Google DeepMind's Hassabis Targets 2030 for AGI as AlphaFold Reshapes Biology

Google DeepMind哈萨比斯：AlphaFold重塑生物学，AGI目标锁定2030年 通用人工智能将在本年代末到来，将长达十年的药物研发过程转变为仅需数天的数字模拟。 Demis Hassabis回顾了他从少年国际象棋神童、游戏开发者到创立DeepMind的历程，其核心在于践行实现通用人工智能（AGI）这一专注且长达20年的使命。他重点阐述了AlphaFold等重大突破，并解释了机器学习为何能成为描述高度复杂、具备涌现特性的生物系统的理想语言。展望未来，Hassabis探讨了将药物研发彻底转向数字模拟、通过“信息优先”范式重新审视宇宙，以及应对围绕AI能动性与意识的深刻哲学问题。 Demis Hassabis 自青少年时期起便确立了构建通用人工智能（AGI）的明确目标，早期在游戏开发和神经科学领域的跨界经历为其积累了关键的硬件与算法认知。 (0:38) DeepMind 在 2009 年顶住学术界质疑，前瞻性押注深度学习与强化学习的结合，并严格遵循 2010 年制定的 20 年规划，Hassabis 预计 AGI 将在 2030 年左右实现。 (4:39) 依托 AlphaFold 的突破与 Isomorphic Labs 的探索，其终极目标是将 99% 的药物研发工作迁移至计算机模拟（in silico），将传统长达 10 年的周期压缩至数天或数周。 (8:51) Hassabis 提出信息与物质、能量同样是宇宙的基础构成，并将机器学习视作比传统数学更能准确描绘生物学等复杂动态系统的描述语言。 (17:18) 关于 AGI 的发展演进，Hassabis 主张分两步走：优先构建极为精准且强大的智能工具，而后再将其作为第二阶段课题来审慎应对自主代理（agency）与意识的本质挑战。 (20:29)

Google DeepMind co-founder Demis Hassabis argues that artificial general intelligence is progressing on a predictable multi-decade timeline toward realization around 2030, functioning primarily as the ultimate instrument for scientific discovery. The Twenty-Year Trajectory to Solve Intelligence (4:39) Hassabis traces his focus on artificial general intelligence (AGI) to a singular plan formed in his teens, combining game development, hardware insights, and neuroscience. His early work on titles such as Theme Park (1994), which sold over 10 million copies, deployed artificial intelligence for economic simulation. However, his startup Elixir Studios illustrated the hazard of outpacing available hardware when attempting to simulate one million citizens on late-1990s Pentium processors, yielding the operating principle to be five years ahead rather than fifty. When founding DeepMind in 2010, Hassabis and his team established a 20-year mission structured in two steps: solve intelligence, then use it to solve everything else. DeepMind pursued the combination of deep learning and reinforcement learning at a time when the academic establishment dismissed AGI and remained focused on 1990s-era expert systems. Hassabis states that the field is currently on schedule to fulfill that 20-year roadmap. Machine Learning as the Language of Complex Science (8:51) DeepMind treated its 2016 AlphaGo victory as algorithmic proof that general methods could address complex real-world systems, immediately prompting the formal establishment of an AI for Science division. The initiative's primary breakthrough, AlphaFold, solved the 50-year-old protein folding problem. Through the spin-out Isomorphic Labs, Hassabis aims to shift 99% of drug discovery into digital simulations (in silico), aiming to compress development cycles from an average of 10 years down to months, weeks, or days. Hassabis posits that machine learning represents the natural descriptive language for biology and emergent phenomena in the same manner mathematics serves physics. Because biological systems, weather patterns (such as DeepMind's Weather Next), and social dynamics contain weak signals and messy correlations that elude traditional hand-coded equations, AI simulators can model them directly. Over time, Hassabis suggests researchers may be able to extract fundamental, explicit equations from these learned models through repeated sampling. Information Primacy, Classical Computation, and Consciousness (17:18) Reversing classical physics paradigms, Hassabis views information as a fundamental building block of the universe equivalent to energy and matter, citing biological systems as mechanisms that actively organize information against entropy. Guided by Alan Turing's thesis that any computable process can run on a simple machine, Hassabis characterizes human brains as approximate Turing machines. He points to AlphaFold as proof that classical neural network architectures can calculate approximate solutions to quantum-scale molecular systems without requiring quantum computers. Regarding the path forward, Hassabis advocates a two-step framework: building AI strictly as an exact, useful tool before addressing autonomous agency and consciousness. While tools are now entering the agent era, Hassabis notes that defining machine consciousness presents a fundamental barrier. Referring to discussions with philosopher Daniel Dennett, Hassabis emphasizes that because artificial systems lack biological substrate equivalence with humans, bridging the experiential and verification gap between human and machine minds remains an enduring challenge. Hassabis maintains an estimated arrival date of 2030 for the widespread distribution of AGI. Open questions remain over whether fundamental governing equations can be extracted from biological AI simulators and how researchers will evaluate consciousness in non-biological substrates.

Google DeepMind联合创始人兼首席执行官Demis Hassabis指出，通用人工智能（AGI）正按其2010年制定的20年路线图稳步推进，预计在2030年左右实现；他主张将AI视为驱动科学突破的终极工具与理解宇宙信息的底层语言，并强调在探索机器意识与自主代理前，应优先开发高精度且安全的工具型系统。 AGI路线图与跨学科演进 (0:02) Hassabis表示自己自15岁起就确立了构建AGI的单一目标，此后在游戏开发与神经科学领域的探索皆服务于该构想。20世纪90年代他在Bullfrog开发《主题公园》（Theme Park，1994年售出超1000万份），借助早期图形引擎和经济AI模型验证了智能系统的吸引力；创立Elixir Studios开发模拟百万人口的《共和国》（Republic）则让他吸取了“领先时代5年而非50年”的教训。随后他转向神经科学研究以汲取大脑算法灵感，并于2010年正式创立DeepMind。 在2009年筹备期间，学术界普遍沉迷于专家系统并视AGI为伪科学，但Hassabis坚信将刚由Geoffrey Hinton等人提出的深度学习与强化学习结合能够扩展至AGI。DeepMind确立了两步走使命：“先解决智能问题（构建AGI），再用它解决其他所有问题”。目前Hassabis认为团队完全符合既定的20年时间表，并维持其对2030年实现AGI分发的一贯预测。 科学突破引擎与计算生物学 (8:51) Hassabis强调AI不仅是商业产品，更是科学发现的终极工具。DeepMind在首尔AlphaGo人机对战获胜后便设立了由Pushmeet Kohli领导的“AI科学”部门，旨在将解决复杂博弈的算法推广至真实世界科学难题。这一努力最显著的成果是AlphaFold攻克了困扰学界50年的蛋白质折叠结构预测难题，这也成为他在DeepMind最引以为傲的里程碑。 以此为基础，Hassabis通过分拆公司Isomorphic Labs推动药物研发革命。其目标是将传统耗时约10年且极度依赖湿实验的药物发现流程，99%转移至数字化模拟（in silico）中进行，仅将实体实验室保留作最终验证。Hassabis预测，未来数年内该模式有望将药物研发周期压缩至数月、数周甚至数天，从而大幅加速重大疾病治疗方案与个性化医疗的落地。 机器学习赋能复杂动态系统模拟 (15:12) 在复杂动态与涌现系统建模方面，传统人工编写特定规则的模拟器往往精度不足。DeepMind已在气象领域通过“Weather Next”实现了超越传统气象学工具的高速、高精度模拟；其正在研发的“虚拟细胞”则进一步拓展至生物学领域。Hassabis提出，机器学习之于生物学犹如数学之于物理学，是描述生物系统的理想语言。面对自然界包含的大量微弱信号、弱相关性和多重因果关系，人脑与传统数学解析表达均难以有效应对。 根据Hassabis的设想，通过对机器学习训练出的内生模拟器进行海量采样，未来科学家或能从中提取出类似“麦克斯韦方程组”的生物学基础显式方程。此外，该模拟范式还将拓展至经济学与社会科学等无法进行现实受控实验的涌现系统，通过运行成千上万次模拟试验，使原本高度不确定的软科学研究转变为数据驱动的精密科学。 信息本体论与工具优先的发展路径 (17:18) 在理论层面，Hassabis提出信息是宇宙最基础的构件，与爱因斯坦质能等价理论（E=mc²）类似，信息与物质、能量同样具备等价关联；抵御熵增的生命结构本质上就是信息处理系统。受图灵计算理论的启发，他认为人脑可被视为近似图灵机。AlphaFold表明，基于经典硬件的神经网络足以逼近原本被认为必须依赖量子计算的微观量子生化系统。 在人机交互与意识议题上，Hassabis主张严格的“工具优先”两步走路径：当前应优先打造极其精确且强大的“工具型”智能，而在进入赋予机器自主代理与意识的下一阶段前必须保持审慎。他认为自我意识包含主客体界限、时间连续性等要素，并指出由于人工系统与生物体缺乏“基质等价性”（借用哲学家Daniel Dennett的观点），机器与人类的经验鸿沟将长期存在；当前应借助高级AI工具反哺神经科学以更准确地解构人类意识本身。 Hassabis将AI定位为破译宇宙运作法则的认知透镜，坚信其将在2030年逼近AGI并在科学领域引发范式转移；然而，该构想将宏观科学革命押注于算法模拟对真实物理规律的完备逼近能力，且在机器越过纯工具属性迈向具备自主意识与能动性的实体时，如何在缺乏生物基质等价性的前提下界定并控制系统风险，仍是未解的技术与哲学难题。

[4:39] Demis Hassabis: At least if we were going to fail, we would fail in a different way than people had failed, you know, to get to AGI in the '90s... we would fail in an original way if it didn't work. [7:25] Demis Hassabis: Step one, solve intelligence, i.e., build AGI. [7:25] Demis Hassabis: Step two, use it to solve everything else. [2:52] Demis Hassabis: you want to be 5 years ahead of your time, not 50 years ahead. [15:12] Demis Hassabis: No, we can't... and I'm not sure that would be a good idea. [15:12] Demis Hassabis: Maybe. If they do exist, I don't see why we won't be able to find them with these methods. [23:33] Demis Hassabis: reading the language of the universe. [20:29] Demis Hassabis: It would be best to build a tool first—an incredibly intelligent and useful and precise tool—and then cross the next sort of Rubicon. [23:33] Immanuel Kant: the mind creates reality [25:04] Demis Hassabis: You want a game theorist, I think. And I think he's the best... I feel he'd be a good teammate.

[0:02] Demis, thank you so much. [0:03] >> Exciting to be here. Um, thanks everyone [0:05] for coming. It's great to be here. [0:06] >> We're we're so honored to have you at [0:08] our chocolate factory. [0:09] >> Yes, I had just heard about that. Yeah, [0:11] looking forward to the chocolate [0:12] afterwards. [0:13] >> Excellent. Well, Demis, we're going to [0:15] jump right in. We have one of the [0:18] OGs in every way. Uh, original thinkers, [0:21] founders, visionaries in all things AI. [0:24] A true believer, true scientist in in [0:27] Demis. [0:28] Uh, we're going to spend the beginning [0:29] of the conversation about the early [0:31] days. [0:32] Then early days of DeepMind, and then [0:34] we'll get into the science and open up [0:35] the room for some questions. So, let's [0:36] jump right in. [0:38] So, Demis, you were a chess prodigy. [0:40] Uh, you also were a founder of a gaming [0:43] company. [0:44] You're a neuroscientist, [0:46] and you're the founder of DeepMind, and [0:48] now leader of a really big consequential [0:50] company. [0:51] Those seem like pretty unrelated pieces, [0:54] but you've said that there's a common [0:55] thread. [0:56] Can you walk us through that? [0:58] >> There is a common thread, and um, [1:01] maybe I made it into a common thread. [1:02] So, it could be a post talk, you know, [1:04] sort of shaping. But, I wanted to do AI [1:08] for a long time. So, I kind of decided [1:10] it was the most important thing I could [1:11] possibly and most interesting I could do [1:13] in my teenage years. [1:16] And then I picked things to study or do [1:18] that I felt would eventually would help [1:20] me [1:21] build a company like DeepMind. So, I I [1:24] had that [1:25] as a plan from about 15, 16 years old. I [1:29] had a detour into games because actually [1:31] that was in the '90s, this is, [1:34] that's where all the most cutting edge [1:35] technology was being done. [1:37] Obviously, not just in AI, but in [1:38] graphics especially, including hardware [1:41] of course. The GPUs we all use today, [1:42] they were designed for graphics engines, [1:44] and we use the I mean, I was using the [1:46] very first GPUs back in the day in the [1:47] late '90s. [1:49] So, there was a lot of really cutting [1:51] edge technology, and then all the games [1:53] I made, including all the games I did [1:55] for Bullfrog, but also my own company, [1:56] Elixir Studios, all involved AI as the [2:00] main gameplay component. So, probably [2:02] the most, you know, well-known game I [2:04] made was Theme Park when I was about 17, [2:07] and that was a simulation of an [2:09] amusement park, and thousands of little [2:11] people came into your theme park and [2:13] played on your rides and decided what to [2:15] buy from the shops. So, there was a [2:16] whole kind of economics AI model [2:17] underneath. And it was one of the first [2:19] games of its type along with SimCity. [2:22] And when I saw it sold, you know, 10 [2:24] million plus copies, and when I saw how [2:25] delighted people were to interact with [2:27] the AI, you know, that was one of the [2:29] things that made me think about spending [2:31] my whole career on it. Um and then, of [2:33] course, neuroscience [2:35] is to get inspiration from how the brain [2:37] works, and [2:39] and and different algorithmic ideas from [2:40] that, and then just bringing all those [2:42] different things together for the start [2:44] of DeepMind when the timing we felt was [2:46] right. And then, of course, we use games [2:48] as the early proving ground for our AI [2:50] ideas. [2:51] >> Mhm. [2:52] So, we've got a room full of founders [2:54] here, [2:55] and you can relate because you're a [2:57] founder not just once, but twice. [2:59] >> Mhm. [2:59] >> Take us back to the first time. Elixir [3:01] Studios. [3:02] >> Mhm. [3:02] >> What was that like? It's not the startup [3:05] that you're most known for, [3:06] >> Mhm. [3:07] >> but it was one that you had incredible [3:08] success with. [3:09] >> Mhm. [3:09] >> How did you lead that, and what did it [3:10] teach you about building? [3:12] >> Well, look, we we we start I started [3:14] Elixir Studios straight out of college, [3:16] um and uh I was lucky enough to work at [3:19] Bullfrog Productions, which, those of [3:21] you know games, it was a kind of [3:22] legendary game studio [3:24] uh in the early days of the game [3:25] industry, probably the best one in the [3:26] UK in in Europe. And uh I wanted to do [3:30] something that combined uh pushed AI. [3:33] So, effectively, I was funding AI back [3:35] in those days through the back door, [3:36] through games development. [3:38] Um and then push the the forefront of [3:41] that and combine it with cutting-edge [3:42] creativity. And I think that's still [3:44] relevant today with the way we do our [3:46] blue sky research, but maybe the biggest [3:48] lesson I learned was you want to be um 5 [3:51] years ahead of your time, not 50 years [3:52] ahead. So, we tried to do a game called [3:55] Republic at Elixir Studios which [3:57] simulated a whole country, and then the [3:59] idea of the game is you could sort of um [4:02] overthrow, I think there was a dictator [4:04] in charge of the country in any number [4:06] of different ways. And we basically [4:07] simulated living, breathing cities. And [4:09] this is Bearing in mind, this is like [4:10] the late '90s on a Pentium. So, we had [4:13] to get all the graphics and all the AI [4:15] for million people [4:16] um working on uh you know, home PC at [4:19] the time. So, it was a little bit [4:20] ambitious. And um [4:23] and maybe it was too ambitious. And it [4:25] caused it caused probably some some [4:26] issues. And I took that lesson with me [4:28] of like, you want to be ahead of your [4:30] time, you don't want to be obviously [4:31] when it's obvious to everyone, it's too [4:33] late. But um if you're 50 years ahead, [4:36] then there's probably no way you can get [4:38] it to be successful. [4:39] >> All right. So, speaking of not being too [4:41] far ahead of your time, it was 2009 and [4:44] you decided there would be AGI. [4:45] >> Mhm. [4:46] Yes. [4:47] >> [laughter] [4:48] >> Maybe it was only you know, 10 years [4:49] ahead of our time that time. Better than [4:51] 50 years. [4:52] >> So, so tell us about again, room full of [4:54] founders here. Tell us about '09. [4:57] How did you convince the first few [4:59] brilliant talents? Cuz you pulled in [5:01] really high-caliber employees, early [5:03] team members. How did you convince them [5:05] to believe in what seemed like total [5:07] sci-fi at the time? [5:09] >> Well, there was some interesting threads [5:11] that I think we picked up on. I I think [5:13] we thought we were 5 years ahead, but [5:15] maybe we were more like 10. But it was, [5:16] you know, deep learning had just been [5:18] invented by Jeff Hinton and colleagues [5:20] sort of in academia, but almost no one [5:22] had really realized it was a big deal. [5:26] Uh we knew a lot about reinforcement [5:28] learning, and we felt there was huge [5:30] progress to be made by combining those [5:32] two techniques which almost had never [5:34] been mixed together really, certainly [5:35] not at any kind of anything other than [5:38] toy problems. Uh in the academic [5:40] subjects, they're quite two quite siloed [5:42] parts of AI. [5:44] Um then we could see the compute uh the [5:46] GPUs at the time were going to be really [5:48] useful. Of course, we use TPUs now, um [5:50] but the accelerated computing industry [5:52] was was was going to be very helpful. Uh [5:55] and then we also felt at the end of my [5:57] PhD and post-doc and some of the other [5:58] people I got together were computational [6:01] neuroscientists that we had enough ideas [6:03] and principles um from the brain that [6:06] could be useful, including the idea that [6:07] reinforcement learning uh could [6:09] eventually scale uh to AGI. So, um so we [6:13] felt we had these ingredients and we [6:15] almost felt like we were keepers of a [6:16] secret because no one uh either in [6:19] academia or industry really believed [6:21] that any big progress was possible. In [6:23] fact, a lot of the people in academia uh [6:25] used to roll out that that, you know, [6:27] literally rolled their eyes up at us [6:29] when we were sort of suggesting we would [6:30] work on AGI or strong AI, as it was [6:33] sometimes called at the time, um because [6:35] it was like, "Well, we know this doesn't [6:36] work. So, you know, everyone tried it in [6:38] the '90s." You know, I did my post-doc [6:40] at MIT, which was the which was the sort [6:42] of center uh [6:44] uh point for expert systems. Um uh and [6:48] first all the logic language systems. I [6:49] mean, it seems amazing to think that [6:51] now, but I was already feeling that that [6:53] was archaic then, but they they, you [6:55] know, that's still how it was done both [6:56] in Cambridge in the UK and also in MIT, [6:59] these big centers of traditional AI. Um [7:02] and it felt like, but actually that [7:03] convinced me even more that we were onto [7:05] something because at least if we were [7:07] going to fail, we would fail in a [7:09] different way than people had failed, [7:11] you know, to get to AGI in the '90s. So, [7:14] that felt like it was worth doing uh no [7:16] matter what, even if obviously it was [7:17] research, we didn't know for sure it [7:19] would be successful, but at least we [7:20] would we would fail in a original way if [7:23] it didn't work. [7:25] >> Was there any common sticking point in [7:27] that early belief? Was there something [7:28] that you had to prove either to yourself [7:30] or to your early followers to get them [7:32] on board? [7:33] >> Well, it was [7:34] I mean, we had Put it this way. I would [7:37] have been spent my life on AI um no [7:39] matter what had happened. So, as it's [7:41] turned out, it's gone [7:43] you know, it's sort of gone on the [7:45] absolutely amazing side of the [7:47] optimistic side of what we thought. [7:48] Still actually within what we were [7:49] predicting in 2010, we thought it would [7:51] be a 20-year mission. [7:53] And I think we're basically exactly on [7:55] track as a field for that, and obviously [7:57] we played our part in that. But even if [7:59] it hadn't transpired that way, and it [8:01] was still now niche subject, that's all [8:04] that's what I would still be doing [8:06] because I felt it was the most important [8:08] technology ever. It feels obvious to me, [8:11] you know, our original state mission [8:12] statement at DeepMind was step one, [8:14] solve solve intelligence, I build AGI. [8:16] Step two, use it to solve everything [8:18] else. So, it was always I always thought [8:20] it was the most important technology [8:22] that could ever be invented, but also [8:24] the most interesting one. So, as a tool [8:27] for science, as an interesting artifact [8:29] in itself, and actually as one of the [8:31] best ways to understand our own minds, [8:33] you know, like the nature of [8:34] consciousness, dreaming, creativity, all [8:36] of these questions I had as a [8:38] neuroscientist, I felt one of the things [8:40] that was missing was an analysis tool [8:42] like AI, but also a comparison, a that [8:45] you could do sort of a controlled [8:46] experiment, study, and compare [8:50] two different systems against each [8:51] other. [8:51] >> Mhm. Let's talk about AI for science. [8:54] You've been early to that, you've been a [8:55] believer, and you've been really a [8:57] purist about this. This is the driving [8:59] mission. What about the way you set up [9:01] DeepMind and set the culture has [9:03] positioned it to be on the constant [9:05] forefront of AI for science? [9:08] >> Well, that was the [9:10] the ultimate goal, at least for me, my [9:12] personal passion. There's one There's my [9:13] own drive to build AI, which was to [9:16] advance science and medicine and and our [9:17] understanding of the world. It's my [9:19] expression of that mission was to sort [9:22] of do it in a meta way, right? Build the [9:24] build the the ultimate tool, and then [9:26] come back when that was ready and use it [9:28] to make breakthroughs in science, things [9:30] like AlphaFold that we've done, and I [9:32] think many more things. So, we've always [9:34] had that at the heart of what we've been [9:36] trying to do at DeepMind. So, actually, [9:37] we've had an AI for science group [9:39] division [9:41] led by Pushmeet Kohli that [9:44] has existed for nearly a decade now. [9:45] Actually, pretty much the day after we [9:48] got back from Seoul in the AlphaGo [9:49] match, which is sort of 10 years to the [9:52] month now, [9:53] is when we started formally started the [9:55] AI for science efforts cuz I was waiting [9:58] for the algorithms to be powerful enough [10:00] and the ideas to be general enough [10:02] and for me, you know, cracking Go was [10:04] that point, that time [10:06] that we thought, "Okay, now we're ready [10:08] to really apply these ideas to important [10:10] real-world problems starting with these [10:12] big scientific challenges." So, we've [10:15] always had that in mind as the the most [10:18] beneficial use of of AI. Like, what [10:20] could be better than using it to, you [10:22] know, cure diseases and give us [10:24] healthier lifespans and and to help with [10:26] medicine, followed, obviously, by other [10:28] really important areas like material [10:30] science and the environment and energy [10:32] and these kinds of topics, which I think [10:34] AI is also going to be have a huge part [10:36] to play in the next few years. [10:37] >> And how does AI break through in [10:40] biology? You're deeply involved with [10:41] Isomorphic. This is an area of deep [10:43] passion. You have been a purist on the [10:46] potential of AI to cure diseases from [10:48] the very beginning. [10:50] When do we have the type of moment that [10:52] we've had in language and coding but in [10:55] biology? [10:56] >> Yeah, well, I I mean, I'd argue we've [10:57] already had one of those moments with [10:59] AlphaFold. So, [11:01] you know, this is a 50-year grand [11:02] challenge, protein folding, and the 3D [11:05] structure of proteins is incredibly [11:07] important thing to know about if you [11:09] want to design medicines or if you want [11:11] to understand biology. Of course, it's [11:13] only one part of the drug discovery [11:15] process, an important part, but it's [11:16] only one part. So, Isomorphic Labs, [11:18] which is our, [11:20] you know, later spin-out, having a lot [11:22] of fun running that as well, and it is [11:25] is to build adjacent technologies in [11:27] more biochemistry and chemistry space [11:29] that can actually design the compounds [11:31] automatically to kind of fit and bind to [11:34] the right part of the protein. So, we [11:35] now know the protein that the shape of [11:37] the protein, we know that what's on the [11:39] on the surface of the protein and what [11:40] we have to target. [11:42] Um and but now we've got to build the [11:43] right compound that of course binds [11:46] strongly to where you want it to bind on [11:47] the target of interest, but doesn't bind [11:49] to anything else ideally cuz that would [11:51] be a a toxic side effect. So, um the [11:54] dream is to do almost all the [11:56] exploration [11:58] uh which is 99% of the of the work and [12:00] the time in silico and then save the the [12:03] the the wet lab step just for the [12:05] validation step. Right? So, that would [12:07] be um you know, I think if we can do [12:09] that and and I think we can get there in [12:11] the next few years, um I think we could [12:14] reduce drug discovery times instead of [12:16] that for taking like, you know, an [12:17] average of 10 years down to months, [12:20] maybe even weeks and perhaps even days [12:22] one day. And um and then I think then [12:25] all disease could be in reach and I [12:26] think things like personalized medicine [12:28] will become possible. Uh you know, like [12:30] personalized variations off of base [12:32] medicines. So, uh I think the whole of [12:36] uh the whole medical area, drug [12:37] discovery areas is going to be [12:39] revolutionized in the next in next few [12:41] years. [12:41] >> Brilliant. You talked a lot about AI for [12:44] science. [12:44] >> Mhm. [12:45] >> Do you think that at some point AI will [12:47] create new sciences? Al uh industrial [12:51] revolution and thermodynamics, will [12:53] there be something new taught [12:56] fundamentally in our education system? [12:58] And if so, what would it be like? [13:01] >> Well, I think there's several things [13:03] uh [13:04] that along those lines that I think we [13:06] is going to happen. So, first of all, um [13:09] the understanding and the analysis of AI [13:12] systems themselves, I think it's going [13:14] to become a whole science, a kind of [13:15] engineering science. These incredible [13:17] incredibly interesting artifacts that we [13:20] are building and um they're incredibly [13:23] complex as well. As complex eventually [13:25] they'll be as complex as the human mind [13:27] and the brain. And so, they'll need to [13:30] be studied so we can understand fully, [13:32] way beyond where we we are today, how [13:34] these systems work. So, I think there's [13:36] a whole kind of field mech and tarp is [13:38] part of that, but there's a lot more, I [13:40] think, that we can do to analyze uh [13:42] these systems. So, that'll be a science. [13:45] But, I think also um [13:47] AI itself will maybe unlock new [13:49] sciences, which is maybe what you're [13:50] getting at. Um I the one I'm [13:52] particularly excited about is AI for [13:55] simulations. So, I I love simulations. [13:58] All my all the games I wrote not only [14:00] had AI, but they were simulations. And I [14:02] think simulations is the way we can [14:04] address some of the um what we maybe [14:06] think of social sciences uh like [14:08] economics um [14:10] and and other more humanistic subjects [14:12] because um it's very difficult to do [14:15] control studies in that, you know, why [14:16] aren't they just sciences like physics [14:18] today? Because the problem is they're [14:19] emergent systems um just like biology, [14:22] actually. And it's very hard to do [14:24] repeated controlled experiments. You [14:26] know, if you raise interest rates by [14:27] half a percent, you have to do it in the [14:28] real world and then see what happens. [14:30] You can have theories, but you can't run [14:31] it thousands of times. But, if you could [14:34] simulate things uh really accurately, [14:36] then maybe there's sort of new sciences [14:39] to be done where you can sort of uh [14:41] rigorously sample uh from a very [14:43] accurate simulator. [14:45] Um and then I think that will turn that [14:47] give allow us to make much better [14:48] decisions in these today what are very [14:51] uncertain [14:52] uh domains. [14:54] >> What will it take to get to those [14:55] extremely accurate simulations? World [14:58] models? What kind of science is [14:59] necessary and engineering to get [15:00] >> Yeah, well, look, I I mean, I'm thinking [15:02] a lot about that in we do a ton of that [15:04] work at AI like learning simulators is [15:07] basically what it'd be. So, you know, [15:08] these are in domains where you can't We [15:10] don't know the mathematics of it well [15:12] enough or it's perhaps too complex. We [15:13] can't just write We can't just write a [15:16] uh uh uh directly down a special case [15:18] simulator. Um Um, it's just not accurate [15:21] enough, doesn't capture all the [15:22] variables. Um, we're doing that it was [15:25] done in with weather. [15:26] Um, we have the most accurate kind of [15:28] weather simulator in the world, Weather [15:30] Next. And it's far faster than what the [15:32] meteorologists use. [15:33] >> weather yet? [15:34] >> No, we can't. We can't, no. And I'm not [15:36] sure that would be a good idea. But the [15:38] the first step is to understand it [15:39] better. [15:41] Um, and but then even biology, you know, [15:42] we we're working on a kind of what I [15:43] call a virtual cell. So, you have hugely [15:47] um, dynamical emergent system. And I [15:49] think biology is [15:51] perfect uh, [15:52] sort of machine learning is perfect [15:54] description language for biology in the [15:56] same way math is for physics. Because I [15:59] think in biology and in lots of these [16:00] natural systems you have loads of weak [16:02] signals, weak correlations, tons of [16:04] data, far too much than any human mind [16:07] can analyze. But there are connections [16:09] and correlations and interesting [16:11] causalities within that mass of data. [16:14] So, I think it's sort of it's always [16:16] struck me that machine learning is the [16:17] perfect tool [16:19] to describe those kinds of systems. [16:21] Where until today, you know, mathematics [16:24] hasn't been able to do that. Either [16:26] because we can't manage it as as top [16:28] mathematicians cuz too complex or the [16:31] expressive power of math is not enough [16:33] for [16:34] to understand these sort of highly [16:37] emergent dynamical systems. [16:38] >> Is it also because of the messiness and [16:40] stochastic nature of Yeah, sure. [16:42] >> And and I mean eventually you could by [16:44] the way, once you learn these [16:45] simulators, it maybe there's maybe [16:47] there's another branch of new branch of [16:49] science. You could maybe extract some [16:51] equations from the once you have the [16:54] simulator. So, you have this sort of [16:56] implicit simulator or intuitive [16:57] simulator and then maybe you could [16:59] extract explicit equations from that. [17:02] Um, because partly cuz you could also [17:05] sample it as many times as you want. [17:07] >> As fundamental as Maxwell's or [17:09] >> Maybe. If if I don't know that exists [17:11] for such emergent systems, um, but if [17:13] they do exist, I don't see we we won't [17:15] be able to find them in this with this [17:17] with this methods. [17:18] >> That would be amazing. You've talked [17:20] about this theory that the basic [17:22] building block of of everything in the [17:24] universe could be information like. This [17:26] is more theoretical. How do you think [17:28] about it that and what does that mean [17:29] for [17:31] a traditional classical Turing computer? [17:33] >> Well, look, I think you can of course [17:35] all the famous, you know, equals mc [17:37] squared and all the stuff Einstein did [17:39] and energy and matter are kind of [17:40] equivalent. [17:41] But I actually think information has a [17:44] kind of equivalence in the same way. So [17:46] you can think of, you know, the [17:47] organization of matter and structure and [17:49] especially things like biology that are [17:51] resisting entropy as basically [17:54] information processing systems at their [17:56] heart. So I think one can convert all of [17:58] those three kind of quantities into each [18:00] other, but I have this feeling [18:02] information is most fundamental. So it's [18:04] a little bit the opposite way around to [18:06] the classic physicist thought in the [18:08] 1920s and things where, you know, it's [18:10] sort of energy and matter primary. [18:13] I actually think it's a better to to [18:14] better way to understand the world the [18:17] universe is to think about it as [18:18] information first. And if that's true, [18:21] and I think there's quite a lot of [18:22] evidence for that, then of course AI is [18:25] even more [18:26] sort of profound in a sense [18:28] >> Totally. [18:28] >> than we think and it's already pretty [18:30] profound because it's it's also about [18:32] organizing information and understanding [18:34] information [18:35] and constructing [18:37] information or objects. So AI in my [18:41] opinion is all about information [18:43] processing. [18:44] So I think there's something sort of [18:46] very deeply connected with with these [18:48] different areas if you think of it [18:49] through the lens of information [18:51] processing as the primary way to think [18:53] about it. And do you think a classical [18:55] Turing machine will be able to compute [18:57] everything? Well, I sometimes think you [19:00] know, I sometimes sort of think about [19:02] what we're doing and refer to ourselves [19:03] as Turing's champion because Turing [19:06] machines, I think Alan Turing is one of [19:08] my all time, you know, favorite [19:09] scientific heroes. [19:12] I think what he did obviously laid the [19:14] foundations for the computers for [19:16] computer science but also AI. [19:19] And I think it's one of the most [19:21] profound results ever is the Turing [19:22] machine result. You know, that [19:24] everything that is computable can be [19:26] computed by relatively simple [19:28] description of a machine. [19:30] So I think our brains are [19:32] likely to be approximate Turing [19:34] machines. [19:35] And I think it's interesting to think [19:38] about the connection between Turing [19:40] machines and quantum computers and [19:41] quantum systems. But I think at least [19:43] what we've shown with things like [19:44] AlphaGo and especially AlphaFold [19:47] is that [19:48] classical Turing machine obviously in [19:51] the guise of a modern neural network [19:53] it can model what was thought to be in [19:56] the case of protein folding is a quantum [19:58] system. You know, in some sense it's [20:00] very you know, it's dealing with very [20:02] small [20:03] particles [20:05] and one might think you'd have to take [20:07] into account all the quantum effects of [20:09] the water bonds and all sorts of things. [20:11] But it turns out you can get to an [20:13] approximate [20:15] optimal sort of solution on a classical [20:18] system. So it may turn out that a lot of [20:21] things that we think are [20:22] would need a quantum system to model or [20:24] run might be modelable on a classical [20:27] system if thought about in the right [20:28] way. [20:29] >> So [20:31] you've talked about AI consistently as a [20:33] tool. Like a telescope or a microscope [20:36] astrolabe [20:38] through the centuries. [20:39] When you think about a machine that can [20:41] model almost anything [20:43] let's say it can't even model quantum [20:45] systems like you pointed out. When does [20:47] it stop becoming a tool and will that [20:49] ever happen? [20:51] >> Well, my my my feeling strong feeling is [20:54] we should in this sort of mission to and [20:56] journey to build AGI [20:59] those of us on that journey many of [21:00] people in this room, you know, I feel [21:02] like it would be best to build a tool [21:04] first [21:05] an incredibly intelligent and useful and [21:08] precise tool and then cross the the the [21:11] next sort of Rubicon. That's already [21:13] profound enough and has, you know, of [21:16] course the tool could start becoming [21:18] more and more autonomous and agent-like [21:19] that we're all seeing. We're in the [21:21] midst of that the agent era now. [21:24] But then there's a further step of like, [21:26] you know, does it have agency? Is it [21:27] conscious? These sorts of questions [21:30] which are also going to be questions [21:31] we're going to need to address, but I [21:33] would um [21:34] I would recommend we do that as a second [21:36] step. Perhaps using the tool in the [21:38] first step to help us [21:41] with those next profound questions and [21:44] ideally also we could understand [21:46] our own brain and and minds better and [21:49] define things like consciousness a lot [21:50] more precisely than we can today. [21:53] >> Do you have estimations of what that [21:54] definition of consciousness might look [21:56] like? [21:56] >> No, I mean I haven't got much to add [21:58] beyond that thousands of years of [21:59] philosophy hasn't said already, but I [22:01] mean it's very clear to me that [22:04] it's obvious some components are going [22:06] to be needed. Probably necessary but not [22:08] sufficient. Like things like [22:09] self-awareness [22:11] and, you know, the idea of self and [22:13] other, [22:14] some kind of continuity over time. So [22:17] some of these things are clearly needed [22:20] for anything that might look like [22:21] consciousness, but I mean obviously it's [22:24] an open question as to as to [22:26] what the full definition is and I've [22:28] talked to many of the great philosophers [22:30] about that. Daniel Dennett [22:31] obviously sadly passed away recently, [22:33] but we had a long conversation a few [22:35] years back about this and I think um [22:38] you know, one of the issues is how does [22:40] a system behave? Does it behave like a [22:42] conscious system? So that's [22:45] you know, you could argue some of the AI [22:47] systems might end up being able to do [22:48] that as they get close to AGI. [22:50] But then there's still the question of [22:51] why do you you know, why do we think [22:53] each other are conscious? One is the way [22:55] we're behaving. We're behaving like [22:56] conscious beings, but the other thing is [22:58] we're we're we're running on the same [22:59] substrate. So I think if both those [23:01] things are true, then it's parsimonious [23:04] to imagine you're you're experiencing [23:05] the same thing I'm experiencing, which [23:07] is why we don't have that debate [23:09] about, you know, normally about as are [23:11] each other conscious. But, I think we'll [23:12] obviously we'll never have the substrate [23:14] equivalence with an artificial system. [23:17] So, there'll there'll always be a I [23:19] think it'll be hard to to completely [23:21] close that gap. So, you can look at it [23:23] behaviorally, but but what about [23:24] experientially? [23:26] There are probably some ways to do that [23:27] post AGI, but let me speak about out of [23:30] scope today even for AI for science [23:32] discussion. [23:33] Yeah. [23:33] >> Brilliant. So, we're going to open the [23:35] room to questions in just a moment. Get [23:37] your questions ready. [23:38] But, you brought up philosophers. You've [23:41] mentioned Kant and Spinoza as two of [23:43] your two of your favorite philosophers. [23:45] Kant is this [23:46] you know, deontological, highly [23:48] duty-driven [23:49] philosopher. Spinoza almost has this [23:51] deterministic view of the universe. How [23:54] do you kind of connect those two [23:55] beliefs? And where where is your [23:57] thinking of how the world looks? [23:59] >> Well, the reason I like those two they [24:01] they stuck out for me is that [24:04] I think Kant when when I was doing my [24:07] PhD in neuroscience, you know, his [24:09] his sort of statements about the mind [24:11] creates reality, right? I think that's [24:13] basically true. And [24:16] and so, another reason to study the [24:17] mind, right? How the brain works. [24:20] And I'm interested ultimately in the [24:21] nature of reality. So, we have to [24:23] understand how the mind is interpreting [24:25] that. [24:25] >> Mhm. [24:25] >> And so, I think that's for me what I [24:28] took from Kant. And then Spinoza is more [24:30] about the [24:32] you could almost call spiritual [24:34] dimension of like, well, if you're [24:36] trying to understand the universe using [24:38] science in my case as the tool, you're [24:41] sort of understanding some deep mystery [24:43] about how the universe works, right? In [24:45] a in a really [24:46] uh uh [24:48] kind of deep way. And [24:50] that's what I feel I'm we're doing and [24:52] I'm doing when, you know, I do my [24:54] science and we work on AI and we're [24:56] building these tools is somehow we're [24:58] kind of [24:59] reading the language of of the universe. [25:03] >> Beautiful. [25:04] What a beautiful way to say what you do [25:05] every day. Demis, scientist, orator, and [25:08] philosopher. [25:09] We will, before we wrap, do a couple of [25:12] rapid-fire questions. [25:14] >> Okay. [25:14] >> Thank you for [25:15] >> [laughter] [25:15] >> finishing He's not seen these yet. [25:18] Over under on distribution year of AGI. [25:22] >> Oh, wow. [25:23] Um [25:23] >> Or reject premise of question. [25:25] >> No, 2030. I've been pretty consistent [25:27] about that. [25:28] >> Okay, 2030. [25:29] Must read book, poem, or paper [25:33] for when we achieve AGI. [25:34] >> Oh, wow. [25:36] Um [25:37] As for when we achieve once we achieve [25:38] it, [25:39] um [25:41] Well, my favorite book is The Fabric of [25:43] Reality by David Deutsch, so I think [25:45] that still holds. I'd hope to answer the [25:46] questions in that book with with the [25:48] AGI. That's my post-AGI work. [25:51] >> Brilliant. [25:51] >> Yeah. [25:52] >> Proudest moments so far in DeepMind. [25:54] >> Oh, wow. [25:56] Um we've been lucky to have a lot. I [25:57] mean, probably AlphaFold. [25:59] >> Okay. [25:59] >> Yeah. [26:00] >> Now, a couple games questions. If you [26:02] were engaged in a high-stakes [26:04] strategy game, turn-based strategy game. [26:07] >> Okay. [26:07] >> Civ, [26:09] Polytopia, strategy games. [26:12] And you can select one scientist from [26:13] history. We're thinking the the [26:15] Einsteins, the Turings, Newtons. Who [26:18] would you select to be on your team? [26:19] >> On my team? [26:21] >> On your team. [26:21] >> Oh, gosh. [26:25] >> [laughter] [26:26] >> Um probably von Neumann, I think. I [26:30] mean, he's Yeah, you want a game [26:31] theorist, I think. And I think I think I [26:33] think I think he's the best. [26:35] >> Makes sense. [26:37] >> I feel he'd be a good teammate. Yeah. [26:40] >> All right. [26:41] >> Yeah. [26:41] >> [snorts] [26:41] >> Well, Demis, you do it all. Thank you so [26:43] much for being with us. Please join me [26:44] in thanking Demis. [26:45] >> [applause]