Reid Hoffman on Why Everyone is Now a Manager of Agents

Added: 2026-05-31

[00:00:00]Thank you everyone. Uh you know I'm really honored uh you know to join now by someone who actually really need little introduction. Uh but but certainly someone really I respect a lot uh you know because certainly how he think about technology networks and the future work for the last two decades. We hoffman here as a fellow pipel myself especially fantastic >> to have two of us here on stage today.

[00:00:28]Thank you Weed. Pleasure.

[00:00:29]>> Weed uh is uh co-founder of LinkedIn which of course many of you guys are power users here sir a partner at great lock and co-founder of menace AI. He's a AIdriven drug discovery company. He started with Dr. Sahhata Mohiji. He was also an early investor in open AI, co-founder of inflection AI and is the author of super agency which makes one of the most thoughtful cases for how AI can expand what each of us is capable of. Few people actually have moved between operator, founder and investor with the range that you have weed and even fewer have thought as publicly and carefully about where this is all heading in the age of AI.

[00:01:19]>> We well thank you for being here.

[00:01:21]>> Pleasure.

[00:01:22]>> Thank you for spending the time with our community here today.

[00:01:26]Well, first of all, let me start with um you know, you said something about AGI, which is really about you make the comment which is the AI we haven't invented yet.

[00:01:37]>> So, I'm going to ask you I guess right now what is the most useful question you should be asking? you know.

[00:01:45]So I think uh probably the most useful question is what are the ways that we take these scale learning systems which on hypothesis tend to be probabilistic systems and get to um acceleration formats and that's part of the reason why the coding um capabilities you know kind of intensely kicked off in December with cloud code and opus 46 and actually codeex with 55 is also uh particularly strong um are ways to look about that look at this for everything that involves knowledge work. And so for example, you know, like one of the the principal um things that was uh kind of you know, I suspect it was kind of an academicish difference of like wow just pure gener only that. But one of the things that you're seeing and happening with these coding tools is the development of skills, recipes, and then compounding loops for you have a probabilistic system that creates some deterministic code that can execute that can be evaluated by the probabistic learned system and then refactored in terms of of so doing. That is essentially one of the kind of accelerations and increase in capabilities that goes across an entire range including science. So that part of the thing is not just like oh you know scientists can use cloud code >> to you know do lab instrumentation to do analysis. Um one of the things that I think is is part of doing this is when you look at how you create kind of agentic um you know kind of uh workflow processes even when you're using the same model you're actually prompting it to different roles in a team like format. And so, you know, as a as a small example, like one of the things that's interesting in progress is so we started I do this possible this podcast called possible. We have read riffs on it. It's p primarily my co-host and I Arya Finger and I talking to each other about you know kind of news of the week and um part of what we um we constructed early like literally this like long time ago January um was uh was a process for translating it into 20 languages. uh we redid that process with the uh GBD 5.5 codecs and uh it now does it in a much more oneshot you know kind of thing with a much simpler set of agents and that's how like part of like designing your thinking in a in strategy for whatever you're doing in an AI native way is to be >> projecting where do you think the AI capabilities are coming >> and part of the thing that I thought was was was misleading about um you know Sutton's thing is actually in fact the loop between these generative AI systems and systems that they work with and gener and produce themselves like code and other kinds of things are part of what creates that acceleration of capability.

[00:04:42]>> Wow. Well, very exciting stuff. I I love to kind of find ways to work with you on that fun by the way. But since you mentioned about you know this amazing about 20 languages with chat codeex and AI for science specifically I do want to talk a little bit about the manace thesis which is you know with menace you said it isn't enough to have the best AI or the best science you want to actually put them together what's the hardest about that in practice though >> well so um you know as as complicated as go is and move 37 is one of the numerous places where there is novel and good in the intersection um the the the kind of the biology and so like for alpha go more moves in go than there are atoms in the universe and it's a way of kind of articulating the difficulty and the complexity of space biology is even more complicated yes uh seriously and so >> yes >> part of what you need to do this is one of the things that's useful about the various forms of predictive AI which come from you know um you know supervised learning and reinforcement learning and other kinds of things in order to generate prediction calculi.

[00:05:52]Some of which are hallucinations. But by the way, you could say move 37 was a hallucination, too. Certainly during the mid part of the game, everyone thought it was the losing move and and was a hallucination. And the difference between a hallucination and a good idea is one that proves out to be good.

[00:06:08]>> Uh when you know when originally looks novel and bad. And so um the the question is how do you generate that in the appropriate way for this much more complicated system in biology because like taking uh curing cancer as a as a thing is you get there's not just one thing cancer there's many different cancers thousands very true >> there is um lots of different targets you have to identify targets knowing targets is one of the key things just like in science knowing what the >> question is and how to formulate the question well is pretty central then you've got to have things that bind that target uh likely don't have um negative toxic other kinds of consequences >> and that you know you can then uh you know manufacture and make in a way test in a way go through animal testing go through human testing go through clinical testing etc and so you're applying predictive AI to all of that but it's an enormously complicated space yes >> so part of the way that um we like like I kind of joke internally internally to what we're doing at the company is we have a group of world-class AI people >> but as opposed to general reinforcement learning we have SID reinforcement learning because Sedar Mukerji goes okay here is some of the you know atmospheric patterns in order to you know focus in this enormously complicated uh space here and then you have to do various kinds of techniques to generate uh data well because one of the things that when you look at the methods to build these kind of predictive algorithms usually have to have a lot of compute and a lot of data.

[00:07:44]>> And while there is a ton of biology in the world, there is actually very sparse forms of biological data. So you have to solve that problem, too. Yes.

[00:07:55]>> Um but anyway, all of those things go into it. Now, we think um you know, we've already started seeing some things that could be move 37s just from the approach we're looking at. But of course, it's it's like, you know, uh genius and and mad people. Um you don't know until you're down the road some which one of those two you are.

[00:08:15]>> Well, I love that actually. I you know, I think this is I agree actually. In fact, I I'm it's very dear to my heart as a matter of fact about cancer because I I've lost my dad because of that. So, I'm also it's a personal mission of mine by the way. Um and and I also want to welcome actually that because small uh going to lead by Nobel laurate Randy Sackman we actually have an entire track on biology but I certainly agree with you on that and I think part of that is also why we're building the seir that integration you know um but but I guess my question is what have you learned at menace that we should know about you know now that you've been building this right you know what what are some of those important things we need to know >> well we haven't announced a lot so I'm not you know even though at a science event. I can't pre-announce what we're doing in the company. Um but what I would say is that the thesis that these forms of modern AI with kind of scaled compute learning of which generative AI is a important part of of of of how it operates do look like they can produce good prediction uh candidates for both targets and drugs. And we've already in our prototyping done I guess I don't know how many people here understand the farm industry but we've already um discovered a number of IND candidates um you know for the initial filings that happen for saying I have a prospective uh you know molecule uh that could make a big difference.

[00:09:49]>> Okay. Well, but but I also have to ask you this this question too because I think it's important is in biology, you know, is of course uh with AI is generally changing the timetable, but I guess what is hype versus not. Oh well, so um look, one of the confusing things in AI is I know very smart people who have been right in a number of predictions themselves who claim that we'll have some version of super intelligence in the next, you know, call it three years.

[00:10:21]>> And you know, it would be nice to to simply say that is just all hype and BS.

[00:10:28]Um uh maybe the closest thing and maybe some of you will know who I'm referring to. I'm not going to call them out by name that I think is BS is um in two years we'll all be uploading ourselves into Vonoyman probes for 15 cents per upload and going out and exploring the universe which is from a very smart person. Um, so I tend to think that the uh that this um this uh there's a set of things here you would say, oh that's all hype in BS. The problem is is that it's kind of unknown in these exponential curves which don't just include the compute but include like the coding acceleration that I mentioned earlier and other kinds of things. What kinds of new capabilities would be there? And maybe these new capabilities because by the way the current AI we have today already has some characteristics of super intelligence right like super intelligence is not like turn on the light switch super intelligence not it's this kind of jagged edge of capability where some things uh come out I think one of the things that's going to be >> I think Terrence Ta did some pretty good maths of is it out yet? Yes, we're working on it. This is my co-founder. In fact, when we actually announced it, I want to bring you over.

[00:11:41]>> Yes. Yes. I didn't think it was out yet, but I know of it.

[00:11:44]>> Yes.

[00:11:44]>> Right. So, but like you know, that's kind of an area where you say, well, with generative AI, there is actually in fact novel results that are coming and so forth.

[00:11:52]>> Yes.

[00:11:52]>> Um, and he doesn't think he could have have done those particular discoveries without partnering with generative AI in terms or using it as a tool in terms of how you're operating. So, so what caus it like these kind of what the limits are are uncertain.

[00:12:06]>> Yes.

[00:12:07]>> Um, and so there is a lot of hype in terms of you know like oh in two years we'll have a white collar blood bath you know blah blah blah blah and the the various things but you can't count out a large range of possible capabilities in the next one to five years that might happen at some probability. Maybe the probability is 0.1% or 0.001% 01% but it's hard to go to zero.

[00:12:32]>> Yes.

[00:12:32]>> Now in terms of things to uh expect is like take for example say stuff we're doing at Manis >> um we've got some very good initial work in saying here are some possible um targets and and and molecules that might work that show possibly very unique characteristics almost like move 37 characteristics. But it will also take us at least two to three years to fully to really understand if we if we have what we think we do. Yes. Because even though we can kind of run the like cross-checking because this is again one of the things that tends to happen like a lot of it is like lot a lot of the claims around generative AI tend to be oneshot versus well what happens if you're in an agentic reinforcement loop and what's happening and when you have you have the same model saying like one of the bizarest things about this that's kind of counterintuitive is you you get a result from a generative model and you say make it better and it makes it better right which is totally unlike human. And so what happens in the kind of workflow and and and and and you know kind of give three hypotheses and argue between them and all the rest of the stuff in terms of doing it. You can do all that in silic but as you know when you get to biology it's like well let's try it.

[00:13:48]>> Yes.

[00:13:48]>> Right. And so that's at least two to three years.

[00:13:50]>> Okay. Well I I said I actually in fact I was just at at Harvard um and also Yale.

[00:13:55]I mentioned that I hope to with AI's help perhaps in 10 years we'll be able to solve the problems most of the problems in cancer. Well I I hope perhaps that may be a good goal too.

[00:14:07]>> I think it's a good goal. I think what you'll we'll my guess is if it's really going fullbore we will have figured out call it you know 10 to 20 maybe 30 of the really major cancers.

[00:14:21]>> Yes. Um, but by the way, you know, the usual thing is once you solve one problem, you figure out there's new problems to solve. Yes. So, we might actually be on shot on goal on some of like the real difficult killers, you know, like those kinds of things.

[00:14:34]But, >> um, like, you know, blood cancers, most people other than chemo, which is, >> right, as a way of doing it, I think we actually will figure out we we'll have some stuff that's really on target there. I don't think it'll be all cancer just because we'll discover that oh there's these little remote ones and all the rest >> but but my point is I love the work that you're doing by the way and we love to me and sir uh the foundation would love to work with you so we'll follow up on that but moving on from founder to investor because you obviously was mentioning you've been legendary from LinkedIn to great now to manice you keep going back to operating why >> um you try to solve the problem >> okay >> so whether the problem is you know like I had founded social net but then uh was recruited into PayPal as an executive because of solving the problem and then went and founded LinkedIn >> um you know lead investor for open AAI again you just like what's the shot here that I can help solve this problem >> and part of the reason why I did Manis was because >> I felt that one of the things that happens in Silicon Valley is you kind of get a um call it a software blind spot which is it's like well we're going to solve this problem entirely in software.

[00:15:49]It'll either be all in silicone and simulation or it will be uh we'll create, you know, novel drug discovery researchers and we just press a button and you can get 3,000 more drug discovery researchers. Both of which are totally good and fine ideas but like I think there's some challenges about where to get there. Um where how do you uh bring in and the reason why we tend to avoid things like um you know the world of biology, the world of atoms because it slows things down. It's very difficult. you got lab work, you got other things. And I was like, no, no, this is we now have the tools to take this shot. And I think that's that's kind of the that that's the reason I went into a co-founder role there to to make that happen. Y >> now um so whether it's it'd be surprising if I went back to an operating job. I mean, I'm >> I don't know. I guess I'm not that old.

[00:16:41]58, but you know, uh >> that's quite young. 58.

[00:16:44]>> Yes. Yes. you know, I hope uh but the uh but I think that the most likely and I think I will probably co-ound another thing or two um because I can do by part of blending investor and and operator is um helping pull that initial team together really get it going >> uh put some real time intensity and then move to more of a you know call it uh exec chair on the board.

[00:17:13]>> That's great. Well, I mean I guess because there are a lot of builders here as well. So when you meet a researcher, obviously many of them, you know, like a researcher is also in our community thinking about starting something.

[00:17:25]>> What what's your advice for those folks, you know?

[00:17:29]>> So a typical thing because obviously anyone who's doing research on these things has a very, you know, high, you know, IQ, uh, learning, etc., and tends to think things in business will be simple and follow and and follow.

[00:17:47]>> Um so that tends to uh scuttle a lot of projects uh run them into minefields.

[00:17:53]And another one is um like this is probably the principal thing I learned when I decided to move from academia into business is that the the business fitness function challenge is the simplest um thing that is valuable, >> not the gold medal for solving the hardest problem, >> right? And so if you're trying to create a business, you're orientating at what's the simplest thing I can do that's high economic value. And that's already hard enough itself. And when you're dealing with because especially because you know like researchers will generally speaking be go well I'm figuring out this new form of you know kind of um chemistry or this new form of quantum or this new form whatever the thing is. You're like okay that's a you got a whole bunch of hard science problems. Let alone when you get to all the business problems which are like how do you develop customers because you in a startup path you generally get how you get a limited amount of capital to prove yourself to get to the next capital you know etc etc and all of that can you know suddenly you're playing a time like you've lit the fuse which you will ultimately blow up >> if you don't succeed in creating it into a business.

[00:19:05]>> Yeah. Well very good. Well thank you so much. I I think that don't think of business is simple and and we shouldn't absolutely not and you're absolutely right about that. Uh moving on to about your book super agency right and certainly you know your book makes the optimist case >> but you've also set public sentiment toward towards AI gets worse in 2026 >> and so how do you hope both and what's AI narrative that we're getting wrong right now you know >> so the general problem is is that um most people call it well north of 67 70% prefer to have certainty of lack of downside versus upside.

[00:19:49]>> And so when you come along and say we have a really disruptive technology, it's going to change jobs, change ways function in industry, change information flows, uh change uh evolve our conception of what it is to be human, etc., etc. Most people say no. Don't don't like I don't want to have the the the possibility of disruptive downside.

[00:20:08]Um and uh and the problem is is by the way, we're there. It's going to happen.

[00:20:13]It's just like the industrial revolution. We'll try to learn from the industrial revolution. Make this cognitive industrial revolution better.

[00:20:20]>> But it's like it's, you know, take as a microcosm like Hollywood kind of going, you know, oh my god, you know, this thing can produce all of these amazing images and movies and and and and other kinds of things. And that totally changes the way that that the entire content industry operates. Um and that's just kind of a microcosm for this. And so that's part of the reason it gets negative. Another part is, you know, there's a obviously a a general um kind of set of opposition to big tech uh in various ways. And so AI is approximated with big tech and then there's you know kind of questions around um you know kind of like okay so what uh what are the positive use cases for me and so part of the thing that got me in Amanis was you know and I what I advise governments to do is get a essentially a medical assistant a legal assistant an educational assistant basically running for free on everyone's smartphone so they start seeing they have some real serious benefits from this that isn't a oh look you're going to learn how to do amazing new things with AI but things that are kind of deliverable and tangible now I think even with that it will be a it will be a contentious transformation um but one of the things is when you're thoughtful and and trying to do stuff you don't avoid bad futures you don't make good futures by avoiding bad futures you have to steer towards good futures and the just kind of like oh let's all slow down is unrealistic in uh human history.

[00:22:02]>> Well, you know, actually on on that note, I think that you know, like Asians has been um you know, a very hot topic, agents, HGI. So my question to you is, you know, like you've said in 2025 most agents in code, what does 2026 look like when agents break out? And I guess also want to follow up on the single person operating with the capacity of a team.

[00:22:26]>> What does that mean for how science gets done?

[00:22:29]>> Right. So one of the ways to conceptualize all kind of information on knowledge work of which we have the capacity for this today is fundamentally no one should really be an individual contributor. They should be a manager of agents. Now, some of that gets down to we're still kind of, you know, lacking a bunch of the pulling together what is an agent like I have a bunch of agent infrastructure doing stuff because, you know, uh, you know, this year I've started playing with code and coding and, uh, using, you know, claude and codeex >> um, to do various things and pull the stuff together now. But um it's basically you should be thinking about any serious problem of how do I deploy an agentic work framework to accelerate what I'm doing and what my capabilities are here because among things when you think about these agents um they are 247 they don't get bored they're okay doing stuff repetitively they're okay uh doing stuff with a kind of a more of a brute force exploration and that you can apply apply that to almost any problem that's kind of interesting, including a bunch of science problems. And that's not to say that, you know, okay, I'm going to sit back on my couch and you're going to discover um the way to do containment for fusion for me. I don't think we're anywhere near that yet. Um but um even when you're like for example doing research on what are the different possible areas of containment fusion, you write something down as a as your working theory and you say okay um what I would like is a detailed report on um everything that plays into what my idea is for it, against it, alongside it, etc., etc. I'd like you to rank it based on things that I think you um would probabilistically assume is good etc etc and there's an example of something that is an acceleration of science >> just in that now talking to some you know folks smarter than me you know one of the things that they predict is that there will be a lot of science acceleration this year and next year >> um we're already seeing it in math and theoretical physics and it will because those areas have less of a you know complexity of a massive data problem, but you'll see it even in chemistry and biology and all the rest. It's obviously one of the things we're trying to do in Manis and it'll be a little bit of like some here and some there and some here and not just a wave and the whole thing depending on what the overall acceleration of capabilities is.

[00:24:55]>> Okay. And my last question uh is uh you know I I really kind of like science takes longer than software and you've been of course uh you know very successful uh in in the software and in tech. How should patient capital actually work for AF for science companies and then what would be the role of organizations such as Seir alongside venture and other labs working with you?

[00:25:19]>> Yeah. So um so one thing is um because of the speed and capability in software I try to simplify the problems into problems that software can accelerate. That's a general hack that I try to do in everything. Uh because it's one of the like when you have a magic wand, you try to apply the magic wand as much as you can, >> right? It doesn't mean it works on everything. Doesn't mean that everything valuable is the end result of you know when you have a when you have a hammer, you know, not everything is a nail as much as you would try to make everything into a nail. And so um but you know, if you have a hammer, you're like, well, what are the kinds of projects the hammer could really accelerate and what could it work for?

[00:25:58]>> So that's one. Two is in terms of patient capital. I think that part of one of the things that is a general rule in financing is to think about what is future capital look like. So part of, you know, what I learned from coming into Silicon Valley and and doing companies both as a founder and as investor here is whenever I'm thinking about one fund raise, I'm thinking about the next two in various ways and thinking about what that pattern looks like and what the course needs to look like with some range of probabilities and also to navigate that >> and include the >> if we're not seeing X, that's where we essentially quote unquote shut this project down. Now in companies that could be sell it, >> could be shut it down, could be other things, but it's like >> uh could be major pivots >> etc. and then have a a kind of a you know kind of have that thought in mind from the beginning and then what you say what that gets to in terms of patient capital is to say like when I'm tackling a problem or looking at a problem where I think that the later capital is is um much um you know kind of more difficult to get. I generally speaking reserve capital for at least you know being able to catalyze those rounds to make it happen if the data shows it. And I do really like one of the things I think is a discovery of kind of the last few decades, four or five of venture capital is kind of the seed series A series B escalating capital commitments which of course includes money and time and people and all the rest relative to you know increased evidence on the uh on your hypothesis about what it is you think your investment will yield >> and how should organizations just s support those efforts. I can think there could be a couple. I mean, one of them is, you know, one of my like for example, if it's like people trying to do like one of the things I've kind of told people who tried to create um you know, the frequent term for this is technology transfer from universities into companies.

[00:28:12]>> Yes, >> it's a little bit of a weird term, but we'll go with it because and it's the term that's used. Um and it's like well connecting those people who are very bright scientists who have an idea etc to people who can talk about like well what would that be if it were a company and it were going through financing cycles and what are the things that you would need to have in order to do that and most often the scientists are not the right people to be managing that organization and one of the things is finding the right kinds of you know kind of business and entrepor reneurial talent but I think that kind of matchmaking for that vector >> is I think useful. I also think obviously you know doing convenings like this where people can meet each other um and potentially because you know part of part of how networks amplify entrepreneurship >> is um kind of sharing signals around you know um who's interested in what what skills are important what problems are not are important to solve all that kind of thing. And so that kind of information sharing, relationship building is also very useful. And I'm sure there's other things too, but I, you know, limited limited knowledge.

[00:29:24]>> I love it. Well, ladies and gentlemen, please help me in thanking Weed here.