Reid Hoffman on Why Everyone is Now a Manager of Agents

Thank you. Pleasure.

Pleasure.

So I think probably the most useful question is what are the ways that we take these scaled learning systems, which on hypothesis tend to be probabilistic systems, and get to acceleration formats. And that's part of the reason why the coding capabilities, you know, kind of intensely kicked off in December with Claude Code and Opus 46, and actually Codex with 5.5 is also particularly strong, are ways to look at this for everything that involves knowledge work. And so for example, one of the principal things that was kind of an academic-ish difference of like wow just pure generative only, but one of the things that you're seeing happening with these coding tools is the development of skills, recipes, and then compounding loops where you have a probabilistic system that creates some deterministic code that can execute, that can be evaluated by the probabilistic learned system, and then refactored in terms 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 Claude Code to do lab instrumentation to do analysis. One of the things that I think is part of doing this is when you look at how you create kind of agentic 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, as a small example, one of the things that's interesting in progress is so we started, I do this podcast called Possible. We have Reid Riffs on it. It's primarily my co-host Aria Finger and I talking to each other about news of the week. And part of what we constructed early, like literally this long time ago January, was a process for translating it into 20 languages. We redid that process with the GPT 5.5 Codex and it now does it in a much more one-shot kind of thing with a much simpler set of agents. And that's how part of designing your thinking in a 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 misleading about Sutton's thing is actually in fact the loop between these generative AI systems and systems that they work with and generate and produce themselves like code and other kinds of things are part of what creates that acceleration of capability.

Well, so, you know, as complicated as Go is and Move 37 is one of the numerous places where there is novel and good in the intersection of the kind of the biology and so like for AlphaGo 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, seriously. And so 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 supervised learning and reinforcement learning and other kinds of things in order to generate prediction calculi. 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 was a hallucination. And the difference between a hallucination and a good idea is one that proves out to be good when you know when it originally looks novel and bad. And so the question is how do you generate that in the appropriate way for this much more complicated system in biology because like taking curing cancer as a thing is you get there's not just one thing cancer, there's many different cancers, thousands, very true. There is 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, likely don't have negative toxic other kinds of consequences, and that you know you can then 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 we like, I kind of joke 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 Sujata reinforcement learning because Sujata Mukerji goes okay here is some of the atmospheric patterns in order to focus in this enormously complicated space here and then you have to do various kinds of techniques to generate 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. 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. But anyway, all of those things go into it. Now, we think, 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 like, you know, genius and mad people. You don't know until you're down the road some which one of those two you are.

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. 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 an important part of how it operates do look like they can produce good prediction 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 pharma industry, but we've already discovered a number of IND candidates for the initial filings that happen for saying I have a prospective molecule that could make a big difference.

Oh well, so 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. And you know, it would be nice to simply say that is just all hype and BS. 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 in two years we'll all be uploading ourselves into Von Neumann probes for 15 cents per upload and going out and exploring the universe which is from a very smart person. So I tend to think that there's a set of things here you would say, oh that's all hype and 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 come out. I think one of the things that's going to be, I think Terence Tao 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. Yes. Yes. I didn't think it was out yet, but I know of it. Yes. 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. Yes. And he doesn't think he could 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 causes it like these kind of what the limits are are uncertain. Yes. 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 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% 0.1% but it's hard to go to zero. Yes. Now in terms of things to expect is like take for example say stuff we're doing at Manas, we've got some very good initial work in saying here are some possible targets 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 have what we think we do. Yes. Because even though we can kind of run the cross-checking because this is again one of the things that tends to happen like a lot of it is like a lot of the claims around generative AI tend to be one-shot versus well what happens if you're in an agentic reinforcement loop and what's happening and when you have the same model saying like one of the bizarrest things about this that's kind of counterintuitive is 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 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 silico but as you know when you get to biology it's like well let's try it. Yes. Right. And so that's at least two to three years.