaborate thing. It's always a very simple thing. And so it's not a question of whether the machine can ideally solve this idealized complex problem. It's going to hit the brakes faster when it's about to crash into the car directly in front of it. So properly containing the expectation here to the actual real world and not having this spin off into fantasy narratives that you can then criticize.
--:-- Vijay Pande: Yeah, so that's the absolute limits. And then, I think just the generalized fear. And what I always have to remind myself is, I'm a software developer. My background is, okay, I can actually... Can I tell you every aspect of how these algorithms work? No. Do I understand how they work? Do I understand the basic foundations? Do I understand the basic math? Yes. This is why I make the comment about regulating math.
--:-- Marc Andreessen: Yes. As somebody who's not a coder, this whole thing is weird. It's like weird math.
--:-- Vijay Pande: Yeah. Right. Yeah. And so there is a... I have to remind myself to be patient and tolerant of people who don't understand the mechanics of what's happening. That said, I think the people who are going to be ready, they also have to get in the academy to try to understand this. And there's always slippage there.
--:-- Marc Andreessen: Yeah, so what's the antidote to fear? Is it optimism? Is it education?
--:-- Vijay Pande: I mean, ideally, it's a cultural orientation towards new technology. And then ideally it's education and people learning, or the CP Snow two cultures coming together and educating each other. Honestly, a big part of it also, I think, is when things become a fait accompli.
--:-- Marc Andreessen: Yeah. And this is what Tesla's done in self-driving cars.
--:-- Marc Andreessen: Like, if it's just happening. Who would want to go back? The system adapts. There was this famous Uber fought all these regulatory wars in all the cities that they were in because it was not technically allowed. Taxi limo charges in the beginning. So one of the things they did early on was they just made sure that there were always lots of Uber cars available around state houses and city halls. So whenever somebody was giving a roaring speech in City Hall about shutting down Uber, they would come out and they'd have to get home really fast and Uber would show up 20 seconds later. At some point, it just was taken for granted. And then if you said, are we going to take Uber away? People would have said, no, we can't. It's over. And that's what happened. And then later they changed the laws to accommodate that behavior. So I actually think part of it here is just having these tools.
--:-- Vijay Pande: Okay, here's the thing. Here's a good news thing. These tools are becoming widely available up front. 50 years ago, a new technology like this would have been deployed in the government first and then into companies and then years later in the form of something individual people could use. The model today is, it's just online.
--:-- Marc Andreessen: Yeah.
--:-- Vijay Pande: GPT is online right now.
--:-- Marc Andreessen: Yes. Well, the future of paint is really intriguing because from an engineer point of view, it's the engineer's dream that if we make it good enough, such that it can get to a point where people just love it and it's helpful and it does what it needs to do, the rest will take care of itself.
--:-- Vijay Pande: Yeah. I mean, I kind of think that's mostly how things happen.
--:-- Marc Andreessen: Yeah.
--:-- Vijay Pande: No, that's a beautiful feature.
--:-- Marc Andreessen: Yeah. So now look, having said that, healthcare is very sophisticated. There's lots of regulations, there's lots of payment, all these things. So I saw this thing on Twitter the other day. It blew my mind. Because this whole time I've been thinking in terms of diagnosis and stuff in my life. So this doctor posted a video and he said, the problem is whatever diagnosis, whenever I do the diagnosis, I do the prescription. Then it's a question of whether or not I can get the insurance company to reimburse, to pay for the thing. To do that for anything even slightly out of the ordinary, I have to write a letter. The doctor has to write a letter to the insurance company, and that letter needs to be in a specific format, and it needs to make the case. It needs to have the scientific citations. And if I do the letter really well, it's going to get paid for, and if I don't do the letter really well, it's not going to get paid for. It could matter to the life of the patient. And so he's like, it turns out GPT is really good at writing those letters.
--:-- Vijay Pande: With the references.
--:-- Marc Andreessen: With the references, yes. With the scientific references, full on. And so, you've got this bureaucratic process, which is legitimate and required and yet to exist, and that data needs to be submitted. And honestly, it does not matter to that process whether that document is written by a human or a machine.
--:-- Vijay Pande: Yeah, yeah, yeah.
--:-- Marc Andreessen: But all of a sudden, if every doctor in the world is really good at writing this letter correctly, then that doctor now has another four hours a week to actually take care of patients. That's the kind of thing that I think is going to happen more uniquely. And what's interesting about that example is you can imagine that example having a big impact on the efficiency of the healthcare system today without any regulatory changes.
--:-- Vijay Pande: Yes. Oh, it's within the system. Within the system. And that was the one where it's just like, oh, in retrospect, that's obvious. I just haven't thought about it. One guy thinks about it. All the other doctors start to do that. The whole system upgrades, step function, one time. That kind of thing, I think, is a real possibility.
--:-- Marc Andreessen: Yeah, and that could be because someone's working within the system, you can have the transformation immediately.
--:-- Vijay Pande: But then eventually someone has to read all those letters. Someone has to validate them. It's probably some sort of NLP on the other side.
--:-- Marc Andreessen: Well, that's right. It's corresponding. So we have this company called DoNotPay. It's an app that acts like a bot. I've used the app. It started to get you out of basically fake BS traffic tickets. And then he did this thing a while ago where it will unsubscribe you. All these consumer subscription services like Comcast, they all make it hard to ever turn off the subscription. So he has this way for the bot to do it for you. He started using AI in the bot. The way a lot of consumer subscription companies work is you can subscribe online, you can't actually unsubscribe online. You have to call an intended number and you have to argue with a person. There's actually this thing in these companies called save teams where they're actually paid specifically to prevent you from unsubscribing. They'll try to cut special deals with you and they'll try to talk you out of it. So he has this thing wired up where now he has AI-generated text with text-to-speech. It talks to the customer service person on the line. And it does it with infinite patience. It will just sit and argue: "No, I am actually going to unsubscribe." "No, I'm not interested in such a special offer." "No, no, no, no, no." Until finally the other guy gives up and says, "Okay, fine, I'll stop charging you." It was a precondition of the system that that worked the way that it did. It was a burden on people to have to deal with that. AI can now step in and equalize the power imbalance between the customer and the company. And presumably that will change the system.
--:-- Vijay Pande: Yeah, well, one would think. One would hope. And to your point, step one for changing the system might invite retaliation, which is all of a sudden the save teams will be bots and so maybe the bots will be arguing with the bots. But at least it gets you out of this Kafka-esque thing you're in today where you're an individual dealing with a giant bureaucracy. At least it equalizes the power.
--:-- Marc Andreessen: Well, that's kind of amazing and that will be the spark for changing things. Because once you're in that sort of system, we got to do better than this. This is crazy. It's having bots argue with each other all day long. It's just clearly stupid.
--:-- Vijay Pande: Yeah, yeah, yeah, yeah. And especially when it's bots on both sides, now we can finally say, well, let's do an API on both sides. So let's do something smart on both sides.
--:-- Marc Andreessen: Yeah.
--:-- Vijay Pande: Well, Mark, that's such a beautiful, optimistic view of how this could go. Because the future we're talking about is actually much more engineer-driven. If an engineer can build this, and it really, really works, it really helps patients, it really changes things, it'll get adopted. As it gets adopted, culture will work around it and will love it and will not want to go back. And then the future will just be right in front of us.
--:-- Marc Andreessen: Yeah. Patients are going to get a vote. Doctors are going to get a vote. It's an industry made up of people, a world made up of people. People will get a vote.
--:-- Vijay Pande: Beautiful. Thank you so much for joining.
--:-- Olivia Webb: Thank you for joining BioEats World. BioEats World is hosted and produced by me, Olivia Webb, with the help of the Bio and Health team at A16Z and edited by Phil Hegseth. BioEats World is part of the A16Z podcast network. If you have questions about the episode or want to suggest topics for a future episode, please email bioeatsworld at a16z.com. Last but not least, if you're enjoying BioEats World, please leave us a rating and review wherever you listen to podcasts. Please note that the content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund. For more details, please see a16z.com slash disclosures.
