Ground Truths

Above is a brief video snippet from our conversation. Full videos of all Ground Truths podcasts can be seen on YouTube here. The audios are also available on Apple and Spotify.Transcript with links to audio and external citationsEric Topol (00:06):Well, hello it's Eric Topol with Ground Truths, and I'm really delighted to welcome Dr. Rachael Bedard, who is a physician geriatrician in New York City, and is actually much more multidimensional, if you will. She's a writer. We're going to go over some of her recent writings. She's actually quite prolific. She writes in the New Yorker, New York Magazine, New York Times, New York Review of Books. If it has New York in front of it, she's probably writing there. She's a teacher. She works on human rights, civil rights, criminal justice in the prison system. She's just done so much that makes her truly unique. That's why I really wanted a chance to meet her and talk with her today. So welcome, Rachael.Rachael Bedard (00:52):Thank you, Dr. Topol. It's an honor to be here.Eric Topol (00:55):Well, please call me Eric and it's such a joy to have a chance to get acquainted with you as a person who is into so many different things and doing all of them so well. So maybe we'd start off with, because you're the first geriatrician we've had on this podcast.Practicing Geriatrics and Internal MedicineEric Topol (01:16):And it’s especially apropos now. I wanted maybe to talk about your practice, how you got into geriatrics, and then we'll talk about the piece you had earlier this summer on aging.Rachael Bedard (01:32):Sure. I went into medicine to do social justice work and I was always on a funny interdisciplinary track. I got into the Mount Sinai School of Medicine through what was then called the Humanities and Medicine program, which was an early acceptance program for people who were humanities focused undergrads, but wanted to go into medicine. So I always was doing a mix of politics and activist focused work, humanities and writing, that was always interested in being a doctor. And then I did my residency at the Cambridge Health Alliance, which is a social medicine program in Cambridge, Massachusetts, and my chief residency there.(02:23):I loved being an internist, but I especially loved taking care of complex illness and I especially loved taking care of complex illness in situations where the decision making, there was no sort of algorithmic decision-making, where you were doing incredibly sort of complex patient-centered shared decision making around how to come up with treatment plans, what the goals of care were. I liked taking care of patients where the whole family system was sort of part of the care team and part of the patient constellation. I loved running family meetings. I was incredibly lucky when I was senior resident and chief resident. I was very close with Andy Billings, who was one of the founders of palliative care and in the field, but also very much started a program at MGH and he had come to work at Cambridge Hospital in his sort of semi-retirement and we got close and he was a very influential figure for me. So all of those things conspired to make me want to go back to New York to go to the Sinai has an integrated geriatrics and palliative care fellowship where you do both fellowships simultaneously. So I came to do that and just really loved that work and loved that medicine so much. There was a second part to your question.Eric Topol (03:52):Is that where you practice geriatrics now?Rachael Bedard (03:55):No, now I have ever since finishing fellowship had very unusual practice settings for a geriatrician. So right out of fellowship, I went to work on Rikers Island and then New York City jail system, and I was the first jail based geriatrician in the country, which is a sort of uncomfortable distinction because people don't really like to think about there being a substantial geriatric population in jails. But there is, and I was incredibly lucky when I was finishing fellowship, there was a lot of energy around jail healthcare in New York City and I wrote the guy who was then the CMO and said, do you think you have an aging problem? And he said, I'm not sure, but if you want to come find out, we'll make you a job to come find out. And so, that was an incredible opportunity for someone right out of fellowship.(04:55):It meant stepping off the sort of academic track. But I went and worked in jail for six years and took care of older folks and people with serious illness in jail and then left Rikers in 2022. And now I work in a safety net clinic in Brooklyn that takes care of homeless people or people who have serious sort of housing instability. And that is attached to Woodhull Hospital, which is one of the public hospitals in New York City. And there I do a mix of regular internal medicine primary care, but I preferentially see the older folks who come through, which is a really interesting, painful, complicated patient population because I see a fair amount of cognitive impairment in folks who are living in the shelter system. And that's a really hard problem to address.Frailty, The Aged, and LongevityEric Topol (05:54):Well, there's a theme across your medical efforts. It seems to me that you look after the neglected folks, the prisoners, the old folks, the homeless people. I mean that's kind of you. It's pretty impressive. And there's not enough of people like you in the medical field. Now, no less do you do that, but of course you are a very impressive author, writer, and of many topics I want to get into with you, these are some recent essays you've written. The one that piqued my interest to start to understand who you were and kind of discover this body of work was the one that you wrote related to aging and President Biden. And that was in New York Times. And I do want to put in a quote because as you know very well, there's so much interest in longevity now.Eric Topol (06:51):Interrupting the aging process, and this one really stuck with me from that op-ed, “Time marches forward, bodies decline, and the growing expectation that we might all live in perfect health until our 100th birthdays reflects a culture that overprizes longevity to the point of delusion.” So maybe if you could tell us, that was a rich piece, you got into frailty, you related it to the issues that were surrounding President Biden who at that time had not withdrawn from the race. But what were you thinking and what are your thoughts about the ability to change the aging process?Rachael Bedard (07:36):I am very interested in, I mean, I'm incredibly interested in the science of it. And so, I guess I think that there are a few things.(07:49):One thing is that the framework that, the part that gives me pause the most is this framework that anything less than perfect health is not a life worth living. So if you're going to have a long life, life should not just be long and sort of healthy in relative terms to your age cohort, but healthy that when you're 80 you should feel like you have the health of a 45-year-old is my understanding of the culture of longevity science. And while I understand why that's aspirational and everybody worry about my body's decline, I think it's a really problematic thing to say that sick bodies are bodies that have disability or people who have cognitive difference are somehow leading lesser lives or lives that are not meaningful or not worth living. I think it's a very, very slippery slope. It puts you in a place where it sort of comes up against another trend or another emerging cultural trend, which is really thinking a lot about physician-assisted suicide and end of life choices.(09:04):And that in some ways that conversation can also be very focused on this idea that there's just no way that it's worth living if you're sick. And that's just not true, I think, and that's not been true for many, many, many of my patients, some of whom have lived with enormous disability and incredible burden of illness, people who are chronically seriously ill and are still leading lives that for them and for the people who love them are filled with meaning. So that's my concern about the longevity stuff. I'm interested in the science around the longevity stuff for sure. I'm interested in, I think we're living in this really interesting moment where there's so much happening across so many of the chronic disease fields where the things that I think have been leading to body decay over the last several decades for the majority of the population, we're sort of seeing a lot of breakthroughs in multiple fronts all at once. And that's really exciting. I mean, that's really exciting. And so, certainly if it's possible to make it to 100 in wonderful health, that's what I'd wish for all of us. But to hold it up as the standard that we have to achieve, I think is both unrealistic and a little myopic.Eric Topol (10:28):Yeah. Well, I certainly agreed with that and I think that that particular essay resonated so well and you really got into frailty and the idea about how it can be potentially prevented or markedly delayed. And I think before we move on to one of those breakthroughs that you were alluding to, any comments about the inevitability of frailty in people who are older, who at some point start to get the dwindles, if you will, what do you have to say about that?Rachael Bedard (11:11):Well, from a clinical standpoint, I guess the caveat versus that not everybody becomes frail and dwindles exactly. Some people are in really strong health up until sort of their final years of life or year of life and then something happens, they dwindle quickly and that's how they die. Or some people die of acute events, but the vast majority of us are going to become more frail in our final decades than we are in our middle decades. And that is the normal sort of pattern of wear and tear on the body. And it is an extraordinary framework, I think frailty because the idea of this sort of syndrome of things where it's really not a disease framework, it is a syndrome framework and it's a framework that says many, many small injuries or stressors add up to create a lot of stress and change in a body and trauma for our body. And once you are sort of past a tipping point of an amount of stress, it's very hard to undo those things because you are not sort of addressing one pathologic process. You're addressing, you're trying to mitigate many processes all at once.(12:31):When I wrote that piece, it was inspired by the conversation surrounding President Biden's health. And I was particularly struck by, there was a huge amount of clinical speculation about what was going on with him, right? I'm sure you remember there were people, there was all of this talk about whether he had Parkinson's and what his cognitive status was. And it felt to me like there was an opportunity to do some public education around the idea that you need not have one single sort of smoking gun illness to explain decline. What happens to most of us is that we're going to decline in many small ways sort of simultaneously, and it's going to impact function when it tips over a little bit. And that pattern of decline is not going to be steady day over day worsening. It's going to be up and down. And if you slept better the night before, you might have a better day the next day. And if you slept badly, you might have a worse day. And without knowing anything specific about his clinical situation, it felt like a framework that could explain so much of what we were seeing in public. And it was important also, I think to say that nothing was necessarily being hidden from anybody and that this is the kind of thing that, this has accumulated stress over time that then presents suddenly all at once after having been submerged.Eric Topol (14:01):Yeah, you reviewed that so well about the wear and tear and everything related to that. And before I move on to the second topic, I want to just circle back to something you alluded to, which is when Peter Attia wrote about this medicine 3.0 and how you would be compressed and you'd have no comorbidities, you'd have no other illnesses and just fall off the cliff. As a geriatrician, do you think that that is even conceivable?Rachael Bedard (14:35):No. Do you think it is?Eric Topol (14:37):No, but I just wanted to check the reality. I did challenge on an earlier podcast and he came up with his pat answer. But no, there's no evidence of that, that maybe you can delay if there ever was a way to do that. But I think there's this kind of natural phenomena that you just described, and I'll refer people also to that excellent piece that you get into it more.Rachael Bedard (15:06):Peter Attia, I mean, he is certainly the sort of standard bearer in my mind of that movement and that science or that framework of thinking about science. And there's stuff in there that's really valuable. The idea of thinking about lifestyle in your middle decades is having meaningful impact on how you will age, what your final years will look like. That seems intuitively true, I think. And so, thinking about his emphasis on exercise, I mean, his emphasis on exercise is particularly intense and not super achievable for the average person, but the idea that you should sort of be thinking about keeping your body strong because it will decline eventually. And so, you want to do that from a higher peak. That makes a lot of sense to me. The idea that where we sort of draw pathologic disease cutoffs is obviously a little bit arbitrary. And so, wanting to think about optimizing pre-disease states and doing prevention, that's obviously, I think pretty appealing and interesting. It's just really in an evidence free zone.Ozempic for the IndigentEric Topol (16:18):Yeah, that's what I confronted him with, of course, he had a different perspective, but you summed that up really well. Now let's switch to a piece you had in New York magazine. It was entitled, What If Ozempic Is Just a Good Thing? And the reason, of course, this ties into the first thing we're discussing. There's even talk now, the whole GLP-1 family of drugs with the dual triple receptors, pills to come that we're going to be able to interrupt a path towards Alzheimer's and Parkinson's. Obviously you've already seen impact in heart disease, liver disease, kidney disease way before that, diabetes and obesity. So what are your thoughts? Because you wrote a very interesting, you provided a very interesting perspective when you wrote that one.Rachael Bedard (17:11):So that piece I wrote because I have this unbelievably privileged, interesting clinical practice. In New York City, there is public health insurance basically available to anybody here, including folks who are undocumented. And the public hospital system has pharmacies that are outpatient pharmacies that have, and New York Medicaid is very generous and they arranged through some kind of brilliant negotiating. I don't quite know how to make Ozempic to make semaglutide available to people who met criteria which meant diabetes plus obesity, but that we could prescribe it even for our very, very poor patients and that they would be able to get it reliably, that we would have it in stock. And I don't know how many other practices in the country are able to reliably provide GLP-1s to marginalized folks like that. I think it feels like a really rare opportunity and a very distinct perspective.(18:23):And it has just been the most amazing thing, I think to have this class of drugs come along that, as you say, addresses so many problems all at once with at least in my prescribing experience, a relatively mild tolerable side effect profile. I have not had patients who have become incredibly sick with it. And for folks where making that kind of impact on their chronic illness is so critical to not just their longevity, but their disease status interacts so much with their social burden. And so, it's a very meaningful intervention I think around poverty actually.(19:17):I really feel that almost all of the popular press about it has focused very much on use amongst the wealthy and who's getting it off label and how are they getting it and which celebrities are taking it, and what are the implications for eating and diet culture and for people who have eating disorders. And that's a set of questions that's obviously sort of interesting, but it's really interesting in a very rarefied space. There's an unbelievable diabetes epidemic in this country, and the majority of people who have diabetes are not the people who are getting written about over and over again in those pieces. It's the patients that I take care of, and those people are at risk of ending up on dialysis or getting amputations. And so, having a tool this effective is really miraculous feeling to me.Eric Topol (20:10):Well, it really gives me some hope because I don't know any program like that one, which is the people who need it the most. It's getting provided for them. And we have been talking about a drug that costs a thousand dollars a month. It may get down to $500 a month, but that's still a huge cost. And of course, there's not much governmental coverage at this point. There might be some more for Medicare, Medicaid, whatever in the future, but it's really the original criteria of diabetes, and it took almost 20 years to get to where we are right now. So what's so refreshing here is to know that there's at least one program that is helping to bridge the inequities and to not make it as was projected, which was, as you say, for celebrities and wealthy people more exclusively, so that's great. And we still don't know about the diverse breadth of these effects, but as you well know, there's trials in Alzheimer's. I spoke to Steve Horvath recently on the podcast and he talked about how it's reset the epigenetic clock, GLP-1.Rachael Bedard (21:24):Does he think so?Eric Topol (21:26):Whoa. Yeah, there was evidence that was just presented about that. I said, well, if that does correspond to aging, the thing that we spoke about first, that would be very exciting.Rachael Bedard (21:37):It’s so wild. I mean, it's so exciting. It's so exciting to me on so many levels. And one of them is it's just exploding my mental model of disease pathogenesis, and it's making me think, oh my goodness, I have zero idea actually how metabolism and the brain and sort of cardiovascular disease, all of those things are obviously, what is happening in the interplay between all of those different systems. It's really so much more complicated and so much more interdependent than I understood it to be. I am really optimistic about the Alzheimer's trial. I am excited for those results, and I think we're going to keep seeing that it prevents different types of tumors.Eric Topol (22:33):Yeah, no, and that's been shown at least certainly in obese people, that there’s cancers that gets way reduced, but we never had a potent anti-inflammatory that works at the brain and systemically like this before anyone loses the weight, you already see evidence.Long Covid and ME/CFS(22:50):It is pretty striking. Now, this goes back to the theme that was introduced earlier about looking after people who are neglected, who aren't respected or generally cared for. And I wanted to now get into Long Covid and the piece you wrote in the New Yorker about listening to patients, called “what would it mean for scientists to listen to patients?” And maybe you can talk about myalgic encephalitis/chronic fatigue (ME/CFS), and of course Long Covid because that's the one that is so pervasive right now as to the fact that these people don't get respect from physicians. They don't want to listen to their ailments. There's no blood tests, so there's no way to objectively make a diagnosis supposedly. And they're basically often dismissed, or their suffering is discounted. Maybe you can tell us again what you wrote about earlier this year and any updated thoughts.Rachael Bedard (24:01):Have you had my friend Harlan Krumholz on the show to talk about the LISTEN study?Eric Topol (24:04):Not yet. I know Harlan very well. Yes.Eric Topol (24:11):I know Akiko Iwasaki very well too. They’re very, very close.Rachael Bedard (24:14):So, Akiko Iwasaki and Harlan Krumholz at Yale have been running this research effort called the LISTEN study. And I first learned about it sometime in maybe late 2021. And I had been really interested in the emerging discourse around chronic illness in Long Covid in the 2021. So when we were past the most acute phase of the pandemic, and we were seeing this long tail of sequelae in patients, and the conversation had really shifted to one that was about sort of trying to define this new syndrome, trying to understand it, trying to figure out how you could diagnose it, what were we seeing sort of emerge, how are we going to draw boxes around it? And I was so interested in the way that this syndrome was really patient created. It came out of patients identifying their own symptoms and then banning together much, much faster than any kind of institutional science can ever work, getting into message boards together or whatever, and doing their own survey work and then coming up with their own descriptive techniques about what they were experiencing.(25:44):And then beyond that, looking into the literature and thinking about the treatments that they wanted to try for themselves. Patients were sort of at the forefront of every step of recognizing, defining, describing this illness presentation and then thinking about what they wanted to be able to do for themselves to address it. And that was really interesting to me. That was incredibly interesting to me. And it was also really interesting because by, I don't know exactly when 2021 or 2022, it was already a really tense landscape where it felt like there were real factions of folks who were in conflict about what was real and what wasn't real, how things ought to be studied, who ought to be studying them, what would count as evidence in this realm. And all of those questions were just really interesting to me. And the LISTEN study was approaching them in this really thoughtful way, which was Harlan and Akiko sort of partnering really closely with patients who enrolled.(26:57):And it's a decentralized study and people could enroll from all over the world. There's a portion of patients who do have their blood work evaluated, but you can also just complete surveys and have that data count towards, and those folks would be from anywhere in the world. Harlan did this amazing, amazing work to figure out how to collect blood samples from all over the country that would be drawn at home for people. So they were doing this decentralized study where people from their homes, from within the sort of circumstances of their lives around their chronic illness could participate, which that was really amazing to me. And then they were partnering really thoughtfully with these patients just to figure out what questions they wanted to ask, how they wanted to ask them, and to try to capture a lot of multimodal data all at once.(27:47):Survey data, journaling so people could write about their own experience in a freeform journal. They were collecting blood samples, and they were holding these town halls. And the town halls were on a regular basis, Harlan and Akiko, and anybody who was in the study could come on, could log onto a Zoom or whatever, and Harlan and Akiko and their research staff would talk about how things were going, what they were working on, what questions they had, what the roadblocks were, and then they would answer questions from their participants as the study was ongoing. And I didn't think that I had ever heard of something quite like that before. Have you ever heard of anything?Eric Topol (28:32):No. I mean, I think this is important to underscore, this was the first condition that was ever patient led, patient named, and basically the whole path was laid by the patient. So yes, and everything you summarize is so well as to the progress that's been made. Certainly, Harlan and Akiko are some of the people that have really helped lead the way to do this properly as opposed to, unfortunately one and a half billion dollars that have been put to the NIH for the RECOVER efforts that haven't yet led to even a significant clinical trial, no less a validated treatment. But I did think it was great that you spotlighted that just because again, it's thematic. And that gets me to the fourth dimension, which is you're the first prison doctor I've ever spoken to. And you also wrote a piece about that called, “the disillusionment of a Rikers Island Doctor” in the New Yorker, I think it was. And I wonder if you could tell us, firstly, now we're four years into Covid, you were for a good part of that at Rikers Island, I guess.The Rikers Island Prison Doctor During CovidRachael Bedard (30:00):I was, yeah.Eric Topol (30:00):Yeah. And what could be a more worrisome spot to be looking after people with Covid in a prison? So maybe you could just give us some insight about all that.Rachael Bedard (30:17):Yeah, it was really, I mean, it was the wildest time, certainly in my career probably that I'll ever have. In the end of February and beginning of March of 2020, it became very apparent to my colleagues and I that it was inevitable that this virus that was in Wuhan and in Italy was coming to the US. And jails are, we sort of jokingly described them as the worst cruise ships in the world. They are closed systems where everybody is eating, sleeping, going to the bathroom, everything on top of each other. There's an incredible amount of excess human contact in jails and prisons because people don't have freedom of movement and they don't get to do things for themselves. So every single, somebody brings you your mail, somebody brings you your meals, somebody brings you your medications. If you're going to move from point A to point B, an officer has to walk you there. So for a virus that was going to spread through what we initially thought was droplets and then found out was not just droplets but airborne, it was an unbelievably high-risk setting. It's also a setting where folks tend to be sicker than average for their age, that people bring in a lot of comorbidity to the setting.(31:55):And it's not a setting that does well under stress. I mean, jails and prisons are places that are sort of constitutionally violent, and they're not systems that adapt easily to emergency conditions. And the way that they do adapt tends to be through repressive measures, which tends to be violence producing rather than violence quelling. And so, it was just an incredibly scary situation. And in mid-March, Rikers Island, the island itself had the highest Covid prevalence of anywhere in the country because New York City was the epicenter, and Rikers was really the epicenter within New York. It was a wild, wild time. Our first seriously ill patient who ended up getting hospitalized. That was at that time when people were, we really didn't understand very much about what Covid looked like. And there was this guy sitting on the floor and he said, I don't know. I can't really get up.(32:59):I don't feel well. And he had an O2 stat of 75 or something. He was just incredibly hypoxic. It's a very scary setting for that kind of thing, right? It's not a hospital, it's not a place where you can't deliver ICU level care in a place like that. So we were also really worried about the fact that we were going to be transferring all of these patients to the city hospitals, which creates a huge amount of extra burden on them because an incarcerated patient is not just the incarcerated patients, the officers who are with that person, and there are special rules around them. They have to be in special rooms and all of these things. So it was just a huge systems crisis and really painful. And we, early on, our system made a bunch of good guesses, and one of our good guesses was that we should just, or one of our good calls that I entirely credit my bosses with is that they understood that we should advocate really hard to get as many people out as we could get out. Because trying to just manage the population internally by moving people around was not going to be effective enough, that we really need to decant the setting.(34:18):And I had done all of this work, this compassionate release work, which is work to get people who are sick out of jail so that they can get treatment and potentially die in a free setting. And so, I was sort of involved in trying to architect getting folks who were sort of low enough security risks out of jail for this period of time because we thought that they would be safer, and 1500 people left Rikers in the matter of about six weeks.Rachael Bedard (34:50):Which was a wild, wild thing. And it was just a very crazy time.Eric Topol (34:56):Yeah. Well, the word compassion and you go together exceptionally well. I think if we learn about you through your writings, that really shines through and what you've devoted your care for people in these different domains. This is just a sampling of your writings, but I think it gives a good cross section. What makes you write about a particular thing? I mean, obviously the Rikers Island, you had personal experience, but why would you pick Ozempic or why would you pick other things? What stimulates you to go after a topic?Rachael Bedard (35:42):Sometimes a lot of what I write about relates to my personal practice experience in some way, either to geriatrics or death and dying or to the criminal justice system. I've written about people in death row. I've written about geriatrics and palliative care in sort of a bunch of different ways. I am interested in topics in medicine where things are not yet settled, and it feels very of the moment. I'm interested in what the discourse is around medicine and healthcare. And I am interested in places where I think the discourse, not just that I'm taking a side in that discourse, but where I think the framework of the discourse is a little bit wrong. And I certainly feel that way about the Ozempic discourse. And I felt that way about the discourse around President Biden, that we're having not just a conversation that I have a strong opinion about, but a conversation that I think is a little bit askew from the way that we ought to be thinking about it.Eric Topol (36:53):And what I love about each of these is that you bring all that in. You have many different points of view and objective support and they're balanced. They're not just trying to be persuasive about one thing. So, as far as I know, you're extraordinarily unique. I mean, we are all unique, but you are huge standard deviations, Rachael. You cover bases that are, as I mentioned, that are new to me in terms of certainly this podcast just going on for now a couple of years, that is covering a field of both geriatrics and having been on the corrections board and in prison, particularly at the most scary time ever to be working in prison as a physician. And I guess the other thing about you is this drive, this humanitarian theme. I take it you came from Canada.Rachael Bedard (37:59):I did.Eric Topol (37:59):You migrated to a country that has no universal health.Rachael Bedard (38:03):That's right.Eric Topol (38:03):Do you ever think about the fact that this is a pretty pathetic situation here?Rachael Bedard (38:08):I do. I do think about it all the time.Eric Topol (38:10):In our lifetime, we'll probably never see universal healthcare. And then if you just go a few miles up north, you pretty much have that.Rachael Bedard (38:18):Yeah, if you've lived in a place that has universal healthcare and you come here, it's really sort of hard to ever get your mind around. And it has been an absolute possessing obsession of my entire experience in the US. I've now been here for over 20 years and still think it is an unbelievably, especially I think if you work with marginalized patients and how much their lack of access compounds the difficulty of their lives and their inability to sort of stabilize and feel well and take care of themselves, it's really frustrating.Advice for Bringing Humanities to Medicine in a CareerEric Topol (39:14):Yeah, yeah. Well, I guess my last question to you, is you have weaved together a career that brings humanities to medicine, that doesn't happen that often. What's your advice to some of the younger folks in healthcare as to how to pull that off? Because you were able to do it and it's not easy.Rachael Bedard (39:39):My main advice when people ask me about this, especially to students and to residents who are often the people who are asking is to write when you can or pursue your humanities interests, your critical interests, whatever it is that you're doing. Do it when you can, but trust that your career is long and that you have a lot of time. Because the thing that I would say is I didn't start publishing until I was in fellowship and before that I was busy because I was learning to become a doctor. And I think it's really important that my concern about being a doctor who's a hybrid, which so many of us are now. A doctor or something else is you really do want to be a good doctor. And becoming a good doctor is really hard. And it's okay if the thing that is preoccupying you for the first 10 years of your training is becoming a great clinician. I think that's a really, really important thing to do. And so, for my first 10 years for med school and residency and chief residency and fellowship, I would write privately on the side a fair amount, but not try to publish it, not polish that work, not be thinking in sort of a careerist way about how I was going to become a doctor writer because I was becoming a doctor. And that was really preoccupying.(41:08):And then later on, I both sort of had more time and mental space to work on writing. But also, I had the maturity, I think, of being a person who was comfortable in my clinical identity to have real ideas and insights about medicine that felt different and unique to me as opposed to, I barely understand what's going on around me and I'm trying to pull it together. And that's how I would've been if I had done it more, I think when I was younger. Some people are real prodigies and can do it right out the gate, but I wasn't like that.Eric Topol (41:42):No, no, I think that's really sound advice because that's kind of the whole foundation for everything else. Is there a book in the works or will there be one someday?Rachael Bedard (41:53):There may be one someday. There is not one now. I think about it all the time. And that same advice applies, which is I believe in being a late bloomer and taking your time and figuring out what it is you really want to do.Eric Topol (42:10):Yeah. Well, that's great. Have I missed anything? And obviously we only can get to know you in what, 40 minutes to some extent, but have I not touched on something that you want to bring up?Rachael Bedard (42:23):No, I don't think so. Thank you for this conversation. It's been lovely.Eric Topol (42:28):No, I really enjoyed it. I'll be following your career. It's extraordinary already and you've got decades ahead to make an impact and obviously thinking of all these patients that you look after and have in the past, it’s just extraordinary. So what a joy to talk with you, Rachael, and I hope we'll have a chance to do that again in the times ahead.Rachael Bedard (42:51):Me as well. Thank you so much for inviting me.**********************************************Thank you for listening, reading or watching!The Ground Truths newsletters and podcasts are all free, open-access, without ads.Please share this post/podcast with your friends and network if you found it informative!Voluntary paid subscriptions all go to support Scripps Research. Many thanks for that—they greatly help fund our education and summer internship programs.Thanks to my producer Jessica Nguyen and Sinjun Balabanoff for audio and video support at Scripps Research.Note: you can select preferences to receive emails about newsletters, podcasts, or all I don’t want to bother you with an email for content that you’re not interested in. Get full access to Ground Truths at erictopol.substack.com/subscribe

Superimposed on an impressive body of work on the blood-brain-barrier and immune system, Prof Akassoglou and her collaborators just published an elegant study in Nature that centered on the direct binding os the SARS-CoV-2 spike protein to fibrin with marked downstream pro-inflammatory effects. The findings and potential treatments have implications beyond Covid, Long Covid to other neurologic diseases.Full videos of all Ground Truths podcasts can be seen on YouTube here. The audios are also available on Apple and Spotify.Transcript with links to audio and to relevant papers, graphicsEric Topol (00:07):Well, hello this is Eric Topol with Ground Truths, and with me today is Katerina Akassoglou. She is at the Gladstone Institute and she is a remarkable neuroimmunologist who has been doing extraordinary work for three decades to unravel the interactions between the brain, blood vessels and the role of inflammation. So Katerina, there's a lot to discuss, so welcome.Katerina Akassoglou (00:40):Thank you. Thank you so much. It's a great pleasure to join.By Way of BackgroundEric Topol (00:43):It's really interesting going back in your career. First of all, we're thankful that you immigrated here from Greece, and you have become one of the leading scientists in this discipline of important discipline of neuroimmunology, which is not just about Covid that we're going to talk about, but Alzheimer's and neurodegenerative diseases. This is a really big hot area and you're definitely one of the leaders. And what I was impressed is that all these years that you've been working on the integrity of the blood-brain barrier, the importance of fibrinogen and fibrin, and then comes along the Covid story. So maybe what we can do is start with that, which is you've made your mark in understanding this whole interaction between what can get into the brain, through the blood-brain barrier and incite inflammation. So this has been something that you've really taken to the extreme knowledge base. So maybe we can start with your work there before we get into the important seminal Nature paper that you recently published.Katerina Akassoglou (01:57):Yes, of course. So since very early on, I was still a graduate student when we made the first discovery and at the time was like mid-90s, so it was really ahead of its time. That dysregulation of cytokine expression in the brain of mice was sufficient to induce the whole cascade of events, triggering neurodegeneration, demyelination in pathological alterations, very reminiscent of multiple sclerosis pathology. And it was really hard to publish that study at the time because it was not yet accepted that this regulation of the immune system modeling the brain can be linked to neurodegeneration. So that was 1995 when we made that discovery, and I became really interested, what are the pathogenic triggers that actually polarized the immune cells in the brain? So with this, of course, this transgenic animal was expressing TNF, it was an artificially made animal that we made, but naturally what were the triggers that would polarize the innate immune cells? So I looked really early on in this mice and what I found was that the very first event was leaks of blood-brain barrier. It was opening of the blood-brain barrier in this mouse before inflammation, before demyelination, before neuronal loss. And this is really what shaped the question that, is it possible that these blood leaks that happened very early in the pathology, could this be the instigators of pathogenic inflammation in the brain?Eric Topol (03:34):Yeah. So in a way, you got at this question because of the chicken-and-egg and what happens first, and you got to the temporal saying, which happened first as you said, the leak before you could see evidence of inflammation and being able to study this of course in the experimental model, which you couldn't really do in people. And what I love about the description of your career, which has been quite extraordinary contributions is connecting the dots between the blood, the inflammatory response and the brain. Perhaps no one has done that like you have. And before we get into the recent paper, a lot of people are not aware that a year ago, a group in the UK known as PHOSP-COVID, they published a really important paper in Nature Medicine of over 1,800 people who were hospitalized with Covid and they found that fibrinogen was the best marker for cognitive deficits at 6 and 12 months (Figure below)(04:40):So that's just one of many papers, but it's a particularly well done study that already before you got into this work that recently published had emphasized fibrinogen. And by the way, again, having spent a lot of years in clots in the arteries, for me, we have to just get it down to fibrinogen plus thrombin gets you to fibrin. Okay, so fibrin is a major player here when fibrinogen is cleaved. So here we have the basis that you established, which is the fibrinogen leakage into the brain, activating inflammation, activating microglia, which like the macrophages of the brain and inciting the whole process. And before we close, I want to not just talk about Covid, but Alzheimer's too. But now let's get into the study that you did, [Fibrin drives thromboinflammation and neuropathology in COVID-19] which is striking, I mean really striking. And can you kind of take us through, because you not only demonstrated the importance of fibrin in inciting neuroinflammation in this model, but also how you could reverse it or prevent it. So this, and you looked at it in many different ways, this was a systematic approach. Maybe you can take us through how you were able to make such compelling evidence.The Multimodal EvidenceKaterina Akassoglou (06:09):Yes, thank you. First of all, thank you for bringing up the human relevance because this was also our inspiration for the work that we did in the Covid study. So as you mentioned in Covid patients, fibrinogen unbiased mass spec analysis was identified as the predictive biomarker for cognitive impairment in Long Covid patients. And this was in addition to also neuropathology data about the abundance of fibrin deposition in the brain. And these were studies that were done by NIH that have found deposition of fibrin in the brain and the reports for the abnormal and puzzling coagulation in Covid that is not setting other infections and also in many cases not always relating with the severity of symptoms. So even mild cases of Covid also had increased coagulation. I was really intrigued by this human, all this evidence in human data, and I thought that maybe the way that we're thinking about this, that it's systemic inflammation that drives the clotting.(07:24):Maybe there's another aspect to this. Maybe there is a direct effect of the virus with the coagulation cascade, and in this way maybe this can be an instigator of inflammation. So this was the original idea to be able to reconcile this data from the clinic about why do we have this prevalence of coagulopathy in Covid. And of course, the second question is, could this also be a driver of the disease? And of course, we're in a unique position because we have been studying this pathway now for over 20 years to have all the toolbox, the genetic toolbox, the pharmacologic toolbox to be able to actually really address these questions with genetic loss of function studies, with a blood innate immunity multiomics pipeline that we have set up in the lab. And of course, with preclinical pharmacology in our ABSL3 facility. So we had the infrastructure in place and the source in place to actually really dissect this question with both genetic tools as well as also technology platforms.Eric Topol (08:29):And you had in vivo imaging, you're the director of in vivo imaging for Gladstone and UCSF. So you do have the tools to do this.Katerina Akassoglou (08:38):Yes. The imaging that you mentioned is really important because this is, we employed that very early in our studies over now 15 years ago. And the reason was sometimes from snapshots of histopathology, you cannot really understand the sequence of events. So by being able to image these processes, both neuronal activity, microglia activation, infiltration of peripheral cells in the brain, this is how we could see the steps that what happens early on and to be able to answer these chicken-and-egg questions that you mentioned. So these were very, they're very important experiments, especially at the beginning because they were hypothesis driving and we were able to ask the right questions to drive our research program.Eric Topol (09:26):Now was the binding of the spike protein to one key site in fibrinogen, was that known before? [See outstanding Figure below from Trends in Immunology]Katerina Akassoglou (09:36):No, this was not known. So there was evidence that there are abnormal clots in Covid, but it was not known whether the spike protein would directly bind to protein to the coagulation cascade. So one of the key discoveries in our study was to use peptide array mapping and be able to identify not only the binding, but exactly the domains on fibrin that spike binds too. And what we found was two key domains, one the inflammatory domain and the other the plasmin binding site, which is important for fibrin degradation. So this suggested a potential dual deleterious role for this interaction, both by maybe affecting inflammation, but also delaying fibrinolysis, which is the degradation of this toxic protein from the brain. And indeed, we found that this interaction was responsible for all these two aspects, including decreased degradation, more inflammation, but also at the same time increased, increased coagulation. So it was a really pathogenic interaction.Eric Topol (10:47):Yeah, actually it's pretty striking. You have these two sites, the plasmin cleavage site of fibrinogen, which as you say, we knew there was a problem with clots. We knew that, but we didn't know exactly the spike protein how exactly it was implicated, particularly with fibrinogen. And then this other site, the CD11b-C18, now that's fancy for surface receptors of macrophages. And basically, this is critical because it's this microglia activation in the brain, and I know you saw it in the lungs as well through this other site that spike protein activated. So you had a twofer here of things that you discovered that the SARS-CoV-2 spike protein was capable of doing. This was a really big revelation. And then you also looked at mice that were genetically manipulated. So maybe you can, because before we get to your antibody monoclonal, the ways that you proved this were, I mean, one thing after another is really systematic. So maybe you can teach us about that.Thanks for reading Ground Truths! This post is public so feel free to share it.Establishing CausalityKaterina Akassoglou (12:08):Yeah, sure. So the first was about chemistry experiment. So this of course, we had to get to the next step to see is there any causality for this pathway. So we employed genetic loss of function studies and we had knockout mice, either fibrinogen knockout mice, this mice have all blood proteins except fibrinogen, and they have a delay in coagulation so they don't clot properly. But we also had a mutant mouse, which is a fibrinogen NK mouse. And this was a mutation only within this inflammatory domain that you mentioned, inflammatory domain that binds to C11b-C18. Other names for this is of course complement receptor 3, Mac-1 (αMβ2). It's the same, many names for this receptor, that as you mentioned, is expressed not only in microglial in the brain, but also peripheral immune cells including macrophages as well as also neutrophils which are CD11b expressing.(13:12):So we now have genetic models to be able to look at both complete depletion of fibrinogen, but also a very specific mutation and very selective mutation that only blocks the inflammatory properties without affecting the properties of fibrin in hemostasis. And these mice were made many years ago by a very close collaborator, Jay Degen at the University of Cincinnati. So what we found is that when we block either the inflammatory domain or we completely deplete fibrinogen, there was this profound protection after infection in internasal infection with the virus in lung inflammation. And this was both suppression of oxidative stress and this pathogenic inflammation in the lung, but also decreasing fibrosis, which has been associated with also Long Covid. And the surprise came from the transcriptomic data. So when we did transcriptomic analysis in this mice in the lungs, we found perhaps the expected decrease in the immune signatures in macrophages. This was in line with our previous work in, as you mentioned, Alzheimer's models, multiple sclerosis models. But what also was really surprising is there was that genes that are associated with activation of NK cells were upregulated. And of course this was the first time we had infected these mice, previously we had not done an infection before. So I think that maybe because of this region we had not seen before in our data this immunomodulatory role of fibrin that not only surprises the macrophage response, but also increases these NK cells that are important for viral clearance.Eric Topol (15:00):So again, the finding another important unique finding is the natural killer (NK) cells and effect there from the activation of this, as you said, the inflammation site or the CD11b-C18 that we've been talking about. So now another layer of this, a dimension of your Nature paper was that you tested an antibody that you already had developed so-called 5B8. A monoclonal that specifically binds to the domain of the one we're talking about this inflammation domain of fibrinogen. So can you tell us about what that showed?Katerina Akassoglou (15:45):Yes, so we tested this antibody in different models of Covid, which were both models with neuroinvasion and models without neuroinvasion. So we used both transgenic mice for hACE2, the human ACE2 infected with Delta, but we also use mouse adapted viruses like Beta that is just in the wild type mice with no transgenic being involved that these are without neuroinvasion. And we wanted to see if the antibody had any potential protective effects. And what we found is that the antibody protected from inflammation in the lung. So the data looked so similar with a genetic mutation of this pathway, protection from inflammation, decreased fibrosis, increased viral clearance, so decreased spike and viral proteins in the lungs. But we also found a protection in the brain. So the brains of this mice, including both the models we used with neuroinvasion and without, they both have had microglia activation in the brain. And we also found neuronal loss in the Delta infected mice and the antibody protected from both neuroinflammation but also improved neuronal survival in the mice. Showing that there can be this despite regardless of which model we used, there was this protective effect suggesting that by blocking fibrin, either the periphery or in the brain, this could be protected for these models.Eric Topol (17:28):Yeah, so I mean this is fascinating because until now, until this report of yours and your colleagues at Gladstone, there was knowledge that there would be neuroinflammation from Covid, both in patients from various biomarkers and imaging as well as in experimental model. But what this did was take it to the fibrin story, and I guess that's one of the questions you nailed that how important fibrin is, but that doesn't necessarily rule out other triggers of neuroinflammation, right?Katerina Akassoglou (18:04):Oh, absolutely not. So I think that this is one of the mechanisms that can be very important, especially in some patients. But we know that there are additional of course mechanisms of neuroinflammation including auto-antibody responses, as well as also endotheliopathy that are persistent endotheliopathy, this can be interacting also with each other. So I think that it's important for future research that we understand how do these mechanisms feed into each other? Are there a positive feedback loops between autoimmune mechanisms and coagulopathy and endothelial dysfunction with inflammation? But I think most importantly, I think that if we're thinking of this in the context of patients, can we identify patients with mechanism that might be more prevalent in specific cases of Long Covid and tailor our potential future clinical trials towards the needs of Long Covid patients?Towards TreatmentEric Topol (19:06):Absolutely. I did interview some months back on Grounds Truths, Michelle Monje at Stanford, who I'm sure and interact with, and she's also works not so much on the fibrin side, but on neuroinflammation and the likeness between this condition in people and chemo brain because of the inflammation that's seen there. So we've talked about the multiple triggers that could contribute to brain inflammation, which I think most people would say in Long Covid this is one of the most, besides obviously the lack of energy, the profound fatigue and disability, but the cognitive function hit, not just brain fog is often profound. And we've just seen some reports about that, and particularly in hospitalized patients, how bad that can be. So that gets us to a potential treatment. Now, one of the things that's out there dangling, there's many things that people have talked about in terms of why can't we have a treatment for Long Covid?(20:13):And now of course this fibrin pathway, if you will, lends itself to many possibilities, whether it's anticoagulants or fibrinolytics like a tPA or things like nattokinase, which is a Japanese food enzyme that you could get at the nutrition centers or whatever. What are your thoughts? Because we don't have any good studies. There are all these little, tiny studies and they don't provide much conclusion, and you have an antibody that could potentially be effective. As I understand it, you set up a company some years ago, Therini Bio and used to be called MedaRed. You're the first woman scientist at Gladstone to develop a spin out company, which is another point of congratulations on that. But could the antibody be tested in patients or what do you think about these other possibilities?Katerina Akassoglou (21:15):Yes, yes. These are great questions. So first of all, the different approaches that you mentioned have very different mechanism of action. So degrading fibrin, the degradation products of fibrin also can have deleterious effects. The dimer, for example, can be very pro-inflammatory. So at the same time, blocking coagulation can also have a diverse effects because this can lead to excessive hemorrhage. So the approach that we took was to selectively block the inflammatory properties of fibrin without affecting beneficial effects of the molecule in normal hemostasis. So the challenge when I made the antibody was to be able to dissect these two functions of fibrin. It's our most important clotting factor, but at the same time, a molecule with profound pro-inflammatory capacity. So the observation that these two domains, the clotting domain and inflammatory domain were not overlapping, was really the foundation of this invention was that we could maybe create this antibody to be able to target them in a selective way.Other Neurologic Conditions (22:31):So the antibody I developed is neutralizing blood toxicity by blocking the inflammatory domain of fibrin without adverse coagulation effects. And it's now completing phase one trials. So it has already completed the single ascending dose at 40 milligram per kilogram. It's interim data were announced already for this trial, with no safety signals. So if the antibody completes this year, the phase one trials, then it should be possible to be tested in different patient populations. You mentioned before chemo brain, and I think it's important that we think that blood-brain barrier disruption occurs among many neurological conditions, and it's an early event associated with early disease onset and worse prognosis in multiple sclerosis, Alzheimer's disease, traumatic injuries. So I think that it's by developing a strategy, therapeutic strategy to neutralize blood toxicity, this can have applications in a wide range of neurological conditions with vascular dysfunction.Eric Topol (23:54):Yeah, no. In your Nature Immunology 2020 piece [Figure below], you started with the 1883 identification of multiple sclerosis (MS) lesions were “engorged with blood”, the first link between blood leaks and brain inflammation. So this has enormous potential. And what I like about this Katerina is that you've dissected the clot component versus the inflammatory trigger of the fibrinogen and fibrin story. And this is so vital because if you keep throwing these things that just going to work on the clot and not deal with the pro-inflammatory consequences, then you're going to get the wrong impression that clots are not that important. And by the way, you did mention, and I want to come back to that too, endothelial inflammation, which is another feature of Long Covid is another kind of interactive part of this because when the lining of the blood vessel is inflamed, it will attract microthrombi and also be a participant in this whole affair. What do you think about Alzheimer's and the prospects of being able to interfere with Alzheimer's? We have 20 years in someone before this process takes hold and meets clinical manifestations. Would an antibody like this ever be useful along the way?Katerina Akassoglou (25:29):Yeah, so well, our antibody was tested first in Alzheimer's, this models when it was originally published, and we performed reversal trials in Alzheimer's models. So we dosed mice when they have established amyloid plaques, microglia activation, neuronal loss, and we could reverse this effect so it could increase cholinergic neurons in mice, reduce inflammation in a very selective way, only the neurotoxic part of inflammation and for genetic depletion of this pathway with akin mice in Alzheimer's disease. Also, improves from cognitive impairment, and we now have a new paper in Cell Press that is showing this effects also with really nice and unbiased machine learning models for behavioral segmentation [Figure below].So I think that there is the data both from genetic studies and the antibody show projection in Alzheimer's disease. And of course, as you might have read the recent Lancet report from the Lancet committee on dementia that identified the vascular risk factors as the key contributors, especially post sporadic cases of Alzheimer's disease that is over 90% of Alzheimer's disease that is not genetically linked.(26:58):So I think that there is a real need in Alzheimer's disease to be able to block this vascular induced pathology. And an antibody like the fibrin neutralizing therapy could be positioned to be protective from the vascular induced immune-mediated neurodegeneration in this disease as well. I mean, ultimately, I think that we need to be thinking the terms of efficacy. So we want to have a drug that is efficacious, but we also want it to be selective. And the selectivity is really important because the immune system has so many protective functions. So if we block phagocytosis, we end up with more debris, decrease of neurorepair, anti-myelination. So by blocking a ligand here and not blocking, not eliminating a cell type or blocking a global pathway in this cell, but biologic a single ligand, I think we have been able to achieve this balance between efficacy, but also safety because we only block this neurotoxic populations and not the entire innate immune response that also has been beneficial for metastatic functions in the brain.Blocking NeuroinflammationEric Topol (28:19):So you're bringing up another critical concept about targeting the inflammation, this kind of goldilocks story of how much you interfere with the immune response and how much you are able to reduce the adverse pro-inflammatory effects. So that gets me to what if we don't know in any given patient how much fibrin is having a role in their Long Covid. Although we know it has to be a prominent feature because we saw it in, not just a hospitalized patient series that I mentioned we reviewed, but other papers as well. But what about if you just try to take on inflammation like through a GLP-1 drug or cGAS–STING or any of these really strong anti-inflammatory pathways. Do you see a difference in a generalized approach versus a specific approach that is really fibrin centered?Katerina Akassoglou (29:22):Yeah, so we have a focus actually on both because we wanted to dissect the downstream intracellular pathways of fibrin, and it's interesting that we can find specific inflammatory mediators that potentially can also be targeted as well, to be able to preserve that specificity, which I think is really important because if we don't preserve the specificity, we'll end up with a lot of adverse effects by eliminating major immune responses. But the point that you raised I think is really important because it's not enough to have an efficacious and selective drug if you don't know the patient population that will benefit from this drug. So I think that in addition to the drug discovery studies, it's important to develop also biomarker programs with both fluid biomarkers, but also imaging biomarkers to be able to identify the patient populations that will benefit from such treatment.(30:25):So if for example, a patient population has a fibrin deposition, blocking only downstream might not be enough, and it might be really important to neutralize this fibrin toxicity in the brain of patients. And with our target engagement studies, we show that at least in animal models, the antibody can be there. So I'm very encouraged by also programs that are going on now in the scientific community to develop noninvasive ligands to be able to image fibrin in the brain that are already tested in different patient populations like multiple sclerosis. Because I think we're going to learn so much from the biology as we start interrogating and asking these questions now in different patient populations.Eric Topol (31:14):I think that's a vital point you're making because the success of a clinical trial here in a clinical syndrome that is mosaic with lots of different types of pathways. If you can nail down the patients that would have the most to stand to benefit from a particular intervention, that the chance of you not missing the benefit that is matching the marker, what image marker or other markers is so vital. Well, we've talked, I think, about some fascinating discoveries that you and your colleagues have made. I mean, it's really extraordinary, and obviously we need this in Long Covid. But you know what, Katerina, it's almost made me think that you were warming up to this for three decades, that somehow or other you were working on all this stuff and then came Covid. Is that how you see it, that somehow or other you didn't know that all the work you were doing was going to wind up in this space?Katerina Akassoglou (32:18):Oh, I never thought I would work in a virology project. This collaboration started over Zoom with Warner Greene. We were both sheltering in place. It was the beginning of the pandemic, and the first reports were coming out about this puzzling coagulopathy. And our labs were hardly operational at the time, as you know, we had to close down our labs for a while. And however, this was a very big problem, and we thought that this is our role as scientists. If we feel that we can contribute and we have the tools to contribute, we felt that it's important that we pivot some part of our research, and even we wouldn't be doing this before, but it was important to pivot a part of our research and collaborate. And I think studies like this, this study would have been impossible without a team of collaborators. As you know, there were over 50 scientists involved at Gladstone, UCSF, UCLA, UCSD, Stanford University. Without collaboration, this study wouldn't be possible. So I'm really grateful to everyone who came together to solve this problem because I think that's what scientists should be doing. We should be solving problems as they arise.Eric Topol (33:41):Well, and also, I think a lot of people don't realize that, for example, when the Covid vaccines came along, people think, oh, well, it all got done in 10 months since the sequence of the virus, when in fact it took 30 years at least between all the factors that went into having an mRNA and sequencing virus and nanoparticles. And in many ways, your arc of this work is like that because it took three decades to have all the tools and the basic understanding, the antibody that you had developed for different reasons and this fascinating unraveling of what's going on in the model and undoubtedly in some patients at least as well. So before we wrap up, have I missed anything about this just remarkable work you've done?Katerina Akassoglou (34:33):Oh, thank you. I just want to thank you for this discussion and thank you for emphasizing the different areas and the different decisions that this pathway can have implications both for our understanding, our basic understanding of the blood brain immune interface, as well as also potential translation. And I think that the curiosity sometimes of how things work, I never thought it would work on Covid, like you mentioned at the beginning, but I think that basic science and curiosity driven science can sometimes lead to discoveries with translational implications that hopefully might benefit patients one day.Eric Topol (35:21):Yeah, well, undoubtedly it will. We're indebted to you, Katerina and all the folks that you have teamed up with, connecting the dots at the neurovascular interface. Phenomenal work and will follow the subsequent with great interest and it will likely not just a story about Long Covid, but other areas as well, so thank you.*********************************Thanks for listening, reading or watching!The Ground Truths newsletters and podcasts are all free, open-access, without ads.Please share this post/podcast with your friends and network if you found it informative!Voluntary paid subscriptions all go to support Scripps Research. Many thanks for that—they greatly help fund our summer internship programs.Thanks to my producer Jessica Nguyen and Sinjun Balabanoff for audio and video support at Scripps Research.Note: you can select preferences to receive emails about newsletters, podcasts, or all I don’t want to bother you with an email for content that you’re not interested in. Get full access to Ground Truths at erictopol.substack.com/subscribe

When I think of digital biology, I think of Patrick Hsu—he’s the prototype, a rarified talent in both life and computer science, who recently led the team that discovered bridge RNAs, what may be considered CRISPR 3.0 for genome editing, and is building new generative A.I. models for life science. You might call them LLLMs-large language of life models. He is Co-Founder and a Core Investigator of the Arc Institute and Assistant Professor of Bioengineering and Deb Faculty Fellow at the University of California, Berkeley.Above is a brief snippet of our conversation. Full videos of all Ground Truths podcasts can be seen on YouTube here. The audios are also available on Apple and Spotify.Here’s the transcript with links to the audio and external links to relevant papers and things we discussed.Eric Topol (00:06):Well hello, it's Eric Topol with Ground Truths and I'm really delighted to have with me today Patrick Hsu. Patrick is a co-founder and core investigator at the Arc Institute and he is also on the faculty at the University of California Berkeley. And he has been lighting things up in the world of genome editing and AI and we have a lot to talk about. So welcome, Patrick.Patrick Hsu (00:29):Thanks so much. I'm looking forward to it. Appreciate you having me on, Eric.The Arc InstituteEric Topol (00:33):Well, the first thing I'd like to get into, because you're into so many important things, but one that stands out of course is this Arc Institute with Patrick Collison who I guess if you can tell us a bit about how you two young guys got to meet and developed something that's really quite unique that I think brings together investigators at Stanford, UCSF, and Berkeley. Is that right? So maybe you can give us the skinny about you and Patrick and how all this got going.Patrick Hsu (01:05):Yeah, sure. That sounds great. So we started Arc with Patrick C and with Silvana Konermann, a longtime colleague and chemistry faculty at Stanford about three years ago now, though we've been physically operational just over two years and we're an independent research institute working at the interface of biomedical science and machine learning. And we have a few different aspects of our model, but our overall mission is to understand and treat complex human diseases. And we have three pillars to our model. We have this PI driven side of the house where we centrally fund our investigators so that they don't have to write grants and work on their very best ideas. We have a technical staff side of the house more like you'd see in a frontier AI lab or in biotech industry where we have professional teams of R&D scientists working cross-functionally on higher level organizational wide goals that we call our institute initiatives.(02:05):One focused on Alzheimer's disease experimentally and one that we call a virtual cell initiative to simulate human biology with AI foundation models. And our third pillar over time is to have things not just end up as academic papers, but really get things out into the real world as products or as medicines that can actually help patients on the translational side. And so, we thought that some really important scientific programs could be unlocked by enabling new organizational models and we are experimenting at the institutional scale with how we can better organize and incentivize and support scientists to reach these long-term capability breakthroughs.Patrick, Patrick and SilvanaEric Topol (02:52):So the two Patrick’s. How did you, one Patrick I guess is a multi-billionaire from Stripe and then there's you who I suspect maybe not quite as wealthy as the other Patrick, how did you guys come together to do this extraordinary thing?Patrick Hsu (03:08):Yeah, no, science is certainly expensive. I met Patrick originally through Silvana actually. They actually met, so funny trivia, all three Arc founders did high school science together. Patrick and Silvana originally met in the European version of the European Young Scientist competition in high school. And Silvana and I met during our PhDs in her case at MIT and I was at Harvard, but we met at the Broad Institute sort of also a collaborative Harvard, MIT and Harvard hospitals Institute based in Kendall Square. And so, we sort of in various pairwise combinations known each other for decades and worked together for decades and have all collectively been really excited about science and technology and its potential to accelerate societal progress. Yet we also felt in our own ways that despite a lot of the tremendous progress, the structures in which we do this work, fund it, incentivize it and roll it out into the real world, seems like it's really possible that we'll undershoot that potential. And if you take 15 years ago, we didn't have the modern transformer that launched the current AI revolution, CRISPR technology, single-cell, mRNA technology or broadly addressable LNPs. That’s a tremendous amount of technologies have developed in the next 15 years. We think there's a real unique opportunity for new institutes in the 2020s to take advantage of all of these breakthroughs and the new ones that are coming to continue to accelerate biological progress but do so in a way that's fast and flexible and really focused.Eric Topol (04:58):Yeah, I did want to talk with you a bit. First of all before I get to the next related topic, I get a kick out of you saying you've worked or known each other for decades because I think you're only in your early thirties. Is that right?Patrick Hsu (05:14):I was lucky to get an early start. I first started doing research at the local university when I was 14 actually, and I was homeschooled actually until college. And so, one of the funny things that you got to do when you're homeschooled is well, you could do whatever you want. And in my case that was work in the lab. And so, I actually worked basically full time as an intern volunteer, cut my teeth in single cell patch clamp, molecular biology, protein biochemistry, two photon and focal imaging and kind of spiraled from there. I loved the lab, I loved doing bench work. It was much more exciting to me than programming computers, which was what I was doing at the time. And I think these sort of two loves have kind of brought me and us to where we are today.Eric Topol (06:07):Before you got to Berkeley and Arc, I know you were at Broad Institute, but did you also pick up formal training in computer science and AI or is that something that was just part of the flow?Patrick Hsu (06:24):So I grew up coding. I used to work through problems sets before dinner growing up. And so, it's just something that you kind of learn natively just like learning French or Mandarin.New Models of Funding Life ScienceEric Topol (06:42):That's what I figured. Okay. Now this model of Arc Institute came along in a kind of similar timeframe as the Arena BioWorks in Boston, where some of the faculty left to go to Arena like my friend Stuart Schreiber and many others. And then of course Priscilla and Mark formed the Chan Zuckerberg Institute and its biohub and its support. So can you contrast for one, these three different models because they’re both very different than of course the traditional NIH pathway, how Arc is similar or different to the others, and obviously the goal here is accelerating things that are going to really make a difference.Patrick Hsu (07:26):Yeah, the first thing I would say is zooming out. There have been lots of efforts to experiment with how we do science, the practice of science itself. And in fact, I've recently been reading this book, the Demon Under the Microscope about the history of infectious disease, and it talks about how in the 1910s through the 1930s, these German industrial dye manufacturing companies like Bayer and BASF actually launched what became essentially an early model for industrial scale science, where they were trying to develop Prontosil, Salvarsan and some of these early anti-infectives that targeted streptococcus. And these were some of the major breakthroughs that led to huge medical advances on tackling infectious disease compared to the more academic university bound model. So these trends of industrial versus academic labs and different structures to optimize breakthroughs and applications has been a through current throughout international science for the last century.(08:38):And so, the way that we do research today, and that's some of our core tenets at Arc is basically it hasn't always been this way. It doesn't need to necessarily be this way. And so, I think organizational experiments should really matter. And so, there's CZI, Altos, Arena, Calico, a variety of other organizational experiments and similarly we had MRC and Bell Labs and Xerox PARCS, NIBRT, GNF, Google Research, and so on. And so, I think there are lots of different ways that you can organize folks. I think at a high level you can think about ways that you can play with for-profit versus nonprofit structures. Whether you want to be a completely independent organization or if you want to be partnered with universities. If you want to be doing application driven science or really blue sky curiosity driven work. And I think also thinking through internally the types of expertise that you bring together.(09:42):You can think of it like a cancer institute maybe as a very vertically integrated model. You have folks working on all kinds of different areas surrounding oncology or immunotherapy and you might call that the Tower of Babel model. The other way that folks have built institutes, you might call the lily pad model where you have coverage of as many areas of biomedical research as possible. Places like the Whitehead or Salk, it will be very broad. You'll have planned epigenetics, folks looking at RNA structural biology, people studying yeast cell cycle, folks doing in vivo melanoma models. It's very broad and I think what we try to do at Arc is think about a model that you might liken more to overlapping Viking shields where there's sort of five core areas that we're deeply investing in, in genetics and genomics, computation, neuroscience, immunology and chemical biology. Now we really think of these as five areas that are maybe the minimal critical mass that you would need to make a dent on something as complicated as complex human diseases. It's certainly not the only thing that you need, but we needed a critical mass of investigators working at least in these areas.Eric Topol (11:05):Well, yeah, and they really converge on where the hottest advances are being made these days. Now can you work at Arc Institute without being one of these three universities or is it really that you maintain your faculty and your part of this other entity?Patrick Hsu (11:24):So we have a few elements to even just the academic side of the house. We have our core investigators. I'm one of them, where we have dually appointed faculty who retain their latter rank or tenured appointment in their home department, but their labs are physically cited at the Arc headquarters where we built out a lab in Stanford Research Park in Palo Alto. And so, folks move their labs there. They continue to train graduate students based on whatever graduate programs they're formally affiliated with through their university affiliation. And so, we have nearly 40 PhD students across our labs that are training on site every day.(12:03):So in addition to our core investigators, we also have what we call our innovation investigators, which is more of a grant program to faculty at our partner universities. They receive unrestricted funding from us to seed a new project or accelerate an existing area in their group and their labs stay at their home campus and they just get that funding to augment their work. The third way is our technical staff model where folks basically just come work at Arc and many of them also are establishing their own research groups focusing on technology R&D areas. And so, we have five of those technology centers working in molecular engineering, multi-omics, complex cellular models, in vivo models, and in machine learning.Discovery of Bridge RNAsEric Topol (12:54):Yeah, that's a great structure. In fact, just a few months ago, Patrick Collison, the other Patrick came to Stanford HAI where I'm on the board and you've summarized it really well and it's very different than the other models and other entities, companies included that you mentioned. It's really very impressive. Now speaking of impressive on June 26, this past few months ago, which incidentally is coincident with the draft genome in the year 2000, the human sequence. You and your colleagues, perhaps the most impressive jump in terms of an Arc Institute contribution published two papers back-to-back in Nature about bridge RNA: [Bridge RNAs direct programmable recombination of target and donor DNA] and [Structural mechanism of bridge RNA-guided recombination.] And before I get you to describe this breakthrough in genome editing, some would call it genome editing 3.0 or CRISPR 3.0, whatever. But what we have today in the clinic with the approval of CRISPR 1.0 for sickle cell and thalassemia is actually quite crude. I think most people will know it's just a double stranded DNA cleavage with all sorts of issues about repair and it's not very precise. And so, CRISPR 2.0 is supposed to be represented by David Liu's contributions and his efforts at Broad like prime and base editing and then comes yours. So maybe you can tell us about it and how it is has to be viewed as quite an important advance.Patrick Hsu (14:39):The first thing I would say before CRISPR, is that we had RNA interference. And so, even before this modern genome editing revolution with programmable CRISPRs, we had this technology that had a lot of the core selling points as well. Any target will now become druggable to us. We simply need to reprogram a guide RNA and we can get genetic access to things that are intracellular. And I think both the discovery of RNA interference by Craig Mello and Andy Fire or the invention or discovery of programmable CRISPR technologies, both depend on the same fundamental biological mechanism. These non-coding guide RNAs that are essentially a short RNA search string that you can easily reprogram to retarget a desired enzyme function, and natively both RNAi and CRISPR are molecular scissors. Their RNA or DNA nucleases that can be reprogrammed to different regions of the genome or the transcriptome to make a cut.(15:48):And as bioengineers, we have come up with all kinds of creative ways to leverage the ability to make site specific cuts to do all kinds of incredible things including genome editing or beyond transcriptional up or down regulation, molecular imaging and so on and so forth. And so, the first thing that we started thinking about in our lab was, why would mother nature have stopped only RNAi and CRISPR? There probably are lots of other non-coding RNAs out there that might be able to be programmable and if they did exist, they probably also do more complicated and interesting things than just guide a molecular scissors. So that was sort of the first core kind of intuition that we had. The second intuition that we had on the technology side, I was just wearing my biology hat, I’ll put on my technology hat, is the thing that we call genome editing today hardly involves the genome.(16:50):It's really you're making a cut to change an individual base or an individual gene or locus. So really you're doing small scale single locus editing, so you might call it gene level or locus level cuts. And what you really want to be able to do is do things at the genome scale at 100 kb, a megabase at the chromosome scale. And I think that's where I think the field will inevitably go if you follow the technology curves of longer and longer range gene sequencing, longer and longer range gene synthesis, and then longer and longer range gene editing. And so, what would that look like? And we started thinking, could there be essentially recombination technologies that allow you to do cut and paste in a single step. Now, the reason for that is the way that we do gene editing today involves a cut and then a multi-step process of cellular DNA repair that resolves the cut to make the exertion or the error prone deletion or the modification that ends up happening.(17:59):And so, it's very complicated and whether that's nucleases or base or prime editing, you're all generally limited to the small-scale single locus changes. However, there are natural mechanisms that have solved this cut and paste problem, right? There are these viruses or bacterial versions of viruses known as phage that have generally been trying to exert their multi kilobase genomes into bacterial hosts and specialize throughout billions of years. So our core thought was, well, if there are these new non-coding RNAs, what kind of functions would we be excited about? Can we look in these mobile genetic elements, these so-called jumping genes for new mechanisms? They're incredibly widespread. Transposons are thought to be some of the most diverse enzyme mechanisms found in nature. And so, we started computationally by asking ourselves a very simple question. If a mobile element inserts itself into foreign DNA and it's able to somehow be programmable, presumably the inside or something encoded in the inside of the element is predictive of some sequence on the outside of the element.(19:15):And so, that was the core insight we took, and we thought let's look across the boundaries of many different mobile genetic elements and we zoomed in on a particular sub family of these MGE known as insertion sequence (IS) elements which are the most autonomous minimal transposons. Normally transposons have all kinds of genes that they use to hitchhike around the genomic galaxy and endow the bacterial host with some fitness advantage like some ability to metabolize some copper and some host or some metal. And these IS elements have only the enzymes that they need to jump around. And if you identify the boundaries of these using modern computational methods, this is actually a really non-trivial problem. But if you solve that problem to figure out with nucleotide resolution where the element boundaries end and then you look for the open reading frame of the transposases enzyme inside of this element, you'll find that it's not just that coding sequence.(20:19):There are also these non-coding flanks inside of the element boundaries. And when we looked across the non-coding, the entire IS family tree, there are hundreds of these different types of elements. We found that this particular family IS110, had the longest non-coding ends of all IS elements. And we started doing experiments in the lab to try to figure out how these work. And what we found was that these elements are cut and paste elements, so they excise themselves into a circular form and paste themselves back in into a target site linearly. But the circularization of this element brings together two distal ends together, which brings together a -35 and a -10 box that create and reconstitute a canonical bacterial transcriptional promoter. This essentially is like plugging a plug into an electrical socket in the wall and it jacks up transcription. Now you would think this transcription would turn on the transposase enzyme so it can jump around more but it transcribes a non-coding RNA out of this non-coding end.(21:30):We're like, holy crap, are these RNAs actually involved in regulating the transposon? Now the boring answer would be, oh, it regulates the expression. It's like an antisense regulate or something. The exciting answer would be, oh, it's a new type of guide RNA and you found an RNA guided integrase. So we started zooming in bound dramatically on this and we undertook a covariation analysis where we were able to show that this cryptic non-coding RNA has a totally novel guide RNA structure, totally distinct from RNAi or CRISPR guide RNAs. And it had a target site that covaried with the target site of the element. And so we're like, oh wow, this could be a programmable transposase. The second thing that we found was even more surprising, there was a second region of complementarity in that same RNA that recognized the donor sequence, which is the circularized element itself. And so, this was the first example of a bispecific guide RNA, and also the first example of RNA guided self-recognition by a mobile genetic element.Eric Topol (22:39):It's pretty extraordinary because basically you did a systematic assessment of jumping genes or transposons and you found that they contain things that previously were not at all recognized. And then you have a way to program these to edit, change the genome without having to do any cuts or nicks, right?Patrick Hsu (23:05):Yeah. So what we showed in a test tube is when we took this, so-called bridge RNA, which we named because it bridges the target and donor together along with the recombinase enzyme. So the two component system, those are the only two things that you need. They're able to cut and paste DNA and recombine them in a test tube without any DNA repair, meaning that it's independent of cellular DNA repair and it does strand nicking, exchange, junction resolution and religation all in a single mechanism. So that's when we got super excited about its potential applications as bioengineering tool.Eric Topol (23:46):Yeah, it's pretty extraordinary. And have you already gone into in vivo assessment?Patrick Hsu (23:54):Yes, in our initial set of papers, what we showed is that these are programmable and functional or recombinases in a test tube and in bacterial cells. And by reprogramming the target and donor the right way, you can use these enzymes not just for insertion, but also for flipping and cutting out DNA. And so, we actually have in a single mechanism the ability to do bridge editing, if you will, for universal DNA recombination, insertion, excision or inversion, similar to what folks have been doing for decades with Cre recombinase, but with fully programmable recognition sequences. The work that we're doing now in the lab as you can imagine is to adapt these into robust tools for mammalian genome editing, including of course, human genomes. We're excited about this, we're making good progress. The CRISPR has had thousands of labs over the last 10, 15 years working on it to make these therapeutic level potency and selectivity. We're going to work and follow that same blueprint for getting bridge systems to get to that level of performance, but we're on the path and we're very optimistic for the future.Exemplar of Digital BiologyEric Topol (25:13):Yeah, I think it's quite extraordinary and it's a whole different look to what we've been seeing in the CRISPR era for over the past decade and how that's been advancing and getting more specific and less need for repair and being able to be more versatile. But this takes it to yet another dimension. Now, this brings me to the field that when I think of this term digital biology, I think of you and now our mutual acquaintance, Jensen Huang, who everybody knows now. Back some months ago, he wrote and said at a conference, “Where do I think the next amazing revolution is going to come? And this is going to be flat out one of the biggest ones ever. There’s no question that digital biology is going to be it. For the first time in human history, biology has the opportunity to be engineering, not science.” So can you critique Jensen? Is he right? And tell us how you conceive the field of digital biology.Patrick Hsu (26:20):If you look at gene therapy today, the core concepts are actually remarkably simple. They're elegant. Of course, you're missing a broken gene, you need to put it back. And that can be curative. Very simple, powerful concept. However, for complex diseases where you don't have just a single gene that goes wrong, in many cases we actually have no idea what to do. And in fact, when you're trying to put in DNA, that's over more than a gene scale. We kind of very quickly run out of ideas. Is it a CAR and a cytokine, a CAR and a cytokine and another thing? And then we're kind of out of ideas. And so, we started thinking in the lab, how can we actually design genomes where it's not just let's reduce the genome into individual Lego blocks, iGem style with promoters and different genes that we just sort of shuffle the Lego blocks around, but actually use AI to design genome sequences.(27:29):So to do that, we thought we would have to first of all, train a model that can learn and decode the foreign language of biology and use that in order to design sequences. And so, we sort of have been training DNA foundation models and virtual cell models at Arc, sort of a major effort of ours where the first thing that we tried was to take a variance of transformer architecture that's used to train ChatGPT from OpenAI, but instead apply this to study the next DNA token, right? Now, the interesting thing about next token prediction in English is that you can actually learn a surprising amount of information by just predicting the next word. You can learn world knowledge is the capital of Azerbaijan, is it Baku or is it London, right? Or if you're walking around in the kitchen, then the next text is, I then left the kitchen or the bathroom, right?(28:33):Now you're learning about spatial reasoning, and so you can also learn translation obviously. And so similarly, I think predicting the next token or the next base and DNA can lead you to learn about molecular biochemistry, is the next amino acid residue, hydrophobic or hydrophilic. And it can teach you about the mechanics of some catalytic binding pocket or something. You can learn about a disease mutation. Is the next base, the sick linked base or the wild type base and so on and so forth. And what we found was that at massive scale, DNA foundation models learn about molecular function, not just at the DNA level, but also at the RNA and the protein. And indeed, we could use these to design molecular systems like CRISPR-Cas systems, where you have a protein and the guide RNA. It could also design new DNA transposons, and we could design sequences that look plausibly like real genomes, where we generate a megabase a million bases of continuous genome sequence. And it really looks and feels like it could be a blurry picture of something that you would actually sequence. This has been a wonderful collaboration with Brian Hie, a PI at Stanford and an Arc investigator, and we're really excited about what we've seen in this work because it promises the better performance with even more scale. And so, simply by scaling up these models, by adding in more compute, more training data or more powerful models, they're going to get sharper and sharper.New A.I. Models in Life ScienceEric Topol (30:25):Yeah. Well, this whole use of large language models for the language of life, whether it's the genome proteins and on and on, actually RNA and even cells has really taken root. And of course, this is really one of the foundations of that field of digital biology, which brings together generative AI, AI tools and trying to push forward our understanding in biology. And also, obviously what's been emphasized in drug discovery, perhaps it's been emphasized even too much because we still have a lot to learn about biology, but that gets me to these models. Like today, AlphaProteo was announced by DeepMind, as we all know, AlphaFold 1, 2, now 3. They were kind of precursors of being able to predict proteins from amino acid 3D structure. And that kind of took the field by a little bit like ChatGPT for life science, but now it's a new model all the time. So you've been working on various models and Arc Institute, how do you see this unfolding? Are we just going to have every aspect of the language of life being approached in all the different interactions? And this is going to help us get to a much more deep level of understanding.Patrick Hsu (31:56):I'll say two things. The first is a lot of models that you just described are what I would call task specific models. A model for de novo design of a binder, a model for protein structure prediction. And there are other models for protein fitness or for RNA structure prediction, et cetera, et cetera. And I think what we're going to move towards are more unifying models where there's different classes of models at different levels of scale. So we will have these atomic level models for looking at generative chemistry or ligand docking. We have models that can unify genomes and their molecules, and then we have models that can unify cells and tissues. And so, for example, if you took an H&E stain of some liver, there are folks building models where you can then predict what the single cell spatial transcriptome will look like of that model. And that's obviously operating at a very different level of abstraction than a de novo protein binder. But in the long run, all of these are going to get, I think unified. I think the reason why this is possible is that biology, unlike physics, actually has this unifying theory of evolution that runs across all of its length scales from atomic, molecular, cellular, organismal to entire ecosystem. And the promise of these models is no short then to make biology a predictive discipline.Patrick Hsu (33:37):In physics, the experimentalists win the big prizes for the theorists when they measure gravitational waves or whatever. But in biology, we're very practical people. You do something three times and do a T-test. And I think my prediction is we can actually gauge the success of these LLMs or whatever in biology by how much we respect theory in this field.The A.I. ScientistEric Topol (34:05):Yeah. Well, that's a really interesting perspective, an important perspective because the proliferation of models, which we're going to get into not just doing the things that you described, but also being able to be “pseudo” scientists, the so-called AI scientist. Maybe you could comment about that concept because that's been the idea that everything from the question that could be asked to the hypothesis and the experiment design and the analysis of data and then the feedback. So what is the role of the scientists, that seems to have been overplayed? And maybe you can put that in context.Patrick Hsu (34:48):So yeah, right now there's a lot of excitement that we can use AI agents not just to do software enterprise workflows, but to be a research assistant. And then over time, itself an autonomous research scientist that can read the literature, come up with an idea, maybe run a bunch of robots in the lab or do a bunch of computational analyses and then potentially even analyze data, conclude what is going on and actually write an entire paper. Now, I think the vision of this is compelling in the long term. I think the question is really about timescale. If you break down the scientific method into its constituent parts, like hypothesis generation, doing an experiment, analyzing experiment and iterating, we're clearly going to use AI of some kind at every single step of this cycle. I think different steps will require different levels of maturity. The way that I would liken this is just wet lab automation, folks have dreamed about having pipetting robots that just do their western blots and do their cell culture for them for generations.(36:01):But of course, today they don't actually really feel fundamentally different from the same ones that we had in the 90s, let's say. Right? And so, obviously they're getting better, but it seems to me one of the trends I'm very bullish about is the explosion of humanoid robots and robot foundation models that have a world model and a sense of physics and proportionate space loaded onto them. Within five years, we're going to have home robots that can fold your clothes, that can organize your kitchen and do all of this while you're sleeping, so you wake up to a clean home every day.Eric Topol (36:40):It’s not going to be just Roomba anymore. There's going to be a lot more, but it isn't just the hardware, it's also the agents playing in software, right?Patrick Hsu (36:50):It's the integrated loop of the hardware and the software where the ability to make the same machine generally intelligent will make it adaptable to a broad array of tasks. Now, what I'm excited about is those generally intelligent humanoid robots coming into the lab, where instead of creating a centrifuge or a new type of pipetter that's optimized for your Beckman or Hamilton device, instead you just have robot arms that you snap onto the edge of the bench and then they just work alongside you. And I do think that's coming, although it'll take a lot of hardware and software and computer vision engineering to make that possible.A Sense of HumorEric Topol (37:32):Yeah, and I think also going back to originating the question, there still is quite a debate about the creativity and the lack of any simulation of AGI, whatever that means anymore. And so, the human in the loop part of this is obviously I think it's still of critical nature. Now, the other thing I learned about you is you have a great sense of humor, which is really important by the way. And recently, which is great that you're active on X or Twitter because that's one way we get to see what you're thinking on a day-to-day basis. But I think you put out a poll which was really quite provocative , and it was about, here's what it said, “do more people in the world *truly* understand transformers or health insurance?” And interestingly, you got 49% for transformers at 51% for health insurance. Can you tell us what you're thinking when you put that poll together? Because obviously a lot of people don't understand either of these.Patrick Hsu (38:44):I think the core question is, there are different ways of looking at the world, some of which are very bottom up and some of which are very top down. And one of the very surprising things about transformers is they're taking something that is in principle, an incredibly simple task, which is if you have a string of text, what is the next letter? And somehow at massive, massive scale, you can unlock something that looks an awful lot like reasoning, and you've got these emergent behaviors. Now the bottoms up theory of just the linear algebra that's going on in these models couldn't possibly really help us predict that we have these emerging capabilities. And I think similarly in healthcare, there's a literal set of parts that are operating in some complex way that at massive scale becomes this incredibly confusing and dynamic system for how we can actually incentivize how we make medicines, how we actually take care of people, and how we actually pay for any of this from an economic point of view. And so, I think it was, in some sense if transformers can actually be an explainable by just linear algebra equations, maybe there will be a way to decompose the seemingly incredibly confusing world of healthcare in order to actually build a better way forward.Computing Power and the GPU Arms RaceEric Topol (40:12):Yeah. Well that's great. Now the other thing I wanted to ask you about, we open source and the arms race of GPUs and this whole kind of idea is you touched on the need for coalescing a lot of these tools to exploit the synergy. But we have an issue because many academic labs like here at Scripps Research and so many others, including as I learned even at Stanford, have limited access to GPUs. So computing power of large language models is a problem. And then the models that exist today that can be adopted like Llama or others, and they're somewhat limited. And then we also have a movement towards trying to make things more open source, like for example, recently OpenCRISPR with Profluent Bio that is basically trying to use AI for CRISPR guides. And so, how do you deal with this arms race, computing power, open source, proprietary models that are not easily accessible without a lot of resources?Patrick Hsu (41:30):So the first thing I would say is, we are in the academic science sphere really unprepared for the level of resources that are required for doing this type of cutting edge computational work. There are top Stanford computer science professors or computational researchers who have a single GPU in their office, and that's actually what their whole lab runs off of.(41:58):The UC Berkeley campus, the grid runs on something like 12 megawatts of power and how are they going to build an on-premises GPU clusters, like a central question that can scale across the entire needs? And these are two of the top computer science universities in the world. And so, I think one of our kind of core beliefs at Arc is, as science both experimentally and computationally has gotten incredibly complex, not just in terms of conceptually, but also just the actual infrastructure and machines and know-how that you need to do things. We actually need to essentially support this. So we have a private GPU cloud that we use to train our models, and we have access to significantly large clusters for large burst kind of train outs as necessary. And I think infrastructurally for running genomics experiments or doing scalable brain organoid screens, right, we're also building out the infrastructure to support that experimentally.Eric Topol (43:01):Yeah, no, I think this is one of the advantages of the new model like the Arc Institute because not many centers have that type of plasticity with access to computing power when needed. So that's where a brilliant mind you and the Arc Institute together makes for a formidable recipe for future advances and of course building on the ones you've already accomplished.The Primacy of Human TalentPatrick Hsu (43:35):I would just say, my main skill, if I have one, is to recruit really, really smart people. And so, everything that you're seeing and hearing about is the work of unbelievable colleagues who are curious, passionate, and incredible scientists.Eric Topol (43:53):But it also takes the person who can judge those who are in that category set as a role model. And you're certainly doing that. I guess just in closing, I mean, it's just such a delight to get to meet you here and kind of get your thoughts on what is the hottest thing in life science without question, which brings together the fields of AI and what's going on, not just obviously in genome editing, but this digital biology era that we're still in the early phases of, I mean, I think you could say that it's just going to continue to accelerate the exponential curve. We're still kind of on the bottom of that, I would imagine where we're headed. Any other things that you want to bring up that I haven't touched on that will round out this conversation?Patrick Hsu (44:50):I mean, I think it's very early days here at Arc.Patrick Hsu (44:53):When we founded Arc, we asked ourselves, how do we measure success? We don't have customers or revenue in the way that a typical startup does. And we felt sort of three things. The first was research institutes live and die by their talent. Can we actually hire incredible people when we make offers to people we want to come, do they come? The second was, when those folks do come to Arc, do they feel like they're able to work on important research programs that they couldn't do sort of at their prior university or company? And then longer term, the third thing was, and there's just no shortcut around this, you need to do important work. And I think we've been really excited that there are early signs that we're able to do all three of these things, and we're still, again, just following the same scaling laws that we're seeing in natural language and vision, but for the domain of biology. And so, we're excited about what's ahead and think if there are folks who are interested in learning more about Arc, just shoot me an email or DM.Eric Topol (46:07):Yeah, well I would just say, congratulations on what you've already achieved. I know you're going to keep rocking it because you already have in a short time. And for anybody who doesn't know about Arc Institute and your work and your team, I hope this is going to be putting them on notice actually what can be accomplished outside of the usual NIH funded model, which is kind of a risk-free zone where you basically have to have your results nailed down before you send in your proposal frequently, and it doesn't do great things for young people. Really, I think you actually qualify in that demographic where it's hard for them to break in for getting NIH grants and also for this type of work that you're doing. So we'll look for the next bridge beyond bridge RNAs of your just fantastic efforts. So Patrick, thanks so much for joining us today, and we'll be checking back with you and following all the great work that you'll be doing in the times ahead.Patrick Hsu (47:14):Thanks so much, Eric. It was such a pleasure to be here today. Appreciate the opportunity.*******************Thanks for listening, reading or watching!The Ground Truths newsletters and podcasts are all free, open-access, without ads.Please share this post/podcast with your friends and network if you found it informative!Voluntary paid subscriptions all go to support Scripps Research. Many thanks for that—they greatly help fund our summer internship programs.Thanks to my producer Jessica Nguyen and Sinjun Balabanoff for audio and video support at Scripps Research.Note: you can select preferences to receive emails about newsletters, podcasts, or all I don’t want to bother you with an email for content that you’re not interested in. Get full access to Ground Truths at erictopol.substack.com/subscribe

Arvind Narayanan and Sayash Kapoor are well regarded computer scientists at Princeton University and have just published a book with a provocative title, AI Snake Oil. Here I’ve interviewed Sayash and challenged him on this dismal title, for which he provides solid examples of predictive AI’s failures. Then we get into the promise of generative AI.Full videos of all Ground Truths podcasts can be seen on YouTube here. The audios are also available on Apple and Spotify.Transcript with links to audio and external links to key publications Eric Topol (00:06):Hello, it's Eric Topol with Ground Truths, and I'm delighted to welcome the co-author of a new book AI SNAKE OIL and it's Sayash Kapoor who has written this book with Arvind Narayanan of Princeton. And so welcome, Sayash. It's wonderful to have you on Ground Truths.Sayash Kapoor (00:28):Thank you so much. It's a pleasure to be here.Eric Topol (00:31):Well, congratulations on this book. What's interesting is how much you've achieved at such a young age. Here you are named in TIME100 AI’s inaugural edition as one of those eminent contributors to the field. And you're currently a PhD candidate at Princeton, is that right?Sayash Kapoor (00:54):That's correct, yes. I work at the Center for Information Technology Policy, which is a joint program between the computer science department and the school of public and international affairs.Eric Topol (01:05):So before you started working on your PhD in computer science, you already were doing this stuff, I guess, right?Sayash Kapoor (01:14):That's right. So before I started my PhD, I used to work at Facebook as a machine learning engineer.Eric Topol (01:20):Yeah, well you're taking it to a more formal level here. Before I get into the book itself, what was the background? I mean you did describe it in the book why you decided to write a book, especially one that was entitled AI Snake Oil: What Artificial Intelligence Can Do, What It Can't, and How to Tell the Difference.Background to Writing the BookSayash Kapoor (01:44):Yeah, absolutely. So I think for the longest time both Arvind and I had been sort of looking at how AI works and how it doesn't work, what are cases where people are somewhat fooled by the potential for this technology and fail to apply it in meaningful ways in their life. As an engineer at Facebook, I had seen how easy it is to slip up or make mistakes when deploying machine learning and AI tools in the real world. And had also seen that, especially when it comes to research, it's really easy to make mistakes even unknowingly that inflate the accuracy of a machine learning model. So as an example, one of the first research projects I did when I started my PhD was to look at the field of political science in the subfield of civil war prediction. This is a field which tries to predict where the next civil war will happen and in order to better be prepared for civil conflict.(02:39):And what we found was that there were a number of papers that claimed almost perfect accuracy at predicting when a civil war will take place. At first this seemed sort of astounding. If AI can really help us predict when a civil war will start like years in advance sometimes, it could be game changing, but when we dug in, it turned out that every single one of these claims where people claim that AI was better than two decades old logistic regression models, every single one of these claims was not reproducible. And so, that sort of set the alarm bells ringing for the both of us and we sort of dug in a little bit deeper and we found that this is pervasive. So this was a pervasive issue across fields that were quickly adopting AI and machine learning. We found, I think over 300 papers and the last time I compiled this list, I think it was over 600 papers that suffer from data leakage. That is when you can sort of train on the sets that you're evaluating your models on. It's sort of like teaching to the test. And so, machine learning model seems like it does much better when you evaluate it on your data compared to how it would really work out in the real world.Eric Topol (03:48):Right. You say in the book, “the goal of this book is to identify AI snake oil - and to distinguish it from AI that can work well if used in the right ways.” Now I have to tell you, it's kind of a downer book if you're an AI enthusiast because there's not a whole lot of positive here. We'll get to that in a minute. But you break down the types of AI, which I'm going to challenge a bit into three discrete areas, the predictive AI, which you take a really harsh stance on, say it will never work. Then there's generative AI, obviously the large language models that took the world by storm, although they were incubating for several years when ChatGPT came along and then content moderation AI. So maybe you could tell us about your breakdown to these three different domains of AI.Three Types of AI: Predictive, Generative, Content ModerationSayash Kapoor (04:49):Absolutely. I think one of our main messages across the book is that when we are talking about AI, often what we are really interested in are deeper questions about society. And so, our breakdown of predictive, generative, and content moderation AI sort of reflects how these tools are being used in the real world today. So for predictive AI, one of the motivations for including this in the book as a separate category was that we found that it often has nothing to do with modern machine learning methods. In some cases it can be as simple as decades old linear regression tools or logistic regression tools. And yet these tools are sold under the package of AI. Advances that are being made in generative AI are sold as if they apply to predictive AI as well. Perhaps as a result, what we are seeing is across dozens of different domains, including insurance, healthcare, education, criminal justice, you name it, companies have been selling predictive AI with the promise that we can use it to replace human decision making.(05:51):And I think that last part is where a lot of our issues really come down to because these tools are being sold as far more than they're actually capable of. These tools are being sold as if they can enable better decision making for criminal justice. And at the same time, when people have tried to interrogate these tools, what we found is these tools essentially often work no better than random, especially when it comes to some consequential decisions such as job automation. So basically deciding who gets to be called on the next level of like a job interview or who is rejected, right as soon as they submit the CV. And so, these are very, very consequential decisions and we felt like there is a lot of snake oil in part because people don't distinguish between applications that have worked really well or where we have seen tremendous advances such as generative AI and applications where essentially we've stalled for a number of decades and these tools don't really work as claimed by the developers.Eric Topol (06:55):I mean the way you partition that, the snake oil, which is a tough metaphor, and you even show the ad from 1905 of snake oil in the book. You're really getting at predictive AI and how it is using old tools and selling itself as some kind of breakthrough. Before I challenge that, are we going to be able to predict things? By the way, using generative AI, not as you described, but I would like to go through a few examples of how bad this has been and since a lot of our listeners and readers are in the medical world or biomedical world, I'll try to get to those. So one of the first ones you mentioned, which I completely agree, is how prediction of Covid from the chest x-ray and there were thousands of these studies that came throughout the pandemic. Maybe you could comment about that one.Some Flagrant ExamplesSayash Kapoor (08:04):Absolutely. Yeah, so this is one of my favorite examples as well. So essentially Michael Roberts and his team at the University of Cambridge a year or so after the pandemic looked back at what had happened. I think at the time there were around 500 studies that they included in the sample. And they looked back to see how many of these would be useful in a clinical setting beyond just the scope of writing a research paper. And they started out by using a simple checklist to see, okay, are these tools well validated? Does the training and the testing data, is it separate? And so on. So they ran through the simple checklist and that excluded all but 60 of these studies from consideration. So apart from 60 studies, none of these other studies even passed a very, very basic criteria for being included in the analysis. Now for these 60, it turns out that if you take a guess about how many were useful, I'm pretty confident most cases would be wrong.(09:03):There were exactly zero studies that were useful in a clinically relevant setting. And the reasons for this, I mean in some cases the reasons were as bizarre as training a machine learning model to predict Covid where all of the positive samples of people who had Covid were from adults. But all of the negative samples of people who didn't have Covid were from children. And so, essentially claiming that the resulting classifier can predict who has Covid is bizarre because all the classifier is doing is looking at the checks history and basically predicting which x-ray belongs to a child versus an adult. And so, this is the sort of error in some cases we saw duplicates in the training and test set. So you have the same person that is being used for training the model and that it is also used for evaluating the model. So simply memorizing a given sample of x-rays would be enough to achieve a very high performance. And so, for issues like these, I think all 60 of these studies prove to be not useful in a clinically relevant setting. And I think this is sort of the type of pattern that we've seen over and over again.Eric Topol (10:14):Yeah, and I agree with you on that point. I mean that was really a flagrant example and that would fulfill your title of your book, which as I said is a very tough title. But on page 29, and we'll have this in the post. You have a figure, the landscape of AI snake oil, hype, and harm. And the problem is there is nothing good in this landscape. So on the y-axis you have works, hype, snake oil going up on the y-axis. And on the x-axis, you have benign and harmful. So the only thing you have that works and that's benign is autocomplete. I wouldn't say that works. And then you have works facial recognition for surveillance is harmful. This is a pretty sobering view of AI. Obviously, there's many things that are working that aren't on this landscape. So I just would like to challenge, are you a bit skewed here and only fixating on bad things? Because this diagram is really rough. I mean, there's so much progress in AI and you have in here you mentioned the predicting civil wars, and obviously we have these cheating detection, criminal risk prediction. I mean a lot of problems, video interviews that are deep fakes, but you don't present any good things.Optimism on Generative AISayash Kapoor (11:51):So to be clear, I think both Arvind and are somewhat paradoxically optimistic about the future of generative AI. And so, the decision to focus on snake oil was a very intentional one from our end. So in particular, I think at various places in the book we outline why we're optimistic, what types of applications we think we're optimistic about as well. And the reason we don't focus on them is that it basically comes down to the fact that no one wants to read a book that has 300 pages about the virtues of spellcheck or AI for code generation or something like that. But I think I completely agree and acknowledge that there are lots of positive applications that didn't make the cut for the book as well. That was because we wanted people to come to this from a place of skepticism so that they're not fooled by the hype.(12:43):Because essentially we see even these positive uses of AI being lost out if people have unrealistic expectations from what an AI tool should do. And so, pointing out snake oil is almost a prerequisite for being able to use AI productively in your work environment. I can give a couple of examples of where or how we've sort of manifested this optimism. One is AI for coding. I think writing code is an application that I do, at least I use AI a lot. I think almost half of the code I write these days is generated, at least the first draft is generated using AI. And yet if I did not know how to program, it would be a completely different question, right? Because for me pointing out that, oh, this syntax looks incorrect or this is not handling the data in the correct way is as simple as looking at a piece of code because I've done this a few times. But if I weren't an expert on programming, it would be completely disastrous because even if the error rate is like 5%, I would have dozens of errors in my code if I'm using AI to generate it.(13:51):Another example of how we've been using it in our daily lives is Arvind has two little kids and he's built a number of applications for his kids using AI. So I think he's a big proponent of incorporating AI into children's lives as a force for good rather than having a completely hands-off approach. And I think both of these are just two examples, but I would say a large amount of our work these days occurs with the assistance of AI. So we are very much optimistic. And at the same time, I think one of the biggest hindrances to actually adopting AI in the real world is not understanding its limitations.Eric Topol (14:31):Right. Yeah, you say in the book quote, “the two of us are enthusiastic users of generative AI, both in our work and our personal lives.” It just doesn't come through as far as the examples. But before I leave the troubles of predictive AI, I liked to get into a few more examples because that's where your book shines in convincing that we got some trouble here and we need to be completely aware. So one of the most famous, well, there's a couple we're going to get into, but one I'd like to review with you, it's in the book, is the prediction of sepsis in the Epic model. So as you know very well, Epic is the most used IT and health systems electronic health records, and they launched never having published an algorithm that would tell when the patient was hospitalized if they actually had sepsis or risk of sepsis. Maybe you could take us through that, what you do in the book, and it truly was a fiasco.The Sepsis DebacleSayash Kapoor (15:43):Absolutely. So I think back in 2016/2017, Epic came up with a system that would help healthcare providers predict which patients are most at risk of sepsis. And I think, again, this is a very important problem. I think sepsis is one of the leading causes of death worldwide and even in the US. And so, if we could fix that, I think it would be a game changer. The problem was that there were no external validations of this algorithm for the next four years. So for four years, between 2017 to 2021, the algorithm wasn't used by hundreds of hospitals in the US. And in 2021, a team from University of Michigan did this study in their own hospital to see what the efficacy of the sepsis prediction model is. They found out that Epic had claimed an AUC of between 0.76 and 0.83, and the actual AUC was closer to 0.6, and AUC of 0.5 is making guesses at random.(16:42):So this was much, much worse than the company's claims. And I think even after that, it still took a year for sepsis to roll back this algorithm. So at first, Epic's claims were that this model works well and that's why hospitals are adopting it. But then it turned out that Epic was actually incentivizing hospitals to adopt sepsis prediction models. I think they were giving credits of hundreds of thousands of dollars in some cases. If a hospital satisfied a certain set of conditions, one of these conditions was using a sepsis prediction model. And so, we couldn't really take their claims at face value. And finally in October 2022, Epic essentially rolled back this algorithm. So they went from this one size fits all sepsis prediction model to a model that each hospital has to train on its own data, an approach which I think is more likely to work because each hospital's data is different. But it's also more time consuming and expensive for the hospitals because all of a sudden you now need your own data analysts to be able to roll out this model to be able to monitor it.(17:47):I think this study also highlights many of the more general issues with predictive AI. These tools are often sold as if they're replacements for an existing system, but then when things go bad, essentially they're replaced with tools that do far less. And companies often go back to the fine print saying that, oh, we should always deploy it with the human in the loop, or oh, it needs to have these extra protections that are not our responsibility, by the way. And I think that gap between what developers claim and how the tool actually works is what is most problematic.Eric Topol (18:21):Yeah, no, I mean it's an egregious example, and again, it fulfills like what we discussed with statistics, but even worse because it was marketed and it was incentivized financially and there's no doubt that some patients were completely miscategorized and potentially hurt. The other one, that's a classic example that went south is the Optum UnitedHealth algorithm. Maybe you could take us through that one as well, because that is yet another just horrible case of how people were discriminated against.The Infamous Optum AlgorithmSayash Kapoor (18:59):Absolutely. So Optum, another health tech company created an algorithm to prioritize high risk patients for preemptive care. So I think it was around when Obamacare was being introduced that insurance networks started looking into how they could reduce costs. And one of the main ways they identified to reduce costs is basically preemptively caring for patients who are extremely high risk. So in this case, they decided to keep 3% of the patients in the high risk category and they built a classifier to decide who's the highest risk, because potentially once you have these patients, you can proactively treat them. There might be fewer emergency room visits, there might be fewer hospitalizations and so on. So that's all fine and good. But what happened when they implemented the algorithm was that every machine learning model needs like the target variable, what is being predicted at the end of the day. What they decided to predict was how much patient would pay, how much would they charge, what cost the hospital would incur if they admitted this patient.(20:07):And they essentially use that to predict who should be prioritized for healthcare. Now unsurprisingly, it turned out that white patients often pay a lot more or are able to pay a lot more when it comes to hospital visits. Maybe it's because of better insurance or better conditions at work that allow them to take leave and so on. But whatever the mechanism is, what ended up happening with this algorithm was I think black patients with the same level of healthcare prognosis were half as likely or about much less likely compared to white ones of getting enrolled in this high risk program. So they were much less likely to get this proactive care. And this was a fantastic study by Obermeyer, et al. It was published in Science in 2019. Now, what I think is the most disappointing part of this is that Optum did not stop using this algorithm after this study was released. And that was because in some sense the algorithm was working precisely as expected. It was an algorithm that was meant to lower healthcare costs. It wasn't an algorithm that was meant to provide better care for patients who need it most. And so, even after this study was rolled out, I think Optum continued using this algorithm as is. And I think as far as I know, even today this is or some version of this algorithm is still in use across the network of hospitals that Optum serves.Eric Topol (21:31):No, it's horrible the fact that it was exposed by Ziad Obermeyer’s paper in Science and that nothing has been done to change it, it's extraordinary. I mean, it's just hard to imagine. Now you do summarize the five reasons predictive AI fails in a nice table, we'll put that up on the post as well. And I think you've kind of reviewed that as these case examples. So now I get to challenge you about predictive AI because I don't know that such a fine line between that and generative AI are large language models. So as you know, the group at DeepMind and now others have done weather forecasting with multimodal large language models and have come up with some of the most accurate weather forecasting we've ever seen. And I've written a piece in Science about medical forecasting. Again, taking all the layers of a person's data and trying to predict if they're high risk for a particular condition, including not just their electronic record, but their genomics, proteomics, their scans and labs and on and on and on exposures, environmental.Multimodal A.I. in Medicine(22:44):So I want to get your sense about that because this is now a coalescence of where you took down predictive AI for good reasons, and then now these much more sophisticated models that are integrating not just large data sets, but truly multimodal. Now, some people think multimodal means only text, audio, speech and video images, but here we're talking about multimodal layers of data as for the weather forecasting model or earthquake prediction or other things. So let's get your views on that because they weren't really presented in the book. I think they're a positive step, but I want to see what you think.Sayash Kapoor (23:37):No, absolutely. I think maybe the two questions are sort of slightly separate in my view. So for things like weather forecasting, I think weather forecasting is a problem that's extremely tenable for generative AI or for making predictions about the future. And I think one of the key differences there is that we don't have the problem of feedback loops with humans. We are not making predictions about individual human beings. We are rather making predictions about what happens with geological outcomes. We have good differential equations that we've used to predict them in the past, and those are already pretty good. But I do think deep learning has taken us one step further. So in that sense, I think that's an extremely good example of what doesn't really fit within the context of the chapter because we are thinking about decisions thinking about individual human beings. And you rightly point out that that's not really covered within the chapter.(24:36):For the second part about incorporating multimodal data, genomics data, everything about an individual, I think that approach is promising. What I will say though is that so far we haven't seen it used for making individual decisions and especially consequential decisions about human beings because oftentimes what ends up happening is we can make very good predictions. That's not in question at all. But even with these good predictions about what will happen to a person, sometimes intervening on the decision is hard because oftentimes we treat prediction as a problem of correlations, but making decisions is a problem of causal estimation. And that's where those two sort of approaches disentangle a little bit. So one of my examples, favorite examples of this is this model that was used to predict who should be released before screening when someone comes in with symptoms of pneumonia. So let's say a patient comes in with symptoms of pneumonia, should you release them on the day of?(25:39):Should you keep them in the hospital or should you transfer them to the ICU? And these ML researchers were basically trying to solve this problem. They found out that the neural network model they developed, this was two decades ago, by the way. The neural network model they developed was extremely accurate at predicting who would basically have a high risk of having complications once they get pneumonia. But it turned out that the model was saying essentially that anyone who comes in who has asthma and who comes in with symptoms of pneumonia is the lowest risk patient. Now, why was this? This was because when in the past training data, when some such patients would come into the hospital, these patients would be transferred directly to the ICU because the healthcare professionals realized that could be a serious condition. And so, it turned out that actually patients who had asthma who came in with symptoms of pneumonia were actually the lowest risk amongst the population because they were taken such good care of.(26:38):But now if you use this prediction that a patient comes in with symptoms of pneumonia and they have asthma, and so they're low risk, if you use this to make a decision to send them back home, that could be catastrophic. And I think that's the danger with using predictive models to make decisions about people. Now, again, I think the scope and consequences of decisions also vary. So you could think of using this to surface interesting patterns in the data, especially at a slightly larger statistical level to see how certain subpopulations behave or how certain groups of people are likely to develop symptoms or whatever. But I think when as soon as it comes to making decisions about people, the paradigm of problem solving changes because as long as we are using correlational models, I think it's very hard to say what will happen if we change the conditions, what will happen if the decision making mechanism is very different from one where the data was collected.Eric Topol (27:37):Right. No, I mean where we agree on this is that at the individual level, using multimodal AI with all these layers of data that have now recently become available or should be available, that has to be compared ideally in a randomized trial with standard of care today, which doesn't use any of that. And to see whether or not that decision's made, does it change the natural history and is it an advantage, that's yet to be done. And I agree, it's a very promising pathway for the future. Now, I think you have done what is a very comprehensive sweep on the predictive AI failures. You've mentioned here in our discussion, your enthusiasm and in the book about generative AI positive features and hope and excitement perhaps even. You didn't really yet, we haven't discussed much on the content moderation AI that you have discreetly categorized. Maybe you could just give us the skinny on your sense of that.Content Moderation AISayash Kapoor (28:46):Absolutely. So content moderation AI is AI that's used to sort of clean up social media feeds. Social media platforms have a number of policies about what's allowed and not allowed on the platforms. Simple things such as spam are obviously not allowed because let's say people start spamming the platform, it becomes useless for everyone. But then there are other things like hate speech or nudity or pornography and things like that, which are also disallowed on most if not all social media platforms today. And I think a lot of the ways in which these policies are enforced today is using AI. So you might have an AI model that runs every single time you upload a photo to Facebook, for instance. And not just one perhaps hundreds of such models to detect if it has nudity or hate speech or any of these other things that might violate the platform's terms of service.(29:40):So content moderation AI is AI that's used to make these decisions. And very often in the last few years we've seen that when something gets taken down, for instance, Facebook deletes a post, people often blame the AI for having a poor understanding. Let's say of satire or not understanding what's in the image to basically say that their post was taken down because of bad AI. Now, there have been many claims that content moderation AI will solve social media's problems. In particular, we've heard claims from Mark Zuckerberg who in a senate testimony I think back in 2018, said that AI is going to solve most if not all of their content moderation problems. So our take on content moderation AI is basically this. AI is very, very useful for solving the simple parts of content moderation. What is a simple part? So basically the simple parts of content moderation are, let's say you have a large training data of the same type of policy violation on a platform like Facebook.(30:44):If you have large data sets, and if these data sets have a clear line in the sand, for instance, with nudity or pornography, it's very easy to create classifiers that will automate this. On the other hand, the hard part of content moderation is not actually just creating these AI models. The hard part is drawing the line. So when it comes to what is allowed and not allowed on platforms, these platforms are essentially making decisions about speech. And that is a topic that's extremely fraught. It's fraught in the US, it's also fraught globally. And essentially these platforms are trying to solve this really hard problem at scale. So they're trying to come up with rules that apply to every single user of the platform, like over 3 billion users in the case of Facebook. And this inevitably has these trade-offs about what speech is allowed versus disallowed that are hard to say one way or the other.(31:42):They're not black and white. And what we think is that AI has no place in this hard part of content moderation, which is essentially human. It's essentially about adjudicating between competing interests. And so, when people claim that AI will solve these many problems of content moderation, I think what they're often missing is that there's this extremely large number of things you need to do to get content moderation right. AI solves one of these dozen or so things, which is detecting and taking down content automatically, but all of the rest of it involves essentially human decisions. And so, this is sort of the brief gist of it. There are also other problems. For example, AI doesn't really work so well for low resource languages. It doesn't really work so well when it comes to nuances and so on that we discussed in the book. But we think some of these challenges are solvable in the medium to long term. But these questions around competing interests of power, I think are beyond the domain of AI even in the medium to long term.Age 28! and Career AdviceEric Topol (32:50):No, I think you nailed that. I think this is an area that you've really aptly characterized and shown the shortcomings of AI and how the human factor is so critically important. So what's extraordinary here is you're just 28 and you are rocking it here with publications all over the place on reproducibility, transparency, evaluating generative AI, AI safety. You have a website on AI snake oil that you're collecting more things, writing more things, and of course you have the experience of having worked in the IT world with Facebook and also I guess also Columbia. So you're kind of off to the races here as one of the really young leaders in the field. And I am struck by that, and maybe you could comment about the inspiration you might provide to other young people. You're the youngest person I've interviewed for Ground Truths, by the way, by a pretty substantial margin, I would say. And this is a field where it attracts so many young people. So maybe you could just talk a bit about your career path and your advice for people. They may be the kids of some of our listeners, but they also may be some of the people listening as well.Sayash Kapoor (34:16):Absolutely. First, thank you so much for the kind words. I think a lot of this work is with collaborators without whom of course, I would never be able to do this. I think Arvind is a great co-author and supporter. I think in terms of my career parts, it was sort of like a zigzag, I would say. It wasn't clear to me when I was an undergrad if I wanted to do grad school or go into the industry, and I sort of on a whim went to work at Facebook, and it was because I'd been working on machine learning for a little bit of time, and I just thought, it's worth seeing what the other side has to offer beyond academia. And I think that experience was very, very helpful. One of the things, I talked to a lot of undergrads here at Princeton, and one of the things I've seen people be very concerned about is, what is the grad school they're going to get into right after undergrad?(35:04):And I think it's not really a question you need to answer now. I mean, in some cases I would say it's even very helpful to have a few years of industry experience before getting into grad school. That has definitely, at least that has been my experience. Beyond that, I think working in a field like AI, I think it's very easy to be caught up with all of the new things that are happening each day. So I'm not sure if you know, but AI has I think over 500-1,000 new archive papers every single day. And with this rush, I think there's this expectation that you might put on yourself on being successful requires a certain number of publications or a certain threshold of things. And I think more often than not, that is counterproductive. So it has been very helpful for me, for example, to have collaborators who are thinking long term, so this book, for instance, is not something that would be very valued within the CS community, I would say. I think the CS community values peer-reviewed papers a lot more than they do books, and yet we chose to write it because I think the staying power of a book or the longevity of a book is much more than any single paper could do. So the other concrete thing I found very helpful is optimizing for a different metric compared to what the rest of the community seems to be doing, especially when it comes to fast moving fields like AI.Eric Topol (36:29):Well, that last piece of advice is important because I think too often people, whether it's computer scientists, life scientists, whoever, they don't realize that their audience is much broader. And that reaching the public with things like a book or op-eds or essays, varied ways that are intended for public consumption, not for, in this case, computer scientists. So that's why I think the book is a nice contribution. I don't like the title because it's so skewed. And also the content is really trying to hammer it at home. I hope you write a sequel book on the positive sides of AI. I did want to ask you, when I read the book, I thought I heard your voice. I thought you had written the book, and Arvind maybe did some editing. You wrote about Arvind this and Arvind that. Did you write the first draft of the book and then he kind of came along?Sayash Kapoor (37:28):No, absolutely not. So the way we wrote the book was we basically started writing it in parallel, and I wrote the first draft of half the chapters and he wrote the first draft of the other half, and that was essentially all the way through. So we would sort of write a draft, pass it to the other person, and then keep doing this until we sent it to our publishers.Eric Topol (37:51):Okay. So I guess I was thinking of the chapters you wrote where it came through. I'm glad that it was a shared piece of work because that's good, because that’s what co-authoring is all about, right? Well, Sayash, it's really been a joy to meet you and congratulations on this book. I obviously have expressed my objections and my disagreements, but that's okay because this book will feed the skeptics of AI. They'll love this. And I hope that the positive side, which I think is under expressed, will not be lost and that you'll continue to work on this and be a conscience. You may know I've interviewed a few other people in the AI space that are similarly like you, trying to assure its safety, its transparency, the ethical issues. And I think we need folks like you. I mean, this is what helps get it on track, keeping it from getting off the rails or what it shouldn't be doing. So keep up the great work and thanks so much for joining.Sayash Kapoor (39:09):Thank you so much. It was a real pleasure.************************************************Thanks for listening, reading or watching!The Ground Truths newsletters and podcasts are all free, open-access, without ads.Please share this post/podcast with your friends and network if you found it informative!Voluntary paid subscriptions all go to support Scripps Research. Many thanks for that—they greatly helped fund our summer internship programs for 2023 and 2024.Thanks to my producer Jessica Nguyen and Sinjun Balabanoff for audio and video support at Scripps Research.Note: you can select preferences to receive emails about newsletters, podcasts, or all I don’t want to bother you with an email for content that you’re not interested in. Get full access to Ground Truths at erictopol.substack.com/subscribe