This Week in Startups - Uber earnings + using AI to revolutionize clinical trials w/ Unlearn.AI’s Charles Fisher | E1734

Episode Date: May 3, 2023

First up, Jason breaks down Uber’s huge Q1 results! (1:21) Then, Unlearn.AI CEO Charles Fisher joins to discuss the advancements his company is making in fast-tracking clinical trials (9:49), how machine learning is used in drug development (16:24), Unlearn’s business model (35:24), and more! (0:00) Jason kicks off the show (1:21) Uber’s Q1 earnings  (8:31) QuickNode - Get one month free by using code TWIST at https://go.quicknode.com/twist (9:49) The foundation of Unlearn.AI (12:42) Unlearn.AI’s impact on the clinical trial process (14:24) Criticisms of the current clinical trial model (16:24) ML’s Impact on drug discovery  (24:16) CacheFly - Get 10 terabytes free by signing up at https://twist.cachefly.com (25:42) The data used in the medical system today (34:05) Microsoft for Startups Founders Hub - Apply in 5 minutes for six figures in discounts at http://aka.ms/thisweekinstartups (35:24) Unlearn.AI’s business model (40:22) Building a foundational model for health (44:06) The pace of AI today vs. the previous decade (46:35) More on Unlearn.AI’s business model (48:15) Creating foundational datasets in health FOLLOW Charles: https://twitter.com/charleskfisher FOLLOW Jason: https://linktr.ee/calacanis Subscribe to our YouTube to watch all full episodes: https://www.youtube.com/channel/UCkkhmBWfS7pILYIk0izkc3A?sub_confirmation=1 FOUNDERS! Subscribe to the Founder University podcast: https://podcasts.apple.com/au/podcast/founder-university/id1648407190

All right, everybody, it's a great day for your boy, J-Cal. I am still long Uber. I still have a large amount of holdings and the stock was up massively today because they had record high free cash flow and they beat on revenue. This week in startups is brought to you by QuickNode gives blockchain developers unparalleled reliability and speed with access to unlimited endpoints across 18 chains and 35 networks. Get one month free by using code twist at go. quicknode.com slash twist. Cashfly is a pure play CDN provider that makes CDN simple, effective, and secure. Deliver content faster than your competitors and get 10 terabytes free forever if you sign up

at twist.cashfly.com. And the Microsoft for Startups Founders Hub helps all founders build a better startup at a lower cost from day one. Startups get up to $150,000 in Azure. credit, access to free OpenAI credits, free dev tools like GitHub, technical advisory, access to mentors and experts, and so much more. There is no funding requirement, and it only takes minutes to join. Sign up today at aka.m.s. This is a great day for me. You can see I'm very enthused. My net worth went up, but also the bet I placed. Over a decade ago, that defined my

career as an investor, I always believe this company would start becoming a money printing machine, and thus that has happened. Q1 revenue 8.8 billion. That's up 29% year over year, year. And we're talking about cloud computing. Had Alex on last week, and we're talking about, oh, it's falling to single digits. Well, you know, it's not falling to single digits. Uber, up 29% year over year. Revenue came in 100 million above analyst estimates, according to CNBC. See, the total trips in Q1, 2.1 billion. I remember when Uber had a total of three riders on the platform, and we'd done maybe a dozen rides,

up 24% year over year. That's incredible. And this includes mobility and delivery, monthly active platform customers. Maps. Maccubs. There's no way to pronounce MAPSys. MAPS.

Actually, MAPSys works. monthly active platform customers. This is Uber's sort of Mao's monthly active users, but what they want to say is these are customers who are active on the platform, not dormant accounts, ones that actually did something. That was $130 million. That is extraordinary. We're talking nine figures worth of customers monthly. And looking at the revenue, 72% revenue growth for mobility, 23% year over year, for delivery. And for delivery, And freight, the freight business dropped 23% year over year. I'm not sure what the details are there.

We'll double click on it in a future episode. Can't win them all. Two out of three ain't bad. But that's a small number. It's an emerging business for them that they're investing in. Total gross bookings, right? This is the top line before they give a large percentage of the money they make to the drivers. And the drivers are doing spectacular.

That's why so many people are driving for Uber. Don't believe the fake news, which keeps saying like Uber drivers are making. $8 or $9 an hour. That's all made up nonsense. The truth is they're making $25, $35 an hour. And the gross bookings were $31.4 billion, up 19% year over year. So, you know, the revenue grew 29% year over year, but gross bookings grew only 19%, which means Uber's, you know, more profitable and charging more for their services. It's just great to see. Bottom line, they had a small net loss of 157 million, and that included a $320 million benefit from net unrealized gains related to Uber's equity investments. But really, cash flow, that's what matters, right? How much cash makes it into

the coffers, as we say? And this excludes capital expenditures, right? So you'll have some things on the books that are capital expenditures, but the actual amount of cash into the business. So this is accounting issues. Record high, free cash flow, $549 million. Cash and cash equivalence and short-term investments, $4.2 billion for the team over at Uber. And I think the really interesting part of all this is Lyft's demise. They are a shrinking amount of this industry. And Uber, I don't want to say it's a monopoly, because it's not a monopoly. You have DoorDash out there.

You have Lyft out there. You have people competing like public transportation and micromobility and people owning their own cars, rent a cars, and taxi and livery drivers, right? So to say Uber has a monopoly or they don't on mobility, nor do they have it on delivering. What they have is they have the majority now of app driven rides and they have a strong presence. I think they're number two in the United States behind DoorDash in delivering groceries and food. But people order from other sources as well, right? They're still Instacart. You still have Amazon delivering groceries and whole food. So it's not quite a monopoly, but it's a strong position in those areas. And the average

MAPSI did 14 rides or food orders in the quarter. That's almost five per month, which is very impressive because you take the 2.1 billion trips and you divide it by 150 million MAPCs, just taking a guess here that maybe they're taking 14 rides or food orders in the quarter, which would be about five per month. That's the average. Now, that means there's people who use Uber a lot. There's people who drop in, but if you're like me, I'm using Uber Eats and taking Uber rides at least 10 times, 15 times a month. So I think I'm probably in the 40 or 50 a quarter, 200 a year kind of group as a family because man my daughters love to they get me every time oh we did our homework can we get boba or you know can we order sushi and yeah i'm a sucker uh because i just take

it out and i'm just like you know what i want you to have a great childhood and enjoy some nice sushi and some boba so just to to dara and the team and dora's coming on the show uh over the summer we'll have a great interview and catch up but the stock is ripping it's up almost 11 percent today, which puts me in a great mood, not just because money, which is nice, but I got enough of that. It's just about being right and betting on a team and really seeing the investment come to fruition. This year at Founder University, I'll make somewhere between 50 and 100, 225K investments as a tribute to that 25K investment I made in Uber as one of the first investors, maybe the third

or fourth, I don't know. So if you want $25,000 for me to start your company. company, get two or three founders, have one technical person, make an MVP, come to founder. Dot University, hang out with me. We're going to actually have our own space in San Mateo soon. And just come hang out with Jacao. Let me give you that lucky 25K first check so you can incorporate. Maybe come to our accelerator, the launch accelerator, we'll give you 100,000, and let us syndicate you on the syndicate.com, share it with other angel investors and get you a millie or two. I think the average is like 700K to the syndicate. So it's been an amazing journey with Uber as my best investment in

history. And I'm trying to hit another one or a bigger one. And so that's not going to be easy. But in the next 10 years, I hope to invest in maybe two or 300 names per year, which would put me at 2,000 more investments in my second decade of investing to go with the 300 in my first, or 250 in my first decade. I'm going to 10x that. And you know, you never know. Maybe I hit a another Uber or two, and maybe that's you. So, founder. dot university or launch.co slash apply to meet with our team. Great job. Team Uber. All right. Next up on the program, Charles Fisher, the CEO of Unlearned AI. Okay, everybody, you know all of the complaints about building apps on the blockchain.

It's slow. Oh, it's less reliable. Oh, there's no support if things go wrong, right? The blockchain is this incredible innovation, and there's some good news here. Execution on the blockchain just got super easy. QuickNode has solved all of these problems. They give blockchain developers unparalleled reliability and speed with access to unlimited endpoints across 18 chains and 35 networks. QuickNode provides amazing response time and a dedicated 24-7 customer support team. They offer consistent performance at any scale, lightning-fast API responses,

that are 2.5x faster on average than competitors, and the most sophisticated and globally balanced cloud and bare metal Web3 architecture. Listen, everybody, this is the AWS or Azure of Web3. If you are building DAPs, decentralized apps, you need to use Quicknote. It's that simple. Find out why companies like Twitter, Adobe, Coinbase,

and OpenC use QuickNode. Get one month free by using the code Twist at go.com slash twist. That's go.com. W-I-C-K-N-O-D-E dot com slash twist. And remember to use the code twist. All right, everybody, welcome back to our special AI series. I've never seen anything move this fast in the 30 years.

I've been in the technology business. So we are having people on the program three, four times a week here at this week in startups to share what they're working on and why so that we can all keep up with this crazy pace. we found an interesting guest today. His name is Charles Fisher. He's the founder and CEO of unlearn.

a.I. Unlearn AI. And he is taking AI models to try to speed up clinical trials for pharma companies, which seems like a really interesting idea, Charles, and welcome to the program,

but also one that has me a bit concerned because using chat GPT, 3.5, 4, and some of the other tools, barred from Google and the CORA AI tool, they're frequently hallucinating and giving wrong data. So maybe we could start with a little bit of how long you've been working on this and then getting right to how this is going to change

clinical trials and the hallucination problem. Yeah, definitely. Yeah, so thanks for having me. How long have I been working on this? That's a really interesting question. I think how long have I been working on like generative modeling as an area? Because I was an academic researcher before starting Unlearn. So I don't know, like 15 years probably since I've been working on sort of generative AI.

But we've been at this with Unlearn now for about six years, not quite, almost six years, working on, yeah, developing generative AI to currently speed up problems in clinical trials. So, like, let's make clinical trials biggest bottleneck. We want to make that faster. But eventually, we want to roll this out to think about how we can really sort of revolutionize the way we think about medicine, turning all of medicine from something that's really today kind of an art form into a real predictive science that's founded on computer science. So let's talk about what is a clinical trial?

what is the state of the art architecture of that? Because I don't invest in this area, but, you know, some of my contemporaries do. And they talk about how incredibly frustrating and humbling it is to beat the placebo, as I've been told, which is, placebo seem to work 10% of the time, 15% of the time. They have some efficacy that is not zero. It's in some cases pretty amazing what the placebo. people can do to people's minds in terms of having an impact versus actually having an impact. So maybe the definition of in 2023, what is a clinical trial? How does it work today?

And then how is your software going to change that? A clinical trial is simply a comparison. So it's really the same thing as an AB test that you would run in any other area. So I have some new experimental treatment. And I want to compare that to usually what is currently available. So placebo usually doesn't mean that you get no treatment at all, usually given whatever you would normally get for that particular disease plus a placebo. So you're still receiving some treatment. And I just want to know which of these two things is better, which is safer, which works better,

like so forth. Typically, clinical trials are staged out, and so we have three different phases. Phase one is usually done in around 10 or so, often healthy people. you give them your experimental drug, you just increase the dose until you see too high of dose. And then that lets you figure out how much is like a safe dosage.

Then after that you move on to a phase two trial. That's usually like around 100 people. And this is just an early signal of does this seem like a drug that is worth continuing to pursue? And then the last thing would be a large phase three clinical trial. That's usually around 1,000 people.

And here you're going to randomly assign half of the people to receive your new experimental treatment. You're going to randomly assign the other half to receive the control. And then at the end of the trial, you're going to compare, see if it was better or worse. And then you can submit those data to the regulators like FDA to help them make a decision about whether or not your drug should be marketed. Got it. And so that process, I've heard a lot of criticism of that process.

objectively, what are the criticisms of that process? And then we'll get on to sort of what you're doing. There are a million criticisms of the process. Well, the major ones, you know, that are valid. Right. So, I mean, the first thing that I think we tend to encounter is just the amount of time and cost that goes into running one of these trials. One of these big phase three clinical trials, just the individual trial itself can cost

hundreds of millions of dollars to run. and they often take more than five years, right? So you're talking about spending five plus years on a single experiment in hundreds of millions of dollars. And most of the time, these trials fail. So the majority of time, actually only about 10% of drugs that enter clinical trials end up being successful. So 90% failure rate in clinical trials. So you're spending like a decade and hundreds of millions of dollars on an experiment with a 90% failure rate. And would that mean if one out of ten actually work, we're talking about billions of dollars

to get a successful drug to market if you were to look at it as a portfolio of, say, 10 drugs? That's right. Yeah. Yeah. So incredibly expensive and incredibly time consuming. I think that there are other things in terms of like we need to get participants to be willing to join and take part in these clinical trials. And then you get into other issues of, you know, certainly issues like placebo control are controversial amongst like patient advocates. Why is it that you're participating typically in a clinical trial? Usually that's because you want access to this new experimental therapy. So I think that what people are thinking about and certainly what we're thinking about

are ways that we can leverage new technologies to alleviate these problems of the speed and cost of clinical trials, but also align them more closely with what patients want. Got it. So how are you using AI machine learning to test drugs? Because it would seem to me that the human body is complex in many ways. In other ways, probably very simple. And interactions are hard. So are you literally running a simulation of,

hey, here is this new drug for, I don't know, lowering your cholesterol? all. And you can model the human body and how it would interact. And does this occur in parallel to phase one, two, and three? Or is this something that is just run in a simulation that informs how you would then run these different trials? Explain to the audience how this works. Sure. So like I said, every clinical trial is a comparison. And what we really want to know is for an individual person, we wish we could tell like what would happen to this person if I could take them and I gave them this new experimental drug and I observe how they respond. And I simultaneously don't give them the drug and I observe how they respond, right?

And so the way to run that experiment is to invent a time machine. So you give them the drug, you take your time machine back in time to the point you did it and then you don't give it to them and you see what happens, right? You do this comparison. We don't have a time machine, but we do have computer models. And so the whole idea kind of behind what we do is that what we're going to do is for every individual person in a trial. We create a digital twin of that person. And it's a computer model that allows us to simulate what would happen to that individual person.

And in our case, in the trials, we're always simulating what would happen if they got the control. So we don't simulate what would happen if they got this brand new experimental treatment. Only what would happen if they got the existing treatment. And the reason is brand new experimental treatment, it's not really a machine learning problem, right? Like machine learning, we learn from data and we make new predictions, right? So what we can do is we'll have data from like 100,000 patients receiving the current treatment. And then our task is given a new patient, how will they respond to this current treatment? And that's kind of a standard machine learning problem, as opposed to here's a brand new molecule.

What will it do to a person? That's a very, that's much harder. Yeah. So, if I were to reflect this back to you in simple plain old English. I wasn't simple enough. Well, I'm going to even try to simplify it for me. Explain it to me like I'm a five-year-old kind of situation here. We have 100,000 people who have taken this current cholesterol lowering drug.

Right. You're in the new trial. You're going to get cholesterol lowering drug 2.0. It's completely new. And then there's people who are going to get the placebo. But hey, since we know these 100,000 people's age, cholesterol level, BMI, heart rate, whatever battery of information, we can say, hey, J-Cal, 52-year-old J-Cal, 1754 pounds, you know, this blood pressure, this cardio fitness level, whatever it happens to be. We take your watch data. I don't know what data is state of the art these days.

And, okay, yeah, look, we have another, out of those 100,000 people. people, we do have 2,000 people who are just like J-Cal, we're going to run a simulation to see what would happen with you on the 1.0 medicine. You're going to take the 2.0, and of course, we've got some other group of people who are taking the placebo. Is that about right? What's happening? Yeah, I mean, well, yes, so we're taking this historic, this data from the people that currently exists. We're training this kind of machine learning model on that data. Yeah, and then exactly. So we would predict J-Cal comes in. We'd say, what would happen to you if you got the placebo? and and version 1.0 cholesterol medicine.

And so the interesting thing is, if you take that to the extreme, and let's imagine this case where that machine learning model we've built is perfect, makes no mistakes at all. That's not true, but that's just imagine that word. If it's 50% correct, it's going to have an extraordinary impact. Right, yeah.

But this interesting scenario, which you can kind of work backwards from, is, well, if that were very true, then for every patient, I can just give them the experiment, the new 2.0 cholesterol medicine, and I can see how they respond to it. And I don't need any compare, I would not need any real patients receiving a placebo. Because for every patient, I'm just predicting exactly perfectly what would happen if they got a placebo. So if you could sort of get to that point where our machine learning models are sufficiently accurate, you get to a world in which you're running clinical trials that don't have placebo

groups. It would be 100% of the patients receiving your new experimental treatment and zero patients receiving a placebo. So that would mean you'd have a clinical trial that's got half as many patients in it, which is way faster and cheaper to run. Also, all of your patients are getting access to this new experimental treatment, which is what they wanted, right? So that future is really great for our customers, the pharma companies, because they get faster trials. Actually great for all of medical research because you basically speed up medical research twice as fast. It's also great for patients. It's not perfectly achievable because our models aren't perfect today. And so then we get at this question about how we still run randomized studies where some patients receive placebo is to guard

against what you were calling earlier with hallucinations, right? So we want to make sure that we can guarantee that the clinical trials that we work in produce the right results, even if our models are not perfect. It seems to me that you, when a new drug comes out, depending on the corpus of data you have about individuals, you could give it a shot and say to the AI, hey, make your best predictions and give me, you know, I mean, depending on compute power available, give me all the possible predictions you could come up with in some reasonable amount that a human could actually compare them and say, predict what will happen.

with these 2,000 people who are joining the trial as best you can and then give them the trial and then see which sets of thinking the AI got correct. That's also a possibility and would also cost nothing because all you're doing is saying just make a simulation and be like

running a simulation on who's going to win the NBA finals based on the data you have from the regular season. Is anybody doing that as well? It seems like it could be a worthy use of time. Yeah, I mean, right? So what we are basically doing, again, is we're simulating how every single patient in this trial is going to respond if they got the new treatment. We are very interested as well into as drugs come out, incorporating.

So this is kind of the future world. So the way I kind of view it, interestingly, is that clinical trials are highly regulated area, super scientifically rigorous, right? But we think they're the easiest area, actually. They're the easier than all of the other areas of medicine. And the reason for this is because treatments are randomly assigned to patients. So basically, the way that this will work is that the model will make mistakes. It will make those mistakes on patients who are randomly assigned to receive the placebo.

And it makes the same mistakes on the patients randomly assigned to receive the treatment. And basically, in the end, the mistakes end up canceling. out. And so because of that, it's like that particular application is really robust to these mistakes that machine learning models make today. So, in everybody, when it comes to the blocking and tackling of running your startup, you don't need to reinvent the wheel. CDNs, aka content delivery networks, are the place where startups can really overcomplicate things. You don't need custom authentications or custom codes. Nope. If you're a startup, You need to just check out Cashfly. It's a pure play CDN. And CDNs are literally all they do. So they're the best in the world at it. They've been doing it for over two decades. That's 20 years. Cashfly makes CDN simple, effective and secure. Let me say that one more time. Simple, secure, effective, effective. Effective, secure. You eventually are going to outgrow the smaller ones that are trying to give you too many discounts. You want the best of breed.

So don't burn your startup credits using a CDN at one of those larger players. Let Cashfly handle it all for you. They will help you deliver your content faster than your competitors. You have to be fast if you want to compete, whether it's videos, your mobile app, games, content, SaaS. The faster you go with your delivery, the more people will use your products. So go check out Cashfly. And Twist listeners are going to get 10 terabytes free forever when you sign up at twist.c.c.c-c-fly.com. That's twist.com.

terabytes are waiting for you for free. Stop what you're doing. Pause the podcast. Go to twist.cashfly.com and get your 10 terabytes for free forever. Talk about the data. Well, the thing I'm curious about is the data you that is currently used for these trials. My understanding is one of the problems is garbage in, garbage out.

You get information from patients if they tell you information, well, have you, how How many drinks do you have a week? Sure. They're like, ah, like two. And, you know, it's really 20. Or, you know, they're, so if they're reporting data, it's obviously going to be flawed. And then if you look at the data that's available in medical records, well, why do we even have medical records today?

It's for billing, right? It has nothing to do with, right? It has more to do? Am I correct with billing than it does with reality? A hundred percent. Is that right? Yeah. Yeah.

Yeah. So, like, what data do we? we actually have that has some truth to it. It feels like wearables, you know, are perhaps the Holy Grah blood tests. You can't fake those. I don't believe. You can tell me if I'm wrong. But blood tests that are historical, maybe body scans, which I just did the pro novo,

body scan and wearables, if I gave you my Fitbit data for 10 years and then my Apple Watch, which I switched to, all of those seem to be like, that's pretty rock solid. So are any of those type of things being currently used in these trials, or is it still just like they go to people's medical records and they give them a survey to fill out? Well, clinical trials are a really unique space when it comes to data and medicine because it's a research study. So one of the problems with medical records, like if you looked at my medical records, you would see that I've been to the doctor like four times in the last 20 years. right? And all of the information in between those dates is not even there. Because I didn't go to the doctor, so it's gone. It doesn't exist, right? But a clinical trial is really different. So in a trial, it's set up ahead of time, and you define this giant battery of exams that you're going to give to patients. And that always includes things like blood tests. Now it includes other new things. There are times where people that's going to be wearables, sometimes it's going to be imaging, maybe people are getting MRIs. It could be full genomic tests. Like you might get a whole genome sequence potentially.

So it could be a huge amount of information. It varies from trial to trial. But everyone is going to get this giant battery of tests. And then they're going to come in like once a month for the next year and a half. And they're going to make the same battery of tests every month. So regardless of what happens to them, whether or not they're feeling good or they're feeling bad, it doesn't make any difference. You enroll in the study and you come in like once a month and you get this giant battery of tests. So there's actually a huge amount of information about these diseases that is being captured in these clinical trials.

So this is an opportunity in two ways. First of all, we run tons of clinical trials every year. Like as a society, we run a ton of them. Actually, the government runs a ton of it. The NIH funds a ton of clinical trials. And all of those data just poof out of, they're collected and they're not used again. They're just like, what? So literally you have this incredible diamond mind

Yeah In a clinical study They collect all the diamonds And then they just throw them in the dumpster They're exactly if they yeah They put them in a day to base somewhere It's like the end of Indiana Jones And raiders all our stocks

It goes into some like warehouse That's right yeah The top men are working on it Exactly you know It's like it's in some warehouse And it's never used again Oh How do we

Now hold on a second Let's pause there for a second If this is paid for by the government in a lot of cases. The government owns it. So, yeah, there's a rule that you have to make the data public two years after your clinical trial has been completed if it was funded by the government. So that is sitting on a server somewhere or is it a public website?

There's no government server where, like, it exists that's like all put together. So, like, we have a group of people. So we aggregate data from lots of sources to train from. And we love clinical trial data because there's this high quality, amazing data sets. So, like, we have a group of people who, like, call up professors at universities and are like, hey, you ran this clinical trial two years ago. So the data must be public now. And then we aggregate the data. Yeah.

It's crazy to be the government doesn't do it. Like, people will do in journalism, this freedom of information acts to, hey, listen, the government arrested this person, their documents available, JFK assassination. You know, give us the information. The government will release some percentage of it or whatever. you can actually start going and getting this information. Yes. And then putting it into AI models, this is something that we should have a Manhattan project on where some organization is paid like yours or another.

It could be private sector, public sector collaboration to make a database of anonymized data of every clinical trial that's gone on. Then you could just set the AI on it. 100%. I 100% agree. Yeah. And right now, the other part of it is the industry. sponsor trials.

So there are pharma companies who are running all of these trials. They own the data from those trials. So that's typically how that's set up. They pay it for it. One could argue that maybe the patients should own their own data, potentially, but individually, they should. Yeah, they don't, but that's another point. They really don't.

They don't own a dual license to it? No, in many cases, they're not given it at all. Yeah. See, that's something that... Some patients never find out. whether or not they got the placebo or the real drug. Okay, so this is somewhere where, like, our government's not going to get this done because they're bought and paid for by pharma.

I said that, not you. But this is something where the EU could pass something where they just say, listen, your data, you get a copy of it. You get to know that. Yeah, I mean, or the regulators, right? The FDA could say that, you know, if you want to submit your drug, you have to also submit anonymized data. It's going to go into a database, right?

Yeah, there's a huge amount of opportunity there because there's data from, right now, every year, about one million patients participate in clinical trials across the board. So if you think about every year, there's one million people participating in that level of experiments. And the data are not really being collected. But that's where what we focus on is learning from those style of data. In a country where you have socialized medicine like Canada, let's say, the Canadian government with a pen stroke, Justin Trudeau and could just

say all this is put into an anonymized database all the data or just give people a choice hey if you want to get free health care

you have to give away some amount of data to the collective good anonymized or it can be opt-in but I think if you're giving

socialized medicine it's not too much to ask that your blood results which the government paid for get to be put anonymously

your name your approximate region you know, maybe ethnicity, DNA, whatever, if you, or some amount of it gets, I gotta think this through because it could get a little dystopian if it gets complicated, yeah. It does get complicated.

The state owns your data. Yeah. But there is some trade of services here. So in a commercial country like the United States, it could be, we'll discount your rate if you put it into this pool for future research or like organ, organ doning.

You could just do it out of the goodness of your heart. And in a socialized medicine, they could say, listen, we just want everybody in the, country to give their blood data to this research. I mean, there could be a way to do it in a very positive way. Is anything like that even being considered these days or no? No.

Why is it so obvious to us and not everybody else? Yeah, I mean, it's, yeah. It's so obvious. Yeah, it's difficult. I think that there are a handful of countries that have better medical records where, you know, if you have a national health system, it's easier to have a national medical record system. But it's still not the case that these are like a repository of people taking part in these kind of research studies where you have a really rich, much more rich

information about those people than you do a normal, like just from your medical records. All right, everybody. Our friends from Microsoft are here. Tom Davis, a senior director at Microsoft for startups. How long has Microsoft been working on this cloud that you've now sort of uncovered and offered two founders. It's been years in the making, so to speak. The evolution of AI has taken many twists and turns in its journey. These large language models have really been the game changer. And that's really thanks to Open AI and the work that they've done.

And it's obviously our partnership there has helped us really get ahead of the game on this. And we're seeing great companies like perplexity. a.I. In six months, they've built out an application that has now got millions of users. that wouldn't have been possible in a more traditional way of working. So it's great to see the innovation that startups are able to bring to the table now and not have to make these huge investments in time, resources, and basically cash as well, which is always at a premium when you're starting off your own company.

The Founders Hub that Microsoft provides offers $150,000 in Azure Cloud credits, all the development tools like GitHub and Teams, Office, all that great stuff. You get all that for free. Five minutes to sign up, six figures and benefits, a.k.a.com. slash this weekend startups. Thanks so much, Tom. Tell me, how does your company make money?

Because you are a startup, you raised a series B, I understand. You've done pretty well for yourself here. PCs are placing a big bet on you. What's the business model? So we actually have a relatively simple to understand a business model. Our value proposition for a pharma company

is that by working with us, your trial can be months, months shorter, many months. So it depends a little bit, let's say six, two months to a year shorter. And if you're a pharmac company, your patent clock actually starts when you start your clinical trials. So every six months shorter or year shorter, that's an extra six months to a year of on patent sales. So it's billions of dollars in revenue for the pharma company,

if the drug's successful. And as we said, that's like one in ten. So how do we do that? Well, the way we're going to do it is we're going to allow these pharma companies to run clinical trials that have smaller control groups than normal trials. So you have fewer patients that you need for your trial. So let's say you need 100 fewer patients in your clinical trial. Well, there's a couple of things. One is that, you know, that, again, it might take six months to find 100 patients who are willing

to participate in your clinical trial. So right there, you've saved a whole bunch of time. But pharma companies also pay about $100,000 per patient in their clinical trials. Whoa. Yeah. Yeah, yeah. So that's where that $100 million number comes from you. Yeah, yeah, yeah.

You're at $100 million. Yep. Yeah, yeah, exactly. And the patents are, I think it's pretty standard, 20 years. 20 years. And so you're talking about a couple of year trial. What did you say? Five years.

Five years. So you're at 15. if you were to get them that extra year. Oh, that's just one trial. That's one trial. You still have your phase one. You have your phase two. You've got your trial.

By the time you get it to market, how many years you got left on the pen? Probably like 10 or 12. So you have 10. If you save them one year, you get 10% more money. Yeah, yeah, that's right. Or 10% more time to exploit the drug.

Exactly. Yeah. And not only, I mean, you know, that's the capitalist way. also frame it and say, well, there's a whole group of patients during that year who needed a drug, who now get access to it, right? Because otherwise, if it was, you wait another year, say you have, you know, groups of patients in a disease where people are dying, right? If that drug's not available, all of those people are dead. I know people with cystic fibrosis, and this is an area

where it's particularly acute, and they've made incredible progress, but these drugs are extraordinarily are really expensive for a very small number of people. And there is some compassionate use of this, but it is really a challenging dynamic. Maybe you could talk about the long tail of diseases and how this should apply. Because that does also seem to be something unique. We have the big four horsemen, you know, of, you know, diabetes and cancer and whatnot that kill people,

Alzheimer's, I think, is in that group. But there's the long tail. So is this going to have a dramatic effect on the long tail as well? We think that this should be used in every single clinical trial period. Okay. Yeah. There are challenges. I would call

technical and data challenges to getting there. For all of these small diseases, that also means there's a small amount of data to learn from. Right. So if we're talking about Alzheimer's, so many people have Alzheimer's, we can be really, really big data sets. But you start talking about, I don't know,

something like cystic fibrosis is a lot smaller population. And so we still need to have enough data to train a machine learning algorithm. But we are working on that all the time, how we can do better with these small populations. And over the next few years, we want to be able to roll things out across everything. Actually, our ultimate goal, right now,

basically the way we do our machine learning is there will be one model per disease. So we have like a model for Alzheimer's. We have a model for ALS. We have a model for multiple sclerosis like that. We want to build one model for everything. One model for like all human health. That's an extraordinary mission.

I mean, this is very hard, but. It's kind of like AGI versus vertical, right? Like you're, we verticalize cholesterol or heart disease. You know, you need a certain data set for that. But Alzheimer's might be. overlap 50%, but not 100%. Am I, my ballpark correct here? Exactly.

You know, that's right now, yeah. So you get these specialized data sets. We build specialized models, but the whole point of it is, I think that in order for us to get into these smaller disease areas, we want to have something that looks sort of like a foundation model for health. So one of the things we talk about these large foundation models today doing is that they can do either zero shot or few shot learning. And what that means is that they can learn to predict things having seen one example.

So instead of having to give it like, oh, we need a million examples for you to figure it out. So here's one example, what would happen in these next few examples. And so we want to probably be able to build something similar where for patients, even with really rare diseases, you can still figure it out from just a couple of examples. That's extraordinary. in a way it's like the foundational models of chat GPT or stable diffusion and in some dolly and all this stuff. My understanding is people think you're going to be able to fit this on your smartphone. And so the model will eventually be on a chip. And when that happens, that's going to be pretty wild that you're just like you have a Wi-Fi chip or, you know, a graphics chip on your phone or computer.

the concept of having an AI chip on there that just yes is the next word in a sentence and it's kind of starting you on third base every time you could do that for health. Then my watch, my Apple Watch might have this built into it and it'd be like, wow, we're seeing something with your heart. We know what you ate. We have your blood sugar level because you have a continuous glucose monitor.

It can be like doing stuff in real time. Forget about trials. You can be doing real time interventions. Yeah, there's all kinds of stuff that are really interesting. And again, hard problems, but I want to know not just what is happening with me today, but what will happen with me in the future. I want to have something that can predict the state of my health in the future, depending on what I do today. If I change my diet in this way, how will that actually really affect my state of health over time? If I take on this different workout plan, how will that affect my state of health over time?

and it's a super duper, duper, duper hard problem. You talked about how complex, like the human body has 37 trillion cells in it. It's actually 100 times the number of stars there are in the galaxy. So it's like a really, really complicated system. It's actually so complicated.

I think that AI is going to be the only way we can tackle it, right? It's too complicated for us to try to build up piece by piece. So I think that AI is really going to be fundamentally the new language of biology in the end. We are going to describe biology in 10 or 20 years entirely in terms of like AI algorithms that are learned to understand and tame this complexity. And once you get to that point, yeah, the idea of you have your own digital twin that's on your computer that talks about your health and maybe your doctor also has that same thing. You don't even need to go to the doctor's office anymore. They just pull up your digital twin and they can see what is happening with you today and what's going to happen with you in the future and they can design treatment implants. Maybe it's even an AI doctor, but I think that's the future.

You already have this happening in again, back to vertical AI versus general AI, which is analogous to what we're talking about here. You already have in X-rays. People are starting to build technology to look at the X-rays or to look at like hard. rate monitors over time and just highlight stuff that then goes to a doctor and that's augmentation. And so what I've been really thinking about in this AI future, because this is moving rapidly. You've been doing this for six years. How would you describe the pace we've seen of the past year compared to the decade before? It's definitely moving faster. It's interesting in like, I think what's happened to more is that

we finally reached a threshold of utility. So things seem like they are moving really fast when you're near a threshold of utility, even if they're moving slow. Because if you just stay a linear line, you just increase by X a little bit every year. But there's some threshold at which you need to pass before people care about it. You will seem like no progress has happened. And all of a sudden, you'll pass that threshold. And everybody like, wow, amazing progress.

And I do kind of think that's where we're at, that the AI research has been pretty steady progress over the past 10 or 15 years to bring us to this point. But it's all of the sudden got good enough that we're willing to use it, right? I mean, if you think about like GPT3, the API to that was released three years ago. Right? So it's not like we're in some exponential speed up of like Terminator world because that was a three years waiting period between three and four, right? So that's actually not that fast.

It's more that what's happened is people have figured out, oh, wow, these models are actually able to solve stuff now. And we can build applications on them. And so now it feels like it's incredibly fast because this is all these applications because before they weren't useful, and now they are. And so it's really more of across that threshold of utility than I think a real like speed up in the research. The research is kind of the same. Yeah. There is something perhaps to humans using it, finding the utility. and then the reinforcement learning

or these GPT starting different language models, different AI instances learning from each other that also is, once you get humans using it, it's like, well, GPS is really interesting

for sending a missile or tracking a plane, but it's also pretty good at finding a bakery or getting an Uber, right? It's like the GP, the street finds it's used for technology is what William Gibson said.

And, you know, it's like that really feels like what's happening once you put language models into a chat format. It's just, or you start building auto GTPs or plugins. It's like with streets figuring out all kinds of interesting use cases for it. Hey, listen, thanks for doing this work. And on the revenue question, since you save them that extra year, you just want to take a percentage of that or take a percentage of how much less people can be in the study. Is that the ultimate goal? Yeah. Yeah.

So that's why I was saying earlier. Yeah, the business model is relatively simple. So if you're paying $100,000 per patient and we remove one patient, you should pay us $100,000. If you remove $200,000. So we just get paid based on how much smaller we can make your clinical trials. That's our main business model. Or maybe you split at 50-50, so they get a little savings. Yeah, we try to take them, which again, they're getting billions of dollars in, say, an additional sales.

So they're getting a great deal. by working with us. It's an example of our work in clinical trials, I think, is a really unusual example of something that kind of everybody wins from because the pharma company, they definitely benefit, right? They can make a huge amount of additional sales. But the patients also very clearly benefit because you have a smaller control group and you have a faster time to market for the drug.

Got it. Even the regulators and people benefit from this because it's a use of AI. Well, actually we can prove that the clinical trials produce the same rigorous, scientifically rigorous results. So kind of everybody benefits from this technology. I think that there's going to be a lot more areas in health where this is the case, where technology is just going to totally benefit everybody. I think the difficult part of building in this space is that, you know, it's kind of this legacy

conservative industry and trying to figure out how to get people to trust and adopt new technologies is hard. Yeah, you know, we have the Wikipedia as an example of a foundational data set and the DBPedia that's been kind of built off of it. It's really helped train these models. Is there an equivalent in your world? And if not, would that not be something noble for the government to work on a way of saying, hey, let's find 10,000 people in the United States and give them a battery of tests for their lives and really get that data set and open source it to the world to learn from? I don't think that there's one. There have been attempts to kind of go in that direction.

The UK Biobank is an example of something that kind of starts to look a little bit more like this, which is the NHS's version of this. So exactly that NHS is like, hey, we have a national health system. We could create a big open source data set for everybody. And there is a big open source data set. Verily also in Google had tried to run something they called Project Baseline. I don't know what its current status is, but the whole idea was enroll 10,000 people

into a big observational study and follow them for a bunch of years and collect all this information and then we'd have this data set to learn from. That was Verily, right? Verily, yeah, yeah, yeah, exactly. That was Google's Live Forever healthcare thing? Well, I think you need something to learn from, right?

Like kind of what you're saying. So baseline is like, hey, let's collect this data set, let's build this thing that we could learn from. Again, I don't know what the status of that is. So there have been a few different options. but I think that I 100% would support the U.S. government trying to build a similar type of data set. Yeah, the ability to reduce suffering, extend health span. Maybe we don't add years to life as Peter and his new book has been talking about.

I guess this is, how you pronounce his last name? You know, he talks a little bit about health span versus lifespan. Hey, you live the same number of years, but you're skiing in your 70s and 80s or riding bikes in your 90s. it feels like we're on the cusp of something very interesting here. And so just on behalf of humanity, thank you for choosing this for your entrepreneurial journey. And that you're doing God's work. Or if you're an atheist, you're doing humanity's work. So pick whichever you like, no judgments either way.

Thanks for coming on the program. And maybe we can catch up in a year and hear how are you doing next year with this. Yeah, sounds great. Thanks for having me. All right. Cheers. Thanks for coming on the program.