Zoë Brammer & Ankur Vora | How a board game helps Google DeepMind plan for AI in science

[00:00] Ankur: We wanna take seven hours of people's time and do an AI war game on AI for science futures.

[00:06] Zoë: For most people, when they think about automation and science, they assume that the role of human scientists will go down as AI becomes more and more powerful. One of the prevailing themes from all the games is the criticality of scientists who understand the technologies and who can connect those technologies to the concerns that governments face in order to drive investment.

[00:28] Beatrice: I'm very glad to be joined today by Zoë Brammer and Ankur Vora from Google DeepMind, who together run the Strategic Foresight department at Google DeepMind, which we're going to dive into what that actually is, and especially your project on Science 2030. So let's start — maybe a bit about your backgrounds to begin with, because I think you both come from quite different backgrounds, but now you're working together to run this. So what are your backgrounds, and how did you end up working at Google DeepMind?

[00:59] Zoë: Yeah, thanks so much for having us. First of all, we're really excited to be here. Yeah, so I'm Zoë. I work with Ankur, as you said, on strategic foresight at DeepMind, and my early career was actually in cybersecurity policy. And when AI hit the scene, it became clear to me that the future of cyber and cyber policy would be really closely linked to the future of AI. At the time, I think my idea about the future was heavily influenced by my background, my research, and I was really, really focused on risk. And it wasn't until I joined DeepMind and started working with Ankur that I think I developed a much more nuanced understanding of the future of AI in all kinds of areas, and the kind of risk and opportunity balance that's presented there. So yeah, I'm really excited that I get to work on these types of topics. Yeah, and as I said, glad to be here.

[01:49] Beatrice: Thank you. 

[01:50] Ankur: Yeah, and I think maybe one thing that we have differently is that I've taken quite a meandering path — I think Zoë's is probably a little bit more linear. But one thing that we both share in common is that we haven't necessarily been intentional and set out a goal at the very start of our career, exactly where we would land up. And for me, that translated as — I actually started off studying medicine. I was kind of the classic, you know, wanted-to-be-the-good-child-of-Indian-parents, go and see medicine or engineering or law, and chose to follow my parents' path. And then decided actually politics and philosophy was more my kind of vibe, and went from that into the civil service. And then joined Google initially to do their kind of policy outreach program. And somehow, while there — it was about 2016 or so — saw Demis Hassabis, CEO of DeepMind, give this talk about AlphaGo and kind of the AI breakthroughs there, and their mission at the time was kind of to build AGI to benefit humanity. And I was just immediately sold on everything that Demis had spoken of. I found it kind of very inspiring, exciting, and thought-provoking in terms of the kinds of issues it raised about the future and the many possible futures it kind of presented. And so I vowed that I would, at some point, get a job at DeepMind. Spent a few years kind of doing that, and the opportunity eventually arose, and I joined DeepMind in around 2019. And yeah, it's been a real journey, because I think at the start we were in quite a different place — a lot of the focus was around both DeepMind's efforts on games, so following things like AlphaGo, and then AlphaFold was kind of the project that I initially spent most of my time on. And then, of course, kind of the direction very much went in November '22, and things around ChatGPT, in that kind of large language model direction. So it's been a really interesting journey. But yeah, at no point do I feel like either of us have necessarily been, "This is a five-year plan of where we'll be in our career," and been intentional about how to get there. It's kind of being a bit more opportunistic.

[03:44] Zoë: I'm opposed to five-year plans, actually.

[03:46] Ankur: Yes.

[03:48] Beatrice: Yeah, interesting. Do you wanna expand?

[03:50] Zoë: So I think the exercise of making the plan is probably worthwhile, but at no point in my life has anything gone the way that I thought it would go. And I think if you commit yourself too much to a plan that goes that far out into the future, you might miss out on other opportunities that come up — like your friend and colleague asking if you wanna start a new team doing something you've never really heard of before. So I think it's important to retain optionality.

[04:17] Beatrice: Yeah, I agree. It's good to have a direction-ish —

[04:23] Ankur: Right.

[04:23] Beatrice: — but then be open to catching the opportunities. So let's talk about what strategic foresight means for people who maybe aren't familiar, and why does a lab like Google DeepMind have one?

[04:34] Ankur: Yeah, you wanna start?

[04:35] Zoë: Yeah, sure. So I think, from my perspective, strategic foresight means thinking through the range of possible ways that the future might go, and trying to identify paths that are resilient across multiple different ways that the future might unfold. I think Ankur and I talk a lot about the importance of being humble about how things might unfold. We don't pretend to know with any degree of certainty how the future will roll out. And so strategic foresight is this methodology, or set of methodologies, for thinking about all of those possible paths and trying to identify actions that we can take now that will move us in a direction towards a future that we might be more interested in pursuing.

[05:21] Ankur: I think what we're interested in — and we sit overall within an area of DeepMind that focuses on public engagement — is that there are actually a lot of different mental models about how the future might unfold, and those mental models involve people taking a lot of different kinds of assumptions about what things might be important, and how those things might play out in society. And what we're interested in is getting people with a number of different mental models into the same room, making those assumptions explicit, and then stress-testing each other's assumptions.

[05:54] Zoë: Yeah, I think, to answer maybe another aspect of your question about why doing foresight around AI trajectories might be important — we were at South by Southwest yesterday talking about some of the different aspects of AI that make it, in our opinion, uniquely important to think about foresight for AI. I think one of those is that AI is a general-purpose technology, and like all general-purpose technologies, it's really difficult to think through all of the potential downstream impacts of the integration of that technology into society. Ankur likes to say that it would be like trying to imagine the impact of electricity from the early 19th century, which is a pretty impossible ask. The other piece, which Ankur touched on as well, is the idea that AI is a sociotechnical system, and like all digital technologies, it is both influenced by and coexists with things like human behavior and cultural norms and the ecosystem in which it is built and in which it operates.

[06:57] Beatrice: Yeah. You mentioned that one of the things is to get different types of stakeholders, or people with different views of how things might unfold — does it also entail, if you're building out these different assumptions about how the future could go, seeing where there's overlap of where people want to go, and trying to steer towards that? Is that also part of it?

[07:19] Ankur: Yeah, definitely. I think you mention two aspects, I guess, to looking towards the future. One is, as you say, what's the full light cone of possibilities — and of those possibilities, which ones are more probable, which ones are more, or less, plausible, and why, and what might some of the factors that go into those futures be? I think then there's a much more normative question of which of those futures are more positive, and that's something where it requires people to add their subjective valence to that, right? And I think that's something that we also try to think about — not in regard to us necessarily attaching our values to what is gonna be good or bad, but more for people to reflect, even in these scenarios, like, "Oh, actually there's gonna be some trade-offs." Okay, we've achieved a world in which we're far better on some elements, but there are second- and third-order effects to that, and actually we've created more risk in these areas, right? And so I think a lot of what it comes down to for us is that positive futures are also about making trade-offs and making decisions about what we want to prioritize. And a lot of the time people have tacit intuitions about that, but they haven't necessarily articulated and tested those with other people.

[08:35] Zoë: Yeah, yeah. And the most positive futures are representative of the widest swath of opinion possible. So a lot of our work involves using public engagement as a tool for this two-directional learning, where we have some knowledge that the rest of the population doesn't have, but there's a huge amount of expertise that we don't have and that we really need in order to think through what a positive future might look like, and point ourselves in that direction.

[09:04] Beatrice: Yeah, I mean, it's really exciting that you guys are doing this — I think the public is probably dying to give some opinions, to some extent. But let's talk about the Science 2030 thing, because basically you guys did a role-playing game to think about what science could look like in 2030, and maybe people aren't familiar with why that's a good idea to do a role-playing game. So maybe you guys could explain — why did you choose science as the theme, and why did you choose to do this game?

[09:35] Ankur: Yeah, absolutely. So maybe I can talk a little bit about why science, why games, and then talk a little bit about what Science 2030 specifically is. I think, why science — like we touched on it earlier — science has always been at the heart of DeepMind's mission, and we have this core belief that AI can accelerate, as Demis says, science at digital speed, and unlock breakthroughs that can be transformative for wide swathes of society. So, one, we just believe in science as a good for society. I think the second is that science has many of the same characteristics as AI, right — it's a sociotechnical system. It involves both the kind of technical breakthroughs, but then also a lot of human intent about what problems we care about and why, and how we measure success against those problems. And so it's one of those areas where both the system and how it's done is quite complex, and so we thought it was ripe for doing more work thinking about the intersection of AI and science, and how futures might evolve there. And then, I guess, very quickly, games feel like a natural methodology for exploring some of those dynamics, because we do ask people to role-play and situate themselves in that future, and by doing so they have to make explicit some of these assumptions and some of these trade-offs, and that makes for a really interesting dynamic in which to simulate how people might make decisions, and why.

[11:01] Beatrice: I had the pleasure of playing — not Science 2030, but AI 2027, the role-playing game.

[11:01] Ankur: Yeah, yeah.

[11:01] Beatrice: I was surprised, on the upside, at how informative it actually was to play a role-playing game — I'd never really done it before. But it was just the intensity of it, that you're in this kind of stressful environment and you have to make decisions. I remember I was representing, like, all of the public — it was a very tough role to play. And things never... because I think when you think of a game, you think you can steer it to some extent, but then you realize that all the other actors are not doing what you expect them to do —

[11:51] Ankur: Right.

[11:51] Beatrice: — or these sorts of things. So it's basically — I believe you that it's actually really informative, especially if you get interesting stakeholders playing.

[11:51] Ankur: Yeah, and actually, maybe just on that very quickly — we do some surveys pre and post, because a few of the things that we wanted to try and achieve with the game were exactly this. Do they adjust people's outlook of what futures might look like in AI for science? Do they feel immersive and informative in ways that maybe more traditional methods of engagement, like policy briefings and roundtables, do — so how do they compare to those? And, you know, something about making meaningful relationships and building some social capital in the room. And on all those measures we see really high, positive feedback from people, for exactly some of the reasons you mentioned. And so I think it's helped really validate for us as well that this methodology can play a really important role in bringing stakeholders together to think about futures in a meaningful way.

[12:39] Beatrice: And were you able to compare it to what — roundtable policy briefings, or —

[12:44] Ankur: The majority of people say that it's more informative, and then the remainder say it's as informative, right? And so we get the —

[12:51] Zoë: Engaging.

[12:51] Ankur: Exactly — much more engaging. And I think it's for the reasons you mentioned, right — you can read a report about the future, but it's a different thing, situating yourself in that intense situation and actually getting other people narrating their decisions, and you responding to that. I think it captures a lot more of the messiness of decision-making in real life.

[13:09] Zoë: Yeah, I also think we are less interested in designing an experience where we, as DeepMind, are broadcasting our view about the future, and more interested in building a methodology that enables participants to co-create a future, and then we have a discussion about that at the end. But there's no right or wrong way to play — there's actually not even a win condition, which people really struggle with. So sorry about that. But yeah, we have all kinds of reasons for that as well. Maybe I can tell you a little bit about how the game works.

[13:41] Beatrice: Yeah, please.

[13:42] Zoë: Yeah, so we built Science 2030 with ARIA, the Advanced Research and Invention Agency, and also Technology Strategy Roleplay, which is a charity organization that builds this kind of role-playing strategy game. And the way it works is we bring in roughly 50 external stakeholders from government, industry, and the science community. We put them on teams associated with each of those types of stakeholders — so there are three government teams, modeled on middle-power countries; we have three industry teams, modeled on a range of industry players; and then five science-community teams. And I think, in the game, we wanted participants to grapple with three main types of trade-offs, which we've codified into three core mechanics in the game. The first of those, Ankur mentioned briefly earlier — they're called concerns. These are huge, systemic challenges that governments tend to face — things like long healthcare waiting lines, climate issues, cyber-related issues. And the reason we codified these in the game is that we think governments often invest in technology as a way to address some of these types of concerns. So concerns is the first trade-off. The second one is how to build a research-and-development ecosystem in order to invest in those types of technologies, and within that there are two different types of trade-offs that teams have to make. The first is around the type of R&D to invest in — so everything from data, to compute, to adoption policy — and then they also have to make a decision about what strategic approach they'd like to pursue: do they want to pursue a centralized approach, or a decentralized approach? And each of those choices has trade-offs — for example, if you pursue a decentralized strategy, it's kind of like let a thousand flowers bloom, but you also might face some risk there. So R&D capacity is the second, and then the third are those AI-for-science technologies — so each turn, five new possible AI-for-science technologies come online, and governments have to select one of those five in the hopes of using that technology to solve a concern that they have. And, as I said earlier, there's no right answer.

[16:01] Ankur: We also do some work to localize that to some of the different places that we go. So, the government teams featured in the game are middle powers, but the exact makeup of those middle powers will differ. So when we were in Singapore, it was Singapore, Japan, and the UAE that were the three countries we modeled. Here in the UK, we did the UK, Germany, and Canada. And so we're doing different groupings, and that reveals interesting dynamics around coalitions and competition as well.

[16:31] Beatrice: Why middle powers?

[16:34] Ankur: Yeah, I think it's because a lot of the ongoing work focuses on what you can do with frontier capacity, and I think, obviously, a lot of that development is grounded in the US and China. And what we made an intentional choice about, at the start of this game, was that if it was about thinking about agency and what more you could be doing to invest in AI for science, science is actually a strength for a lot of these different middle powers, right — so the UK, life sciences. And I think a lot of the time their decision-making gets drowned out by what you should do if you're one of these AI superpowers instead. And so we saw a lot of room for development and thinking about what specifically middle powers could do, if you treated US and China actions as exogenous to their decision-making — and especially because we think a lot of the benefit to middle powers can come from identifying the right coalitions to build, to unlock some of these benefits and manage the risks collaboratively as well.

[17:37] Beatrice: Is there anything else on the game design that we should touch on before we go into discussing results?

[17:42] Ankur: Maybe just one quick thing on the technologies — the breadth of what we mean by AI for science is actually kind of huge, right? A lot of people instinctively go for things, understandably, in areas like the natural sciences — what will this do? But actually, for example, in the game, one of the technologies we have is more fine-grained measures of inflation and GDP, because you can do AI simulations for economic measures. So I think one is expanding people's definition of what we actually mean by AI for science — that was one really interesting thing. And then the other is that, to Zoë's point earlier, it forces people to make trade-offs even within areas that they think they place equal weight on, right? For me, a really interesting choice — I said, you know, I'd grown up with a medicine background, so instinctively, investing in health outcomes has been my priority. But actually, as we played this game a few times, investing in verified computing infrastructure and security, especially with some of the cyber concerns that we see emerging, feels actually maybe an area that you might wanna prioritize even ahead of some areas of medicine.

[18:50] Zoë: Feels like a real win for me.

[18:51] Ankur: Yeah, exactly — and with Zoë's background, something she's advocated for for so long. But that was just an interesting insight to me — it's not one that I thought I would have had before.

[19:00] Zoë: And one thing that I didn't anticipate as a learning, which seems, in hindsight, extremely obvious to me, is that different domains of science have different kinds of real-world constraints. So even if you put a huge amount of resource into accelerating climate science, for example, the climate is an enormous system that will take a very, very long time to update, regardless of what you do. Whereas in areas like cyber, the same amount of input in terms of resources might lend you a more immediate upswing of capabilities in that space. So I think it was also interesting to figure out what we could make into a game mechanic that could be influenced by participants, versus what are the real-world constraints that we had to maintain in order to make the game as realistic as possible.

[19:49] Beatrice: So maybe we can dive into what have been the findings — like, what has come out when you've played this game?

[19:59] Ankur: Yeah, I think we touched on some of them earlier — things like people updating their outlook on AI for science. I think a large part of it for us is, we think a critical ingredient to building some shared readiness for advancing AI capabilities is building some shared social capital as well, and by that we mean trust and relationships and shared context, and collective capacity to respond to quite dramatic shifts in capability, or diffusion of those capabilities. And so one thing we like to measure in the game is how many new relationships did people make — how many new collaboration opportunities surfaced. So we've been really excited to see, for example, ARIA spinning up new streams of work with other partners who they've met at these events. Similarly, we have DeepMind researchers there, and they're thinking about areas like the future of mathematics, and might suddenly meet people in other disciplines, and so think about more of that interdisciplinary collaboration. So that's certainly one element of it. And then I think it's been interesting to us — different countries have different incentives for investing in things. So in the UK you see a lot more investment into medicine and bio; in Singapore, people were going much harder into smart cities and materials. And so it's just been interesting to learn what drives people's decisions and what drives people's prioritization as well.

[21:24] Zoë: Yeah, I think, on that point, one of the big learnings — or maybe it's more of a confirmation of a thing we already thought was probably true — is the importance of the narrative around investment in science. I think we've had a lot of conversations about the statistics around the decrease in public trust in science, especially since the pandemic. And I think what's been confirmed through the game is that it's very difficult for governments to have a narrative that works with the public about why they might be investing in foundational or basic scientific research, if it isn't closely linked to one of those concerns that we're talking about — because it's very expensive to do science, especially to do AI-for-science applications. And so, even with something like AlphaFold, for example, it's one thing to try to explain to somebody what the protein-folding problem is, or why they should care about that at all — but if you can link it to the possibility of new medicines or climate resilience, then there's much greater appetite for that type of investment. I think we also saw that, for governments, so much of the prioritization of investment is the result of public opinion, and public opinion changes all the time. And so it's really tricky, even within the game, for governments to make long-term, sustained investment into any area of science, because there are constantly new concerns coming up, and all of those concerns are very valid and very tricky problems that affect real people. And so it's just very difficult to have that kind of sustained investment, which makes those types of breakthroughs really, really difficult.

[23:07] Ankur: And something within that is just how important science communication becomes. So one example there is — early on, in some of the games that we were developing, one of the technologies we featured was animal-language translation. I don't know if you've heard of projects like using AI for decoding whale noises, or — DeepMind had a project called DolphinGemma, where you use these open-source models in the water to decode animals' signals. And so what you find is, in the messiness of the real world, there's a lot of information asymmetries, and there's a lot of misunderstandings of technical capabilities. And so a lot of the time you might be able to do something that's beneficial for bioscience and environmental science, in terms of understanding animal populations, but the way that gets translated in the public's view is like, "Oh, we can now buy collars that will let me understand my pet, my dogs, and let me talk to them."

[24:05] Zoë: I have a very small build on that, which I think is — for most people, when they think about automation and science, they assume that the role of human scientists will go down as AI becomes more and more powerful. And I think, to Ankur's point about science communication, one of the prevailing themes from all the games is the criticality of scientists who understand the technologies and who can connect those technologies to the concerns that government faces, in order to drive investment. I know that Pushmeet, our VP of science, said at some point in a talk — I loved it, and I'm gonna misquote him here — but that humans are the only kind of entities who know what humans care about, at the end of the day. And so you can point AI in any direction that you would like, but at the end of the day, the scientists who have an understanding of the science, working with government and industry folks who have great knowledge about the concerns of the population, and the technology to help out — that's a uniquely human task.

[25:07] Beatrice: Yeah, so before this interview you guys shared this post, Science 2030, but I think that was when you had only done maybe one game or so, in the beginning. And there, one of the key findings that you mentioned was that scientists felt a bit powerless. Is this — yeah, what was your —

[25:23] Ankur: Yeah, I think one thing that we capture in the game is the way in which the scientific production function might change, and what are the changing inputs that go into that. And I think, to Zoë's point, I think a lot of us index in terms of our fears on the way things currently work, and it's difficult to imagine what our role or purpose might be in a system that therefore looks quite different. The way I think about it is — if I was to forecast what my purpose would be if I had been a hunter-gatherer, and my purpose at that time was just gathering food —

[25:56] Zoë: Gatherer.

[25:57] Ankur: — but I wouldn't be able to forecast, in some way, where I would find meaning if you took that purpose away from me, right? And I think it's really interesting, therefore, when you present the way in which science is gonna be done changes, that can trigger this response of, "Oh, therefore scientists themselves don't have agency." Whereas I think, instead, it's about what that agency looks like, and what that role can be, that changes, and what the purpose of that role is. And I think some of those areas, it's very difficult to forecast, because the change is happening so quickly, and it's unclear what the collective intent, I guess, behind that area is. But in other areas I do think it's very clear where the role of humans is, and that's — the "why" is still gonna be a very human endeavor, right? Like, why do we care about these things, what measures do we have for whether we're doing these things well — I think are inherently human questions that require human value judgments and subjectivity. I also think, maybe the last interesting thing I'll say on it is, it differs by domain, right — if you look at something like AlphaFold, it's actually democratized access to this capability that previously required hundreds of thousands of dollars, if not more — like years of a PhD, really expensive equipment to do this. And now you have three million-plus researchers around the world, a lot of them in places like Asia, where some of those resources might not have been as readily available, now using those tools and using them to do all manner of things. I think that's maybe different than in an area like mathematics, where the change has come quite quickly and, all of a sudden, the role of a mathematician has changed quite dramatically because of what can be done end-to-end by the AI system. So I think that's partly very domain-dependent as well.

[27:40] Beatrice: And one thing that I wanna ask you guys is — you started doing this within an AI lab, Google DeepMind. I know that there are other external efforts, like RAND, who has done war games, as they're called. What do you think is the specific function of you guys running it within a lab?

[27:40] Ankur: Yeah, I think maybe it's partly what you touched on earlier — as a lab, I think we just see a different slice of the picture than other people, and that means that we are closer to the development of the capabilities, some of the kind of evaluations around those capabilities. And so I think that's a really important part of the picture that we try to therefore externalize as part of the methodologies that we develop. We try to feature a set of lightning talks around each event, from people like ARIA researchers, DeepMind researchers, to give a little glimpse into what the future of their research looks like. I think what we also then try to do is recognize that, as a lab, there are all these gaps — areas where we're not as familiar with where things will go. And so, how can we maybe lend a bit of our convening power to bringing people together from all these different places?

[28:56] Beatrice: Yeah, do you have a best-case scenario in mind for when you think about AI for science?

[29:03] Zoë: You first.

[29:04] Ankur: You go.

[29:05] Beatrice: Feel free to dream.

[29:06] Zoë: I mean, we've talked about this a little bit — we've talked a lot about abundance as a concept, and access to essentially basic needs would be my, as the more pessimistic of the two of us, happy outcome.

[29:22] Ankur: Yeah, I think, at minimum, expanded access, participation, and abundance of resources themselves. I think, perhaps, my most optimistic vision is a world in which anyone can ask more questions of the world around them and achieve a deeper understanding of that world, and, with that understanding, pursue their own vision of what good might look like. And so that's everything from understanding our own bodies, our minds, and being able to do more with those, in ways that I think feel very personal and subjective, each to their own — as well as in these areas like nature and climate, right? And being able to have — I think if we get some of this right, you can actually have a more human relationship with those things, because we can understand them better. We can situate ourselves within those ecosystems better, understand our impacts on them. And so I would hope for a world in which we can just be a little bit more intentional and a little bit more thoughtful about how we exist, and why. And, in doing so, maybe turn that on ourselves a little bit, right? I think there's so much about ourselves that we just don't understand — people talk about AI being black boxes, but other people are, and I think — don't look at me — we have to work together more. So I just think I'm most excited in areas like social science and behavior as well, that maybe we can achieve better forms of deliberation and working together.

[30:48] Beatrice: Yeah, so I think last question — if someone young is listening to this, and they're feeling like they wanna try to have an impact on how AI develops, do you have any advice for them?

[31:00] Zoë: Yeah, so I think, as we started the conversation with, Ankur and I have had very, very different paths into this space, and I think my first piece of advice is just that you don't have to work at a leading lab in order to make a difference in AI. I think we are surer and surer of the fact that AI is a whole-of-society effort — there's plenty of work to be done in government, there's plenty of work to be done in civil society, in academia. If what you want is to work at a lab, that's also great, but I think there are tons of ways into the space. I think what I said earlier about five-year plans — I think it's really important to be opportunistic, to do things that interest you, to not wait for the perfect job. I think I did a lot of work, even before I started my career, on little side projects that were mostly interesting just because I felt like it, and I think that's a great thing to do — just learn about things you're interested in, do the work, and just say yes to things.

[32:05] Ankur: No, I think that's great advice. I think, similarly, I've spent a long time feeling — you know, when you work at something like a lab, you constantly feel this sense of imposter syndrome, and I think that's partly a good thing, something that I enjoy, because you're always learning something. But at the same time, I think it's important not to feel intimidated by expertise, as if it means there's nothing for you to offer. And I think we're in a phase right now, especially, where the future is so uncertain and the range of possibilities are so vast, that actually it is for each of us to ask questions that feel salient and important to us, because those questions are, in some ways, more important than the answers right now, where we don't necessarily have all of them, and we'll need to work those out together. So I think feeling like you can contribute, even by just thinking about what the gaps might be, what questions other people aren't considering. I think, secondly, to Zoë's point, there's so much alpha right now in just looking at some of those questions yourself and trying to become more of an expert in them, even with all these offline resources.

[33:15] Zoë: Yeah, there's no such thing as AI expertise, just to be very clear.

[33:18] Ankur: And you can achieve a depth of expertise that is astounding, relative to most people, if you just find something that interests you as your niche and go deep in it — even in conversation, as Zoë said, with some of these AI agents, and then doing your own desk research. So I think you can add a lot more to the conversation than people might assume right now. And then the third thing, I think, is just being a bit more experimental — I think we have tried to just throw a lot of things at the wall and seen what sticks.

[33:45] Zoë: Yeah, you can imagine how pitching a game went initially.

[33:49] Ankur: Yeah, exactly, right — we...

[33:50] Zoë: We were like, "So we have a crazy idea."

[33:53] Ankur: We wanna take seven hours of people's time and do an AI war game on AI for science futures.

[33:58] Zoë: We're not gonna have a game for a year — but bear with us, it's gonna be great.

[34:02] Ankur: And so, I do think, though, that it is possible to have much quicker feedback loops now, and to try things out and learn from them very quickly. And so, yeah, just get out there and do things. It's a trite saying, but you can just do things, and I think that's more true than ever — so just try it.

[34:18] Zoë: And also, it's always the right thing to be kind.

[34:21] Ankur: Yeah.

[34:22] Beatrice: I agree with that.

[34:23] Ankur: I agree with that one very much.

[34:24] Beatrice: I agree.

[34:24] Zoë: It's free, and it always works, and it's always the right thing to do.

[34:27] Ankur: Yeah.

[34:28] Beatrice: I agree. Yeah, I think that's a very nice note to end on. Let's be kind and go out there. Thank you so much for joining, thank you.

[34:36] Ankur: Thank you.

[34:36] Beatrice: This was so nice.

[34:36] Zoë: Yay.

‍