Google DeepMind CEO Says Scaling Alone Will Fail to Yield AGI

Google DeepMind首席执行官称仅靠扩展无法实现AGI Demis Hassabis认为，达到人类水平的通用人工智能需要在规划和符号逻辑方面取得数项明确的架构突破，而不仅仅是扩充算力和数据规模。 Demis Hassabis回顾了DeepMind的发展历程：从2010年带着20年路线图秘密起步的初创公司，成长为推动Project Astra、AlphaFold 3和AlphaGeometry等突破性成果的全球AI重镇。他指出，仅靠模型扩展不足以实现真正的AGI，并强调了规划、行动以及将神经网络与符号推理相结合的混合架构的必要性。Hassabis还反思了英国科技生态系统的演变、Alan Turing留下的遗产，以及在AI安全上采取“大胆且负责任”的态度而非盲目进行商业加速的至关重要性。 DeepMind成立于2010年，制定了打造AI领域“阿波罗计划”的20年路线图，将深度学习、强化学习和神经科学相结合。 (0:12) Project Astra代表了DeepMind对具备广阔上下文窗口的原生多模态、实时通用AI智能体的愿景。 (0:43) 真正的规划与行动执行仍是未解决的挑战，无法仅通过扩展next-token预测来解决。 (7:22) AlphaFold 3在AlphaFold 2蛋白质结构预测的基础上进一步扩展，可对蛋白质、DNA、RNA和化合物之间的动态相互作用进行建模。 (10:30) Isomorphic Labs旨在利用AI实现药物发现的虚拟化，将时间周期从数年缩短至数月。 (10:30)

Google DeepMind CEO Demis Hassabis argues that reaching artificial general intelligence requires bridging large multimodal models with specialized planning and scientific engines, cautioning that scale alone cannot solve critical reasoning gaps. The Limits of Scaling and Missing Planning Capabilities (7:22) While acknowledging that scaling compute and datasets is necessary for advanced AI, Hassabis contends that scale alone is insufficient to achieve human-level intelligence. Current large language models excel at next-token prediction but struggle with planning, which requires decomposing abstract goals into interim steps, executing tree search over possible trajectories, and maintaining accurate representations of state. This architectural shortfall is evident in chess, where models hallucinate illegal moves once pushed off standard opening books because they fail to track the board. Multimodal Assistants and Hybrid Scientific Architectures (0:43) DeepMind's near-term development focuses on multimodal agents and hybrid AI systems. Prototypes like Project Astra utilize Gemini's context window—ranging from 1 to 2 million tokens—to maintain ambient memory over live video and audio, enabling spatial awareness and practical assistance like finding misplaced items. In complex reasoning domains, Hassabis expects general-purpose models to operate as controllers that query specialized domain systems, mirroring DeepMind's AlphaGeometry, which combines a neural language model with a symbolic deduction engine to solve Olympiad geometry problems. Transforming Biology Through Computational Modeling (10:30) Applying AI directly to fundamental science remains a primary focus for DeepMind. Following AlphaFold 2's prediction of 200 million protein structures within one year, AlphaFold 3 expands into dynamic biological interactions involving proteins, DNA, RNA, and chemical ligands. Hassabis envisions modeling an entire virtual cell computationally to simulate experiments before wet-lab validation, while DeepMind spin-out Isomorphic Labs uses generative AI to accelerate targeted drug discovery timelines from years down to months. Geopolitics, Governance, and Defensive Safety Research (25:10) Hassabis links technical leadership directly to international influence, noting that domestic technological capability gives nations the standing required to establish safety baselines, ethical norms, and diplomatic engagement. DeepMind rejects rapid commercial iteration in favor of a proactive safety philosophy that combines stringent capability evaluations, neuroscience-inspired interpretability tools, and secure sandboxing to prevent containment failures. Hassabis warns against uncritical accelerationism driven by recent industry arrivals, maintaining that societal benefits in climate and medicine will only materialize over the next decade if foundational safety benchmarks and structural reasoning breakthroughs are solved first.

Google DeepMind 首席执行官 Demis Hassabis 指出，实现通用人工智能（AGI）不能单凭算力与数据的规模扩展，必须攻克规划与行动等核心难题。他主张将原生多模态与专用科学系统相结合，并以严谨、负责的科学方法替代盲目的商业狂热，确保前沿技术在受控环境中造福社会。 规划能力为核心瓶颈，规模扩展难达 AGI (7:22) 针对业内“规模即一切”（scale is all you need）的论调，Hassabis 指出，增加算力与数据虽然必不可少，但单纯扩大模型规模无法自发产生复杂的规划能力。现存模型虽具备初步推理能力，但无法将高阶抽象目标拆解为可执行的步骤。规划在本质上是与“下一个 Token 预测”截然不同的计算范式，需要类似 AlphaGo 的轨迹树搜索算法以及在“心智中进行三维想象与推演”的评估能力。 这一缺陷在棋类任务中尤为显著：大型语言模型因缺乏对棋盘状态的稳定内部表征，脱离预设开局谱后便会失去跟踪能力，甚至产生吃掉己方棋子的非法走子“幻觉”。Hassabis 评估，实现人类水平智能仍需数项关键算法突破，核心在于将 AlphaGo 式的规划技术与语言模型的语义及多模态理解深度融合。 原生多模态与长上下文赋能通用智能助手 (0:43) 在产品与架构演进方面，Hassabis 将 Project Astra 视为实现日常“通用 AI 代理”的原型。他强调，传统纯文本模型缺乏对物理环境与空间维度的感知，而 Google 自底层将 Gemini 设计为原生多模态架构，能够同步处理视觉、音频等感官输入。 此外，Gemini 1.5 引入的 100 万至 200 万 token 超长上下文窗口，使得模型能够一次性处理一小时的视频或海量文本。在 Astra 的实时视频流交互中，这一长窗口赋予系统“常驻环境记忆”，使其能够被动记住长久之前发生的生活细节，例如协助用户在室内定位遗忘的钥匙或药物。Hassabis 认为，这种长程多模态理解能力是智能代理胜任日常协作与科研辅助的关键基石。 科学 AI 突破：从蛋白质折叠到虚拟细胞 (10:30) 在“AI for Science”领域，DeepMind 正在推动从静态结构预测向量子级生物动态交互演进。AlphaFold 2 曾解决困扰学界 50 年的蛋白质折叠问题，在一年内预测出 2 亿种已知蛋白质的结构，节省了约 10 亿年的人力研究时间。最新发布的 AlphaFold 3 则进一步模拟蛋白质与 DNA、RNA 及化学配体的相互作用，Hassabis 的终极目标是在计算机内高保真模拟整个“虚拟细胞”，使实验在进入物理湿实验室前即可在数字环境中完成。衍生公司 Isomorphic Labs 正致力于运用这些技术将新药研发周期缩短至数月。 Hassabis 同时强调，AI 的未来并非大语言模型独占。AlphaGeometry 通过结合神经网络与符号逻辑推理，在国际数学奥林匹克中达到接近金牌水平。他预测未来 AI 架构将趋于“混合范式”：由通用模型担任控制器，根据任务调用擅长数学、代码或生化计算的专用系统。 二十年路线图、人才套利与国家软实力 (0:12) DeepMind 创立于 2010 年，Hassabis 与联合创始人 Shane Legg 设定的二十年技术路线图在深度科技领域罕见地按期推进。尽管早期在英国募集资本十分艰难，只能依靠极少数美国创业者的天使投资，但 Hassabis 坚持将总部留在伦敦以实施“人才套利”战略——吸纳顶尖高校中除投行外缺乏前沿科研去处的数理与计算人才，构建起高壁垒的技术团队。 结合在英国高级研究与发明局（ARIA）的任职经历，Hassabis 与主持人强调，保持科技前沿地位不仅关乎经济利益，更是掌握全球技术伦理、安全标准和国际话语权的前提。例如在筹备首届英国 AI 安全峰会时，中国科技界对英国计算先驱 Alan Turing 的崇敬成为双边沟通的重要纽带，印证了核心科技积淀所能转化出的深层国家软实力。 “大胆且负责”的治理哲学与防御性技术 (29:30) 针对技术安全与风险治理，Hassabis 反对硅谷传统的“快速行动、打破常规”信条，倡导“大胆且负责”（Bold and Responsible）的科学理念。鉴于创业初期便确信 AGI 将在数十年内成为现实，DeepMind 将安全准则早早写入公司章程，后续演化为 Google 的公开 AI 原则。Hassabis 提出亟需优先发展的防御性研究方向：建立能敏锐捕捉危险突现能力的评测体系、开发用于解析神经网络内部表征机制的“虚拟大脑分析”工具，以及构建防止外部渗透与模型失控逃逸的高安全沙箱环境。 Hassabis 批评了近期从加密货币等风口迅速转向 AI、自封为“加速主义者”的盲目投机群体，指出其低估了该技术的严肃性。他坚称乐观必须建立在严谨审慎的前提之上；唯有通过扎实的安全防御研发，社会才能在未来十年内真正迎来 AI 赋能的繁荣局面。 Hassabis 的核心判断表明，人工智能向人类水平智能的迈进是一场严肃的长期科学探索，而非简单的工程参数竞赛。尽管当前模型在多模态交互与特定科学计算上展现出巨大潜力，但在长程规划、状态表征与底层可解释性上仍面临本质挑战；唯有坚持科学审慎与主动安全防御，才能确保技术的长远演进不偏离人类福祉。

[0:12] Demis Hassabis: He would have just definitely, at best laughed us out the room, at worst, sort of thought, we'd sold out. [0:12] Demis Hassabis: Apollo program effort really for AI. [17:55] Demis Hassabis: It always surprises me that the language models are terrible at playing even a simple game of chess. That's really annoying actually, and needs to be fixed. [25:10] Demis Hassabis: If you're not at the forefront of the technology, then you sort of lose your moral authority to lead and influence. [25:10] Demis Hassabis: We could have had Silicon Valley in Manchester... Turing should have been at the head of something—mega startup. [35:28] Demis Hassabis: in 10 years' time we'll be in an amazing flourishing society powered by AI.

[0:12] Welcome everyone. [0:14] I am really excited to be doing this with Sir Demis, [0:18] who as you may know, recently got his knighthood. [0:20] Have you gone to be dubbed yet by the way? [0:22] -No yet. Some time this summer. -Okay, that's still to come. [0:27] There is so much that I want to talk about [0:28] and it's really great that we're doing this here in London, [0:31] where you grew up, where you founded DeepMind [0:34] and where you are now CEO of Google DeepMind, [0:37] running Google's AI efforts around the world. [0:39] Maybe let's start with that. [0:40] So obviously, it must be quite a weird experience for you, [0:45] having started DeepMind in 2010 when, to put it mildly, AI was not obvious. [0:50] There were not international summits on AI. [0:54] There were not trillions of dollars pouring into it. [0:57] And now all those things are true. [1:01] When you look back on the last, [1:03] I guess it's now 14 years since you started it, [1:05] what's been most surprising both about what has worked and what hasn't? [1:11] -It's great to be here and thanks everyone for coming. [1:16] It's great to be here in the heart of London [1:19] thinking about these kinds of things. [1:20] And you're right, when we started 15 years ago, [1:24] virtually no one was talking about AI. [1:27] In fact, certainly no one in industry and even in academia. [1:30] I remember Shane Legg and myself, [1:33] we were both postdocs at Gatsby at UCL at the time. [1:36] We couldn't really tell our professor because [1:39] he would have just definitely, at best laughed us out the room, [1:42] at worst, sort of thought, we'd sold out and gone away from research. [1:47] And what I was trying to do was show that cutting edge research [1:51] could be done in a company... deep research. [1:55] And also we felt that there were enough new ideas in AI [2:00] that we could make very fast progress. [2:02] Kind of like an Apollo program effort really for AI. [2:06] And I'd been thinking about it for at least 15/20 years, even before that, [2:11] but I felt that around 2010 I could see we had the right ingredients. [2:18] Deep learning, or they're called deep belief networks-- [2:21] actually, Geoff Hinton and co had just invented [2:24] some foundational papers in 2006/2007, [2:28] but it hadn't been really widely adopted yet. [2:30] GPU's were becoming widely available [2:34] and we had some understanding of neuroscience [2:37] and how the brain's algorithms worked. [2:39] And bringing that all together was what made us feel like [2:42] really fast progress could be made. [2:45] -What was your starting point back then? Because today there's lots of people, [2:49] not that many people because it's so expensive, [2:51] but there's some people that are like, "We're going to build an LLM", [2:53] "We're going to raise..." [2:54] But that wasn't on the table then. Where did you start? [2:57] -Yeah, we were just trying to scrape together, frankly, our first seed round. [3:03] It was almost impossible. [3:06] I didn't pay myself a salary for the first couple of years, [3:09] things like that. [3:11] But we knew this was our life's work and certainly been my life's work. [3:16] And I felt that finally, maybe the time had come. [3:20] So it was so exciting to try and get that going [3:22] and we managed to find enough people to just get the seed round done [3:26] and actually just start making progress towards this overall goal, [3:29] which seemed monumental at the time. [3:31] We had a kind of 20-year roadmap. [3:33] Everyone has a 20-year roadmap for a deep tech company. [3:36] The interesting thing is usually the joke is like, [3:39] it stays 20 years away as the years progress, [3:42] but actually, we're more or less on track, unbelievably. [3:46] Perhaps that's the most surprising thing [3:47] is that all of these hypotheses we have, I guess, came out to be true. [3:52] So certainly on the research side and the algorithm side, [3:56] betting on learning and generality very early on. [4:00] Things like reinforcement learning and deep learning. [4:03] Obviously the core of our very first work, in our Atari work, [4:06] and we sort of invented the field of deep reinforcement learning [4:11] turned out to be the right approach. [4:15] And then we also knew that if it did work [4:18] it should become the biggest thing ever [4:20] because you should be able to apply this general intelligence [4:24] to almost any field of human endeavor. [4:28] And I think what's happened in the last couple of years, [4:30] is everyone has realized that now, [4:33] what we knew maybe 20 plus years ago. [4:37] -So, a lot of people will have seen or heard about [4:40] some of the stuff you demoed last week at Google I/O, [4:44] but from your perspective, what were the highlights? [4:46] What is AI capable of today? [4:49] -Well, it was fantastic event and it was very exciting. [4:54] For me, the highlight was probably this thing called Project Astra that we showed, [4:59] which is our vision for what I'm calling a kind of universal assistant [5:05] or universal AI agent could be like, [5:08] and how it could help in your everyday life. [5:10] And that the main key thing about it is it's multi modality. [5:15] So, its ability to understand all the different modalities we operate in [5:19] as human beings in the context we're in. [5:22] So I think that was what's missing from the language agencies. [5:24] They didn't understand the spatial context you were in, [5:27] and the environment you are in. [5:29] So that limits their usefulness in the end. [5:32] And we always had this vision of something that could understand [5:37] the world around you through vision, [5:39] eventually audio and all the other senses as well. [5:42] And that's why we built Gemini, our large series of models, [5:47] to be multimodal from the beginning. [5:49] Able to cope with any kind of input [5:51] and eventually produce any kind of output. [5:54] That's the vision that we have for maybe the next year or two [5:59] of having truly smart assistants, [6:02] which I think will be a pretty big game changer. [6:06] -Yeah. If you haven't seen the Astra demo video, [6:08] I do encourage you to watch it. [6:10] I don't want to spoil it, but there's a bit right at the end [6:13] which I think shows that these systems-- shows the power of [6:17] actually having a very long context stock on the video side. [6:20] -That's right. [6:21] That was one of the innovations we came out with with Gemini 1.5 series [6:25] is this million token long context. [6:28] Actually, we extended it to 2 million tokens now [6:31] and maybe we can go a lot further than that. [6:33] And that means you can now ingest, [6:37] all of War and Peace or even an hour-long video [6:42] and then ask questions of the system about it [6:45] and it can answer back intelligently. [6:46] And that opens up whole new possibilities [6:49] as an analysis tool or a sort of research assistant. [6:53] But also then with Astra, [6:56] streaming live video obviously [6:59] and then doing things like remembering. [7:01] I think the thing you're referring to is [7:03] remembering something that happened a long time ago, [7:05] ambiently in the video stream. [7:07] And it could be super useful, [7:09] for an elderly person to remember where their medicine is [7:13] or where your car keys are or something like that. [7:16] So, I just think we've barely scratched the surface [7:19] of what I can imagine these things will be useful. [7:22] -So, one thing I think a lot of people will be interested to get your take on is, [7:26] you already mentioned that over the next couple years [7:28] we can expect to have things like more powerful personal assistants, [7:32] artificially intelligent assistants. [7:36] How do you think about-- [7:37] what are the big remaining unsolved challenges [7:41] that we might see progress in over the next couple of years? [7:43] -Yeah, I mean, it's pretty clear to me, and I've said this publicly a few times [7:46] that what's missing from the current systems-- [7:49] so, obviously multimodality, memory, [7:51] these are things that we're working on and making good progress, [7:54] but it's missing planning and actions. [7:58] And it has some form of reasoning, but not really. [8:00] -What does planning mean in this context? [8:02] -Well, sort of the ability to break down, [8:05] the ability to achieve a goal that you might set a system. [8:08] So, it has to be able to act in the world, take actions. [8:11] So, more like the game agents we're famous for, we built in the past. [8:16] I mean, like AlphaGo, there's some objective that you give it as the user. [8:20] And then obviously in games, it's normally to win the game [8:23] or to maximize the points. [8:25] And then it's got to, if the objective is pretty complicated or abstract, [8:31] it has to have the ability to break that down into a set of steps [8:35] that are achievable, set of interim goals, interim plans, [8:39] and then that to build up overall into this bigger plan [8:43] and that involves needing to do things like, [8:46] maybe tree search over a planning trajectory, [8:50] perhaps imagine in 3D in the mind's eye, [8:56] like what might happen in a certain scenario in order to evaluate it. [8:59] So, it implies a lot of extra capabilities [9:01] that the current systems don't have. [9:04] And I think what we need, and I think the next big advance will be, [9:08] combining the sorts of things we did in planning [9:10] for programs like AlphaGo. [9:13] But instead of it just being limited to a game environment, [9:17] obviously instead, you connect that to all of language [9:21] and perhaps all of multimodal context around you and that model, [9:26] and then do planning over that. [9:29] I think that's the next big advance we're going to see. [9:32] -Well, this is quite important, isn't it? [9:33] Because I think clearly one of the reasons that, [9:37] particularly large language and now multimodal models, [9:39] have got so good over the last few years, [9:41] is just the enormous amounts of compute that are being invested [9:44] in training and also in inference. [9:47] And so some people like to say, "Scale is all you need." [9:50] But it sounds like from what you're saying there [9:51] that actually there are some new ideas [9:53] that won't just come from scaling these. [9:55] -I think that's right. [9:56] I mean, look, I think scaling is definitely required [10:00] and these large models are going to be part of an overall solution to AGI. [10:04] So, I think there's not much doubt there. [10:07] Now, the question is, is scaling enough? [10:10] And it could be that some of these properties [10:12] that we're talking about might just emerge, [10:15] but I find it hard-- like some of these other things we're discussing, [10:18] like planning, feel like slightly different in kind computations, [10:21] So, my guess is that there's still maybe a small handful of breakthroughs [10:26] still required to get us all the way to human-level AI. [10:30] -Maybe that's a good segue into talking about [10:32] a very different sort of paradigm of AI, if you like, [10:34] from the large multimodal models, which is AI for science, [10:38] which I think Google DeepMind probably has a unique strength in. [10:44] So, a couple of weeks ago-- seems like forever in AI-- [10:47] it feels like in AI every week there's a generation of progress [10:51] but two or three weeks ago, you released AlphaFold 3, [10:55] which for a lot of scientists that I know, they describe as revolutionary. [11:00] Can you describe what it is for people that don't know biology? [11:02] -Yeah. So I mean, maybe we start with AlphaFold 2, [11:05] which was the program we released a few years ago now, [11:08] which essentially solved one of the grand challenges in biology, [11:13] 50-year grand challenge, called protein folding. [11:16] And essentially what it is, is a protein [11:18] they're the workhorses of biology. [11:20] Everything in your body, everything in biology, [11:22] depends on proteins for their function. [11:25] And a protein is described by its amino acid sequence. [11:30] So, you can think of it roughly as the genetic sequence for the protein. [11:33] And the protein folding problem is can you predict the 3D shape [11:38] that that protein, that sort of string of beads, if you like, [11:41] folds up into in the body just from the genetic sequence, [11:45] the amino acid sequence. [11:47] And the reason that's important is the 3D shape [11:51] tells you a lot about the function that that protein does in the body. [11:56] So, if you want to understand disease [11:58] or you want to design a drug to target something, [12:01] you need to understand the 3D shape of the protein. [12:04] And then usually in the past, experimentally, [12:07] it's been extraordinarily difficult to get that experimentally. [12:11] It takes like five years. [12:12] The rule of thumb is it takes one PhD student [12:14] their entire PhD to do one structure. [12:17] But there's 20,000 proteins in humans. [12:21] And then there's 200 million known to science in all of nature. [12:25] So, it would just take forever to do it like that. [12:27] So, we want a computational solution to it. [12:30] And that's what AlphaFold 2 was. [12:31] And it sort of essentially solved that problem. [12:34] And then we spent a year folding all 200 million proteins. [12:37] So, that would have taken the equivalent of a billion years of PhD time, [12:42] which is incredible done in one year. [12:45] So, that's the power of computational methods. [12:47] And then AlphaFold 3, which we just released, as you said, [12:50] last couple of weeks takes that one step further now. [12:53] And biology is not a static system, right. [12:56] It's a dynamic emerging system. [13:00] And so to really understand biology, you want to understand the dynamics. [13:03] So in effect, AlphaFold 2 is the static picture of a protein. [13:06] And then AlphaFold 3 now allows you to understand [13:10] how proteins interact with other biomolecules. [13:14] So of course, other proteins, but also strands of DNA, RNA, [13:18] and also things like ligands, chemical compounds effectively [13:22] useful for drug discovery, [13:24] So, it's sort of another huge leap. [13:27] And where we're going with this is, eventually, [13:30] I would like to model an entire virtual cell, [13:34] the workings of an entire virtual cell really accurately [13:37] so that one could do experiments in silico on the cell, the virtual cell. [13:41] And that would be informative [13:43] of what would actually happen in the real lab. [13:46] -And, so, what do you want to do with this? [13:48] Because I think everyone here will have obviously heard of DeepMind, [13:50] but they may be less familiar with Isomorphic Labs. [13:53] What is Isomorphic and what's your plans for it? [13:55] -Right, so we span out Isomorphic after AlphaFold 2 was done [13:57] and the idea was always to-- the reason we were interested in protein folding [14:01] is eventually to revolutionize the way we design drugs. [14:06] And the first step is understanding the surface of the protein. [14:09] The next step, which Isomorphic is doing, [14:12] is designing a drug compound that will bind to the right part of the protein [14:17] and ideally not bind to anything else [14:21] because that would be a side effect, right. [14:22] So, you kind of want to minimize the off-target binding [14:26] and so Isomorphic is moving into-- [14:28] we're improving AlphaFold all the time to make it more dynamic [14:31] but also building other modules like chemistry modules. [14:36] More going into biochemistry [14:38] which is also another perfect area for generative AI [14:42] to design molecules that will bind to the right place. [14:46] So, I think the hope is that, well, of course, [14:50] AlphaFold is being used for millions of things now, [14:52] fundamental biology and disease understanding and so on, [14:54] but also especially revolutionizing drug design and drug discovery [14:59] and taking it down from years to maybe months. [15:02] -I think what's particularly cool, [15:04] apart from the sort of impact that it can have on people's lives, [15:08] one of the things from an AI perspective [15:09] I really like about the stuff that you and your team have done [15:13] in the science side, is that it does show that [15:16] AI isn't this monolithic thing. It's not just this large language models. [15:20] One of the things that you put out I think earlier this year, [15:23] maybe it was the very end of last year, [15:24] was this paper AlphaGeometry [15:26] about solving problems from The International Mathematical Olympiad. [15:30] Which if you are not familiar with, [15:31] extraordinarily difficult mathematical problems. [15:34] And people often talked about this as one of the problems [15:37] that would be the last things before we got AGI. [15:40] But AlphaGeometry solves basically [15:43] to almost gold medal standard geometry problems. [15:45] -Yeah, no, it does. And obviously the geometry problems [15:49] are only around one third of the IMO [15:52] usually in an average Olympiad set of problems. [15:56] But we're making progress also on the non-geometry side, too. [15:59] So yeah, I think that maths is actually quite well suited [16:03] to these sorts of systems and exploring-- [16:07] You almost can think of it as moves in a game, right. [16:10] If you convert a function into another function, [16:14] you do a mathematical operation that's almost like making a move in a game. [16:16] And then you have to assess whether the new state of the equation [16:20] is making progress towards solving something. [16:25] So, I think that it's just almost limitless what we can apply it to. [16:28] And I think we will be able to be gold level soon at IMO. [16:33] And then that can probably translate to other things like reasoning, coding... [16:38] You often find with these systems is [16:40] if you build it to be better in a certain direction, [16:43] that actually improves, you can kind of port that back into the main system [16:46] and improve that. [16:48] But you're right, there's a lot more going on than [16:49] just large language models. [16:53] And I think for a long while yet, the specialized systems [16:56] will be able to do very interesting things [16:58] like AlphaFold for specific domains. [17:00] Eventually when we have AGI, [17:02] we'll have one system that can do everything to a very high level. [17:06] But for now, I think that the next few years, [17:10] I think there's going to be needs for both. [17:13] The interesting thing is that the general system can call the specialized system. [17:16] -Right. -So there's this sort of hybrid notion [17:20] that we're investigating and others are too. [17:22] Like whether the general system learns to do tool use. [17:26] And then can use the specialized systems to improve [17:29] whatever that is, it's maths ability, [17:31] it's coding ability or playing things like chess. [17:34] -Yeah. Well, I think that was one of the things I found cool [17:36] about that paper was that it combined a language model [17:39] and, kind of more traditional symbolic reasoning engine. Yeah. [17:42] Well, you talked about games a lot, so let's talk about games. [17:44] -Sure. -So you were a very serious chess player. [17:50] You then obviously very famously "beat Go" with AlphaGo. [17:55] And lots of people have seen the documentary, [17:58] I think less well-known, although it's in the public domain, [18:02] is that Demis is also a champion Diplomacy player. [18:04] I don't know if you know what Diplomacy is, [18:06] but it is a game of alliances and betrayal. [18:09] It's very open-ended. [18:10] There's no sort of like-- it's not like chess [18:14] where there's a limited number of moves. [18:17] And slightly, well, hugely to my surprise, [18:20] that is also an area where AI has made a lot of progress. [18:26] So, when you think about games, what is AI not yet able to win [18:30] and how do you think about that? [18:32] -Yeah, I think we've got-- Games have always been-- [18:35] So that was part of the founding thesis of DeepMind [18:38] was to use games as convenient testing ground. [18:43] So I think that's the issue always actually with [18:46] if you're starting on a kind of very ambitious deep tech company [18:50] and maybe many of you in the audience [18:52] are working on deep tech companies. [18:55] You have this sort of enormous 20-year ambitious goal, [18:58] but you've got to be able to break that down. [19:00] One of the key things is actually choosing the right interim milestones [19:03] to prove to obviously your backers and investors [19:06] that you're making progress on your team, [19:08] but also your team and yourself that you're actually in the right direction. [19:11] And I think that's one thing we did well with DeepMind [19:15] is pick the right kind of interim goals, [19:17] whether that's Atari game, Go, or protein folding [19:21] and games was always the thing I had in mind, [19:24] partly because of my background, obviously, [19:26] playing a lot of games and using games. [19:28] And I also training my own mind on games, [19:30] I guess you could call it, [19:32] playing chess from a very young age was very formative for me. [19:37] And I think that it has all the right attributes [19:39] to test out your AI algorithms on. [19:42] But I think we have got to the point in the last few years [19:44] where AI can pretty much play almost any game now [19:47] to very high standard. [19:48] And for us Alpha 0 was the pinnacle of that, [19:50] which was a evolution of AlphaGo. [19:53] But that could play any two-player perfect information game. [19:55] Of course there are other challenges, like Poker and Diplomacy [19:59] where there are other things that come into it. [20:01] In Diplomacy there's language, [20:03] in Poker, there's hidden information. [20:05] But even in those domains, I think the current AI systems [20:09] are getting pretty good, [20:10] although it always surprises me that the language models [20:12] are terrible at playing even a simple game of chess. [20:15] That's really annoying actually, and needs to be fixed, [20:17] and we're working on, but it's interesting that it has that issue. [20:22] It's interesting what it says about those kind of language models, right? [20:25] They don't really have a board state, [20:29] an active board state that they can maintain. [20:32] That's certainly not implicitly. [20:34] So it's interesting actually that they fail at that currently. [20:36] -Do you think that will go away with scale? [20:39] Because I remember reading that paper about Othello, [20:41] which is a much simpler game [20:42] and there is some evidence that it has some representation [20:44] of what's going on in Othello. [20:46] -It does. It does have some representation, [20:48] but the hallucinations are especially bad in something as precise as a game. [20:51] So if you've ever tried to play chess against it, [20:54] then it will-- if you, get it sort of out of the normal openings, [20:58] so it's not pattern matching anymore-- [21:00] It starts taking its own pieces and things like that. [21:02] It just forgets where things are. [21:04] So there's something clearly missing. [21:06] I think it's quite important actually, probably, what that's telling us. [21:10] So I think that's-- going back to your earlier question, [21:13] that's a big hole in the capabilities of our current systems. [21:17] I'm sure it'll get fixed, but it suggests to me something's missing. [21:20] Not just scale. Yeah. [21:24] -You touched on a topic I wanted to dive into a bit, [21:26] which is entrepreneurship, and especially entrepreneurship in the UK. [21:30] I mean, when you started DeepMind, [21:34] there were very few people actually starting companies in the UK [21:37] relative to today and certainly deep tech companies like that. [21:42] You ended up having to raise a lot of your capital from outside the UK. [21:47] What's your take on where the UK is today as an entrepreneur? [21:50] -My feeling is it's like night and day [21:54] from how it was back in 2010. It was hard to start anything, [21:58] but especially it was really hard to start a deep tech company. [22:02] Very little understanding of that in the UK investment community [22:07] and also appetite for funding something [22:09] that potentially would have that long a payoff. [22:14] And even though obviously the payoff would be enormous. [22:17] So, I did end up going to the US to find-- [22:19] actually even there it wasn't really the VCs, [22:21] it was more individual entrepreneurs [22:22] who had made it themselves and understood [22:25] what a journey like that would be like, that were backing it. [22:30] But the one reason I stayed in the UK was, [22:33] I mean, firstly, I obviously I love it here, growing up here. [22:36] But I do think we have an enormous talent here, always have. [22:39] I mean, it's like three or four of the top 10 universities in the world [22:44] are in the UK. And it's always been that way. [22:48] And so we produce amazing talent. [22:50] And then we used to think, and why I started DeepMind in London, [22:54] was that back then, if you didn't want to go and work for hedge funds, [22:58] and say you had a PhD in physics from Cambridge or something, [23:02] then there weren't that many options and you wanted to stay in the UK. [23:04] -Totally. [23:05] -So I felt there was a sort of competitive advantage [23:08] by being here and actually making the most of that talent pool [23:12] almost unopposed in the UK, and actually for most of Europe. [23:17] if you look at the big technical universities [23:19] in Switzerland, in Germany, and so on. [23:21] And it worked fantastically well for the first [23:24] five, six, seven years, I would say. [23:26] Then people caught on and now it's great to see [23:28] all of the multinationals have got their London bases here. [23:31] And also all the VCs stood up and noticed. [23:35] So, I like to think that DeepMind had a non-negligible part to play in that. [23:41] Putting the UK on the map for deep tech, [23:44] but also showing that it could be successful here. [23:46] There was the talent base to the VC world. [23:49] And hopefully that's made it a lot easier for companies [23:52] that have followed since. [23:55] And we've got so many strengths here in the UK [23:57] in biotech and fintech and other places where I think quantum-- [24:00] where we've got great university departments [24:04] and great talent and we just need to have that ambition. [24:07] And then I think there is now a much better funding environment. [24:11] -Yeah. I mean, we first got to know each other [24:14] now quite a long time ago, back in 2017 [24:16] when our mutual friend Reid Hoffman introduced us. [24:20] And you invested alongside Reid in Entrepreneur First Series A. [24:23] And two things I'd say on what you just said. [24:27] One, it's amazing the impact, because no one had heard of DeepMind [24:30] because you were in stealth as well, weren't you? [24:32] -Yeah. -I remember when you sold to Google [24:35] the impact that had on particularly graduate students in UK universities [24:41] saying, "I don't actually have to do this 10, 20, 30-year journey [24:47] to being a principal investigator in a lab, I could actually start something." [24:50] And we saw an explosion of AI startups around that time. [24:54] But also this point that it really was the case [24:58] that nearly all the computer science grads that we met, [25:00] and our job was to go into universities and find these people. [25:02] They wanted to go work in banks. [25:04] -Yeah. -I mean, that was back in 2011, 2012. [25:06] And I do think that's changed beyond recognition. [25:10] Something that we're both involved in is ARIA, [25:15] the Advanced Research and Invention Agency, [25:17] which Emily mentioned in her introduction [25:18] and you kind of joined the advisory board, [25:20] I'm the chair of the board. [25:22] I think that is a great example of something [25:24] where the UK is actually sticking its hand up [25:27] and saying we're going to go all in [25:30] on not just being a software capital, [25:34] but actually trying to do the really hard stuff [25:36] that's going to change the world. [25:37] -How do you think about that stuff? -Yeah, I think it's fantastic. [25:39] And I think it's amazing what you've done with EF [25:41] and encouraging these technical founders to start things. [25:47] I think ARIA is amazing too. [25:49] And actually what the UK government's doing in general [25:51] to encourage science and technology, [25:55] including things like hosting the AI summits, safety summits, [26:00] and I know, of course, you've been heavily involved with that [26:03] and Ian earlier, as well, just encouraging that technology base. [26:08] And I really think that Britain needs to lean into that [26:11] from several viewpoints, really. [26:14] From an economic perspective to-- I think that's one of our superpowers [26:17] is that we have this really deep technical base. [26:20] And I think now is the time to start deep technical companies [26:24] that produce really amazing new technologies. [26:29] And we can build not just a company around that, [26:31] but whole industries around that. [26:32] And I think that's what we're seeing with AI. [26:34] That's what we may see in the years to come in many other areas too. [26:38] And I think it's great to see that and us embracing that as a nation. [26:42] And what it also means is that gives us more influence on the global stage, [26:47] if you're to then influence the societal impacts of this [26:51] and how it will go, the ethics of it and so on. [26:53] And I think if you're not at the forefront of the technology, [26:57] then you sort of lose your moral authority [27:00] to lead an influence in that way [27:02] and have a seat around the table. [27:04] And also you'll be blindsided by the advances. [27:09] The only way to not be blindsided is to be at the forefront, [27:12] or have advisers or companies and things and universities [27:16] that are at the forefront that can then advise government. [27:20] And I think it's fantastic experiments like ARIA [27:23] to pursue that even further and in this case change academia a bit too, [27:27] to be more forward thinking, more risk taking, [27:29] more interdisciplinary, go more ambitious [27:33] perhaps like the old DARPA, IARPA, or what ARPA was. [27:37] I think that's what it was modeled on, right? [27:39] -And of course, DARPA had that line in their founding mission, [27:43] which was to avoid and create technological surprise. [27:47] And this point that if you're not doing the science, [27:50] you're not going to know what's coming. [27:51] -I mean, look, I always imagine an alternate universe timeline where-- [27:57] and Alan Turing is one of my all-time scientific heroes. [28:01] But imagine instead of persecuting him, [28:03] the government had unlocked [28:07] all of the secrets [28:11] that were kind of way past their sell-by-day. [28:14] Frankly, after the war with the Enigma machine numbers, [28:16] we could have had Silicon Valley in Manchester. [28:19] Right, that's what should have happened. [28:21] Turing should have been at the head of something-- mega startup. [28:26] -I will say just as a side note, [28:28] because you've already mentioned the safety summits. [28:31] When I was co-leading the work on the preparation for that last year, [28:34] I went to China and went to do the diplomatic work with with them [28:39] about them coming, which they did do. [28:41] And honestly, I think one of the reasons that they were so engaged [28:44] is that their Minister of Science and Technology [28:47] also had Alan Turing as one of his heroes. [28:49] And we did this whole meeting in Mandarin in translation. [28:52] And then he broke into English to tell me that he'd been to Alan Turing's house [28:56] and to the Manchester Industrial Museum. [28:58] So if that's not soft power, I don't know what it is. [29:00] -Well, exactly. I mean, look, I've always said this about the UK, [29:04] obviously we've got world class universities today. [29:07] I hope DeepMind has helped to stimulate an AI industry here, [29:11] and the work people like you have done with the funding environment. [29:14] But we've also got an incredible heritage [29:15] with people like Turing and Babbage even before that. [29:19] And we've always been at the forefront [29:20] of this kind of technology information theory. [29:23] So I think it's fitting that we're helping to be in the vanguard of that. [29:30] -So as you mentioned, there's now a lot of international interest in AI. [29:35] Just today, I think you were on a call with a group of world leaders [29:37] for the second AI Safety Summit, [29:39] which is being hosted in the Republic of Korea. [29:44] One thing that I've always admired about you and your co-founder Shane, [29:48] is that you've been very, very consistent [29:50] in how you've talked about AI safety for long before [29:55] anyone was talking about AI safety. [29:58] How do you think about it today? [29:59] -Yeah, the reason we were thinking that [30:01] from the very inception of DeepMind and even before, [30:05] and then we've infused that into Google, [30:06] and you can see our original ethics charter [30:11] became the first draft really for Google's AI principles, [30:14] which then became public several years ago is that we plan for success. [30:20] So we had this mad, crazy ambitious scheme, roadmap, [30:25] for how AGI could be reached. [30:28] But then if we really believed that it was possible in 20 years, [30:32] then we also knew the impact that would have [30:35] and therefore it would need to be done very responsibly [30:39] and with ethics and safety at the forefront of our minds at all times. [30:43] And I think that's also transpired, [30:46] with the way the technology's gone [30:49] and the impact it's already had, [30:51] which I think is just a fraction of what we're going to see. [30:55] And of course, I spent my whole life working on AI [30:59] because I believe it's going to be the most beneficial technology ever invented, [31:03] but only if we apply it in the right way and build it in the right way. [31:07] And that means the next 10 years, let's say, [31:10] 5-10 years is going to be very critical, what happens from here. [31:14] And I think it's great that maybe the good part of this, [31:21] all the language models being super mainstream, [31:24] is that the governments have woken up to that and I think that's useful, [31:30] and I've been pleasantly surprised by their receptiveness, [31:35] but also their smartness around getting their heads around [31:38] both the opportunities and the challenges for these technologies. [31:41] And I think it's great that we are having that debate now with these summits. [31:45] And you obviously helped organize the UK, [31:48] played a major part in helping organize the UK one. [31:51] It's good that we're doing that now ahead of time [31:54] of when these things become extraordinary powerful the next few years. [31:58] -Yeah. And I think one thing that you and I have talked about in the past, [32:00] that, personally, I find really important to emphasize, [32:04] is that caring about this topic doesn't come from being anti-tech. [32:11] It comes from a position of really wanting to see amazing tech built. [32:16] And something that I know we both agree on [32:19] is the way you deal with the risks of any technology, not just AI, [32:26] isn't just about, banning things and regulating things. [32:28] It's also about building safety into what you do. [32:33] You've done a ton of that from the beginning at DeepMind. [32:38] As you and I were discussing the other day, [32:39] I just started a program at EF dedicated to building defensive technologies, [32:43] which is the way I see acceleration and safety going hand-in-hand. [32:48] What do you think is missing? [32:50] What would you like to see people build? [32:51] -Yeah, I mean that's awesome that you're doing that, [32:53] sort of emphasizing the defensive capabilities. [32:55] I think it's great to try and make that asymmetric. Would be amazing. [32:59] I think we need a lot more research on benchmarking and testing, [33:05] because what we really want to be testing is when new capabilities arrive. [33:10] And then evaluating if those are dangerous or not capabilities [33:13] and that is something the UK Safety Institute is working on [33:17] and we're collaborating on and many others. [33:21] But it's a very hard space, in some ways, coming up with the right benchmarks [33:25] is maybe even harder than coming up with the right algorithms, surprisingly. [33:29] I think we need a lot more analysis tools [33:32] and maybe analogous to neuroscience analysis tools [33:35] for real brains to analyze these virtual brains. [33:38] And there's great work folks like Anthropic do [33:42] on understanding the representations [33:44] that are built up in these neural networks. [33:47] And we do a lot of work on that as well [33:48] with a lot of computational neuroscientists on our team. [33:52] But that needs accelerating. [33:54] I just think beyond that, we need to think about [33:59] building things like safe sandboxes [34:01] that are safe from cyber security perspective from outside breaking in, [34:06] but also the AI systems potentially breaking out. [34:10] And so we need to evaluate. A lot of these things are uncertain. [34:13] It may turn out that a lot of these things are much easier to deal with [34:16] than we expected, and that would be great. [34:19] But some of them may be much more challenging [34:21] or much more risky than we know today. [34:24] So, I think in a situation where there's a high degree of uncertainty, [34:28] then one must proceed with cautious optimism, in my view, [34:32] is the correct way to proceed until you get more evidence. [34:36] That's the scientific method. [34:37] That's the thoughtful and intentional way [34:40] we've always tried to advance the frontier. [34:43] It gets hard when you're in the middle of a commercial frenzy [34:47] like we seem to be at the moment with AI. [34:49] But at the core of it, at least for us, we try to be bold and responsible, [34:56] is the way we call it, rather than say, moving fast and breaking things, [35:00] which I think obviously is amazing for making the fastest progress [35:04] and that's the sort of value mantra. [35:05] But I think is not appropriate for something as transformative, [35:09] as powerful as AI is potentially going to be. [35:12] I think something like the scientific method is a much more suitable approach. [35:16] And the funny thing we talk about is, [35:19] you probably couldn't meet two more techno-optimists than-- [35:22] I mean I'm unbelievably techno-optimist [35:24] in the sense that I've always thought AI would be the ultimate tool for science, [35:28] scientific discovery, and thereby advance all of humanity [35:31] and solve many, many of the big grand challenges that are facing [35:35] society and humanity today, [35:37] from climate to disease to economic output. [35:40] But what I maybe find amusing about the people [35:45] who come to this space a lot later, maybe in the last year or so, [35:49] perhaps they were doing crypto last year or something. [35:51] And then this year it's AI, [35:52] and they're like accelerationists, or whatever it is. [35:56] And what is amusing is they don't actually fully understand [36:00] the enormity of what's coming. [36:02] Because if they did, they would understand [36:04] that I'm very optimistic we can get this right, [36:07] but only if we do it carefully and take the time needed to do it [36:11] and don't rush headlong, blindly into it. [36:14] I think we've got to treat it with the respect that the technology deserves. [36:19] And if we do that, I have no doubt in 10 years' time [36:23] we'll be in an amazing flourishing society powered by AI. [36:27] -One minute left, I want to ask you a completely different question. [36:31] Something again, people might not know about you is [36:35] you are a very avid reader. [36:38] You read a lot of things that have nothing to do with AI. [36:40] If you're going to recommend one non-AI related book to the audience, [36:43] what would it be? [36:45] -Oh wow, there's so many to choose from. [36:49] Look, I think one of my all time favorite books, [36:54] it would be The Fabric of Reality by David Deutsch. [36:57] I think that's the best book on trying to sort of understand [37:00] the nature of reality, which ultimately is what I'd like to try and do with AI. [37:04] And then maybe one of the books that [37:08] in the last year I've read, that I've enjoyed most, is [37:11] When We Cease To Understand The World by Labatut. [37:14] He's an amazing writer, it's an unbelievable book. [37:16] -One of the best insights to what genius is. [37:20] -Yeah and scientific discovery and how amazing it is [37:23] but also what it's like to be at the frontiers of knowledge. [37:27] -Fantastic, well Demis this has been great fun. [37:30] Thank you for being here. [37:32] And I think to wrap up the day we have Marc Fryer, [37:34] who leads Stripe in the UK. [37:36] -Thanks, great. Cheers. [37:43] -Thank you so much. That was wonderful. Thank you.