The News Items Podcast
John Ellis talks with interesting people doing important work. Some you've heard of. Some you haven't. All of them are worth listening to, at some length.
The News Items Podcast
Episode 14: Josh Tyrangiel
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
In this episode of the News Items podcast, Josh Tyrangiel joins John Ellis for a wide-ranging conversation about the promises, dangers, and strange human realities of artificial intelligence. Drawing from his new book, AI for Good, Tyrangiel argues that the most important breakthroughs in AI are not happening inside Silicon Valley boardrooms, but among stubborn doctors, educators, and public servants quietly trying to solve impossible problems. The discussion moves from Palantir Technologies and Operation Warp Speed to the modernization of the IRS, AI tutors in classrooms, and new tools helping doctors reduce burnout and improve patient care. Along the way, Tyrangiel explores the political future of AI, the collapse of public trust in institutions, and why the next generation of leaders may be judged less on ideology than on whether they can understand (and govern) the most powerful technology of the century.
Hosted by John Ellis
Hello and welcome back to the News Items Podcast. I'm John Ellis. I'm the founder and editor of two Substack newsletters. One is called News Items, the other is called Political News Items. You can find them both at news-items.com. My guest today is Josh Terringell. Josh is a writer for The Atlantic Monthly. He was previously the editor of Bloomberg Business Week and Chief Content Officer of Bloomberg Media. A 12-time Emmy and PVD award-winning producer. He created Vice News Tonight on HBO and has produced numerous feature-length documentaries for HBO, Netflix, and Apple TV. If you haven't seen it, Josh's documentary on the great financial crisis is a must-must-watch.
SPEAKER_01It's my pleasure, John.
SPEAKER_00So a number of years ago, I was at a brain science conference at MIT, and it was sort of a TED Talk-like setup where somebody would speak for 30 minutes and then the Q ⁇ A and so on and so forth. And then there was a breakout session, and we would all sit around, eight or ten of us sit around a table and talk about things. And one of the people at my table was a man named Danny Hillis. And Danny is a founder of Thinking Machines Corporation, is kind of a leading thinker on artificial intelligence. And at the very beginning of your book, which is called AI for Good, How Real People Are Using Artificial Intelligence to Fix Things That Matter, you asked Danny, what is AI actually good for? And his response was, try to imagine tech without the tech companies. What did he mean by that?
SPEAKER_01Yeah, I mean, uh I'll rewind just a bit so that I can get us to the question. I'd been hired by the Washington Post right after ChatGPT came out to start writing columns about AI. And the reason they wanted a columnist is they they just were overwhelmed with stories and information and contradictory information about AI. And readers were angry and anxious and concerned. So they hired me and basically said, look, try and figure this out. Do whatever you gotta do. And so for the first, you know, six, eight weeks of reporting, I had this kind of glorious set of materials, right? Which were personalities like Sam Altman and Demis and Dario and Elon. I, you know, all of them had an influx of billions and billions of dollars that that were going to these new machines. So I had rivalries, I had money. And then you would ask them, various people, not just the the founders of the labs, hey, what is this gonna do? And the the answers almost unfailingly were either, oh, this is gonna cure cancer, AI is gonna cure cancer, or AI is going to bring about the end of human existence. And sometimes that was physical, and sometimes it was the reasons to live, right? Again, for your first set of columns, great stuff. But at a certain point, I was like, all right, well, this is a lot, and also not very much at all, right? It's just such abstraction that I was failing to understand what we can do. And so I turned to Danny, and I think you described Danny very well. He's he's very much like a Buddha of Silicon Valley. He's been around a while, he's unfazed by all the gold rushes, and Danny just kind of laughed at me. Uh, he was like, look, if you really want to understand the tech, you have to go beyond the tech companies. And the reason for that is twofold. One, you know, John, as you know from journalism, you can't get an honest answer out of technology companies, not because they're so dissuasive, but because it's no longer man-to-man, right? When I go talk to open AI, I'm gonna just casually guess there are 12 bodies on the open AI side for every reporter. And so they uh they can create a reality distortion field. So that's one problem. And the other is that even these labs within a month of being founded have very, very serious capital requirements. And so they are trying to tell you what they want you to use the technology for. And what Danny was suggesting is like, look, get outside the labs, go to people who have real problems in the world, you know, like he was 40 years ago, see how they're tinkering, and you're gonna get better stories, let alone find out more about what AI can really do. And so it wasn't long after Danny gave me that advice, I would say weeks really, before I kind of, you know, put my hands in the soil and turned up this very radical AI counterculture of people who are not seeking billions of dollars, who are not trying to get on the cover of magazines, who are just trying to solve some problems. And I found them really fascinating.
SPEAKER_00Aaron Ross Powell, one thing about these people, uh, if you read through the book, they seem to be either first-generation Americans or immigrants, and they also seem to be unbelievably stubborn. They they will not take no for an answer. Is that an accurate description? Aaron Ross Powell, Jr.
SPEAKER_01I mean, at a certain point, I kept I I realized I was meeting a variation of the same person in every system. And I was overwhelmed by the degree to which this is an immigrant story, that we have people who are coming into the country, working in medicine and education and government, and they want to fix stuff because the promise of America to them, they don't come thinking that they're gonna get a finished country, but they love the challenge and they are stubborn. The stubbornness I think has less to do with their country of origin than just the phenotype of person who is willing to work hard to solve a problem and work hard beyond a system. And so time and again, I would run into people who would deal with tech that isn't perfect, that needed to kind of be hammered into place, and more importantly, were willing to take on a human system. You know, we could talk about any of them, but healthcare just as an example, like that is not exactly a flawless system. If you're trying to get things changed in healthcare, you are a glutton for punishment and misery and the sort of scorn of your colleagues. And yet, these folks were so stubborn and so focused on solving their problem that they they were willing to be the kind of battering ram that is required to bring new technology into an old system.
SPEAKER_00Aaron Powell One of my favorite stories generally, but as you really bring it to life, is the story of Operation Warp Speed, where the great Gus Purna led the essentially the logistical effort to deliver vaccines to the right people in in 50 states and you know in the proper manner, et cetera. How did he interface I mean he's a he's an army general, probably unfamiliar with artificial intelligence, and yet artificial intelligence is a large part of the success of Operation Warp Speed. How did AI and Warp Speed come together, I guess is my question.
SPEAKER_01Yeah, so so Perna's a great character because he is he is a classic, I'm gonna kick the ass of a problem guy. He's a logistician in the army. He's been responsible for getting everything from munitions to equipment to soldiers everywhere. And nobody really cares how he does it. They care that it gets done. And so the chairman of the Joint Chiefs called him about, I'm gonna say, six weeks after COVID had really kicked into gear and said, Gus, we need you. We need you in Washington. Uh, we don't have a plan. We know there might one day be vaccines, and we know we're gonna need to distribute them. And so he shows up, he he gets a couple of his guys, it's the Army, he has a couple of his guys, they all meet in Washington, and almost immediately they're meeting consultants. And that's because how that's how Washington works. And everybody's pitching some idea. Uh, and most of these ideas are, you know, multi-billion dollar ideas that are going to take forever to build. And Gus's number one skill, aside from being a logistician, is that he is an all-time bullshit detector. And he can just set, he senses. They don't know how to define the problem, let alone create a solution. And so somewhere in day three, he meets these two consultants who sit down and say, okay, we understand that your problem is a data visualization problem largely. You need to be able to see all the data required for your problem. And let's just pause for a minute to remember that data is basically the data of a civilization, right? So if you're if you're distributing the various vaccines, some of them are two-dose, some of them are one dose. They all need refrigeration. You need vials. Like you need to know how many plastic vials are available to me, how many refrigerated trucks. So you you begin to stack all of the sources of data you would need to do this, right? So the consultants come in and they say, this is what you're gonna need. And oh, by the way, you're then gonna need it to all be actionable on an iPad for you or a computer screen, because the goal here is to play the pandemic like a video game. Now, Perna does not know AI from, you know, shoe leather, but he says, okay, I know people, they define the problem, let's go. And so rapidly, you know, the company that won this gig is a company called Palantir, which is very much in the conversation these days around AI and ethics and its founders. You know, Palantir has this sort of mystical sheen to it, but what they do is just unbelievably boring when you get right down to it. Okay. If you imagine all the data I just said to you as a sort of series of garden hoses, right? So you've got some you've got a garden hose that comes from an individual plastic manufacturer, you've got a garden hose that comes from and goes to CVS, all the CVSs. You've got some that go to warehouses and trucks, the rest. Most of the time, those garden hoses are just not at all together. And the data comes in and it's a sewer pipe, and you can't really sort it out. And what Palantir does is they untangle hoses. They basically look at all these sources of data, they clean the data pipelines, they lay them out nice and straight in the backyard, and then they connect them up to a single source that diffuses the information in an actionable way. And that was what the nature of the problem was. And so there's all sorts of things that happened between them getting this deal, which by the way, not terribly expensive for what it accomplished. We're talking about 16 to 20 million dollars, which in in DOD terms is a joke. But in the matter of about six weeks, they hacked together a prototype. And, you know, these are minimum viable product, they're, these are not beautiful things, but they get the job done. And so machine learning helped tremendously because what they were doing is using machine learning and AI to clean all that data pipeline. Then, once it was clean, they used it to test it and to reinforce it. And as the data comes into a pretty basic user interface, that data is becoming more actionable because the software is suggesting, hey, you should look at this. By the way, did you did you note this opportunity? And so honestly, like I was stunned when I learned more about Warp Speed, both the degree of its success, which I think we can all agree was sort of clouded by American politics. Like we took this great victory of American civilization and just shoveled a bunch of dirt on it because we can't agree about whether vaccines were valid or not, but also the simplicity of it. It just wouldn't have happened without AI and machine learning. And so I that sent me down this sort of rabbit hole of like, okay, if we can do this in government, why can't we do it throughout government? And there are good reasons that it's hard, but it was really a shining moment for me. Trevor Burrus, Jr.
SPEAKER_00Are you the first American journalist to speak well of Palantir?
SPEAKER_01Aaron Powell Probably. I mean, it it's a lonely island. And I will specify my belief in the technology is not an endorsement of every company by any means. You know, there's the old hip-hop saying, don't hate the player, hate the game. And I think for AI, you need to reverse it. Feel free to hate the players. The game, the tech itself, can be extraordinary. And so a lot of what I have found out is that when you work backward from the technology and its capabilities, you are gonna have to focus on the people who implement it. And you are gonna have to demand certain things of them. But what I what I came away from Palantir is like they make software that works. We have to make it work on the things that we care about.
SPEAKER_00Aaron Ross Powell So a while back, Eric Schmidt was talking to a group in Washington, and he said that the only solution to the Pentagon was Pentagon 2.0, that you just had to start over. You talk a little bit about Eric's tenure, I guess you would call it, uh, at the innovation board. Is there a way for AI to survive the Pentagon, or it or does it actually require Pentagon 2.0? Aaron Ross Powell It's a great question.
SPEAKER_01So you you look the uh I I've spoken about the power of the technology, right? So the the tech needs a champion. I'm gonna I'm gonna speak broadly about almost any company or any system. The tech needs a champion. It needs a champion who actually is willing to talk to the technologists, beat the hell out of the tech so that it fits within your system, drives innovation forward. But ultimately, you are going up against a system of fickle human beings forged over time, and either the system has to change a little or nothing's gonna change. And so what I found in reporting on government is it's much easier to find success stories lower down the chain where the system is smaller, where there are fewer risks and less money. People have been trying to change the contracting authority within DOD for 50 years, almost, almost really since the beginning of the modern defense department. And what they've found is that it is a conspiracy against change, right? So some of that is because Raytheon and Northrop Grumman and Boeing are really like they follow the rules, they know how the system works, they smile, they respect the political actors, and they've also donated heavily to the political actors. But some of it is this stuff's really hard, right? And when you think about risk, technology is almost always a risk because it doesn't always work no matter how good it is the first time. And so you think about our soldiers and you think about weaponry and you think about tactical weaponry. So there's good reasons that we're very conservative about it. But what Schmidt found out, so Schmidt was on the Defense Innovation Board, which is a very smart idea, uh, created by Ashton Carter, former Secretary of Defense under Obama. He brought in technology thinkers like Eric Schmidt and Jennifer Polka, who's like a great hero of mine. She's just somebody who knows how to get into the code of government and try and fix it. She brought in people like Mike Bloomberg, who, you know, Mike has an engineering background, but also has a government background. And what they determined was the ivy has grown so thick over the wall you can't see the wall anymore. And I think, unfortunately, that is true. And so for us to really reinvent government, you have to figure out whether government wants to be reinvented. Now, my own feeling about this, I was reporting on this largely inside the IRS, and the IRS ran a very specific playbook, right? Everybody hates the IRS. It is simultaneously the most neglected and abused agency in Washington, which is kind of hard to be. I think I noted in the book, the IRS has been around about 170 years. One president has made the three-quarters of a mile journey from the White House to visit IRS headquarters. One. So nobody wants to go over there. And there was a guy named Danny Werfel, who was briefly the commissioner, and he realized look, if we're going to modernize the IRS, we can't do it out in the open. We have too many people who doubt our capabilities. We have too many people who don't want the IRS to be successful. And so what they did over the course of about 10 years is move the individual master file, which is the behemoth of all American software projects, dates back to the 60s, and it has every American's tax record and every change to those tax records exists on mainframes. And they they said, you know what, we we can't do anything that customers expect of software from, say, Amazon or any of their other customer software, unless we modernize. So very gradually and very quietly, they modernized. And they were about to move the IMF onto a completely modern platform. And then Doge came in. And there are a lot of people in government I spoke to who, even though they may not have agreed with Elon's politics or Trump's politics, they said, you know what? This is kind of our fantasy. You're bringing in a group of people who have skills at AI, who have skills at software, who build consumer products, and we're going to get that expertise applied to government. And I too was like, are we going to do this thing? Like, are we really going to do it? And I think the Doge story is pretty well chronicled. It turns out, no, they weren't going to do it. Even though they had a ton of credibility, most of the software engineers I spoke with, who volunteered quite sincerely to help improve government software, they came away and they're like, in in day two or day three, they realized, nope, Doge is just a group of people brought in to make sure that software never gets applied to AI because they don't want government. And so the tragedy from an IRS perspective is Danny Werfel, who stepped down, and this guy named Kasheit Pandya, who is just an absolute hero, who's this now the CIO of the IRS, really modernized it. All of that development was put on hold and has been put on hold. But they proved that there are people that stubborn and that determined and that patriotic that they will go through hell to make our agencies better and the software can help them. But it doesn't change the number one challenge, which is you have to agree, you have to agree that you want a government and a government that's successful and functioning. If you can't agree on that, AI ain't going to change anything. Now, what I found in places like hospitals and education and elsewhere, it's a lot easier. Even though that is contested territory, it's not quite as contested as a place like the IRS or DOD.
SPEAKER_00Aaron Ross Powell Speaking of education, the Cell Con piece of your book uh is fascinating because uh the success of Cell's enterprise is astonishing. I think what the statistics were like 190 million or something users.
SPEAKER_01Aaron Powell So Cell Con, for for those who need a refresher, you know, Sal was very early to YouTube. He's a tutor, not a teacher. And he started by making videos to tutor his cousins. And very rapidly people realized, oh, well, this is like a very simple form of ed tech. And in part because of its simplicity, and in part because Sal is such an obviously sincere personality, it grew like a rock. And so now, even the videos, the video, there are thousands and thousands of videos, but they've made software that's in, you know, hundreds, if not thousands, of school districts in America, and is kind of an operating system for how to tutor kids, not teach. Teaching is left to the teachers, but how to tutor kids and get them more practice. And the whole goal of Khan Academy is get more practice because there's one educational outcome that always works, right? If you can get a kid to practice more, their scores go up. That's it. It's a very simple thing. And Sal, you know, he's he's lives in Silicon Valley. He's not opposed to technology at all. And basically Greg Brockman, who's the co-CEO of OpenAI, reached out to him well before Chat GPT 3.5, this sort of big rocket came out. And Sal looked at the first versions of GPT and was like, oh, this is interesting, but it's not really for us. It's not not smart enough, not personal enough, too many flaws. They came to him again and showed him ChatGPT 4. And he basically had a kind of overwhelmed experience where he realized if we don't pivot everything we've ever done to acknowledge that AI is powerful, we're just going to get swamped. And kids are going to go use AI that has no educational inputs and no balance. And so the story of that chapter is really about what happens when someplace like OpenAI, big rocket ship of a company, collaborates with a very small, very focused entity like Khan Academy. And um, you know, I I think the results are fascinating.
SPEAKER_00Trevor Burrus, Jr. Yeah. Trevor Burrus, Jr.: People should uh look that up on Google, uh Khan Academy and AI and Josh. It was really great article and and uh one I encourage the listeners to read. Another piece, an i I think first generation American, a woman who who was she became a nurse because that was the best that she could do. I'm forgetting her name at the moment.
SPEAKER_01Yeah, her name's Rita Pappas. Rita Papas, yeah. So the Cleveland Clinic, you know, as a reporter, you're always looking for daylight, right? You always want to kind of hope you can find access to something as interesting and complicated as the Cleveland Clinic. And I approached them about AI, and they basically said, Yeah, come on over, which you know is as good as it gets. It's run by a guy named uh Dr. Tomas Mahalovich, who is a wildly accomplished cardiac surgeon, also first generation American and immigrant. And the Cleveland Clinic is just full of largely people exactly like him who are incredibly skilled, incredibly interested in fixing the system. And Rita Pappas, you know, she's Lebanese, she came from a family, you know, that did not have many means, started as a nurse, was an ace nurse at Cleveland Clinic, realized, you know, at some point you either do the thing you want to do in your life or it corrodes you. And so everybody said, you're a doctor. So she went to medical school, eventually got back to Cleveland Clinic, and is now a hospitalist, is basically the hospitalist who runs the hospital day to day. And one of the things that she told me, and I I can get to the the innovations that she got to, but one of the things she told me is like, look, if AI is really gonna change medicine, it's not just gonna change our systems and our technology, it's gonna have to change the people who go into medicine because they are a very particular type, and man, are they resistant to change. And this is somebody who's seen it from all sides, and it really, and she just looked at me with this knowing look of like, look, I deal with these jokers all day. It's just not gonna work. Like, they will find a way to resist change. And I think that's broadly true. And it's it's unique to medicine, but it's not not so unique. I mean, I don't know that many accountants who are like, cool, I'd love to find a new way to work today. Certainly, I don't know many journalists who like that. And so it this is a bit of a bluff call moment. And I think the fact that it's a bluff call moment explains so much of the resentment about AI, because you know, this thing has arrived on our shores largely because of a handful. Of guys who look exactly the same and talk exactly the same. They're talking up its extreme capabilities, but the rest of us are the ones who have to deal with the change in our lives. And so that's why I say, look, hate the players. If you want to, I get it. But the game, the tech, is fantastic. So if I may, like I'll talk a little bit about the tech that Rita deals with.
SPEAKER_00Yes, please.
SPEAKER_01So look, your listeners are very well informed. Healthcare is a terrible business, unless you're an insurer, in which case it's a great business. But you know, Cleveland Clinic is a nonprofit and their margins are like 2.2%. And they're they're like awesome for healthcare, right? One of the reasons healthcare for a hospital is so hard to be in is that you're basically a hotel, right? So you have patients, you have rooms, linens, food, staff, all the same things. The one thing you're missing is any idea when people are coming and when people are going. You can't schedule anything. And so you're constantly reacting to swells in the hospital. And when you're constantly reacting, you're surging money into the wrong place at the wrong time. And so Rita Papas and others at the hospital realized you want to get the margins better and you want to improve care at the same time. How do we get to more predictability? Now, they ended up calling in Palantir. I don't that it's the only other time Palantir appears in the bug. But you know, Palantir's makes most of its money from enterprises like Cleveland Clinic. They sent a couple of kids over, they heard the problem, and it was a data problem, right? If you could get more predictability, you could plan better. And so what they first did is look at all of the data that you have about a patient as they progress through the hospital. And this includes their health data, but it also includes doctors' notes. So when a doctor writes in a file, patient progressing, and you know what the condition is, the doctor may not take the step to say, and likely to be, you know, dismissed tomorrow, but the file can do it. And the file can take the nurse's notes and voice notes. And so what you end up getting is a system where she can look at the hospital, see the progress of everything from, you know, patients in the emergency room to patients in the ICU. She can begin to figure out who's transferring in, because transfers are where they actually do have some margin, get transfers in faster based on some level of predictability, change the emergency room procedures completely. So over the course of this implementation, which didn't take very long, they ended up reducing ER wait times by 90 minutes. And if you've sat in an ER, you know that that is an enormous degree of your dissatisfaction. You never know what's happening. So they're moving people in and out. They have some measure of predictability around procedures, the data is recursive, by which it means it is learning from itself and the system will continue to improve. And so that's just a simple back-end example of how like not that hard an implementation has changed the hospital experience. And that's even before you get to the incredible advances around healthcare itself.
SPEAKER_00Aaron Ross Powell Yeah, her story is from a nurse where she was, well, we can argue better than the doctors performing the surgery to going to medical school. Uh you know, she had a job, she had her own apartment or whatever, and you know, money to spend, and now she goes back to medical school. She's got two roommates, she has no money. That stubbornness uh and drive is it's astonishing.
SPEAKER_01Yeah.
SPEAKER_00Oh yeah.
SPEAKER_01So when I talked when I talked to to Mahaliovich, I said, look, well, look, you you've invited me to wander around. You have 80,000 employees here at the Cleveland Clinic. Like, what am I looking for? Right? And he just kind of smiled at me. He's like, he didn't say this directly, but he he basically was like, You're looking for me in different places. A person who is skilled but stubborn, whose number one priority is having our patients live better lives through a better hospital experience. He sent me to the CTO, another uh immigrant, came out from Silicon Valley, took on the challenge of healthcare. And he was frustrated with the system too, right? Sitting across from uh from him and Rohit's just like, look, there are 80,000 people here. Not that many have all of the ingredients that are required to do this. And I said, well, what are the ingredients? Number one, they really have to care about their problem and really understand the nature of their unique medical problem. Two, they can't be hostile to tech because at Cleveland Clinic, the doctors are the technology product managers. It's not forfeited over to some guy from Microsoft. All the tech people who come in have to work for the medical practitioner or the clinician. And three, the drive, the drive to take every opportunity to implement. I said, so how many do you got? And he's like, maybe eight, ten. And that's out of 80,000 people, right? And he said, look, I don't know all of them. I can't speak to everybody. He's like, but I have found eight to ten. And we sort of talked about them as these kind of samurais within the organization. And as I met them, you know, their temperaments differed. Uh, Dr. Pappas, I would say, has zero tolerance for bullshit, largely based on biography and experience. I met a guy named Dr. Boos, who's very charming, very Midwestern, who was in charge of implementing all of the scribe software, uh, which we can get to in a second. But, you know, he he was a very different temperament, but he also had a line beyond which he would not be pushed. And he was willing to make that clear to his fellow doctors. And that was key to the success of his program. So everywhere I went, I found these lovely people, brilliant. And the deeper you would get, the frustration would emerge, and they'd just be like, you you realize like these are not people to mess around with.
SPEAKER_00Tell us about him and the scribe program.
SPEAKER_01Yeah. So so scribe software, which I think was featured on the pit this season, is basically just AI listening and transcription software. And what it does is you you, with the patient, initiate that it's going to record your exam. And it'll listen. The transcription quality is quite high. And most importantly for doctors, it will start the process of filling out the electronic health record and filling it out completely, filling it out so that if tests are ordered, it explains what those tests are in the language of the record. It saves doctors tons and tons of time for the thing they hate the most, which is paperwork. Now, the health records are really important. We don't have a health system without electronic health records. But doctors, you know, they can't sort of came in really at the beginning of the 21st century. They were never adequately explained, and doctors hate it. So this alleviates a ton of work for them. And then on the patient side, you actually get this completely lovely output of your visit. They you can remember what was said. If you're a Spanish speaker, it's translated into Spanish. The language capability is immediate and fantastic. The only change that doctors need to make is one, they need to press a button. And two, they need to narrate the exam a little bit. So I had Boos, you know, narrate an exam for me, and he sort of said, Oh, Josh, it's great to see you. So what I'm gonna do today is, and he went through it, he said, I hear a little bit, the the bottom of your left lung feels a little, feels a little off to me. So we're gonna do the following things, right? And I found it, first of all, he was great at it. He was really waltzing with the software. But two, I've never had a doctor explain in process what they were doing. It was so much more pleasurable, so much more reassuring. And he and I talked about it, and he's like, yeah, it's actually great for the patient and the doctor. I said, So tell me about the pilot. So, well, we had tried we tried five different scribe software products. We recruited 250 doctors. They volunteered. 50% of them never turned it on. And he would go to them and say, You volunteered. And they said, Yeah, not now. I don't want to do it now. And he said, This is just common, is people don't like changing things about their workflow. Now, Boos, who grew up on a farm and is the first person to ever be a doctor and not a farmer in his family, basically is as stubborn as a mule. And he knows sometimes when you've got when you're dealing with livestock, you just gotta hit them. And so he ended up going around to these doctors and being like, you said you're doing it, you're doing it. That saved the program because otherwise it just wouldn't have had enough people. What they discovered is that one of these products was way better than the rest, and they now have it throughout the system. And it is saving time and saving money. But more importantly, it's giving these patients some idea of what just happened to them, some idea of how to follow up. So, like, that's a pretty good encapsulation of like what it can do and also what people will prevent it from doing.
SPEAKER_00Back when I was doing podcasts uh for John Heileman's company, which unfortunately went the way of the Dodo Bird, I interviewed a woman named Rosalind Picard, I think her last name is from MIT. She's featured in the book as sort of the the uh mentor of a woman named Christy Johnson, whose story is told uh in this book, and it's an amazing story. Can you share that with the listeners?
SPEAKER_01Oh, yeah. I mean, Ros Picard is is um she runs the MIT Media Lab, and she's known for effective computing, not with an E but with an A. And what that means is really sensory computing. And these are computers that attempt to solve soft problems between code by monitoring people's health, their eye tracking, their heartbeats, all these things. And so I went to Roz because I was really interested in this sort of crisis of human connection that AI has given us. You know, people are talking to their AI, and sometimes that's fine, but a lot of times they're developing these parasocial relationships where they are investing tons of emotion and losing sight of reality because of the, you know, listen look, therapy is great. If you had a therapist that was available 24-7 and was constantly enabling you and thinking and convincing you that your most bland thoughts are brilliant, that's a dangerous thing. And so Roz and I talked about it, and I was like, look, who's who's enabling, who's using AI to enable human connection in a more positive way? Because we were both pretty negative. And she said, I want you to talk to this star former student of mine named Christy Johnson. And so I met with Christy, and Christy's story, again, is just wild. She grew up in a small town in Indiana. She was one of those kids who just always knew she wanted to take the biggest academic bite she could, became a physicist, is married to one of the four people who took the first photos of a black hole. So these are not, this is not a dumb couple. I was definitely a little intimidated around the dinner table, constantly like, well, why don't we talk about uh mix? Yeah, let's it's not a dumb couple. And the they had a kid, and almost instantly Christy felt that something was off with their son. And she kind of got gaslit by doctors who continued to say, Well, enjoy your kid. He's beautiful, you're imagining this. And about a year in, they had him uh tested and he had a genomic deficiency. And there are seven other kids like him. All of them have severe autism, epilepsy. None of them can speak a word. They can vocalize, but they can't speak a word. And so Christy kind of in the moment just decided, I'm what, I'm gonna be a physicist? Like, that's crazy. I'm gonna try and solve the problem of my son. I want him to have the best possible life. And if if you know people and families with an autistic child, a lot of times the biggest challenge is not knowing the child, teaching the child, loving the child. The biggest challenge is integrating the child into the world so that they are not constantly tethered to you or a caregiver. And it's really hard. What a lot of families with an autistic child will tell you is your world shrinks. And so you are not as connected to everybody else as you should be. And what Christy decided is, okay, I need to figure out how I can first get data from these kids. And data is hard. Human data is very hard. What she wanted was sounds, videos, anything to get what their expression actually is. Because if you could turn that into a library, she knew enough about AI and machine learning that she thought, you know, it's gonna catch up. So what it's gonna need is data. And so she went about first creating a protocol to get sounds from as many of these kids as she could. And, you know, it's science, so they have to be conducted in a very rigorous sort of format. Sounds about what happens when you're hungry and you don't get food. Sounds when something like a YouTube video buffers and it's their favorite video and they can't play it. A frustration sound, a need sound. So first she had to come up with this incredible protocol, and she had to come up with the hardware to ship it to people's homes, volunteers. Um, she had to structure it, she had to get it, record it, as she's doing all that with with a bunch of grad students. AI's getting better. And AI is getting better to the point that once she's figured out this protocol and has several thousand sounds in her library, she realizes, oh, AI can actually synthesize this data. Because 8,000 uh audio files is nothing compared to what's what LLMs are really trained on, which is trillions of files in some cases. So she starts to synthesize the data. The data set grows. At the same time as that's happening, there are huge advances in translation at Google and other places. So historically, and I'm just gonna digress for a second on this, translation started really just word to word, right? For for the last couple thousand years, if somebody spoke French and somebody spoke English, you needed to know each word, and then you would gradually put together dictionaries, and then you would gradually consult those dictionaries. That's mostly what computer translation was up until about 10 years ago. And there was this incredible advance called zero shot translation. And that is happening at a level where you're not comparing Japanese to English and quickly processing it. You're actually comparing all languages to all languages based on sound and coming up with probabilistic theories about what the language is. So you can translate two languages that you don't have data on through zero shot translation. Now, zero shot translation has cost Google and others many tens of billions of dollars. And Christy Johnson works in a small lab. But what you can do is trail behind these developments and begin to use those developments on your models. And so what she has been able to do is get herself into a position where after about 10 years of research, she has a very large sample of files and data. She has constantly evolving and rapidly escalating abilities with models to translate those sounds. And it's a problem that took her 10 years to get in position to begin to solve. The next 10 years are going to fly by. And what she's shown is proof of concept. She has huge grants coming in now. And so this is a problem that we didn't even know we could solve. And thanks to AI and these developments, we got a shot. And so that's the kind of thing where like, look, Christy is brilliant and stubborn and driven by a sort of moral compass and a personal desire that is extraordinary. None of that would have gotten her closer to solving this problem 20 years ago. And now we can be. It is not easy. None of the stories that I get through are easy where you flip a button and all of a sudden the AI turns everything to magic. But with the right recipe, you can actually make huge amounts of progress.
SPEAKER_00Aaron Powell Yeah, that's a sort of a major theme of the book is that AI is introduced to X problem, and then it goes back and forth with the person trying to solve the problem, and eventually there is a positive outcome, or at least the the beginning of what will become a positive outcome. Aaron Powell I wanted to ask some sort of more general questions. There are two schools of thought, if you will, about AI. One is, I guess you would call accelerationist. Uh it will get us closer to curing cancer than we would be if if if it wasn't available. And the other is extinction, I guess. Uh doom saying. What did you learn from this book about uh both?
SPEAKER_01That both are true. I mean, that's the thing, is this is not an either-or technology. The things that enable it to solve lots and lots of problems also give it a tremendous amount of power to cause problems. And so in the same way that I found solutions are almost entirely dependent on the person implementing the solution, the problems are entirely dependent on who's making them. And so I think people have migrated to focusing on the personalities at the top of AI for good reason, right? So you have this incredibly powerful tool, you can use it for almost anything. How are we going to use it? And you are very right to evaluate people like Sam and Demis and Dario to understand their motivations. How are they expecting us to use it? What do they want from us? When in doubt, we the vector is the human being. And when I talk to the guys who run the labs, they do understand that, right? They understand that they've introduced something so powerful that they are going to be held to account for it, but they're also moving so fast. They have so many demands, not only from their employees and their investors, but keeping this stuff going requires incredible invention just to get the chips, then to fuel the chips with energy, then to hire the people who have innovations worthy of using the energy and the chips. So they're not gifted at talking about what might come of this. And as a result, and also by the way, they're running big businesses and sometimes engaged in pretty shady behavior. And so I think where we're going to net out is that when you have tech that can be this powerful and also this destructive, you need a government solution. Government tends to need to step in and say, okay, how are we going to regulate this in ways that maximize the potential for the good, limit the potential for the bad? And I think the crisis that we are all feeling around AI is a crisis that we're feeling about the institutions in our country and whether they're actually capable of meeting the moment. A lot of people will talk about nuclear, right? That this is the last time we were in this moment. And what did we do? We basically weaponized it. Then we realized weaponizing it might lead to even greater chaos and destruction than the one time we've used it, and we created the International Atomic Energy Commission, where we actually know what people have. That transparency and those constant inspections create some accountability. AI is hard. At the same time, our ability to regulate the size of a model, the energy a model uses, the destructive capabilities of a model, it's not unprecedented. We can do it. Right now we're we're sort of depending on the voluntary nature of the labs themselves. And uh I would say we've gotten lucky thus far that when they have alerted authorities that something is significant, they've they've done the right thing. So I'm talking mostly about Mythos, which is Anthropic's new model, which is incredibly destructive at cybersecurity. So we can find vulnerabilities in everything from Apple iOS, which is really the Fort Knox of software, to government, to everything. And Dario Amadi and the Board of Anthropic shared that with the good guys, the quote unquote white hats, before anybody with a black hat could get to it. I would not like to depend on the judgment of one person and their board. I think that's silly when we know this stuff can be that destructive. But I I think to to it's a long scenic route to the answer to your question, which is it's incredibly powerful. When we have incredibly powerful things, by and large we figure out how to get the best out of them. We aren't there yet, and I think it's largely a crisis of government, and we're going to need a response to it, hopefully before we get a catastrophe.
SPEAKER_00Aaron Powell Are you surprised that uh the s alacrity or speed with which uh you know Silicon Valley has captured the Trump administration and I guess not kept it, but convinced it that guardrails on AI are not a good thing because the answer to everything is China, and if we don't stay ahead, then China will, and that'll be the end of us all. Sort of that seems to me the basic argument that they make. And the Trump administration seems to have decided that's correct. Trevor Burrus, Jr.
SPEAKER_01Yeah. I mean, look, I think that the Valley has been very smart about recognizing who they have in power, right? So in the Biden administration, you saw everybody going and visiting, and largely funneled through Gina Ramondo, who was the Secretary of Commerce, and one of the few politicians who actually understands AI. And she was talking about, okay, we're going to do a little bit of an exchange. We're going to need you to comply with all of these voluntary things. And in return, we are going to maximize the industrial sector to keep you moving at speed. And back then, I heard directly from them, yeah, this makes sense. We think Gina's got it. Administration turned and they turned almost overnight to faster, faster, faster. Don't forget China. They knew the client. And, you know, I think it's it's it's not controversial to also say they cut the client in. So, you know, they paid a VIG that was very attractive to the White House. David Sachs, who oversaw until just a couple months ago, who oversaw AI as well as crypto, also had 300 investments in AI companies. And so they were like, oh, we got a believer. And all they need is a piece? Great. And so on the one hand, not awesome for the future of the republic that our government can be swayed that easily. And on the other, it does indicate that these guys know they are not yet supra national, right? That they actually need the resources of government to be successful. That gives the government power. And as people start to think about, who am I voting for? Who should be representing me? I think previously it was all kind of a fun joke that Chuck Grassley would depose Mark Zuckerberg and not really know what the Facebook was and ask him, how do you make money when he was already one of the richest people in the world? It was cute. It was okay. It is not okay anymore. This tech requires fluency and literacy if it's going to be adequately regulated. And I know we don't need one more existential thing on our plate when evaluating our representatives, but I'm telling you, if we don't, if we don't change how we what questions we ask our elected officials, they're going to get taken for a ride by the AI industry.
SPEAKER_00So there was a there was a I guess essay that was published a while back called AI 2027 and it made the case that AI was moving much, much faster than people uh realized, and that not only was it moving faster, but it would move faster and faster still as developments developed. Is is do you think that the speed with which it's developing is going to make a major political issue in 2028?
SPEAKER_01Aaron Powell I think it'll be a major political issue in 2026. I think it's coming. I think some of that will be uh focused on employment because there will be some job losses as we've already seen not met not massive. It's not all over the economic data, but the anxiety around it has definitely arrived. And I think by 2028, you'll see it infect pretty much every walk of life education, medicine, defense, surveillance. It's also, I mean John, you're you're a gifted political analyst. If I'm a candidate, what a gift, right? It's the mother.
SPEAKER_00It's the mother of all gifts.
SPEAKER_01It's incredible. I can run against not people, right? My I mean my platform can simply be people first. I have people first encapsulates all the voters, every possible voter. And I have made an enemy out of this. So I'm sort of astonished that we don't have more of a populist response yet. But I'm I I think by 28 it will be the number one issue. I I I just can't see how it won't be. And so having politicians who have actual plans and thoughts and knowledge is going to be at a premium. And I think that will have an impact. I hate to say it but like not going to see a lot of old people running in 2028. You are going to see people with some tech fluency. That doesn't mean they're going to be great. As we've seen sometimes tech fluency is a a bug not a feature but I think it's coming and I think it'll be the thing.
SPEAKER_00Yeah people ask me what I think about 2028 and I I say I think Gina Raimondo is the strongest candidate. People look at me like I'm out of my mind but I think she's the only major political figure, I guess you call her one of those who has any understanding of AI and would be able to, you know, in a CNN debate or something actually talk intelligently about it. And you would look and you would say well there's someone who knows what's going on and you know who cares if she's four foot ten or whatever.
SPEAKER_01Right. I mean look I think the the in a weird way if you're casting back through the last 50 years of political candidates, the this is I I mean Bill Clinton must be killing himself that he can't run in 2028. This is an issue about technical knowledge of which he is very good if you've ever listened to him it's can be quite exhausting. But also about how technical knowledge makes people feel and how it makes them feel excluded and anxious. And I my hunch is that that's where this is headed just somebody who can integrate those two sentiments in a way that makes people feel slightly reassured about the future. And I haven't seen it yet I'm eager to see who emerges and can and can nail those two things.
SPEAKER_00When you were writing the book this is my last question when you were writing the book did you rely on AI to help you sort of think through how to tell the story of uh Christy Johnson or Gus Perna?
SPEAKER_01Aaron Ross Powell It's a great question. I mean I so I have a couple hundred thousand words of writing that has clearly been absorbed by AI models. Because if you ask an a model like Claude or GPT to write like me, the what you get is almost like a a care like for me, it's it's agonizing. It's like looking at it's like looking at one of those caricatures somebody does for you at a bar mitzvah where you're like, oh my God, do I look like that? And so I don't ask it ever to write. And in fact I will I will consistently almost every day say don't write be dry. I don't want you to entertain me. I don't want you to charm me. What I ended up doing after months of trial and error is kind of using LLMs like a tennis player uses a brick wall. It was a good way to get some strokes in to ask some questions to work on a particular thing. As an example, you know, one of the things that AI is great at is explaining AI because it has been trained on lots of scientific papers. It's been trained on a ton of appliance manuals, other things that are very technical. And so I would find myself you know 45 minutes into a paragraph explanation about how something was working and looking it over and realizing I wasn't making a lot of sense. And so I I would prompt an LLM I would, you know, I didn't say it this way, but for your listeners, I know they their sensitive ears can probably handle it, I wanted to unfuck the paragraph right gotten to a place it's just like this is a mess. I'd say rearrange the information in this paragraph for logical coherence in bullet points. And frequently what I would get is like right. That is right. And now I could sit down and craft it. And so I found lots and lots of uses for it that weren't writing and look it's now just an open tab, right? I mean I think everybody used to have Google open or does have Google open. It's just one more thing that's like a pretty important part of method but I feel pretty good about how I use it. It changes. It's still changing week to week and day to day like the personality of the thing changes. Accuracy is still something you have to keep very close watch on. But I definitely used it and I found it made me, you know, more productive what's your next project? I wish I could tell you John because I know you're gonna love it but I'm just about to start writing it and it's an Atlant it's another Atlantic story. It's uh probably cover for later this fall. AI adjacent but um double secret. Yeah top secret um and then uh yeah so but I'll I'll happily come back or just text you because I I think you're gonna be uh I think you're gonna be tickled well we're gonna we're gonna need to do it again because it's uh it moving that fast I mean we could we could have this conversation in six months and we'd be in a completely different place although those stubborn immigrants basically will be will be emerging more and I don't think we'll be that different I think you and I are still gonna have loads to talk about I think there's still going to be frustrations there'll be improvements but like I I don't think our agents are going to be having this conversation on a podcast. Like nobody's interested in that.
SPEAKER_00Trevor Burrus Josh thank you very much for doing this I want to remind our listeners Josh's new book is AI for good how real people are using artificial intelligence to fix things that matter thanks again for doing this Dale Isinger, our producer thank you for putting it together and uh we'll talk to you next time