This Week in Startups - Next Unicorns: Bringing AI-powered solutions to medical analysis with PathAI CEO Andy Beck | E1780

Episode Date: July 19, 2023

This Week in Startups is brought to you by… Eight Sleep. Good sleep is the ultimate game changer. Now you can add the Pod Pro Cover to any mattress! Go to eightsleep.com/twist to check out the Pod Pro Cover and get $150 off at checkout! Carta now lets you launch and administer SPVs for your syndicate. Share your knowledge, capital, and network to launch your syndicate SPVs through Carta. Get 10% off your first SPV at carta.com with promo code TWIST. LinkedIn Marketing. To redeem a $100 LinkedIn ad credit and launch your first campaign, go to linkedin.com/thisweekinstartups * Today’s show: PathAI CEO Andy Beck joins Jason to discuss AI in the medical field (1:58), demos his product, which uses AI to analyze millions of cells (15:37), and much more! * Time stamps: (00:00) PathAI CEO Andy Beck joins Jason (1:58) How AI aids with health diagnosis (5:46) Analyzing biopsy images and labs switching to digital (7:43) How images and annotations from Pathologists are used to train the AI model (9:58) Eight Sleep - Go to https://eightsleep.com/twist to check out the Pod Cover and get $150 off at checkout! (11:30) Resistance to machine learning in the medical field (15:37) Andy demos PathAI’s AIM-PD-L1 algorithm (21:01) The measure it manage it phase and the road to more effective diagnosis (26:01) Carta - Get 10% off your first SPV at https://carta.com with promo code TWIST (27:27) The past 50 years in cancer diagnosis and looking into the next 10 (33:49) LinkedIn Marketing - Get a $100 LinkedIn ad credit at https://linkedin.com/thisweekinstartups (35:05) The use and scale of datasets and creation of systems in a capitalistic society (45:40) How PathAI is able to minimize hardware constraints (48:22) Democratizing access to health expertise * Check out PathAI: https://www.pathai.com/ Follow Andy: https://twitter.com/andybeck * Read LAUNCH Fund 4 Deal Memo: https://www.launch.co/four Apply for Funding: https://www.launch.co/apply Buy ANGEL: https://www.angelthebook.com Great recent interviews: Steve Huffman, Brian Chesky, Aaron Levie, Sophia Amoruso, Reid Hoffman, Frank Slootman, Billy McFarland, PrayingForExits, Jenny Lefcourt Check out Jason’s suite of newsletters: https://substack.com/@calacanis * Follow Jason: Twitter: https://twitter.com/jason Instagram: https://www.instagram.com/jason LinkedIn: https://www.linkedin.com/in/jasoncalacanis * Follow TWiST: Substack: https://twistartups.substack.com Twitter: https://twitter.com/TWiStartups YouTube: https://www.youtube.com/thisweekin * Subscribe to the Founder University Podcast: https://www.founder.university/podcast

It's just super exciting the way this whole field will be transformed. As you mentioned, it's kind of like two big platform shifts essentially at once because most of it is not yet, like you mentioned, microscopes aren't even connected. So just the basic advances of digitization, creating an inexhaustible digital resource that you continue learning from without destroying any tissue over time. Connecting all of these brains together across the world is just a huge platform advance in itself. The advantages of cloud. and digital being connected.

And then at the same time, you know, as that's being transformed, we're getting the substrate for training all these new models. Amazing. That can do things at lower cost, more accurate, more reproducible and more predictive. So like this is like, you know, where things are going. And I agree with you within five to ten years, you know, the world of pathology and diagnostics for drug development as well as for clinical care, I think we'll be totally transformed by this.

This week in startups is brought to you by eight sleep, good, Sleep is the ultimate game changer. Now you can add the pod cover to any mattress. Go to 8Sleep.com slash Twist to check out the pod cover and get $150 off at checkout. Carta now lets you launch and administer SPVs for your syndicate. Share your knowledge, capital, and network to launch your syndicate SPVs through Carta. Get 10% off your first SPV at Carta.com with promo code twist. and LinkedIn jobs. A business is only as strong as its people, and every hire matters. Post your first job for free at LinkedIn.com slash unicorn.

All right, everybody, welcome back to this week in startups. I'm really excited for our next presenter here, our next founder. His name is Andy Beck. He's a CEO and co-founder of Path AI. And Path AI is a research platform. It's designed to improve the speed and accuracy of cancer diagnosis and treatment. Every time we have a conversation about AI, the past decade, the incredible example people like to give is doctors during cancer or identifying cancers or looking at scans and getting better at it. So we thought we'd find that person. After a bunch of research, we knew somebody was working on it. The name of that company's Path AI. And we're featuring them today on our next Unicorn series. Andy, welcome to the program. Thank you.

Great to be here. You heard my little preamble there about somebody has got to be working on this because they always give the example of using AI to help with cancer diagnosis. Cancer is, I don't know if it's the leading source of death here in the United States. I know it's up there with heart failure and obesity, right? Those are the top three. I'm not sure where it stands today. But how are you using AI to help with? diagnosis and treatment. Yeah, no, it's a great question. And definitely part of the reason we sort of

formed this company about seven years ago was, you know, really thinking about what are the, the killer applications of AI and the big advance back then was deep learning as it still is. You know, they could really make a difference for people. And really the first big area of deep learning where it made a huge impact was the interpretation of images, kind of imagery, deep convoluting, Neural Nets really created this incredible advance in the field. And then images play a big role in medicine in a couple different areas. So if you picture your patient, you go into the doctor, you might have a cough and have some risk factors.

So they might say, you know, I don't know what's causing this cough. It's either something totally that's going to get better on its own or it could be a malignancy that, you know, you want to deal very aggressively with to cure the patient. So the next step would be go to radiology and they'll take images. And in the bottom of the radiologist's report, they'll say, you know, impression, I'm concerned this is something bad like cancer or impression. This looks totally benign. Let's just track it. But if they really want to get the definitive diagnosis to guide definitive therapeutic decisions after, what's going to kind of be considered

the ground truth for what the diagnosis of that patient is, they actually stick in a needle to take a piece of tissue and they send that tissue to pathology. And then the way pathologists examine the tissue is a process where it's embedded in a wax block. a thin five micron section of tissue is cut off the block. It's put on a glass slide. And that's put under the microscope. And the definitive diagnosis of cancer really is defined by bad malignant cancer cells invading into normal cells. And that really is an image-based diagnosis. So no matter what's happening at the gene mutation level or the art expression level or even the protein expression level, it's kind of like cancer is defined by bad cells invading into good cells. And you need to actually

take a picture of that or look at an image of that through a microscope. So those are the types of images that we analyze at PathAI. So the input to our system are what are called whole site images, very, very large pictures of tissue biopsies that have been stained, stained in such a way that we can see the cancer cells, see the normal cells. And then the AI's job is to look through these very large images with hundreds of thousands to millions of cells, identify what every cell is, identify all the normal cells, all the malignant cells, to really accurately and reproducibly diagnosed diseases like cancer.

And in certain cases, if patients have cancer, to try to make predictions on how aggressive the disease is and which treatments it may respond best to. So cancer is a big majority of what we work on, although we also work in other diseases that involve tissue biopsies like liver diseases, as well as inflammatory bowel disease. So when you get these images, there are giant images. They're created by what type of machines are best for doing cancer scanning? Curious? Yeah.

So for these images, so I would just say just so, you know, kind of where the field is largely in the diagnostic side, many labs are still using microscopes to generate images and they're not captured digitally. So someone's just sitting, they've got on one side of them a big stack of glass slides,

and then they're just sitting at their desk and looking at each slide into the microscope, and then, you know, interpreting this is what I think I see in the image, putting that in the report, and that's it. But increasingly, labs are becoming digital, And to answer your question, there's whole site imaging systems that essentially are like microscopes in an imaging solution where they can create these microscopic images, capture them. And then, you know, the digital file can then be sent to the cloud and analyzed.

Got it. With image processing. So this is not, we're talking about the biopsies where they actually take a little piece of material from your, let's say, lungs or liver, I guess, for cancer. And then they analyze it, put it on that swab or whatever. those typically have all been 100% manual and those aren't even online microscopes. Wow, that's crazy to think in 2023 that microscopes are not digital

and hooked up to a cloud server somewhere. What percentage of this is done offline versus online today? It depends by country, but I think... Go with the U.S., yeah. Yeah. So I think overall about 90% today is done offline and about 10%. Oh my God.

This is not even with algorithm. This is just digital. Plain old digital. Wow. That is extraordinary when you think about it. It's almost like we're going to look at this in five years and think we were in the stone ages. Just like having to go send your, when you get an x-ray, like going to the dentist, remember in the old days they had to send them out? Like when's the last time they send out an x-ray?

They're all done right on the spot with digital imaging now. So how many samples are in the cloud? and is there like a national repository of lung cancer? And you could look at it and say, I have a smoker who's 50 years old, report smoking since the age of 15, so the 35-year smoking history, two packs a day, and their Norwegian and Irish descent,

and their BMI's 30. Here they are. Do we have that kind of fidelity yet in any kind of database like that on a national or international basis so we can actually start to do some analysis of cancer across some reasonable number of people? Not much.

I mean, I would say there are different initiatives to try to do that, but it's such a massive problem. I think we're only at the very, very beginning of putting those sorts of resources together. It's something we very much invested in building ourselves. So I think in terms of another sort of the AI theme, like what are some of these areas that are not yet on the internet and not yet captured digitally and aren't being incorporated

into some of these massive models that are being built off public data? Because this is an area where there's very, little digital data to begin with, relatively to the scale of the problem. And then, you know, a very small part of that that's in the public domain. So there is a need for companies to do a lot of this work themselves and to work with great partners who are making investments in digital. And, you know, 10% of a very large health care system is still a pretty big percentage. So, you know, we're able to partner and generate data ourselves to create these very, very large sets of data. The good thing about pathology, though, is each image is so data rich. These are massive images that contain on the

order of hundreds of thousands to millions of cells per image. So even with, you know, data sets that aren't super massive, we can use deep learning and obtain annotations from expert pathologists to train models that work extremely well. So we've invested heavily in not just the whole site images, but also creating a community of almost 500 pathologists across the country who are tied into our platform for helping to train these models, which are now trained on the order of tens of millions of annotations provided by both the images, but also by the expert pathologists. If you want to get ahead in your career, you need to sleep well. And if you want to be a stallion, if you want to be a workhorse, you need to get a great night's sleep. It's that easy.

So do what I do. And that's getting an eight sleep, right? Especially over the summer, right? Temperatures are going to rise. You want to stay cool. And eight sleep, mattresses allow you to set the temperature. And they're selling more than just mattresses right now. They have something called the pod cover by eight sleep. And this fits on any bed. It's just like a fitted sheet. So if you don't want to get a new mattress, but you do want eight sleep, you just get the pod cover. It's going to keep you cool all night, all the way down to 55 degrees. And it's going to improve your sleep by automatically adjusting to the temperature on your side of the bed. And, you know, your partner or spouse, they get to set their own temperature. So no more battle with your partner or your

spouse over the thermostat. Nope, it's going to adjust based on you and your partner's preferences. And you get personalized sleep reports. Eight sleep, I love it. In the winter, oh, I keep it nice and toasty. But in the summer, I like those cool crisp sheets. And I don't want to have to blast the AC in my house and burn all that extra fuel. No, I just use my eight sleep. I dial it to exactly where I like. And, you know, I get a better night's sleep. Temperature really does correlate with a good night sleep. So go to eightsleep.com slash twist to get 150 bucks off the pod cover. Stay cool this summer with Eight Sleep, please. And they're now shipping, not only in the U.S., but Canada and the UK, and some other countries in the EU and Australia. So, 8Sleep.com slash twist for $150 off the pod

cover. Can't go wrong. I would suppose one of the solutions here would be to give your 500 people participating. I'm curious if you would give 10 of them the ability to mark up the same patients, or if you do this, the same patients slides, as it were, and then see if you can find some sort of trends of 10 doctors, because this is what we all want as patients. If I had some loved one, God forbid, with lung cancer or my dad had prostate cancer, my mom had breast cancer, both survivors, which is extraordinary when you think about it, 30 years and 15 years for each of them, 45 years combined, wild. But imagine I could have had 10 folks look at my dad's samples and then give their opinions

and I can say, well, five are saying this and three are saying this and two are saying this. When you come as a technologist, and listen, you're, I believe, a PhD yourself on this, and you talk to medical professionals, do they have a God complex or do not want to be one of ten people giving blind advice on something

and then having you as a technology company second-guessing it in their minds? Obviously, you're not trying to second-guess it. You're trying to save people's lives. a spirit and culture of this, hey, we should get better at our jobs. What's the resistance like in this field? Yeah, it's a good question. And I kind of had a bit of both. I trained first as a pathologist and then did the machine learning after. So at least I've worked very closely with pathologists for a long time now. And I think pathologists truly come to work every day

wanting what's best for patients. And the one thing that would keep a pathologist up at night is is giving the wrong diagnosis and not having the best diagnosis for, you know, their own parents or your parents. I mean, that is what, why everyone comes to work every day. So I don't think that's the big barrier. I think it's more technology hasn't existed for that long to really move all these images around to have this network of pathologists connected through technology through a platform is very, very new. Of course, the deep learning and image analysis is very, very, very new and in and just the overall incentives for making that transformation happen in the health care system is also a challenge. So I don't think it's individual pathologists who are our barriers

to this. Yeah. I think it's really, you know, getting the technology out, getting into the hands of users. And I think, you know, I think the future you outline, we could do that today, even without algorithms. And we do do this in clinical trials in terms of having consensus of experts who support a diagnosis. And that really has been unlocked by this technology platform. Because literally before people were doing this, and even still today, people are sending around glass slides. So imagine how long it would take to send around glass slides to 10 people,

you know, versus distributing the images. Yeah, I mean, it basically kills it immediately because it's just the cost and getting them to town. And then also your 10 experts might be in 10 different cities in three different countries. So there's a real opportunity here. I believe that if people could, you could upsell people on this as a, like there are a lot of affluent people. Uber started with black cars, not UberX.

I mean, there are rich people. We'd be like, I can get three more opinions for $10,000 each. I'll do it, right? So there is a, I mean, as perverse as it is when we talk about healthcare, we can talk about the profit incentive here. There is a profit incentive here.

People go for second opinions. But the people go for second opinions, correct, but I'm wrong, tend to be rich people. who are paying for their own health care and don't care as opposed to poor people who have to stay in network,

etc. So maybe we could explain or maybe we should do the demo first but then I want you to explain economics to me of how the industry works. Let's do a demo. Sure. I think that's the most interesting part of this.

And just looking at my notes here, 2000, 2021, heart disease, 700,000 deaths, cancer, 600,000, COVID-450,000. And who knows how many of those were with COVID or from COVID, obviously. Disclaimer. That's the United States.

data. All right. So, and here's the, the demo from, you know, you just mentioned the cancer statistics, the cancer with the highest, responsible, I believe, for the highest number of deaths each year is still non-small lung cancer, despite all the really major advances that have been made, particularly in the last five years, one area of huge advance has been these new therapies called immunoncology therapies that target either the PD1 or the PDL1 protein that have really had massive advances in increasing five-year survival for patients with lung cancer.

So one of the most important factors for deciding which immunotherapy regimen a patient should get is their protein expression of PDL1. So this first algorithm that I'm going to show in this demo is AIMPD-L-1, AI-based measurement of PDL1 and non-small lung cancer. And this is actually a demo of our AI site platform. So here you can see sort of the opening page with a case list in the dashboard. So if we pick one of these cases, is that's awaiting review. We start out with shown on the right here in this display is a whole slide image in the center here of a resection of a portion of a patient's lung tumor here in the middle,

as well as we have a sort of guide for where you're looking at on the top right. And for people who are watching, I look like I'm in my Gmail inbox, certain extent. Standard navigation on the left and the top and then the main body in the top right and the main body of where your email would be. And we're looking at a sample of somebody's lung, I guess, lung tissue. Yep. Lung tissue is including lung cancer in it. And this would be for a patient who's already been diagnosed with lung cancer. And the pathologist's job now is to look through this sample, which if you just look on the top right, where you can actually see where I'm looking.

That's like your heads-up display in a video game. That's like the map view. Yeah, the map view. So you can see we're already looking at dozens of cells. And this little red dot here indicates how much of the sample we're looking at. Wow. So we've zoomed in. This is equivalent.

It looks like Australia, the sample. If you imagine, if you're listening, like Australia is a map. And then a little dot that would be Sydney. It looks like there's a little city. And then when you look at it, you're seeing a bunch of little blue dots on an orange and white background. Exactly. And then there's clusters of. darkness. I'm assuming that's not good, the darkness, but I'm taking the guess.

So the pathologist job here, yeah, is to estimate what proportion, so you outline these little blue dots. So there's a few different types of little blue dots. Most of these are cancer cells. Some of these are immune cells in the middle. And the pathologist's job is to estimate what proportion of the cancer cells are brown are expressing this brown protein. And literally the clinically, there's a few clinically important. cutoffs, but they include 1% and 50%. So just so you know how difficult this manual task is... Oh my God.

Really, the job is to enumerate how many positive cells there are, how many negative cells there are, and then what's the overall percentage... And they're doing that by eyeballing and counting with a pen on their screen or something? Yeah. You're kidding me. This is the state of medicine today? With computer visualization where it is, this is something that AI could just immediately give you a heat map of.

This would be like using the topography view to plan a drive in Google Maps or something and then having to look at each house to find the address on the front of it using Google Street View. It's barbaric. It's extremely difficult. Whereas our system has been trained now on millions of annotations from hundreds of thousands of slides, now can be forced to exhaustively classify every single cell based on, is it a cancer cell? Is it an immune cell? Is it positive?

Is it negative? And does it come from the cancer tissue here outlined in red or from the supporting stroma here outlined in this kind of yellowish color? And then we can get an exact count of what's going on. So you can just see how this would be an impossible task to do manually. Like here, we're still just looking at this tiny box versus, you know, we can actually have the machine learning system do this for the whole image, do it behind the scene. to when the pathologist sits down at their desk,

unbelievable. They basically can say, is this slide acceptable? Sure. You know, what do I think the results are? You know, this says, okay, 1.5% cancer cells are positive, which is really important because it's more than 1%, but it's less than 50%. And then you can actually see the supporting results

in terms of, you know, looking inside the black box of, well, how did it come to that number? There's almost 12% positive immune cells. there's about 80, almost 82,000 cancer cells, almost 380,000 immune cells. And then you get this exact proportion of positive cancer cells, positive immune cells. You know, you can spot check a few of the areas and then, you know, accept the score. So this is, you know, our goal is for all of our algorithms to make pathologists more accurate, more reproducible as well as more efficient.

So there's this quote, if you measure it, you can manage it. You know, like if you actually know your weight and you weigh yourself every day, you know, you might be able to manage it a little better. Or if you're Steph Curry hitting threes, you can look at the arcs now. There's a three-point system that basketball players use. The Knicks famously used it three or four years ago and the everybody on the team got better at shooting threes because they were looking at their arc and their stance and y'all. all the best practices there of measuring stuff.

So cancer diagnosis and then the treatments that come after it, we are in the measure it so you can manage it phase of this solution. And that is in and of itself extraordinary to me as a neophyte that we're getting off of essentially people doing this manually on pen and paper and writing down on a legal pad and typing this stuff out. It's just crazy. But if we can measure it and then you put a, if you have this level of fidelity when you are measuring stuff,

well then if you do another, is it called a biopsy when you take the sample? It's like the proper term. We do another biopsy and you do biopsies every year. I don't know how invasive they are. But if you did multiple biopsies over multiple years for somebody with lung cancer or multiple months, would you not be able to then measure the efficacy much better of treatments? Absolutely, 100%. And even if you just do a pretreatment and post-treatment.

So for many serious cancers, as well as another, I can talk about liver disease, there's a biopsy that makes the diagnosis and then you can actually give treatment. Like if you're diagnosed with early stage lung cancer, a real option now is to get a biopsy, to get a treatment and then to analyze the resection specimen. So in a single patient, you can analyze what's changed from the biopsy to the resection and what does that mean about how effective this treatment's going to be for curing this patient of their disease. and should we stick with the treatment we used in the so-called neo-adjuvant setting, the prior to surgery setting, or should we switch to a new treatment? And you can't really inform that analysis with manual scoring in the way that you can with AI, because now we can measure hundreds of important quantitative features of the image before treatment and after treatment in a highly reproducible, scalable way

that would just be impossible to do manually. So absolutely seeing the effective treatment. The other disease is Nash, where actually the primary endpoint, Non-alcoholic steatohepatitis. It's a very fatty liver disease. It's increasing in prevalence significantly. And the way that patients are either enrolled in clinical trials or we assess whether treatments have worked is a pathologist manually scoring things like inflammation,

amount of fat, amount of ballooning hepatocytes, amount of fibrosis. And there's a lot of noise in those manual measurements. And then you're trying to say, did it get better or worse after treatment? Well, if there's a lot of noise in the measurement, you don't know if it changed due to the treatment or due to the noise, that increases noise in the placebo group as well. And it makes it harder to assess whether new drugs are effective. Whereas applying a system like this, you know that, you know, at least the interpretation was highly reproducible between prior and post-treatment. So we think this will be very important for approving new medicines for this. So increasing

common disease. These cancer cells, these are PDL1, you called them? This example I showed, yeah, these were, this was a case of lung. cancer. And the way we subtype lung cancer is based on how much is it expressing this protein called PDL1. And a lung cancer patient could be PDL1 negative, which would have certain implications for what treatment they should get. They could be sort of positive at the 1 to 50% level, or they could have greater than 50% of their cells expressing PDL1. And there's different treatment recommendations for patients based on their level of expression of this protein. Got it. And the PDL1 protein kind of obscured.

purifies the cancer is there, it kind of hides it? Is that the... Yeah, it does. Yeah. So PDL1, it's in the pathway that's targeted by major drugs like Ketruda or Optivo, which are two of the major, relatively new cancer drugs from the past few years. And those uncloak it, I guess. They kind of make it appear. I know just enough about this to be dangerous. Yeah, exactly. So PDL1 sort of inhibits the immune response. to the tumor, which is kind of the body's way of trying to defeat the cancer. And if you give these drugs, which are antibodies against PD1 or PDL1, they inhibit the activity of either PD1 or PDL1.

And that enables the immune system of the patient to really be unleashed and then to really work hard in anti-cancer immunity and has led to significant improvements in survival for patients and curates for patients with serious cancers. It's like a shock grenade or something. You shock the cancer cells. They stop growing and then the immune system can do its thing and kill the cancer cells or get them out of there. Yeah. Listen, if you're in the tech industry, you've heard of Carter, right?

The leading venture capital and equity management platform, you probably hear about them all the time because they manage cap tables for your startup or others. But they have some huge news to share. Carter now lets you syndicate an SPV, a special purpose vehicle. You can create your own syndicate through Carter. What's an SPV? hey, you got 20 people, they want to invest in a startup, but they don't all want to put in 250K. So everybody puts in 5, 10, 15K, whatever they're comfortable with under one line item

on the cap table of that startup. Or maybe it's your startup and you want to do an SPV because you got all these friends, family, other founders, angels, but you want to have one signature to roll them all up, right? Well, Carter is used by more than 4,500 funds representing over $120 billion. and assets under administration. And they support you at every stage of the fundraising journey. From your first syndicate to building your own global VC firm, you can raise and deploy from anywhere because Carter offers US and international SPVs and fund locations. Carter provides an automated back office solution for you so you can focus on finding a great startup and building

those relationships. So here's a very simple call to actions. Go to cardi.com, C-A-R-T-A-com, and use the code twist. You get 10% off your first SPV. That's carda.com and use the code twist. This is extraordinary. In your estimation now, working on this, a lot of us like to say, oh, yeah, cancer's going to be solved in our lifetime.

There's a lot of reasons, treatments, measurements, early detection. Do you think it's 20, 30 years before most cancers? Because most cancers are treatable now. That was something crazy to think of in 1980. In 1980 or before 1980, you got cancer. You basically started your goodbyes and did your trust, your will, whatever. Now here we are in 2020. You get cancer.

They say, okay, here's your options. Here's your plans. And it's very rare that somebody gets a cancer diagnosis and they say, hey, you know, now it's starting to start your will and say goodbyes. It's that advanced. So amazing what progress we made in, what, 50 years, I would say, 40 years maybe, 40? you tell me, what is the next 10 years having stories? Is it possible the idea of dying from lung cancer and brain cancer will become very obscure,

like dying in a plane crash kind of situation? I think it will become more and more manageable over time. There's going to be more early detection. There's going to be fewer cancer cases as we treat more risk factors effectively for developing cancer, smoking being an example for lung cancer, that's quite important. I don't think it's going to be as rare as a plane crash. I mean, I think there's something about how fundamental making errors during cellular replication are and how important cellular replication is as you keep living life and dealing

with all of these insults. We're constantly getting on a daily basis from the environment, including sunlight and things we're ingesting down our GI track and things we're breathing in and hitting our lungs. So cells get injured. need to proliferate and recover. Errors are made because you're trying to copy this enormous genome and you're going to make mistakes. The machinery is going to start aging for finding those mistakes.

And cancers will develop. And the question is going to be. So it's like, it's way more fundamental than plane flight, which you don't do that often. And I don't know, there's probably technical ways to get around plane flights. I think cancer is sort of fundamentally tied to being alive for a long time. However, I think we can, you know, make the rates of it go down, start to prevent it, start to identify it early and manage it. We're already making headway there. And there's some really bad cancers that we have made less headway on that we need to do even more. So I think we're

still at the very beginning. And I do think it will become much more manageable over time with these advances. It's pretty amazing. I literally just was at my doctors and she said, oh, you know, there's this, you like all kinds of new stuff. There's a new test. You can take a blood draw and I will tell you if you have cancer. And I'm like, do we respect I have cancer? I'm like, no, no, it's just a blood test that tells you if you have cancer at this moment, not predicting the future of it.

And so I did it. And I get my results back at some point. I mean, I don't know. I'm 52. I don't know if I just, and I just did the pre-nova body scan. Nothing came up on that. This is the stuff that could really, I think, maybe you could speak to body scans. and the impact of these general tests, like a blood test?

I mean, at every age, if everybody did a blood test yearly and a body scan yearly, what would cancer diagnoses look like in that world? Yeah. So, I mean, I think there's a few things. One is just the things we already know about that aren't going to change that are almost like universally good for many things, including reducing cancer. Like as much as we can preventing obesity, encouraging, you know, exercising. even though I feel like the exercise cancer link is potentially not huge.

But I think there's just general like healthfulness stuff that doesn't cost money and has no adverse side effects and isn't going to result in a wild goose chase that you found something. You don't know what it is and you get unnecessary surgery, unnecessary biopsies. So to me it's like there's all this really basic essentially not technology stuff that can have. Yeah, exercise, vegetables. They can have a huge, huge impact. So I'd say that's one, just whole big area that we should not. forget about because I think it could just have a massive impact. In terms of screening, screening historically has always been tougher to prove the benefits in patient outcomes than we

would hope. So I just think there's a lot of excitement about screening, but we're still kind of waiting to see a lot of the data about improvements in overall survival from screening alone. And we do see it in some areas, but even that the impact size is relatively small. And the reason to think about it is you need to be finding things that would have otherwise, you know, killed you, but you're finding them in an early enough stage where they're curable. And those things do exist, but they're not as large a proportion as we might like. So the positive and why I am in favor and think we're not doing enough of this is you do find those things early.

The downside is, let's say you, you're, you know, only 52. You had a healthy scan. So your pre-test odds of having a cancer that's likely to kill you might be extremely low. and something shows up in your blood. And then you are in this kind of wild goose chase where you're getting more and more imaging, more and more biopsies, a lot of anxiety. And if nothing showed up on imaging,

yeah, I don't really know what the next step would be for you. So I don't know. I think it's complicated. I love it. I got to say, I know that I'm going to die. I'm convinced it's going to happen.

So since I've already accepted that I'm going to die, if I could get, even if they told me there's something here, we're not sure what it was, I would rather, because I guess I'm not like a prone to anxiety or like catastrophizing. I would rather go on the goose chase than not. I'm okay with a little goose chasing, you know.

But I think you're thinking holistically about the entirety of the population and like this impact it could have. Yeah. I think I took Gallery. G-A-L-E-R-I was a test I took. And pre-nova was the body skin I did. And then I don't know. Did you see that Daniel Eck, the Spotify founder?

has invested in a scanning company that only costs $400 or something, or he thinks he can get it to $400 bucks, which is pretty interesting in and of itself. All right, listen, you're listening to the next unicorns, right? That's table stakes. Well, if you want to start on the path to becoming a unicorn yourself, you're going to need to find and hire great candidates, and how do you do that?

LinkedIn, of course, you're on LinkedIn, I'm on LinkedIn. Over 900 million users now, the March to a billion continues for the amazing team over at LinkedIn. then you can attract both the passive and the active job seekers. And all you have to do is put that purple hiring ring on your profile. Then you post some interesting content. When they see that purple hiring ring, it says, whoa, wait a second. I know that founder.

I know that CEO. I know that VC. Well, they're hiring. Let me click and check it out. And then all of a sudden you start getting this great inbound, right? We love LinkedIn. I mean, I've gotten some of the most amazing people. In fact, I just got a new personal assistant because things are going so well.

And I've got the new accelerator that we have coming in. San Mateo, and I need somebody to help me run that and set it all up. And I said, you know what? I need to have an executive assistant again. So we found somebody amazing on LinkedIn. Of course. And so here's your call to action. LinkedIn jobs helps you find qualified candidates you want to talk to fast. Post your job for free at LinkedIn.com slash unicorn.

That's right. LinkedIn.com slash unicorn to post your first job for free. Terms and conditions apply because LinkedIn is so generous. Maybe you could talk to me about what you think. all of this data is going to do and if there's anything with general AI that's starting to happen that inspires you.

Because pools of data are starting to emerge. I don't know if you have an Apple Watch on, but I used Fitbit for 10 years, Apple Watch for five. I would happily give every bit and I have eight sleep for my bed. I would give every piece of data I have to the cloud,

including every blood test I've ever done. taken and I would love to be part of a national project to get all this data together. And I think there are some going on, but I don't think you need that many people either. I don't know what the statistical number would be, 500 people, 1,000 people doing this in order for y'all to have incredible data. Why is there not some national project? Why is it also piecemeal, I guess, is what I'm getting at. I talk to smart people like you. You have all the solutions. You need only implement them. Why isn't there like a national effort, like we did with the Manhattan project, to just understand the human body in relation to the new tools that are available,

i.e. AI, computer visualization technology, and make the proper data set for y'all to really have a go at this. Why is that not being done, or is it? Well, I think different groups, I think Gallard is a great example. That was largely funded by industry. They've done massive trials, incredible data, I think, and I'm very bullish in general on this technology that, like, this is the future and it's better than certainly what we're currently doing, which is much more reactive and less likely. I was just saying every patient, like you, instead of the, anxiety might not be a concern, but you have to like very much think about what's the potential risk of the biopsy I might have to get to confirm something versus the risk of having cancer

and just compare those risks. Just that, you know, everything has some risk. associated. So one just has to be really cognizant of what those tradeoffs are. But so I think the industry when they see an opportunity, you know, is going after this. I mean, pharma is extremely interested in non-invasive tests. There's large companies like Illumina with Grail who funded these major trials around gallery that I think you're going to provide very good evidence around that test. And I think it's been, yeah, largely, you know, many of these efforts, large-scale efforts have been private. There have been some great federally funded efforts like TCGA, the Cancer Genome Atlas. So I think it's a whole slew of different problems.

And, you know, many of them are being funded through industry or through the government. And it's not just one problem. One that actually sounds a lot like what you said is another privately funded, the project baseline by Verily, which is very much, I think, similar to what you're describing of just capturing lots of data from essentially healthy people and following them over time to be able to use machine learning over time. So I think there is a ton of activity in this area

and there's many different problems. And for some problems, it makes sense to put it all together, but for many, it sort of makes sense to try to solve that problem. Whether it's through industry or through government-funded efforts. It's really fascinating

when you think about ecosystems, you know, or just systems theory in general, which system will provide the most innovation and solutions, the capitalistic system, where people are like, you know what, so many people die of cancer, or these are the top three reasons people die, obesity, cancer, obesity, cancer related, obviously, to some extent. We just got to get Ozempic and Wagovi going in the country.

And the free-for-all of market-based capitalism is going to cure obesity, right? There's enough people who want to lose weight. That 800 bucks a month for Wagovi or OZEPIC, people are happy to pay $10,000 a year for it out of their pocket to not be fat, right? I did it and it worked. So it's fascinating that system versus a top-down system where maybe Europe is doing that brain project. I guess they're trying to study the brain or some of these larger projects.

And then which system ultimately, or maybe it's just many different systems running. Yeah. I mean, and the irony, too, is often these top-down systems sort of break into these fragmented projects anyway. So that's why I think there's $3 billion, but it's being distributed among, I don't know, thousands of labs. So, yeah,

so I, what do I think works? I think lucky for us, we've got a diversity and a really pluralistic set of people attacking this problem from, you know, government fund initiatives,

big pharma, biotech, and startup. So I think, you know, we're all going after it. Certainly, though, in terms of the data piece, you know, single, single payer systems or you know, top-down approaches for organizing patient data at scale where there's no obvious short-term economic incentive for doing it, it makes sense for that to be top-down. And I do think

we could learn from some other countries that have done an incredible job of that, of having, for every patient, they know their ID, they know what treatments they've gotten throughout their life, they know what diseases they've have, and they know their outcome. I 100% agree that that should not be so fragmented, that should be top-down, and it could have a huge impact. And you mentioned sort of scale and like what are we seeing. I mean, we're absolutely seeing on the pathology images. Like, once we crossed a certain scale of data, which we've been building now for about seven years, like these models have gotten so good and so generalizable across the full domain that we started in a very fragmented way. We need this model for this antibody, for this tissue indication,

whereas now we've built such a big data that we can build these foundation models that generalize extremely well across broad, diverse of tasks, and that was only accomplished through really this sort of very centralized vertical effort within pathology, where we generated enough data at scale to build that. So there's definitely like a scale effect where you really do need to invest in this large, harmonized, high-quality data. And it's probably in the low thousands of samples, I'm taking a guess, to get to statistical, you know, accuracy? It really depends on the application, hard, hard to say. I mean, for certain things, yeah.

Well, I was just to say, think about verticalized AI, you know, programming languages and, and, you know, GitHub, co-pilot, those kind of things. They were the first out of the gate to prove that, you know, hey, these language models are very powerful. So yours is a similar situation where it's a narrow data set with a, you know, if you get a certain amount of it. I mean, at some point, you'll have more than you need. And it's kind of like self-driving.

I was talking to Elon about this. And he feels they're really. close. And he's felt that way for a long time. But there is a tipping point where when you got a million cars on the road collecting data, he told me they have too much data now. And that the issue is managing how much data they have. And then they only need the data.

And he's talked publicly about this. So I'm not speaking out of school here. But they only need the data when the autopilot disengages or particularly challenging intersections or edge cases. They don't need the data on going up and down the 280 or the five-free. way to L.A., San Francisco. Like, we got it. It's a straight line.

The lines are painted really well, stay in your lane, go to a certain speed. It's when you're at some weird backroad or some funky, you know, going to the tender line there's people in the middle of the street, right?

The edge cases become that. Totally. Yeah. So we think a couple things on that. It's interesting. So we probably have on the order of tens of millions of annotations, hundreds of thousands of slides.

And a lot of it's the scale and the point in time of the thing you're predicting. So for cells and tissues, yes, we have a ton of data. I think it's similar to the self-driving car example. We can cover the vast majority of things every now and then, you know, there's an outlier where if something's for a clinical use case, that's really where you want the human judgment to make sure there aren't some catastrophic mistakes.

But like the vast majority of it's done. But what if you're trying to predict things that aren't in the image, but you're using the image to predict the molecular underpinnings of the tumor or even the outcome of the patient? So while there's, you know, on the order of hundreds of thousands to millions of cells, per slide, there's one treatment and one outcome, say, per patient. So the instances of patients are much fewer. So you actually need a lot more data that we're certainly not there yet to predict what's

going to happen at the patient level. And then as you go further and further out in time, you need even more data because you're trying to make predictions further into the future. I imagine self-driving is probably similar. Like, well, can I predict the risk that this person will get into an accident? Not in this minute, but five years from now. You know, and then you're going to keep capturing data to predict over time. You can look at all miles driven under autopilot.

And how many disengagement's happen per thousand miles? You know, and I think that number, and I don't know what number they use internally, but I know like the number of disengagement. Wow, if I'm going from the Bay Area to Tahoe as an example, it does not disengage on the highway. I disengage it when I get off the highway. And I don't use it on surface streets all that often,

although I've been testing it and it does work really well. And I guess in your model, it's like the surface streets or like your driveway, when you arrive at your location, there's like a private driveway or you're, you know, on some back road and trucky,

which I am sometimes. And it's like, yeah, a lot of people haven't been down this road. That's the example of, oh, this is like a 25 year old who got,

God forbid lung cancer, and it's like, and they don't smoke. Okay, we haven't been down this road before. How many 25 year olds with lung cancer are we going to find? It's not,

it's not an average condition. It's really fascinating. I think we're going to wrap all this up this decade. I know it sounds crazy, but I feel like at the advances that are happening, especially on the hardware level and then optimizing the hardware and the amount of data that's being poured into those two things,

which are being relentlessly innovated on by open source people at Hugging Face and Invidia, you know, whatever. If everybody is doing their best in this crazy capitalistic society we talked about in different systems, Nvidia's got a profit incentive and, you know, everybody competing with Nvidia is now in an existential crisis for their lives, in Tally, for example, etc. they're all now just solving that hardware level. And then all these open source people are trying to build software and products to keep their businesses going on hugging face and GitHub.

Everybody's doing their part to be rabid, innovative capitalists. And then it comes together on this crazy orchestra that you benefit from. What's the hardware platform that you use? Are you hardware constrained like the large language models are? Or do you just need like three H-100s and you're done? Or you just use cloud to, you know, I'm assuming you just use the cloud. We use cloud for deployment.

And we use, we have our own compute center that we also use for training. And I think we're less hardware constrained than many of these companies trying to train massive LLMs. And yeah, and I think to your point, I mean, we are absolutely benefiting from what's going on in the overall ecosystem. All of these things that are sort of pushing in only one direction. And it's just super exciting the way this whole field will be transformed. As you mentioned, it's kind of like two big platform shifts essentially at once because most of it is not yet, like you mentioned, microscopes aren't even connected. So just the basic advances of digitization, creating an inexhaustible digital resource that you can continue learning from without destroying any tissue over time. Connecting all of these brains together across the world is just a huge platform advance in itself.

the advantages of cloud and digital being connected. And then at the same time, you know, as that's being transformed, we're getting the substrate for training all these new models. Amazing. They can do things at lower cost, more accurate, more reproducible and more predictive. So like this is like, you know, where things are going. And I agree with you within five to ten years, you know, the world of pathology and diagnostics for drug development as well as for clinical care, I think we'll be totally transformed

by this. Yeah. I mean, it's mind-blowing. really is like there's something about compounding innovation occurring like you said is very insightful these things are happening in parallel there's some group working on microscopes somewhere i don't know where that is is it germany with carl's ice lenses or is it china somewhere people are working on the lenses inside of those microscopes somebody's working on the electronics in those microscopes and then you're working on your platform and all this comes together

as people buy increasingly powerful smartphones and laptops and play more video games, that's the precursor to your company existing is video game addiction and the smartphone. I mean, if those things hadn't become mass-produced products, we would not have, you know, Nvidia or, you know, the stack that's been growing. It's wild to think about why investment in technology and iteration is compounding, compounding iteration is so worth it. And just the impacts. Like even in the U.S., we have far too few pathologists, but it's significantly

worse around the world in terms of access to expert, you know, diagnosis, expert interpretation. Yeah, like somebody in Africa or the front, let's say frontier markets. Yeah. What did they have? They don't even get biopsies, I bet. Yeah. Well, increasingly they do. So in certain cases, not.

I would say the number of biopsies are going up, the number of people with chronic diseases, we make advances against acute infectious diseases increases over time. But what that comes with is needing to be able to diagnose and manage serious chronic diseases like cancer even better. So many of these very, very large areas have, you know, very, very few pathologists. And, you know, where we are today, where there's like a very small number of experts who aren't at all scalable is not going to address that global problem. Whereas once we've built these very high quality models that are deployable through the

cloud, it should really democratize access to this expertise and enable, you know, the best diagnoses to be distributed everywhere. And of course, that needs to be placed within infrastructure for also using that to guide the best therapies. But we think this is a really key part of that process that will be, you know, unlocked by both digital as well as AI on top, becoming more ubiquitous. Listen, I could talk to you for hours. And I did. we talked for an hour. Andy back, he's Andy back on Twitter. And the website, pathaI.com, they're hiring. So go ahead and go to pathaI.com. If you want to save lives and push the human species to even greater life extension, which I think is noble. I think people living longer,

adds more wisdom and people having longer health span, I guess, as Peter Attila keeps saying, is this is all good stuff. And so if you want to do something meaningful with your quit Facebook and optimizing to get people to click on ads and invade their privacy and go walk on something noble like what Andy is working on at pathaI.com. Life is short. If you're an elite developer, don't go for the quick ad money at some Facebook. Go work on something meaningful. It's much more important and you'll sleep better in that.

Andy, thank you for doing this work on behalf of the human race. I think it's awesome that smart people are choosing to do really important work like you're doing into your team. I thank you on behalf of the human species. Thank you, Jason. Thanks so much for your time. Awesome to get to meet. And I love the podcast. I'm a big fan.

So I'm a real honor to be on here. You know, after doing 1,800 of these and having people like tell me they listened to high school and college and now they're on their second startup, it is the great joy and legacy of my life that so many people have told me they heard this interview or that interview and it inspired them to start a company. And then some of those companies want to change the world, it's really important that founders, because there's so few of us out there, and, you know,

it's like less than 1% of the population who are willing to do what we do. And it's, it's filled with suffering and failure and sleepless nights, but it moves the human species forward. And founders are a unique breed. No, and I love the connection you make across these different areas, like how consumers playing video games and buying cell phones, you know, helps Nvidia. And then Nvidia becomes the core of a lot of what we do. in terms of building and deploying models at scale and how it all works together. And definitely you're both this and all in have been very inspiring, just thinking about what's possible with building companies today.

Like it just gets me really excited about, you know, what we can do more with less going forward. Never have been a more exciting time to be in a company, and I think probably to start companies because, yeah, the world is just so different now. I mean, I started working on this problem probably like 20 years ago. And where we are today is just so. so amazing in terms of technology. They overestimate what they can do in the short term and they underestimate what they can do in the long term.

And I am in the very unique position where I spend my life talking to people about innovation and founders. And my circle is exclusively people who start companies and people who back companies. And so I just get this very weird position where imagine having, you know, 10,000 hours, every couple of years of talking to people like yourselves and the people who back people like, your company and companies like yours, all of a sudden you become this like super connector. And yeah, things start clicking in and you're like, ah, this reminds me in the conversation I had with this person making video games to remind me this company who's making enterprise software.

And then you just, you kind of see like, kind of like that scene where in a, what was the movie The Hangover where Zach Gallifanax is just like doing math and the, all of the math goes like in his brain. he's doing math and like, do you see all the symbols flying over his head? It's pretty funny. All right, listen, Andy back, everybody. PathaI.com,

and we'll see you all next time in this weekend startups.