Scott Penberthy & Google AI for Healthcare

The Harry Glorikian Show 

Scott Penberthy – for September 26, 2023 

Final Transcript  

Harry Glorikian: Hello. Welcome to The Harry Glorikian Show, where we dive into the tech-driven future of healthcare. 

You’ve heard me say many times before that Google AI for healthcare is going to change almost everything about the way drugs get developed and the way healthcare gets delivered. 

If anything, I’m even more convinced of that now than I was a year ago, or five years ago. 

And there’s probably nobody better placed to see how this transformation is already happening than today’s guest, Scott Penberthy, who will talk about the Google healthcare AI.

Scott works at Google Cloud, where he’s the director of Applied AI in the Office of the CTO. 

He and his team work with Google’s big corporate customers to help them solve business problems that require large-scale computing and deep learning. 

That includes a variety of customers in healthcare and pharmaceutical R&D. 

Scott compares Google’s cloud computing capabilities to a racecar that can be adapted to any type of race, whether that’s a customer like Ginkgo Bioworks that leans on computation to reprogram bacterial cells to pump out pharmaceuticals and other products, or a giant health network like Anthem that uses AI to deliver personalized services to members. 

Because Scott helps set up these partnerships—and because he gets the first look at the Google’s emerging products and services—he has a unique picture of how computing is changing the everyday practice of doing R&D and running a healthcare company.  

As he himself puts it, he’s in the catbird seat. Which is what made this conversation so fun. 

Throughout the interview you’ll hear Scott and me basically geeking out about the latest advances in AI. Because both of us share the same amazement about how far things have come, and the same excitement about how much more is just around the corner.  

So without further ado, here’s my full conversation with Scott.  

Harry Glorikian: Hey, Scott, welcome to the show. 

Scott Penberthy: Hey, Harry. Great to be here. 

Harry Glorikian: I mean, you know, I’ve been talking to you off and on for a while now, but it’s good to have you on the show. And I you know, I’ve got a whole list of sort of questions and thoughts I want to go through with you. And I mean, it was funny because I know when we just started this, I was like, this must be the most exciting time on earth for you guys, considering everything that’s going on in Google AI for healthcare right now. 

Scott Penberthy: Yeah, it’s the kind of thing where, you know, I jump out of bed like at five and I knowing my kids dad go, Why do you get up at five for? I’m like, There’s so much going on. I mean, it’s crazy. It’s like, you know, I’m just a student of this. I’ve been studying this for a few decades, and I’m barely a student now. I mean, it’s a hundred papers a day. And so, like, how the heck do you keep up with that? So what are we doing at Google? And I have some companies who do the same thing. We just swap like, Hey, what’s cool and just to figure that out, to figure out like, where’s this thing going? I mean, it is wild. And so, you know, you know, we were, you know, the deep learning crazies, you know, just a decade ago. And then we’re a small team just focusing on it for a while now. Everybody’s that like. Everybody is on the Internet. Everybody’s mobile, everybody’s now, it’s great, you know, And you all are seeing what we were so fascinated with for a few decades. You’re like this is cool stuff. Yeah, it is. And so, yeah, I’m, it’s the highlight of my career, that’s for sure. It’s suddenly the things you were studying were esoteric back way back in school and like, what are you doing? What’s this? That’s not going to work. And now it’s the coolest thing. And so, yeah, I’m having so much fun, Harry. And a lot more people like me, they’re all just like everybody. Like, I don’t know math. You can contribute. You can design, you can help, you do legal, you can help. It’s just all hands on deck. It’s just it’s a wild time here at Google. 

Harry Glorikian: Well, look, let’s step back for a second here because, you know, listeners are like, who are you talking to? What does he do? So it’s like, okay, you’re the director of Applied AI in the CTO’s office at Google now, you know. What does that mean? And what kind of people are on your team, what kind of projects do you work on? And, you know, I’m going to beg you to sort of slant a little life sciences and health care. But, you know. How would you answer those questions?  

Scott Penberthy: Yeah, so I love how these things start. It started in 2016. There’s a there’s a great guy. He’s a special teams wizard. His name is Will Grant. It’s actually and I talked to him in 2016 and he goes, we got this new thing we’re trying to start. I said, Well, what is it? And he said, Well, you know, we got to talk to all these customers now, you know, and we like to build things. We don’t have time to show up to talk to the customers, but we can’t. We’ve got to send somebody in there talking about. And I had spent my career hanging out with a couple of CEOs. First Lou Gerstner at IBM when he was doing that turnaround and Bob Moritz at Price Waterhouse when he was doing his tech transformation. And so I used to hang out with those guys and have some experience of trying to figure out how to apply the latest tech and turn a big ship, right? And they said, Why don’t you come to Google and help us do that? And I said, Well, I want to do AI. Oh, that’s great, You know? And so I joined in 2016 and he goes and I sit down ready like, okay, what accounts are they going to be? You know, where is he going to send me? Like, what? What do you want me to work on with projects? And he sits across the table and goes, Glad you’re here. And I said, Well, thanks. And he goes, What are you going to do? And what? And I go, Well, you know, I like AI. 

Scott Penberthy: And he goes, It’s good. We can do with it. I said, you know, try to figure out how to make a difference with AI. And he goes, Oh, better get started. And that’s how it started. And I first was trying to figure out as a tool of science, that’s my passion. And I wanted to I always wanted to go to the moon and NASA’s right next door. So I did a bunch of stuff with NASA and that was fun. And then 2020, I got smarter. Harry, I guess you knew before I did, but because of Covid, we all got focused on health care and learned about next generation sequencing. And I’m like, Oh my God, this is the source code of life. Right. And I was talking to people like Oxford Nanopore and people who build these things and like, oh, no, 16 year old kids at Harvard, they’re sequencing strawberries. You’re what? Yeah, Yeah. We’re sequencing, reading the source code to strawberries. Now, there’s some errors in it because, you know, long form sequencing versus short form, I’m like, that’s incredible. And so I just started chasing like, could we use Google AI for healthcare to understand health at a very like a first-principles level, kind of like a physics level. And people are like, that’s crazy. You know, how do you know, if you look at the technology, there’s like 3 billion base pairs, 3.2 billion base pairs, right. How do you understand that language? And, you know, PhDs and Nobel Prizes were simple correlations. 

Scott Penberthy: So good gracious, trying to understand that language and said, well, I’ve got to start somewhere. And so that’s what I do now, Harry, is I try to figure out and work with a number of top companies. Can you use these techniques of Google AI for healthcare to understand the language of nature and use it to drive better health outcomes for people? And so I focus a lot of that. My personal thing is on cancer because I lost my mom to cancer 20 years ago and she always wanted me to help and said, You know, my brother’s an oncologist. He’s much smarter than I am and said, I’m a programmer. And now it’s a software problem. And we’re starting to see more and more evidence that if we could just understand the code and then start to debug it, could we then start to figure out what actions could we take to actually avoid getting sick? And that’s where I’ve been at. People like Lee Hood, who’s been focusing on this, he calls it phenomics. And that’s just where I spend my time now. And what I do in summary is that so I’m at Google Cloud and we’re in the office and we work on our largest deals. And what I’m what I work on is trying to figure out where can I make a difference largely in health care. And so the way that works is I bring in partners who really have a big problem that helps their business, helps our business, but we’re really doing it to advance the science. 

Harry Glorikian: Yeah. I mean, what you just said has been most of my career is like just trying to do those things and just having so much fun doing it. Oh, my goodness. Ridiculous, right? And you get paid to do it right on top of it, which is the best part.  

Scott Penberthy: Yeah. And there’s that commercial angle. And I think that’s what’s nice about the commercial angle, is that we have a lot of world class research and other people on the cover of Nature all over the place and think, what I, what I focus on is, well, how do you have material impact in a commercial sense? Because, you know, the pricing philosophy at Google is, you know, if we are pricing something for a dollar, we want to give at least $10 of value, if not more. So if we see flow, we’re like, oh, you’re making a difference now, right? And so that’s what we’re trying to do is how do you get that flow in health care and do it look at it in particular for Google AI for healthcare and making a difference for people. And it’s great. As I said, now it’s the hottest thing and I’m just trying to keep up and just learn it all and as much as I can and help our customers do the same thing. 

Harry Glorikian: So when you’re thinking about this or you’re you’re interacting with a new customer or you’re, you know, you’re sitting down in a meeting and they’re like, you know, what is a typical conversation look like? I mean. Uh, are you. Are you saying. Yeah, we’ve already got something for that and you’re sort of presenting it that way, or you’re like, Hey, tell me what your problem is. And we’ve got this cloud set up that can help you satisfy this problem. How does that how does the how does it go when you’re when you’re in the room? 

Scott Penberthy: Yeah. What’s what’s nice about, I mean, where I’m sitting now is, we’re a crew, a few dozen people in the office, and we all have different specialties, Right. And different depths and, you know, have all been there and have built, you know, large teams and have lots of scars to prove it. Right. And so what I typically engage with, so we’re a small team. We sit right close to our CTO of the cloud business, and we engage with our field teams a lot or the research teams. And it’s often like they’ll be working on a deal with a customer and like, does anybody know anything about omics? How do you spell omics? No, it’s omics. Let’s call this omics guy in the office. Let’s just see if he’s interested and or you know, we’ll, we’re trying to get into a new area. So I have a friend of mine, Bill Fitzgerald, out of Boston and I’m all I’m fascinated with this space as you are. And I’m like, we got to get the word out. And he’s like, all right, let’s just host an event. So you get event money. Yeah, I got free beer. And so we went to Boston as like in May, and then we met amazing people we’re all learning from. 

Scott Penberthy: I’m just sharing what I’m doing. They’re sharing what they’re doing. Like, wow, there’s a lot of overlap and they’re like, We should do something together. And that is where our Ginkgo deal started. Right did we just announced because we were in the area and I’m like, I’m really fascinated with the space. They’re they’re far ahead doing amazing things. And we we’re a platform company and they’re doing a lot of molecule discovery and protein discovery and great database. I’m like, Oh, that’s a great marriage, Let’s do something cool together. And so that’s an example. Or Mayo Clinic is another one or doing things with Anthem, now Elevance, trying to help them talk to their members. So that’s how it starts, Harry, is really a conversation of the customer has a deep problem trying to solve and they dig around. Does anybody Google know what this is? You know, And they go, oh, I raised my hand. And then that’s where it starts, typically the C-suite. And then we have teams that then say, is there something here commercially interesting? And that turns into deals of all shapes and sizes.  

Harry Glorikian: Yeah, I’m a big proponent of sort of events because, I mean, most of the business I’ve started is you get on stage, you talk, you meet like-minded people. The next thing you know, you’re walking out with, you know, projects. That’s historically been the way that it works. 

Scott Penberthy: Exactly. 

Harry Glorikian: Rather than calling people up one by one, trying to. 

Scott Penberthy: Yeah, I’m not I’m not I’m not like 1-800-DIAL-OMICS, you know, hey, you want some omics stuff like. No it’s it’s just more, it’s more of a natural conversation, typically very senior level and they have hard problems that they’ve been working on for years and and sometimes a fresh pair of eyes or a new platform might be helpful. And that’s what we try to do is like, can we be helpful? And a lot of times when we start this, we’re not sure we can. Um, but that’s where the office comes in, is saying, you know, can you work across multiple teams and research and see pull, pull off something interesting, you know, like, like the thing we just did, a peer of mine did a great job with Wendy’s, for example, you know, drive up and talk to a talk to an AI. Super cool. 

Harry Glorikian: I was just reading about that. I was like, whoa. And you don’t even have to get the the name of the order or what you’re ordering for the system to be able to understand what you’re ordering. And I’m like, Oooh. 

Scott Penberthy: Isn’t that the neatest thing? Yeah, we were at this. Yeah, I was just I was just in our team meeting and they’re talking about we just had our trade show. It’s called Next. And this guy was like, his name’s Adrian. He’s like, Oh, it was great. I said, What? And he goes, Well, I’m walking him through. And they’re used to a bot. And so customers would come up and they’re like, Well, this is going to be just like, I know what bots are, I know how they are. And you talk slowly and you think and they said, he’s like, No, encourage them to just talk more naturally. And he said, There’s this moment when they’re talking to the Wendy’s drive through and they’re like, Oh my goodness. Right? Because it’s like, this is much different. Like, can you have fries with that? Oh, change that. Change the burger to a double and add ketchup. No problem. Like how to do that and that that moment of delight he goes, and I would end my hours because you know you have tthe man in a booth right? Kicking in. And he goes, I would still show up just to see that moment of delight in a customer’s eye. And he goes, It’s just it’s like a warm fuzzy every single time. And then Isn’t that interesting? It’s just like this delight. Like, wow, that’s why I is now getting the point where it is just and we’re just getting started. But it’s like doing astonishing things now. 

Harry Glorikian: So I want to say, like, I think it was 18 months ago I sat down with there’s a robot called Moxie and oh, and it wasn’t even as sophisticated as what you’re talking about. And when I was talking to it and the way that it would adapt and the way that it would ask a question and the way that it would interact, even back then, I was like, Oh my God, if this just gets this much better, like you would want this on your desk and just converse with it, right? 

Scott Penberthy: Or just think of all the times when I need to call the dry cleaners or let’s say, uh, you know, here’s, you know, you get a bill and you got to call the DMV. Like I got to call them, you know, I just, I’m busy. Wouldn’t it be so cool just to say I’m going to send my bot, you know, and it calls and he goes, Hey, Harry, I got them on the phone. Everything’s all set. There’s one last question. You hop in, you hop in. You do like that’s where it’s going. Like all that stuff we have to do, you know, like, I want to call my kid, but, you know, she’s at college. It’s really hard to reach her. But can you just dial every now and then just to make sure and then pick up when she’s there? Right. It’d be awesome. Just basic stuff.  

Harry Glorikian: It’s funny because the other day I was trying to make a reservation and the Google system said, Would you like us to call for you? And I was thinking, yes, that person, is that person going to be pissed? Right. Because the machine is calling. It’s, you got to get used to it. You got to get I mean, you got to step into this world where everything is going to be different. But let me jump back here for a minute because I want to go in to see see if we can, you know, pick apart maybe an example where, I think you you probably know Freenome. Right. And if I’m not mistaken, they use Google Cloud to manage terabytes of multi-omics data they get from different blood samples. I don’t know if you’re familiar with how you guys are playing the role. Like are they doing everything self serve? Are you guys working with them to get more out of your cloud storage, your, you know, your Google Kubernetes engine, you know, things like that? Can you go through and it doesn’t have to be them. I’ll just I’ll take any example that you feel comfortable sharing.  

Scott Penberthy: Yeah we’re like. We’re basically it’s a platform play, right? You know, like a Porsche 911. Right. It’s like one of my favorite cars. Right. But yeah, but basically, you know, so you think of a platform, right? So and these companies like the race car drivers. Right. And so what we do is we have to understand, well, what kind of kind of car do you need, Right? What kind of platform do you need? And so what I’m so fascinated with and you are I think you are, too, this whole space of biology is becoming computational for the first time. And I can remember, I mean, we had statistics for years and correlations and that drove all kinds of stuff. But now it’s becoming where we can. Now it’s a simple idea. So powerful is can we have a computer guess a molecule and then go test it. Mean, it’s called in silico. It’s such a genius simple observation. But my gosh, is that powerful? You know, imagine shrinking drug delivery times from seven years to n months. Right. That’s cool. And so if you do that now, you’ve got you’ve got a lot of data. And for years we you know, as I’m learning, you probably teach me some more, too, is that a lot of this stuff is biomarkers and you get a blood test, you get some measurements. Basically that’s a that’s like a spreadsheet of values, right? And we used to we used to use statistics to look at correlations like, hey, when your temperature is high, it means you basically might be sick. Well, Captain Obvious, right? But there’s more sophisticated versions. 

Scott Penberthy: But now it’s like, how about all these pathology slides? Oh, those are huge, right? How about the X rays? How about the sonogram? How about. And you just keep going on and on and you get all these different modes. And these things are big and fat right there. There’s a lot of data. We didn’t have the technology to even reason with that stuff. Humans will look at it and they turn it into observations. They put it into clinical record. They turn into text. I can now deal with that directly. And that’s cool, right? So now if you take the X ray, you take the health care record, you take the these, these different omics, you can now all reduce these things to the way that the reasons and ask it questions. And that and that is what these customers are doing is they’re like, you know, some will say, I’m really good on pathology slides. I’ll take that. I’m really good. If you look at what’s happening at Ginkgo, I’m really good on proteins. We’re going to take that, but they reduce it to a space where now I can reason with it as a tool of science and that’s the new new. And so if you’re going to do that, here’s the basic thing. Let’s say you got a file that comes from a, you know, it’s a variant calling format file that’s like a billion variants. I think it’s huge, right? It’s gigabytes. And that’s not even the FASTQ File, which is. That’s just a that’s Goliath, right? How do you move that thing across the Internet? It takes a while.  

Scott Penberthy: And so we have a, Google is, since we had to move multiple copies of the Internet around, we built our own fiber. And so I’m like, Oh, if I got an application for you, right? We have YouTube, those billions of hours every, you know, they’re being served every day. Oh, that’s fantastic for biology. Right. So it’s it’s just it’s like a Porsche on a racetrack. It just it just flies through the network. And that’s the kind of things that they use it for, is saying your network’s fantastic. You’re built for large data. You’ve got a computational engine. You now have a tensor stack with tensor processing units, because that’s what these things are as tensors because I want in and that’s why I’m seeing a lot of companies use it saying you can you can always use VMs and other things. But we’ve been doing tensors with consumers for what, 20 years? And I’d say actually more like 12 years. I would say maybe we’ve been looking at analyzing consumers for that that long, 2 to 3 times a year, doing the searches. That’s all tensor math now, and we can now use it for biology. And so what we’re doing, what we’re doing at cloud is like, how do you enable that as fast as you can? And and we learn from them saying, what kind of pipelines do you need? What kind of data are you moving around? What’s it look like? You know, because, you know, these are big, lumpy pieces of data you’ve got to analyze. And that’s where the platform for that. 

Harry Glorikian: Well, but you know this, because, you know, we built something internally ourselves like, you know, actually trained a transformer from the ground up for a particular application. I mean, this is not falling off a log, right? I mean if you really want to get something out of this, somebody’s really got to know what they’re doing. I mean, I can tell you, like we were trying to get something out of it and we we had it focused on one body of of knowledge. And, you know, what we got out of it was just honestly sounded like a bunch of CEOs talking about their companies. Right? To a certain degree, because we were focused on an area of finance. We’re like, no, no, no. We need to get a broader set of data because because this thing is not, it’s it does what you tell it to do, but it’s, you know, and even then you. Yeah. This is not falling off a log just yet. 

Scott Penberthy: Not yet. But I think it’s going to be right for a lot of the common cases. And what I love about, you know, sitting in the hyperscale and where we see is, I see I see the patterns from hundreds, if not thousands of customers. Right. Right. And boy, is it is it iterating quickly and some and some interesting things to still out of that, like everyone’s doing document summarization. It’s the hottest thing, you know, because like and here’s the pain point I don’t have time to read all that stuff. Can you just summarize it and said, We’re supposed to read it all? Well, you know, don’t read it all. What I do know is I do a triage, I scan, and then I read the things I really need to know for a case. I’ll deal that. But the green light stuff, let it fly through that. That pain point is everywhere in business, right? And so we’re like, okay, let’s make that easy. Let’s make that super easy, right? So how do you take a model and basically train it on a document core? How do you make search to search across these documents? Two clicks in five minutes. Mayo Clinic did it. That, right. So. So now is it for every person? Is it every special. No, no, no, no, no. You’ll still have that. But the idea is that how do you make it as easy as search? Right. And so you’re doing prompt engineering and everything else. But are there common tasks you’re trying to do? Like I want to generate a document from a corpora. We’re basically I’ve got some special documents. I want to summarize it right or I need to find something in there. We’ll make that super easy. And so we’re just learning from you and everyone else what do you need? And then we put people together who just like. Make it easy, and that’s hard. That is really hard. It’s so much easier to ship a product with a thousand features than to ship the the one with three. Which three? That’s the hard part.  

Harry Glorikian: But it’s interesting you say that. But I think it’s gotten okay. Again, this is maybe not everybody, but it’s gotten easy enough and it’s, hell, have gotten cheap enough depending on what it is you’re trying to do. Right? If you’re not going to train a transformer from the ground up, okay, that that that’s harder. But if you’re going to fine tune a dataset or, um, you know, do some of the things that you’re talking about, I’m on a chat group with like, I don’t know, a thousand CEOs where everybody’s coming up with, you know, what do they need to do, right? And. The CEOs who are not tech guys ae getting things done and implementing them and pushing it through their own organization in. You know, I’m sure you’ve see it every day. Oh, yeah, right, and I’ll say fool around and find out for the show. But you know, you know, you know the actual term. Right. And and that’s what they’re doing. And it’s like, well, how much did that cost you to just play with it and sort of get it far enough along, you know, a hundred bucks, 200 bucks? 500 bucks. It’s not $50,000 to run something. 

Scott Penberthy: Right, isn’t that amazing to see? I mean, I’m getting so many meetings now where the senior executives she’s just come off just using the bot at night right. Or using in between lunch and they have direct experience with these things. And it’s just, you know, I love seeing it because I’ve been doing it for years now. Now everyone’s seeing the same stuff that used to get me all excited and what we’re doing, man, guess the question is, How do you make that easy for them? Or what was the question about trying it for them for the adoption? Yeah. So I think it’s what we’re seeing is I have this adage, you got to think about it as the mullet. Which is business in front, party in the back. Right so because it’s business in front, party in the back. So the idea is that you need to focus because it’s so exciting. You get so easy to play with the technology. You get excited about it. Right? Because it’s it’s remarkable. Like when you first see the browser, right? That’s cool. But then they’re trying to figure out like, where do I use this thing? Right. And how do I use this to make my people smarter or to make my business more efficient, to make the quality of life of my company better, my customers better? How do I use that? And we’re finding that, you know, we call it like Duet, but, for bringing Google AI for healthcare as an assistant to a workflow seems to be predominant, and especially with workflows that generate documents and read documents, right? Everybody has that, and that’s like a huge chunk of health care, right? The science piece is interesting, but a lot of it. But the money is in the boring billions of pajama time, right? It’s like. 

Harry Glorikian: Yes, claims data, you know, all that. Submitting a request. Get it? Yes. That stuff. 

Scott Penberthy: Yeah. And that’s the money because it’s it’s a pain. Right. And and I think that’s where everyone’s going to use it first. It’s as simple as like using, you know, using docs. It just starts to help you generate. But it’s like making it more specific for your workflow. There’s hundreds of those today. And then we’re also seeing more advanced science of like. Wait a second. Can use this like playing the game of Go. And my friends at NASA call it an idiot savant. I said, Well, what do you mean by that? They said, Well, Scott, if we’re using AI, it’s amazing what it generates. But we, like anything else, with a new idea we want to test it with the tried and true methods we’ve done for decades, but we’ll take the idea. I’m like, That’s cool. Right. So think of it this way. It generates something for you, Harry, and then you’ll use that, but you’ll add your own insights to say, This is from Harry, right? It just makes it so much easier for you to produce that. And we’re seeing that a lot in professionals in science and business today.  

Harry Glorikian: Well, I was trying to explain it to someone the other day. I’m like, this is the first time where you can actually have a sparring partner. Intellectual play. Ask questions. You can go back and forth with this thing. I mean, you got to prompt it the right way to get, you know, the right back and forth. But. You don’t have to just look at a blank wall anymore. Like you can actually do a few iterations to flush a few of your own ideas out and move things forward, which is really cool. Now, once it’s conversational, I think that will be a whole other level that that will. 

Scott Penberthy: And and I think everyone’s going to have their own assistant and a very short period of time it’s going to be, I think it’s going to astonish all of us how fast it’s going to that’s going to come. And it’s the kind of thing where, you know, as your programmer today, the old adage, if you can’t do it in five minutes, you better Google for it. And either you typically go to Stack Overflow. And then we used to say, good programmers copy, great programmers paste. Right. And what we’re finding now though, is that we all Google and we grew up doing that, the next generation, they’re going to have this I they’re ask questions. They get instant answers. With better information than any of us have. It will generate answers that will assist us in our processes better than anybody ever before. And that’s going to be for everybody. Age six up. I mean, think of that. And then that’s and now the question is, what questions do you ask? And that has been the hallmark of genius for centuries, is knowing what question to ask, because you’ll get an answer. 

Harry Glorikian: But here’s the interesting part, right? When you talk to people about or when I talk to people about this, right? I mean, if you talk more than one shock value above where they are or where they are, like I can already see like the eyes roll back, like the glaze shows up. They’re like, I have no idea what I’m talking about. Or they think it’s simplistic in some way, right? They’re just grasping the impact of what’s coming and how quickly it’s coming, how how quickly it is improving. I’m not seeing a lot of people grasp what’s going on? 

Scott Penberthy: Well, it’s… 

Harry Glorikian: It’s a minority.  

Scott Penberthy: I think so. But it’s it’s I mean, the Internet progressed pretty quickly and, you know, over, you know, 30 years ago. But now we’re all connected. We all have supercomputers in our hands. Most people, you know, have these iPhones or Androids. Right. Um, and so the word spreads so much faster than the internet came out. And this is much more powerful than that. And you’re seeing we’re seeing it now where these medical language models, they’re starting to approximate what some neuroscientists and others called the human language function. This thing doesn’t exist. But the idea is you can imagine us as a function, which is given some utterance, some music that you hear, smoke signals, some math you see. If that’s the input to your function, the output is, well, what’s the most probabilistically correct, best answer to follow that? And in conversations, what would you say next to be a good conversation in a different cost functions? I’m like, Well, that’s cool. I think as a community we’re getting close and we feel it. We can taste it to approximating that. And if we get a good approximation for that, oh my goodness, we’re going to have a tool that essentially is this amazing neocortex for everybody that has more information than all of us, and everybody is going to have access to it for pennies. That’s amazing.  

Harry Glorikian: And I feel like depending on the application, we’re already there and there’s probably people working on it right now that would, you know, depending on how narrow the whole thing is. 

Scott Penberthy: The thing is, in every domain it is going to feel that way. But I think have a general purpose and that’s what’s coming here. I was listening to a podcast and um, Lex Friedman had Manolis Kellis on and he said, AI’s not a tool. I was like, What? No. It’s a tool. No. He says no, it’s a partner. Now that that was really interesting and thinks what you’re saying too Harry, which is like you collaborate with these things and and if you take the approach of an idiot savant, it’s a fantastic brilliant, you know, creator of ideas and of content that you can then use to accelerate yourself so much more than everybody that you ever had before. You know, typewriter was amazing. I was in college. That was amazing. I have to handwrite this stuff. And then he had things would generate. That’s really cool. Now, the actual content itself, you can you can write books in a weekend. 

Harry Glorikian: And there’s people already doing it, right? 

Scott Penberthy: So they outline. So let’s say, you know, topic really well and you know, truth be told, I’m helping write a book on Gen Z. Well, guess how we’re going to do that? Well, I have the outline. I know the facts I want to do and my chapter’s due like in the 20th and the editors all over me. Where’s it doing? I’m not worried. He’s like, what? I’m forcing myself to use for this area because I’m going to give myself, like, a day. And that used to take me like a couple of months. And I know the content. I know what I need to do, but I did it. I did a book over a weekend that’s not that great, but I did my first one like, Oh, this is so doable because I think that’s where it’s headed, Harry. Where now it’s like, How’s your homework today? Nancy is your daughter comes in, she’s seventh grade. Not too much. Dad, what do you have to do? I have two 20 page papers. I’ve got a pictorial essay and I’ve got to do a three minute video. How long will it take? I don’t know, 15, 20 minutes.  

Scott Penberthy: I mean, wow. Right? And then they’re going to go in the teacher that teacher is going to use to judge it, and they’re going have questions and answers around the content. And the question is not the thesis format, cause it’s going to be perfect. It’s not like the pixel perfect. It’ll be good. It’s like, why did you choose that topic? Why did you choose those questions? And it’s much deeper conversation than you have. You’re missing a comma here. Minus one like that were so beyond that. But now it’s like, imagine the conversation in seventh grade. Your daughter is like thinking about why she thought that way. Why is she perceiving this? Why did she guide it that way? So cool. 

Harry Glorikian: Such a different…. Individuals like the people that have got to teach this, the people that have got to understand how to implement this. I mean, and even some of the stuff you brought up from a philosophical perspective, like just having these conversations, I my biggest worry is I’m not hearing enough people have these conversations and it’s like the locomotive is coming and it’s coming orders of magnitude faster than I think most people because we’re all connected. 

Scott Penberthy: We’re all connected. Yeah, No one can. It’s a hundred a day and we’re all like the researchers think they’re chasing this human language function. It’s like chasing spaceflight. It feels like Sputnik. It really does. And I think it’s going to come so much faster than we all expect. And some of my favorite programmers in the world are now spending full time on this. You know, John Carmack, he’s all over this. George Hotz is all over this. Um, we see it. It’s coming. When? I don’t know. I, you know, product managers are like, you know, we’re trying to turn into for planning three year plan and like, don’t know when we’re shipping six months, you know, it’s that fast. 

Harry Glorikian: We used to do five year plans and I already thought that was way too long. 

Scott Penberthy: On my God, how you do a five year plan? And I don’t know. We can see directionally where it’s going. But I think what we’re finding is pragmatically today, Harry, they’re looking at like, what’s a pain in the neck? Like where are we doing administrative stuff that we could just use to just to like a hot knife through butter, melt through it, right. Where where do we do the things where I don’t I want to call my customer back, but I can’t reach them. I don’t have enough time in the day. Right. I want to be able to answer the questions instantly and not put them on, not let them wait. Right. I want to have a great drive through experience, even if the kids get sick or it’s snowy. And see. See, I’m saying all those missing things can come in and it starts to take that pain out of business. 

Harry Glorikian: And I mean, it’s funny, I’ve got to give a talk in three or four weeks to a group of physicians. Fifteen minute. And I’m like. My you know, on one hand, I’m like, I should go through like 15 examples in 15 minutes just to give them an idea of the breadth. Or do you pick like 2 or 3? But I feel like a month from now there’s going to be a whole bunch of brand new examples to bring to them as opposed to what was a month ago. 

Scott Penberthy: Yeah. So I’m on this, there’s a new journal called Precision Oncology run by an oncologist. Doug Flora. My brother is on it. My brother called me. He goes, Hey, we’re doing this thing at Iona College. Do you want to help? I’m like, Well, sure, you know. And now they put me to work. But what’s interesting is what we’re finding is I went to one of their annual conferences, you know, American, ACCC. It’s basically community cancer centers. And we’re all sitting around or standing around the area and started talking and saying, Well, what’s the guess? My brother asked this What’s the number one technology last ten years that’s really changed cancer care? I’m thinking, oh, it’s next generation sequencing. It’s like, you know, man, new treatments. It’s, you know, Car-t therapy. And they’re like, oh, without a doubt, video conferencing. I’m like, Wow, didn’t see that one coming. I said what? No, no, because I can zoom when I can’t make it in. I can do a Google meet with my customer and with my patient who has has a question immediately and provide online patient care. And I can do it from my house when when the weather’s bad. It’s changed my practice more than anything else. And it’s like, wow, I didn’t see that coming. And then, yeah, and it’s like and then I said, What? Said Just their daily practice. These are, these are practicing oncologists. And then and they’re all excited about short form video too. 

Scott Penberthy: And I’m like, Now that’s interesting. Why? And they said, I have 70 patients a day, Scott, and she runs, you know, cancer in a remote state. On my way to get a sandwich out of the machine. I have two minutes. I use it to find new information about, you know, like a share and a market. And there’s a guy called ONC Doc. He’s got almost a million followers now. Sanjay Taneja, it’s a great doctor. They use that to share among themselves, best practices. That’s cool. And so and then they said, now here’s what here’s what I’ll help with. I said, So. So we’ve got video. That’s that’s great. I’m like, Wow. Like the Internet. No video. They said, the next thing we need help with is like pajama time. 

 

Scott Penberthy: And I’m, what’s that? Now, these are like top countries of the world, and they’re like, Scott, I love my patients. I care for them deeply. And when I come home, I have dinner with my family. And then afterwards, I pull out the patients for the next day and do charting. I’m in my pajamas and I typically have a cup of coffee there or you know, if if guest you know, and I’m trying to figure out the plan for the next day and I have to enter the stuff administratively so we can inform everybody. I can fly through that with a Google AI for healthcare that can help me summarize quickly and basically provide rough drafts that I can refine. That’s what I want to do and that’s the summarization thing I talked about. So it’s really just think about the practice of oncology. If you’re a brilliant oncologist and everything else, like what makes your daily life easier? Well, I can connect with my patients by remotely. So they have a question I can immediately so they don’t have to worry about scheduling. They love that. And two, how do I start to document this to make so much better? And that’s what I’m finding is like, the onus of a lot of this care is like, how do you help the practitioner? And then there’s even more sophisticated things coming in cancer research. Now physician scientists will give me a different answer. The practicing oncologist was like the daily practice. 

 

Harry Glorikian: Well, it’s funny because, like, you know, I’m always thinking about how do we use this technology to, you know, diagnose and then determine a therapeutic approach. You know, some therapeutic approach better, right? I mean, my brain goes there, not… 

 

Speaker4: We all do. 

 

Harry Glorikian: It doesn’t go to, you know, pushing buttons on a you know, to to enter the data into the system. Right. Although I know that’s a pain point. It’s definitely not the first thing that comes to my mind when I think about this space. 

 

Scott Penberthy: Now think of… I went to MIT many years ago, you know. I’m Dating myself. We did rule based systems and everybody’s fascinated with Google AI for healthcare as a diagnostician, as a doctor, right? Since the dawn of AI in the 60s, it’s really it doesn’t go very far because I talked to a doctor and and she goes, Well, there you go again, Scott. I said, What do you mean? She goes, I don’t need a diagnostician. I don’t need help. Like, but that’s you see the cover of nature. You see this? He goes, That’s interesting science. It’s cool, but not for practice. Well, what do you need? She goes, Just help me get the data. I said, What do you mean? She goes, When I sit down, I’ve got to go through multiple systems. I need the records. I’ve got different tests, I’m doing different interfaces. Can you just be smart, pull it together, present it to me, summarize it to me, and then let me make the decision. Now, you might say this looks like pneumonia. I’ll take that as consideration, but it’s my I’ve been doing this for 20 years. I’ll say as a GI, I can do this in my sleep, most basic things. And if I can’t, right, I have a consult with experts to figure it out. I don’t need a Google AI for healthcare there. We do need Google AI for healthcare to gather the data, paper the data and help us with the business side of the business. Huge help. I’m like Wow. 

 

Scott Penberthy: So think this is the dichotomy where for years it will continue to because it’s so fascinating. Can we understand the source code? Like what’s happening? Like, I have some friends in India, like we called Apollo 365, where they’re going to do that kind of thing, where in India some care is better than no care. And they’re happy to have any make suggestions for common ailments. That’s much harder than the Western world because there’s a chain of liability that doesn’t exist in India, right, for when it goes wrong. Um, so I think that I think we will see that over time. But I think in the next 3 to 5 years, most that’s just going to be allowing doctors to spend more time with their patients and getting patients the answers to take the oxygen out of the system, to be honest. 

 

Harry Glorikian: I mean, that would, that would be that would be fantastic. But I do believe, like once you get this information in in one place, right, that next step of what could it be and what would I do if it’s this that, you know, it brings it together a lot faster. 

 

Scott Penberthy: Oh, yeah. And that’s just like imagine empowering them with an assistant. So the assistant sits there and she, you know, this, this guy has has a complex case. There’s all kinds of tests, everything else. And she sees the patient. She’s got a couple minutes to do the charting. That’s how it works, right? Seven minute visit, three minutes of chart, and the assistant comes in and goes very tightly organized. Here’s a summary. Here’s the last test. Here’s what I’ve seen before. Here’s relevant research. Okay, doc, here’s the information for the patient. Here’s it all. All the data is there. You can click it, just summarize it for you in the way that you like to get it summarized with references. She’s like, Thank you. And she can take that through the diagnostic and I can say, you know, maybe suggest some maybe refinements and say, We’d like me to write that up for the chart for you. No problem. 

 

Harry Glorikian: And, you know, I don’t I assume Google AI for healthcare didn’t help with any of these other books that you’ve written in the past back then, but you’ve written a book called On Healthcare, and I believe another one called The 42 Genes, which actually came out in May of this year, if I’m correct What was the first one about. And what was the second one about for everybody that’s listening? 

 

Scott Penberthy: Yeah. So The 42 genes was a weekend experiment. And what I wanted to do is say, is this possible? Because I saw some people online that said I wrote a book over the weekend. Like, really? Let’s try that. And so what was I’ve been studying lately is trying to understand the source code of life. And I was looking for the FDA and other ones like what genes are most studied and what genes are used for cancer for for health care, for your fitness, for skin, that sort of thing. So and I found a subset of them about 42. And all the research was pretty, you know, pretty comprehensive and complex. And I said, I want to make this like an eighth or 10th grade level and can I find 42? Because that’s the answer to life in the universe. So let’s let’s pick 42. And can we write a summary of what these genes are as just like quick little handy reference manual? Could I do that in two days from a standing start and publish it on Amazon using Kindle? I’m like, Well, and I’ve never used Kindle before. I’ve never used other other tools and said, Let’s go. And it was two long days and I barely made it at like 9:00 at night and I got the darn thing out. Now you can download it. 

 

Scott Penberthy: I don’t know how to price it. So I’m learning how all that works. But it was interesting to me that I learned a lot in like, Oh, this is powerful. This is super powerful. Now, is it a great book? It’s an okay book. I think it’s more about the going through that and saying what’s interesting now, I’ve been using that in studying a lot since then, but now you’re going to see when we have a Gen AI book coming out, Antonio Gulli is editing this. A lot of Googlers and others are helping him. I’m doing a chapter on enterprise search. That chapter will be largely co-written with an AI. I’m going to use Duet to write it. And I’m like, I did at MIT, I’m forcing myself to do it in a very short period of time and the bar of quality is much higher. That was just me kind of do something. It’s not embarrassing to ship on just a book, but this is like peer reviewed and I’ll find out. And I have I’m getting high confidence we’re going to pull this off. Now, my editors all over me, he’s like, Where’s the draft? But right, but but I think it’s very meta. If you’re writing a book on Gen AI You should use Gen AI with your smarts to write the chapters. 

 

Harry Glorikian: What did you have to do special, if anything, when you were… I’m assuming there’s a whole body of literature that you had to somehow have the system absorb and then help you summarize for the book or at least give you the high points. 

 

Scott Penberthy: Yeah, so the nice thing about being in the C2 office is we get access to things before they ship, right? So I had access to early tools, so that’s kind of cool. Is that cheating? A little, Yeah, probably. But I use those tools. You know, I used to have 2000 GPUs of my own. I don’t have them anymore, by the way. They were taken away. But I was using those and what I was doing is I’d been studying this to try to understand the source code life for about two years, and I’d assembled a list of my favorite genes and I just kept looking at these things like the gene. That’s pretty interesting. The there’s another gene that’s called the warrior or the warrior gene, right? There’s genes that look at, Hey, how’s the cilantro taste? Right? And I’m just fascinated how that all works. And other genes like there’s one gene called DPD. Never heard of it. Well, if you don’t know, you have a particular variant. And if you take a medicine, it can be lethal. And I met a woman whose husband died because he had this variant and he didn’t know the test and he passed away. And through the lawsuit, she’s now trying to change that. It’s called pharmacogenomics. I have I personally have some called factor five, which is a blood clotting thing. And for years and six, you have the same thing. So I was six, they said, Oh, it says croup. When I was a teenager. Oh, he has recurrent pneumonia. And I’m like, What? No, it wasn’t. When finally landed in some place in in in Colorado, they did a test like, oh my God, you got factor five. I Was clotting for 20 years. 

 

Scott Penberthy: Right? And I’m like, I never want people to go through this. Right? So I started studying like, where are these things come from? And that’s where the genes. So I’d already assembled all these documents and I’m like, What the heck? I fed these things to the AI. And then it would ask questions. It would give me detailed scientific explanations. I figured out the prompts to give me something that’s shorter and and I would do the right as it finally iterated. And this is the collaboration to build a chapter. Then said prompts that work. And I started to replicate for the genes. That’s how I got the 42 done, and I was just literally grinding through all that research. And that comes from PubMed, from, you know, online medical journals, everything else. And what you see in that is a distillation at a 10th grade level. So what I chose of that using Google AI for healthcare is pretty good. And I’m like, Oh, that’s cool. So when I go talk to customers like you ever, you know, you just the only thing said, Have you seen my research? It’s like this pile, like this, right? And I’ve gone through that task. So now I know, I have some experience with this that I can share with others. 

 

Harry Glorikian: Do you know how many times I’ve had to iterate in my book because my wife said, it’s too high level, you got to bring it down? And I’m like, What do you mean bring it down? I don’t know how to bring it down more than and you go back and you, you know, you rewrite and you rewrite. I mean, if I could have fed that to something and had it make the language more accessible, Oh, my God, it would have saved me six months. 

 

Scott Penberthy: Oh, at least. A lot of time. And so that so that was an exercise in just, you know, have you actually run a race? You know, could you do this over a weekend? You know, I recommend if you have an idea, want to get it out, you should do it. It’s just fun. Um, but I find now, Harry, is that those tools now, they were very early. They didn’t have all the polish of, you know, tools that ship with the pretty UI. I mean, I was using command line stuff, right? Um, but that alone was enough to pull this off. And I’m like, Oh, this is good. This is really good. And now I think everyone’s going to have that. And so these tasks we have of writing your quarterly business review, writing your weekly summary, writing long letters to get admitted to a school, to justify an expense. Uh, that’s all going to be co-author, like, this year. 

 

Harry Glorikian: Yeah, I mean, so I want to go back to health care, but yeah, we can’t. Yeah. Like, yeah, I mean, it’s. I just believe that this is happening so fast that, first of all, humans are not designed to adapt that quickly. But what happens to people’s jobs. I mean, I’m I’m I’m pretty sure that there’s going to be like a bunch of cuts quickly until the system figures out what else people can do, right? I mean. That’s that’s what I sort of see happening in some of the companies that I you know, I have, you know, friends running or people that I know that are doing. It’s like, oh, I can take customer service down to a third of what it what it is, you know, today, because it’s more efficient. 

 

Scott Penberthy: Well, I think, you know, jury’s still out how that’s going to happen. I mean, there’s no question that the the nature of work will change. And it’s going to I think it could be a beautiful thing. I mean, it’s much like a few hundred years ago, you know, I wasn’t alive, but I’ve read the books, right? I mean, most of us. And it was only not even a billion of us on the planet. Right? 8 billion now. Several hundred million of us would spend our whole day taking care of food. We would build it, you know, plants, vegetables, livestock, everything else. We take care of food. And when we eat it and go back to sleep or then we do some other things to then go raise money for equipment to do that. And bartering that was an agrarian lifestyle for many years. Thank goodness I don’t have to wake up every day and go, you know, I have amazing produce in California. I don’t have to grow it. All right. I’ll just go to the store. I can order it on on Instacart. What am I doing with all that spare time? You know, if you’re not taking care of the farm, where are you going to do now that you don’t have a job to do? Go take go slop the pigs. I’m like, Oh, there’s so much more I can do now. And so will we all be doing slop in the pigs? No. Think a lot of that job will be taken away from that. You know the equivalent of that. But there’d be so many new jobs and so much more higher level powerful jobs than we can even imagine today. And that that is going to be, I think, a like a renaissance period for the human race in terms of the creativity, because now everyone’s gonna have access to it. And I think I think we’re just we get nervous about what. But I know how to take care of my food. That’s what I’ve been doing for my whole life. Well, there’s a there’s so much more ahead and it’s going to be exciting. But if you’re latching on to I’m really good at, like, digging up row from my corn. Yeah, you got to learn something new. 

 

Harry Glorikian: Right, right, right. So, yeah, I think it’s going to be. A bumpy transition, let’s put it that way, until we figure out what’s next. Only because it’s just moving so fast. And I just don’t think people are designed to absorb. That level of or speed of change?  

Scott Penberthy: Yeah. I mean, we still have I mean, the tool is going to iterate rapidly and it’s even more powerful, much like the internet did or mobile did and where as humans figuring out how do how does this really change your daily life? I mean, it’s just until you get there and you you know, I encourage anyone watching this to figure out don’t use it every day, find something. Seriously. Think of a side hustle. Think of a project. You know, like I wrote a little book on 42 genes. Why? Did the world need that? No, but said, How do I do this right? Or, you know, help help out for, you know, OncoAI, which is this, you know, this basically some oncologists work with on for cancer. Same thing with Lustgarten Foundation. So I help a lot of that and I use I every day now for 2 to 3 hours. And I couldn’t do the load work I’m doing without it. 

Harry Glorikian: It’s funny, because I’ve been I’ve been taking this, you know, a few classes that, you know, eight hours straight where we’re sitting in a room playing with every tool or playing with every tool that’s out there and just, you know, going through different examples. And different CEOs in the room will want to see it work in a different way. And because if you don’t lock yourself in the room to do it you’re never going to have time to just…. 

Scott Penberthy: Yeah, you don’t want to be that executive. That said, remember this many years ago, hey, I heard about the internet. Can you download that and print out the website and send it to me as an attachment? It’s just like I was I was 26. I’m like, what? What in the world? And, you know, that’s when email was new, right? And so you don’t want to be that person, right? What you want to do is say, no, you see last night, what are you using it for? Well, you know, we have a we throw a there’s a party for our block every every fall around Halloween. I threw a party. So I wrote the thank you letters using AI. I generated the images for the website using an AI using Midjourney and I use stuff on Google. I use that and then I’m using the AI to help me write my emails for me. Like there you go. Right. So the point is like, pick a project you don’t have time for and force yourself. And then they don’t say, This is pretty cool. And that’s how we’re going to learn. 

[musical interlude] 

Harry Glorikian: Let’s pause the conversation for a minute to talk about one small but important thing you can do, to help keep the podcast going. And that’s to make it easier for other listeners discover the show by leaving a rating and a review on Apple Podcasts. 

All you have to do is open the Apple Podcasts app on your smartphone, search for The Harry Glorikian Show, and scroll down to the Ratings & Reviews section. Tap the stars to rate the show, and then tap the link that says Write a Review to leave your comments. It’ll only take a minute, but you’ll be doing us a huge favor. 

And one more thing. If you like the interviews we do here on the show I know you’ll like my new book, The Future You: How Artificial Intelligence Can Help You Get Healthier, Stress Less, and Live Longer. It’s a friendly and accessible tour of all the ways today’s information technologies are helping us diagnose diseases faster, treat them more precisely, and create personalized diet and exercise programs to prevent them in the first place. 

The book is now available in Kindle format. Just go to Amazon and search for The Future You by Harry Glorikian. 

And now, back to the show. 

[musical interlude] 

Harry Glorikian: So I was looking at your LinkedIn bio and you’re obviously pretty enthusiastic about tensors. 

Scott Penberthy: Like, it’s my license plate. Yeah. 

Harry Glorikian: If you had to explain to a non-expert who might be listening to this show, you know, like, like my father in law or somebody, how do you explain to them what a tensor is and how did tensors make machine learning models work? 

Scott Penberthy: Yeah, so. So the physics folks who are listening to this, they may combat me in some of this, but basically, if you look at a tensor, a tensor is how I represents a concept in the world. Okay? Read this book called A Thousand Brains. And in there, if you look inside your head, inside our brains, in the neocortex, the gray matter at the top, it’s all chemicals and electrical activations, these things called neurons. Right. And so imagine if I were to take a spreadsheet and say, hey, hey, Harry, one second. Freeze! Okay, great. You’re frozen. Now I’m gonna take a spreadsheet out. I’m going to write down the charge level for every single neuron in your brain for that split second. Zero it’s off. Not thinking. One, it’s charged and some variation in between that’s gray, right? I’m going to get a few, you know, a few hundred billion numbers, right? That’s a tensor. Now I will say something else. Those are going to fluctuate rapidly over time. This is the concept that flows through your brain. And so in some respect, it’s a mental model I have for AI, which is a tensor processing machine. We take modalities into our brains, which like the thousand brains theory, we turn that into these activation levels of our neurons. We reason and that goes back out into utterances or actions that we take. And if you look at how Google AI for healthcare works, it’s much simpler than that. It’s not as beautiful as that. Not as complex. 

Scott Penberthy: But it’s effective. So they represent words now as just a tensor, which is much smaller. It’s not 100 billion, it’s maybe 768 numbers wide. Represents the activation of 768. Images, the same thing. It’s a little square, right? Audio, the same thing. And so a concept inside a model is just like when you say Bill Clinton, that’ll activate something in the model, the weights and how they’re set and basically the activation levels. That’s a tensor. Is that cool? And so that’s what tensor. So basically a tensor is how you represent it. And I’m saying the biggest shift we’re going to see from me and enterprises is we now have SQL databases, NoSQL databases and docs largely right, and SharePoint, everything. That’s how the enterprise represents stuff. I think most of it’s going to be tensors in ten years. Right. And it’s a much more sophisticated, more capable way to represent and reason about the world. 

Harry Glorikian: I. Yeah. Pardon me. Well, I think it depends on the company, but I think some of are going to move much faster than that. Ten years seems like a lifetime.  

Scott Penberthy: No, the startups are doing it now. I mean, and what finally these new startups, they’re using it like I am for my side projects. We all have them. They’re using AI or Google AI for healthcare in every process. It’s amazing. 

Harry Glorikian: Every process. Yes. There’s like three kids, kids. They’re younger than me. But basically, like everything they do is I first compared to the way that I would think about approaching the problem and. I mean, one of the companies I know they’re buying a company not for the product, but for the customer service system that they designed internally. That’s completely AI-first. They think it’s going to completely revolutionize the way they do their own customer support.  

Scott Penberthy: Oh, yeah. And that’s what I love about these startups. They have no resources by design. Right? And that’s think of this like when did the little 42 genes book? I had no resources by design, right? I had to do a book on the weekend. And if you do that project, like, how do I do this? You have to reach out to other tools, right? Otherwise it’s just not humanly possible. And so I think they’re learning so much by just saying I’ve got to do marketing, I’ve got to do sales, I got to do customer support, I’ve got to do HR, I’ve got to do employee reviews, I’ve got to do OKRs, I’ve got to do board reports. AI, AI, AI, through the whole thing.  

Harry Glorikian: Well, otherwise you couldn’t. I mean, you know, my consulting firm, I would honestly say that if we had all these tools that I have today, I wouldn’t have needed to hire 40% of the people that I hired.  

Scott Penberthy: Think about what great consultants do, right, is that they say, okay, what are the questions you want to ask? Let me go get the let me go get the temperature of the world’s best. I’ll get you the best opinions of a thousand people or 500 people. You do a sample, you sit down with them and do an interview. Oh, you can imagine doing all that in one day with one person.  

Harry Glorikian: Well, I told you that system that we put together that, you know, reads all the scientific papers. I mean, before I always said non-exhaustive right on every slide, not exhaustive, because it was impossible to be exhaustive. Now, you can use these systems to say, look, we’ve looked at every paper in this category. This is what’s going on. Right?  

Scott Penberthy: I mean, just think of a really different answer. Think of it does for like for management, what it does for governance. Right? When you have the ability to go ask a thousand people and get an answer and synthesize it and summarize it and have a conversation with a thousand people in parallel, if you mandate it right, if scheduling might probably take you 30 days just to land all the meetings. Um, but you can all do with AI or Google AI for healthcare now.  

Speaker4: Amazing. Yeah. I mean, the power of this thing, these things and what’s coming, I just believe, like, there’s. Work as we know it, science as we approach it. Governments, the way that we run the economy. There’s there’s there’s a profound shift coming. It just has to it just cannot stay the way that it is. It’s just impossible. 

Speaker4: I think you’re right, Harry. It’s going to be like, you know, most people a lot of people alive today don’t know the world before the Internet. Right. A lot of people do find that interesting. A lot of people don’t know the world before electricity. And so that’s how profound I think you’re right, is that it’s going to be so pervasive and everywhere and really lift a lot of us up. I think it’ll be like so pervasive and helpful for so many people.  

Harry Glorikian: The only difference between all those other things in my mind is, they were fundamental tools like you needed a human to do something. Push a button. Flip a switch. Something. This is not. This is not that. This is so far, this is you know, I can set up the I have it communicate with Zapier. Zapier goes and does something. I mean, I can have it do an entire process. And set it up to do it in an automated way, like no, no human needed in this process. That’s a different, profoundly different dynamic than I think than any of these other innovations we’ve had before.  

Scott Penberthy: You sound like Henry Ford, right? In other words, how do you do? How do you make a manufacturing line use tools to do jobs that you know you’ve done by hand? I think a lot of the back office stuff we’re doing is going to be done exactly that way. I think you’re right. It’s amazing.  

Harry Glorikian: So so. When you look at life sciences landscape right now. Right. And I’m sure you guys are talking to everybody. How do you see I mean, where are the places where you’re seeing AI or Google AI for healthcare and other forms of computing have the most profound effect on health care and drug development. I mean, what what are the things that sort of make you pessimistic and what are the things that make you optimistic? 

Scott Penberthy: Let’s focus on the, I guess, the optimistic side. So we’ve talked about, you know, the administrative side. There’s a lot of paperwork. That’s good to basically make sure there’s an audit trail and, you know, to document the process. It’s just a good scientific method, right? Um, there’s a lot of write ups that need to get done, so it is going to help with that for now. Absolutely. Submitting to the FDA, doing the FDA reviews, that’s all going to be assisted by AI. Same for documenting clinical trials, all that. I think the more interesting science piece that gets like the Nature magazine level of things, right? We start to see this Scientific American, you know, New England Journal of Medicine, that sort of stuff where that’s happening. I’m seeing a lot of effort and they’re all doing it, which is a simple observation. If I could use an AI to help me pick a molecule, could that dramatically shorten my time to explore which molecules are the right ones to ship for this particular antigen or whatnot? That’s pretty cool. In other words, could an AI or Google AI for healthcare suggest targets which are basically what part of a metabolic process you want to do to basically help someone through a bad condition? Right. Could it suggest a target that reduces my search space? Right. Um, and then can I run the tests in parallel, can I have it then highlight for me which tests I should look at first? 

Harry Glorikian: Right. 

Scott Penberthy: And so what they’re using it is they call it in silico, so can start to take processes that are now do in vitro or in the lab and use AI or Google AI for healthcare to reduce the search space to dramatically reduce the time it takes to to find the cure, if you will. Everyone’s thinking about that. And then so that’s like in silico versus in vitro and there’s a loop. So, like they’re doing it like NASA. In silico? Good suggestion. Let me test it. And then you do in vitro. Oh, this is pretty cool, right?  

Harry Glorikian: And then it feeds in the silico.  

Scott Penberthy: And then what happens is that it makes it in silico that much smarter. And now if you do enough of these things in silico starts to build a neural representation, the tensors we talked aboutm for molecules, for patients, everything else. Now it’s like, okay, don’t take it to trial. Let me run it by these million, you know, profiles I’ve created of different profiles I’ve seen from the clinical trials. Okay, this is the type of profile we should test first. That’s amazing. So now when you go out for the clinical trial, you’d be more specific. Here’s what I really want to test because I’ve tested everything it seems. Okay, now validate that with a genetic test. But here’s what I want from my trial. And it’s the idea that who do you want? The clinical trial, The ten right people. Right. So who are the right people? Don’t know. So do a much larger trial right as helping there so think it’s that that the observation is one help with the paperwork, number one, when that goes aside that’s like going from agrarian to industrial society just help me write all that stuff and review it.  

Scott Penberthy: Second one is I want to do it in silico plus in vitro and make that really tight. And that’s very much like the Ginkgo deal. And so they’re all thinking about that. How do I use AI or Google AI for healthcare to really understand what’s going on? It’s really fast. And then the next thing is to talk about digital twins. For years, digital twins for humans were based off of markers, right? Right now, we can do a much more sophisticated one to basically say, imagine if the digital twin goes, oh, bad allergic reaction. You probably want to test it on a human right? If it says it’s okay, everything else, like, okay, I was going to test anyway. It seems to be a green light. I’m going to go ahead. So what I’m saying, it just helps me that much better. And so now we talk about digital trials. Now that just helps you reduce which trials you need to run to validate it.  

Harry Glorikian: I just you know, in all the, and I’ve been thinking about this even before I started the show because, you know, the books I’ve written and stuff like that is there’s no place along the value chain where you’re not, you cannot by an order of magnitude improve your speed, your accuracy, make it much easier for people to interpret and understand and set up. And it’s just I feel like it’s whack a mole. Like as soon as we’ve improved one part of it, now you can focus on the other part of it and technology sort of moving along at the right pace to to bring that whole thing. I think, in five years the bigger companies that we know are not going to be the big dogs. At least that’s what I think. I think that there are some smaller companies that are well funded and have the right people and the right technology that are going to blow past some of the larger groups. That’s just my hypothesis.  

Scott Penberthy: What’s interesting, Harry, AI or Google AI for healthcare is definitely becoming a tool of science. It’s a very crude microscope today. Right. It’s so effective. Right. Or, you know, like I said, you know, you think of oncology, what’s the what’s the number one technology like? Didn’t expect that answer. Right. But that’s like the like the fax was 30 years ago. But now video, I think just having this thing in so that they can focus more on the science and the healing and less on the administrative burden that it takes to make sure that everything is done correctly. So imagine if you have a sonographer with you the whole time documenting it, you review it. Thank you. Get back to my science and that’s going to make the job, I think, that much more compelling and interesting. And the rate and pace of drug discovery, I think is going to go through the roof. And that’s why, you know, like FDA, I mean, FDA and others are looking at platform approvals. So think of like, you know, what’s happening now. You know, I think my mom passed away from her Her2 positive cancer. There’s trials now where you can do an mRNA based drug, take the toll of 253 nucleotides, put it into the same harness that was used for things like Covid vaccines and cure them. Now, cure. This means your immune system activates and takes away the cancer. Right? THat’s coming. It’s in the lab now. And so imagine platforms getting approved. And then personal. And then you go in and you get a personal drug and we’re going to see that. Yeah. And that that to me now that what’s nice about that is it allows you to prove something with the process. As long as you just change this piece of the code, right? Yes. And then that’s the thing. And so there’s and that’s like Car-T therapy and other therapies I think are so interesting because that’s based on the software of life. 

Harry Glorikian: It’d be interesting to see how the FDA tackles all that. But, you know, we’d love to see that in my lifetime because I’m not getting any younger. 

Scott Penberthy: No, but interesting, you know, as they work through that, think they’ll do it very thoughtfully and they’re thinking about, you know, if that approves it because it’s basically the same process. You’re just changing some of the code that’s used to affect, you know, how these antigens are detected. I think that is a really interesting thread that I’d love to see come to fruition. 

Harry Glorikian: Well, I can hardly wait till, you know, most of this stuff. I mean, it’s happening like, again. Like you and I said I can barely keep up with, you know, the latest piece that comes out, it seems. But I feel like we’re in a massive tectonic change, like nothing that… I thought sequencing was the big thing. Right. But this is going to bring sequencing to life in a different way that we couldn’t have done before or as quickly. 

Scott Penberthy: Yeah, I mean, so, you know, the cost of sequencing now, that’s what was so interesting about these sensors. And, you know, we’re going to get it below 100 bucks, which is really interesting. Right? You need to, SNP tests are still around that. But that’s just a new data set. But now we finally have a tool, Harry, that can actually reason with that stuff. Right. And it’s going to help with ancestry. It’s going to help with so many different things. And so now, you know, as it iterates, the next thing I want to work on is how do you get the efficiency of the AI or Google AI for healthcare itself to come down. Or to go up. What you mean by that is that we probably I’d love to improve by ten to the fifth, ten to the seventh, much like we did the human genome. Right. Because the amount of what we now need for doing these inferences in the big models dwarf what our brain needs. You think really, really hard and it gives you a headache for a few hours. That’s 25 watts. It’s like a little light bulb, right? And, you know, you train a model for a few months. That’s like all the output for one car for a year, right. Like, okay, there’s room for improvement. Right. But we need radically new approaches to modeling. I think we’re going to see that, too. 

Harry Glorikian: But I’m seeing people try. They’re doing the experiments. They’re writing the papers. They’re doing the math. I mean, I’m not saying we’re going to get there tomorrow, but yeah, you are seeing good ideas like that. People are coming up with and trying to test out. And, you know, I think it’s just inch by inch, but we’re moving a lot faster than I would have ever expected. I mean, I think, you know. The computational power is going to be 50 times greater in the next five years than it is today, based on the trajectory that I’m seeing from, you know, the organizations working on chip sets. I mean, you probably know much better than I do where things are headed because of where you are. 

Scott Penberthy: Yeah. And it’s really, you know, from the catbird seat, there are no people going at this because I think the grand challenge is how do you get the watts per inference, like how much energy does it take to actually do an inference. You know, do it to think, reduce it by ten to the seventh. That seems crazy. Well, that’s what happened. The Human Genome Project, you know, these phones we have were supercomputers in 1997. Right. We can get ten to the seventh. How do you do it? We need something radically new. And it’s going to be a 22 year old. And she’s going to come up with this idea because she’s going to ask a simple question we’ve never asked like, that is so genius and she’ll find it. I’m convinced. It’s not going to be me. My brain’s too wired. But I think that’s what’s coming. And we’re also seeing people now thinking about there’s so much tech debt because we’ve been chasing this for so long. We have this large stack. It’s amazing what it does. But there’s people like Chris Lattner and others who are looking at this saying, let’s rethink this. Can we look at a simple language instead of programming in C plus plus and Python and Cuda and accelerators and there’s like multiple layers of Turing equivalence all the way in the stack. He’s like, How about just doing this with a simpler language all the way to the bottom? Because nature’s a lot simpler. 

Speaker4: He’s at, I think it was $100 million. You know Godspeed, want him to do it right. Reminds me of the Lisp machine many years ago and think that’s what’s going to happen, I think, in the next decade. And you’re right. It’s going to bring the cost of compute way down. And people were thinking that the first ones that have come out, that’ll be, you know, oh my gosh, it’s so intelligent. Probably very expensive and big, maybe $1,000 an hour or something. But over time, a couple of bucks. You know, and you’re going to and you’re going to use these in every day and think we’re going to have a bill, like a cable bill, like it’s going to be like our car payment. Because why are you paying for every month? Well, that’s everybody does because life is so much better and I can do it right. And there’ll be the other programs. So that I think that’s what’s going is that, you know, we do that for entertainment, but we we put on suits and dresses and that kind of stuff to go to work. And and we have cars take us to work and think in the future we’ll have our own little bevy of AIs that we train and that’s who we become. Is that plus AIs and that’s that’s what people will hire is you and your team. Isn’t that neat? 

Harry Glorikian: You know, if there’s one thing I miss when I was at a place like Applied Biosystems was we got to ask the crazy questions that come up with the moonshot sort of ideas and think it through and work on it and see if we could actually like make it happen. And you’re at Google, so you get to do just that. And if I had to be say that I miss anything, it’s that, right? Because because doing what I’m doing now is you invest in a startup and you got to make sure that that guy makes it to the finish line. Yeah, right. For for all your LPs. So it’s a different. You’re not you’re not always blowing out that that that insane idea. Like, I have to do that in my spare time.  

Harry Glorikian: Yeah, but now it’ll be the kind of thing where you have these collaborators and they’ll get to know you and think we’re all going to have that. And it’s just it’ll be second nature. Much like searching Google is today. You know, we all we all use Google and think about it. We don’t think about the Internet. It was magical, just that kind of distributed computing we do at Google. That was dreams 30 years ago. Now it’s real. And 18 year old programmers like, oh, of course, you program a data center. That’s just what you do, Like what? Of course machine machines can read and speak and understand human language. Of course they can. That’s what think of that for a 12 year old taking a computer programming course on a weekend today. That’s what she thinks computers could always do. 

Harry Glorikian: It’s amazing to think about, like the tutoring system. I mean, I would have killed for this when I was a kid. Is this thing knows me really well, knows exactly where I’m making the mistake, can push and prod on me to sort of get over that hump. And I can move at any speed I want to. 

Speaker4: Yeah, right. And so I think we’re all figuring this out together. Um, we’re going to figure that, I’m pretty convinced we’re going to figure out this human language function or a darn good approximation to it. Much like the rocket flight or plane flight approximates a bird. Is it as elegant as a bird? Is it as efficient as a bird? No. No. Birds never go to the moon. Right.  

Harry Glorikian: Elon. Elon will help us get to Mars.  

Scott Penberthy: Yeah, absolutely. So I think that’s where we’re at, is that now we have what’s that rocket ship? And, and I think the community is just chasing that with, you know, and we’re going to we’re going to get it. And I’m just so excited to be alive and see it, to see the transition, you know?  

Harry Glorikian: Yeah. I mean, I think right now between you guys and other groups that are like just moving at the speed of light, it feels like like iterating and moving quickly. I think a year from now I think we’re going to see a big profound change. So yeah. Well. I wish you guys incredible luck. Yeah, I mean, that’s all I can say. 

Scott Penberthy: We’re just one participant, I think, Harry. And what I love about what’s happening now is that most people run into at Google. And my customers we’re all people. Right. And we’re all contributing. There’s a legal there’s a legal and ethics and policy and privacy. They’re weighing in every day. Right. There’s people that engineers how do you actually wire up these new machines and they’re weighing it. Logistics people. How do you get the chips to the right location? They’re weighing in. It’s it’s a full court press now. And I think every hyperscaler is into this saying this is really exciting and all the talk is about what’s that next turn of the crank for the cloud? How are you going to build that and what’s that platform we are all going to need? If everybody has one of these things and multiple of them. 

Harry Glorikian: Amazing. I can see the good and the bad, but we’ll deal with all of that as it’s evolving. So it’s been great having you on the show. 

Scott Penberthy: Well, thanks for the conversation. Enjoyed our morning.  

Harry Glorikian: Yeah. Thank you.  

Scott Penberthy: Yeah, we’re trying. It’s and it’s a good team and we’re all doing our best to help. 

Harry Glorikian: Excellent. Thanks. 

Scott Penberthy: Thanks, Harry. It was great seeing you. 

Harry Glorikian: That’s it for this week’s episode.  

You can find a full transcript of this episode as well as the full archive of episodes of The Harry Glorikian Show and MoneyBall Medicine at our website.  

Just go to glorikian.com and click on the tab Podcasts. 

I’d to thank our listeners for boosting The Harry Glorikian Show into the top two and a half percent of global podcasts. 

To make sure you’ll never miss an episode, just open Apple Podcasts or your favorite podcast player and hit follow or subscribe.  

Don’t forget to leave us a rating and review on Apple Podcasts.  

And we always love to hear from listeners on Twitter, where you can find me at hglorikian. 

Thanks for listening, stay healthy, and be sure to tune in two weeks from now for our next interview.