Transcription
Here to make the case for AI, please welcome Reed Hoffman, co-founder of LinkedIn and Manis AI with Atlantic staff writer Josh Tango.
[Music] [Applause]
Yes, I apologize. I'm back. Reed, it's great to have you here.
It's great to be here. Um, so there are many interesting things about Reed Hoffman, but one that I envy is that you're in the sort of position to divide your time and attention into a variety of pursuits. And you've really focused on AI uh in a general sense for the last, I don't know, eight years, maybe even 10.
Okay. Um, and my undergraduate major was an AI too, but you know.
Yeah, mine too. Mine too. Um, but is it fair to say uh that you've spent a lot of 2025 focusing in a way that you haven't before?
Yes. And where have you focused?
So, um, one of the things as you know, um, I, uh, with actually a local New Yorker, Sedarth Muker GI, um, you know, the awesome author, uh, celebrated oncology researcher, co-founded a company called Manis AI, which is based here in New York. Um, it's actually one of the things that, um, you know, I will spend a couple of days working at because I'm a co-founder of this. And the basic idea is to create a to use AI to accelerate drug discovery with a particular focus on cancer. But as you're creating a drug discovery factory, it also allows you to get um diseases that are, you know, significant but not as you know kind of the economic model doesn't work for them because they're fewer, they're rare, orphaned, etc. And the the idea that we you know kind of uh came uh brought brought about for this was that neither the purely uh typical biological pharma approach will work with a little sprinkle of AI nor will the only AI work that you have to reinvent it kind of fusing the two together because um there's some places where AI is a huge accelerant and helps you create that factory of understanding targets possible um, you know bindings for those targets uh ways of understanding how to you know test them before you start getting into clinical testing as a way uh to make that work and uh the AI has to be kind of in that process but also it isn't like just purely simulated and so um it was really funny because the way the company came about is um I had this idea about what are the kinds of things that AI could accelerate that most people uh in Silicon Valley aren't paying attention to and uh drug discover was one of them. And so I called Sid because we've known each other for a number of time and I said, "Hey, coming to New York. I'd like to have dinner with you." Um he said, "Great." And we had dinner and I so I was I was running through all this stuff and he said, "Well, you know I've brought cancer drugs to market, right?" I'm like, "No, I didn't know that." Right? So So then that's that's and then it becomes the company.
So um one thing that I think we were just talking about quantum and uh there's a high degree of pretending when it comes to quantum. Um, I do think there's also a high degree of pretending.
There's a super position of pretending.
Yes. Yes. Um, yes. There's a degree to which we've all accepted that AI and healthcare are sort of like the Reese's peanut butter cup of hard problems and advanced technology. Can you actually just take a minute and explain why they pair pretty well?
Yeah. Yeah. Yeah. So, by the way, there's uh medical as a whole range, and we can go to any of these things, but there's everything from one of the things I think we want as a human, as a human race, as a society, as urgently as possible as a 24/7 medical assistant running on every smartphone. That is already, you know, like like it's it's it's essentially doable. Like, if for example, if you get a serious diagnosis and you don't consult Chat GBT or Frontier Model as a second opinion, you're making a mistake. I actually know people's lives who've been saved by that consultation. Not first opinion, second opinion, but still extremely important. So that's at the kind of very front end. There's also of course um like for example if you're doing medical practice and you're not actually in fact uploading your case notes and doing a cross check second opinion again like like at some year that's a that's malpractice just because what are these things? they have ingested a trillion words and have said I I have all this knowledge. We don't have that. I mean that's part of you know that that's part of the amplification we get. And then you get into um, you know, a whole bunch of other like, you know, kind of how do you um, you know, kind of operate hospitals medical care. You can imagine for example with the medical assistant it goes oh you should get to the ER right now. You can imagine scheduling and things like you know single-payer systems like NHS of oh as opposed to just well, you set an appointment and it shows up in eight weeks. It's the hey you talk to the assistant the assistant actually in fact triages you in so you get all that to hear. Then you get to the stuff that um, you know, as a tech entrepreneur and and investor is um part of the thing that you can see what's happening with AI is language and biology actually in fact uh closely pro approximates it's not exactly a language but there's a lot of language and computation in it and so the notion for example like the the the the typical cancer therapeutic chemo is we don't know how to do something specific with the one of thousand cancers you have. So, what we're going to do is poison you and we're going to hope that we kill the cancer before we kill you. And we have this little toxicology window that we're like, "Oh, let's make sure we don't kill you, but let's try to kill it." And that's what we're trying to do. Well, that's because we we haven't uh crafted the specific molecules that would say your particular kind of cancer, how do we um, you know, figure out a molecule that will bind to that only bind to that? So, not kill you, right? Kill the cancerous cells but not kill you. And and then that that becomes a huge search space problem in language. And so how do you use AI which is search based language to accelerate the understanding of what are the things that would bind that because by the way it's not part of the binding question for the folks who you know uh may not uh I've now learned a whole bunch of drug biology. Um, it's not just binding with the specific cell itself but also what else does it bind with? What else might it kill? And you have to make sure you don't have those negative effects. Right.
So you as a serial entrepreneur, you you sort of launched companies, you've entered different industries. Tell me a little bit about what it's like to enter healthcare. Um, and how is your approach different from big pharma and from interests that have been there for a really long time and are super well capitalized?
Well, so um uh I would say that my normal thing as a you know kind of Silicon Valley entrepreneur is avoid the regulated industries because anytime you have this kind of regulation, it slows down innovation, massively increases complexity, makes it more likely you're going to fail as a startup, you know, harder to get financing, a whole bunch of other things. Part of my uh kind of I wouldn't have ventured into this except with Sid because Sid knows all that part of the world um extremely well. Uh, but it's oh I have some differential knowledge about AI that could make a difference here and this makes a difference at a human like five-year-olds die from cancer. I mean it's like it it's it makes a difference at a humanity scale. I will navigate all of what looks to me like uh overly ornate craziness around regulatory industry, how economics flow, how uh clinical trials work, you know, how all this stuff works. But we'll we will get through that because the outcome is potentially worth it. But that's the
Okay. What what about the competitive set though?
Oh, and then pharma. Yes. So um look um uh look all companies uh learn how to be very professional at the thing they're doing uh but tend to be bad at new technological trends um and sometimes the those companies themselves um, you know, embrace and extend the new technological trends. So for example uh Microsoft's doing very well at AI but by the way didn't do so well at mobile right? So it's like it even on the the most elite most competent tech companies you don't always hit that that curve the right way. Sometimes you do there everyone else the new technology is usually like they just don't know how to build and embrace farm is the same way. So in terms of their biological science in terms of their understanding of the uh the kind of the regulatory process the therapeutic process bunch you know world class AI don't understand it at all. So it's sort of hammers and nails every and and is your suspicion I mean obviously you're making a bet um is your suspicion that AI is transformative enough that approaching it from the other side will yield gains faster more furious than than they can do?
Uh, that is 1,000% the bet and I'm certain that the bet is true. The question is can we do it?
Got it. Um, so uh you work very closely with Microsoft you're on the board. Um, and they are giving you some tools as well.
Oh yeah. So I mean look yeah describe some of those tools and let us know a little bit what may not be on the market now that you might get to use in this fight.
Well so one of the things that um Microsoft uh has is kind of Microsoft research which is um a whole bunch of extremely smart people who basically are paid to kind of just do research and they research on a number of things including quantum. For there's a quantum program that's come out of out of Microsoft and um and some of it is by the way because there's a natural thing between software and now AI and biology like it's like some of the world's best computational chemists and like I wasn't even aware that there was a discipline of computational chemistry until we started going down this path and then Sid's like well you know there's this guy who works at MSR. like well we can go meet with MSR that's very easy right and so and so the the there's a number of different of the MSR folks have have uh constructed uh very impressive computational uh chemistry tools and it's again part of the thing that we're doing at Manis is we're not trying to to only say we're going to build all of it ourselves if someone else has a piece of technology that's really helpful we're just going to use it right.
Um, you touched on this a little bit but I want to um project forward, you know, it's 10 years out. Um, someone's diagnosed with stage 2 lung cancer, which means that that's a tumor of 4 cm or less. Um, we know what would happen today. Why don't you just take us step by step what would happen 10 years from now if you're successful and how that's different?
Well, my uh hope would be uh especially for something major like lung cancer that we would have discovered a molecule that um for that particular cancer would literally all you would essentially need to do is get it into your bloodstream and would bind with the right cells, kill those cells, right? And nothing else. And it, you know, obviously the the the gold standard is could it be something you a pill you swallow? If it's an injection, fine. Sure. Right. And if it's an injection that has to be injected in a specific place, you have to go to a clinic. Fine. These are a million times better than chemo, right? So, um, you know, and and by the way, of course, if it's still super small, then you might still do a surgery, right? It depends a little bit on what the efficacies are, but like it's it's kind of the question of where you where you where you fit between, you know, on that spectrum. But once you get out of the the surgery thing, the idea would be a molecule that essentially solves your cancer.
And you think that's feasible in 10 years?
Less it's certainly feasible in 10 years. I mean again, can we do it? But but it's but someone will do it.
Yes. Okay. Um, so you work this kind of portfolio life which is great. One of the joys of which is you have really spent a lot of time with AI been an enthusiast. um what's changed the most in the last six months from because you've had a lot of pretty strongly held feelings but this thing is a roller coaster. So where have you shifted in the last six months?
Well I continue to be super positive as you know because of super agency and all the rest and we've talked about that before. Um, I'd say that um probably let's see the key things to track and updates are um like this year um a bunch of the Chinese efforts have demonstrated that they are actually in fact very much in the game um that there is a set of of different things that they have now there were a bunch of spurious claims like oh we could train a frontier model for much cheaper than you are and that's just all fiction but the but but they're training good models Um and there's at least four Chinese models that I'm tracking right now is as interest having interesting capabilities and things that you do. That's one. Um two is it's it was started earlier than this but it's one of the things that most people and general audience here aren't tracking is uh how AI's reasoning capabilities are improving. And if for example uh speaking to everyone here and online or whatever else um if you haven't tried deep research on uh chat GBT or Gemini or Claude or you know pick your favorite frontier model do and what specifically is to do like uh is you want it it's it's the kind of instruction you would give kind of like a research assistant. um like you'd say okay um I'd like you to you know analy I'd like to get a detailed analysis of the last uh 30 years of what advancements have been made in cancer medicine and d and I would like to include these particular things and you know and then it will go and compute for call it 10 minutes and then come back with something that's pretty amazing doesn't mean it's always uh doesn't mean there aren't sometimes hallucinations and other things but it's it's like stunning how good it is and it's getting better each time.
Yeah. And I would add just for people who haven't used it, um as a journalist, it is very easy to put in include citations and then fact check those citations. So the the hallucination problem, which is still real, is um much less problematic than it was even six months ago.
Yeah. Um, but but by the way, so so the chain of thought thing, the reasoning the reason I mention it is is that reasoning capability is another vector that's that's increasing improvement every quarter. And so part of the thing that you know the journalist cycle as you know likes to go, "Oh my god, it's going to be everything. Oh my god, it's oversold. Oh my god, it's going to be everything. Oh my god, it's oversold." It's not actually oversold, right? Uh, because like there's vectors for improvement on it. And then the very last one is a way to think about kind of what the future you're going to be living in is um is we are all going to anytime we're solving any serious problem deploy one or more agents. And what's more in the creation of software all of us will be in some number of years be creating custom software for the things we want.
Um, does your optimism extend to the transition in the labor markets.
Um, long-term yes. Uh, short medium-term uh, it's transition is massive difficulties with transitions like I anticipate massive difficulties. The reason I say as you know cognitive industrial revolution is both the outcome being really really important but the transition the the the transitions that happen being also quite difficult. So one of the challenges is actually just measuring where we are in that transition and what's happening.
If you were Secretary of Labor, what would you be looking at to get early signals? And if there weren't the right signals, what would you fund or create as a metric that would help us know what's going on?
Well, um, it's a good question. Uh, you know, um, the uh tempted to say something snarky about the current administration, but I will pass on that.
Oh, no. Yeah, hold it because I have other opportunities. Let's just So, um, but the I guess the thing I would be looking for is Okay. So you you want to take take a look at um the key place where you you will see a lot of different transitions are two forms. One is uh where will be the jobs of replacement be customer service etc will be part of that and then where will the jobs of amplification be and then what happens in the patterns of amplification. Now in the first patterns of amplification that will actually also look like job uh reduction um because as as the transition goes and you want people who are doing the amplification but actually in fact I think in a lot of those cases that will be a temporary effect versus a enduring effect because say for example you know you and I are company one and company two and we are uh competing with each other and marketing is one of the things we compete on. Sure, maybe um like my current our current marketing departments at 20 people could now do that same way work with four people, but we're competing with each other. And so part of how we got to 20 people right now is because we're still competing with each other and doing it. Right. Right. So we'll probably get back to a different set of configuration of 20 people and competing. That's the that's the kind of the amplification story and there's a lot of places where it actually has that initial kind of dip and then figuring out how you can figure whether it's product development because I actually think there's infinite demand for creating you know new software and product development. So the whole like, oh, software engineers are going away, I think is is literally just like silly, right? But customer service, um, I don't think that the same number of customer service jobs will exist on the other side. And and what's more, like one of the things that I find is kind of entertaining. And I will get back to your Secretary of Labor question. Um, is kind of entertaining is one of the things we're seeing is people are calling a human uh has a human on the other side and say, "No, no, put the human on." Right? Because they're so frustrated with a human who's trying to follow a script that they think it's a robot. Yes. Right. So, so it's like okay um that's part of the reason why I think like the very first calls, but that will transform what is customer service experience because right now it's how do we uh hire someone for the least possible wage anywhere in the world, have them follow a script and get you off the phone is essentially the customer service thing. Now, if you're running AI, it's how do I help you? Hey, are you having a good day? Is there anything else? you know like it's fine right to do all that stuff. So it it'll be a transformed experience but there will be jobs. Now on the secretary of labor side it would be okay take a look at these two kind of categories of jobs and then be tracking what current volume both entry level Stanford did some really good work here, you know, entry coming in and current volume is and then if there are uh repeated changes going on think about like what kinds of things you want to do to try to help with job transitions um and actually getting a set of you know kind of good thinking and ideas because one of the things we know is important for the functioning society. It's okay if jobs changes. Okay. If like the you know I know there's still horse and buggy drivers here in New York. Yes. And I still have a I still have my footman and he's great. Yes. But Yeah. But you know like not so much anymore. Yeah.
So that that really leads into the next question which is that for better or worse uh we currently live in a civilization. Yes. And that civilization has different sectors that have to collaborate in order to manage these changes.
Oh, you're reminding me of Mahatma Gandhi's quote on Western civilization. It it would be a good idea. It's true. Um, so let's say one day we get one. Um, h where are you on the uh role of politics in helping with this transition? Do you have any faith in the current political environment? Um, can we accomplish something as large as transitioning a significant sector of the workforce through this uncertainty?
Well, I think obviously, you know, um, you know, I worked pretty hard to try to not have the current administration. Um, and you know I have a whole list of things whether it's from you know insanity around vaccines to um, you know, kind of challenges around um, you know, like tariffs and volatility and what it does for business markets and prices and everyday living and jobs and all this like and it we could spend three hours just listing all this stuff. So that intrinsically makes you less optimistic. Um, now that being said, I think that the the responsibility for all citizens is to try to make it work. And so, you know, that's part of the reason why I went, okay, um, you know, this this administration doesn't want to hear advice from me. That's fine. Um, I will go build stuff and make it happen. But I think that um that you know we are not uh look if if if we're having a discussion about whether or not vaccines make sense, right? Or you know other things where there's extremely clear scientific evidence on these things. It makes it much harder to have faith that other kinds of good call it intelligent governance um play out well.
Nice. Um, so uh you have uh I I I omitted all the exploitives in I thought I thought you did a great job there. Um, the conversation around AI and safety has shifted fairly dramatically and in large part I you think I'm pivoting off this subject. I'm not. Um, in part because we now understand the power of authoritarian regimes and AI. So let me ask you a bit of a hypothetical that I hope we don't uh have as a practical. Um, in what ways can an authoritarian government use AI currently as currently exists in the marketplace um to increase its control over society and how would we know that they're using it?
Oh um the second part is a little bit more challenging. Um although part of the question okay so the first one is it's already happening it's happening in China there's there is like you know a bunch of things in terms of weaguers and other things that it's already happening that I've um years back not many not within the 10-year window but a um an investor friend of mine went to a uh company in in Beijing and for an investor pitch and they showed him. Here's where everywhere you've been in Beijing because they had all the the the footage and he's like, "Okay, thanks. I'm leaving now." Um, and so, uh, so that's like there's none a number of different things that could easily be done now and are being done in in some places. Now tracking that it is happening um I think is probably mostly a function of you know well-ordered societies which we're having some trouble with right now. Um, the um there's on the most cutting edge you have to be working with the companies and one of the things that the companies are good at is they have alternative like they have a board of directors and they have shareholders and they have a bunch of other things and and most of those companies um, you know, are pretty explicit about uh we refuse to have our technology used for mass surveillance of citizens. Right. And so far, yes.
Um, okay, we're going to end it on a slightly more optimistic note. Um, you are famously worked. Um, this is a networking event of some kind. Um, I'm curious, what conversation have you had over the last year or so that's just completely blown your mind, changed the way you think about the world?
Um, well, probably the good news is I probably encounter those um reasonably often. Um, that's that's that's the the the one of the things I love about, you know, the kind of place where I end up. But maybe this would be the kind of because it's something like I know really well, but I still learn something really amazing. So I was sitting with Ethan Mullik at Wharton um who is like one of the great like you know kind of leading thinkers and like follow his dreams and so forth and he was like oh people are undercutting they don't understand all the stuff that can you can already do with multimodal AI and I was like sure I think I know everything tell me what you're what you're what you're thinking and he said well um one of the the big problems in construction projects is it's very hard to know if you're on track if there's a long pole uh tent if if if um if there's something you can do to try to accelerate it and all the rest. And so you have inspectors go by, but the inspectors are like one time slice and very expensive to do and and you know kind of manage and all the rest. So what he did is on a construction project he set up 24 cameras. He loaded into the AI the construction plan and they asked the AI to do a report at at each day where the construction was because it was reading in multimodal for the cameras and got a detailed like and this is like a guy who's not even really a coder got a detailed like here's the things that are working well here's the things that are at risk here's the things that are falling behind here's the things you would need to do in order to try to stay on schedule etc. Wow. And at the end that his contractor was still like, "You can have it fast, you can have it cheap."
Yes. Exactly. Cool. AI. Um, all right, Reed. Always a pleasure. Thank you so much for being here.
Thanks.
[Music]