love

VIP

I'm bob

May 2025

17 Following 21 Followers 299 Post

Paid Articles (2)

Bitcoin’s Silent IPO: Why This Consolidation Isn’t What You Think

The quiet liquidity event reshaping Bitcoin’s ownership base and its future stability

0.02 USDC 6 purchases

A payments protocol for the internet. Built on HTTP

Payments on the internet are fundamentally flawed. Credit Cards are high friction, hard to accept, have minimum payments that are far too high, and don't fit into the programmatic nature of the internet. It's time for an open, internet-native form of payments. A payment rail that doesn't have high minimums + % based fee. Payments that are amazing for humans and AI agents.

0.07 USDC 2 purchases

Post Media

love

· 3d

剪辑功能（Clipping）上线了 🎬

从任意视频里截取你最喜欢的片段，加上你的看法，不用离开 X 就能直接发布。

创作者：每个片段都会标注你的原视频并链接回去，新观众能顺着找到你的作品继续看。

剪辑者：现在有了一个简单的原生方式，可以直接在 X 上分享你喜欢的创作者的精彩片段。

更好的是：通过「原创内容奖励」（Original Content Rewards），原创作者和剪辑者都能从片段的播放量中获得收益。

好的片段传得越广，每个人都受益。剪辑功能是让 X 成为创作、成长和赚钱最佳平台的又一步。

x.com/XCreators

love

· 3d

Shopify 把移动端 App 从 React Native 重写成原生，原因之一是想改善启动时间这类性能指标。

这里把重写后、优化过的原生版 Shopify Shop App，和用 Expo 做的同一个 App 放在一起对比。Expo 版本调用的是同一套 GraphQL API。

x.com/ccheever

love

· 3d

已经有超过 15 万人开了账户，让 AI 智能体替他们交易。

这些智能体可以：

执行他们的策略
再平衡投资组合
寻找交易机会

Robinhood CEO Vlad Tenev 做客 No Priors 播客。

x.com/NoPriorsPod

中文 English

0:00 我们称之为“代理交易”，就像有个代理帮你执行交易并执行策略。

0:05 我们看到非常强劲的采用率，发布几个月来已有超过15万客户开设了代理交易账户。

0:15 你可以把它看作是云代码或Codex用户的体验。

0:19 它几乎像是一个为交易优化的类似工具。

0:22 你可以设置循环，比如每天执行或运行我的策略，重新平衡组合，寻找机会。

0:32 我认为未来会整合所有编程工具的能力，提供简洁、专为交易者设计的版本，让主动交易变得非常简单。

0:00 We call it "agentic trading," like having an agent execute your trades and do a

0:04 strategy for you.

0:05 We're seeing really strong adoption, like over 150,000 customers have opened

0:11 Agentic trading accounts with us in the few months since we've launched it.

0:15 You can kind of think of it as if you're a cloud code or a codex user.

0:19 It's almost like a tool like that that's optimized for trading.

0:22 So you can run loops and say, okay, every day, you know,

0:27 execute or run of my strategy, rebalance my portfolio, look at opportunities.

0:32 And the way that I think that'll evolve is we're going to look at all the

0:37 capabilities that

0:38 the coding harnesses have and see how we can provide really nice, clean, trader

0:45 -native versions

0:46 of that that just make it really, really easy to be an active trader.

love

· 4d

软件业 90% 毛利率的时代过去了：Modal CEO 谈 AI 基础设施的生意

一年估值涨十几倍的 Modal，CEO 却说 AI 公司的毛利率回不到 90%：用户付的一块钱要被应用、模型、基础设施、云厂商和 NVIDIA 层层分走，Modal 长期毛利率在 40%–60%。

love

· 4d

据说拉里·佩奇心里有个数字：未来有多少算力会跑到太空里？答案是「五个九」，也就是 99.999%。

帕特里克·奥肖内西是辗转听来的：「有人问拉里·佩奇，将来有多少算力会在轨道上，他说：五个九。」

马丁·卡萨多认为这可能成真的原因：「基本上就是监管问题。」他认为美国目前有 500 项提案要阻止建数据中心，连得州也在其中。

出自播客《Invest Like the Best》。

x.com/hosun_chung

中文 English

0:00 我最近听了个故事，有人问拉里·佩奇未来多少计算会在轨道上，他说是99.999%，原因是这纯属物理问题，最终所有计算都会在太空中完成。

0:11 你知道，这太有趣了，实际的太空方案确实行得通，但它最合理的原因只是监管问题。

0:16 因为它规避了监管。是的，是的，基本上就是监管问题。

0:20 我觉得美国光是有500条提案想阻止数据中心建设，包括德州。如果你连德州的数据中心都建不了，那我真不知道我们还能干什么，所以我觉得如果这种监管压力持续，你真的没法动工

0:34 我听过很多顶级数据中心人士说，现在开工非常难，那你不是去海外就是去太空了。

0:00 I heard this story recently that someone asked Larry Page how much of

0:03 compute will be orbital and he said five nines and the reason was it's just

0:08 physics

0:08 like eventually it's just like all going to be in space.

0:11 You know it's so funny like the actual space thing actually pencils out

0:13 but the reason it makes the most sense is just regulatory.

0:16 Just because it skirts it. Yeah yeah it's just basically regulatory.

0:20 I think there's 500 proposed legislations to stop data center building the

0:24 United States alone

0:24 including Texas right like if you can't build a data center in Texas I don't

0:28 know what we're

0:28 doing anymore and so I think like if that regulatory pressure continues and you

0:33 actually can't break

0:34 ground and I've heard from many of the top data center people like it's very

0:38 tough to break

0:38 ground right now then either you go international or you go into space.

love

· 4d

DeepMind CEO 戴密斯·哈萨比斯：

「一个真正懂 AI 的人，能胜过一整个创业团队。

5 年后，用 AI 做事的人，只需要十分之一的人手就能运营一整家公司。」

x.com/RoundtableSpace

中文 English

0:00 [掌声] 谢谢，Alastair，如此精彩的介绍。

0:05 很高兴能回到剑桥。

0:08 每次回到剑桥，我总有一种温暖的归属感。

0:13 Alastair提醒我，特别是这间讲堂。

0:15 我想这是我第一次上课的讲堂。

0:17 它一直是我最喜欢的讲堂。

0:19 我记得那天告诉过很多剑桥老朋友，Aaron，我想有一天我会回到这里讲座，谈论宣布通用人工智能，或许机器人会走上台，吓到所有人。

0:35 今天我不会这么做，让你们失望了。

0:39 但也许几年后我会回来，做那场讲座。

0:43 这里真是个了不起的地方。

0:45 充满了激励，我将谈谈剑桥如何激励了我整个职业生涯，也希望能激励在座各位和学生们。

0:56 我的人工智能之旅始于游戏，特别是国际象棋，就像Alastair提到的。

1:04 我从四岁开始下棋，并为英格兰青少年队认真比赛。

1:10 这让我开始思考“思考”本身。

1:13 我们的头脑是如何想出这些计划和想法的？

1:17 我们是如何解决问题的？

1:18 我们如何能改进？

1:19 显然，当你年轻时下棋并竞争，你会努力提升这个过程。

1:25 比游戏本身更吸引我的是背后的思维过程。

1:33 事实上，我第一次接触人工智能和计算机就是在国际象棋环境中，用早期的国际象棋计算机，如右边这台。

1:44 我想这是我拥有的第一台国际象棋计算机。

1:46 它有物理棋盘，需要按压方格来移动棋子。

1:51 当然，我们应该用它们来训练开局理论，学习更多国际象棋知识。

1:58 但我记得我很着迷于有人竟能编程让这块无生命的塑料玩国际象棋并且相当厉害。

2:07 我对这种实现方式和编程过程感到非常着迷。

2:13 到了少年时期，我开始用Amiga 500电脑尝试自己编写这类人工智能程序，那是80年代末90年代初很棒的家用电脑。

2:23 我编写了玩奥赛罗这类游戏的AI程序。

2:27 那是我第一次真正接触到人工智能。

2:30 从那时起，我就被深深吸引住了。

2:32 我很早就决定，将整个职业生涯都投入推动这项技术的前沿发展。

2:41 这也把我带到了剑桥，我在这里度过的三年对我影响深远。

2:49 我来自伦敦北部一所补习学校。

2:54 在我的记忆里，没有人去过牛剑。

2:59 我想来剑桥的原因，是听到过太多激励人心的故事，关于这里发生的事，和我曾读过的很多伟人传记，特别是左上角的Crick和Watson。

3:14 我特别记得一部电影，《双螺旋之争》。

3:18 这是一部80年代的精彩电影，如果没看过，推荐看，有Jeff Goldblum早期的表演，他扮演Watson，充满热情。

3:28 他们在剑桥探索DNA，度过了令人难忘的时光。

3:35 我想，我也想参与其中。

3:38 我想感受站在发现前沿的感觉，还有比这更令人振奋的吗？

3:44 那部电影让我真切感受到这种体验。

3:50 当然，我所有的英雄，我的科学英雄们，很多人都曾在剑桥求学。

3:56 人们会想到艾伦·图灵和查尔斯·巴贝奇，自然是在我们现在坐的这个讲堂里。

4:02 甚至像鹰巢酒吧这样的地方，如果你从皇后学院开始，他们会在入学的第一天带你参观，展示他们讨论DNA结构时使用的那个老式桌子。

4:17 走在国王大道上，真让人忍不住被激励。

4:22 我几乎觉得过去那些知识巨人为你从石头中低语。

4:29 我去Gardeniers深夜吃汉堡时，就是有这种感觉，这些伟人们百年来走过的同样路激励着我。

4:41 这就是剑桥无与伦比的历史，我相信我们至今仍能汲取灵感。

4:49 这里是我和亚伦的合影，他是我在皇后学院最好的朋友之一，显然是在数学桥上拍的。

4:57 最后，正如Alissa提到的诺贝尔奖，十二月去斯德哥尔摩领奖是我一生的荣幸。

5:07 那是一周精彩纷呈的活动。

5:10 但我最喜欢的活动是签署诺贝尔基金会的那本诺贝尔书。

5:15 那本书就在这里，是我拍的一张照片，我开始翻阅那本书。

5:20 你签下名字，然后往回翻。

5:22 你会想那是不是克里克，果然是。

5:25 继续往回翻，爱因斯坦的签名就在那，真是令人震撼。

5:29 我花了一个小时拍下书中的每一页。

5:35 这张照片让事情回到了原点。

5:39 然后真正看到那个80年代末的电影。

5:45 2010年，我们在伦敦创办了DeepMind。

5:50 当时，我们将其视为一种“阿波罗计划”的努力，试图打造通用人工智能。

5:58 一种真正通用，能完成所有人类认知能力的人工智能。

6:06 也就是说，它将是真正的通用AI系统。

6:09 这个想法实际上来源于图灵和图灵机。

6:13 正如图灵用他的图灵机展示的，能够计算任何可计算问题的机器。

6:20 这其实是我在剑桥上课时带走的最主要基础之一——计算机科学和计算理论的理论基础，正是图灵和香农在40、50年代所开创的。

6:39 所以我们在2010年开始。

6:40 真是令人惊叹。

6:41 那已经是15年前，某种程度上也不算太久。

6:44 但当我们开始DeepMind时，几乎没人研究AI，现在几乎每个人都在做，非常难以置信。

6:53 短短十多年间，进展极为迅速。

6:57 显然，我们参与了这段激动人心的旅程。

7:02 DeepMind的使命从一开始就是负责任地构建AI，造福人类。

7:09 刚开始我们把它表述为两步走，第一步，解决智能问题；第二步，用智能解决其他所有问题。

7:18 2010年时这个目标看起来很疯狂。

7:21 你可以想象向投资人推销这个使命时的情景。

7:26 看起来相当疯狂。

7:27 但我至今仍坚定地相信这一点。

7:30 我觉得越来越多人认识到，这种通用AI能够对几乎所有领域产生深远的变革性影响，这正是使命的第二部分。

7:43 对我来说，这意味着加速科学发现、医学进步，以及推动我们对宇宙的理解。

7:56 刚开始时，基本上只有两种方式。

7:59 事实上，当我在90年代这里学习时，广义上讲，有两种构建人工智能的方法。

8:04 一种是专家系统方式，就是直接预先编程解决方案，比如90年代我在这里学习时，深蓝著名地击败了卡斯帕罗夫。

8:17 这大概是专家系统的巅峰例子。

8:21 但专家系统的问题是，为什么它们从未真正扩展到通用智能，是因为它们无法应对意外情况。

8:30 如果发生了你未预料到的意外，系统内没有任何机制能处理它。

8:38 它们受逻辑系统启发，因此非常僵硬、脆弱和易碎。

8:45 而现代方法则建立在学习系统之上。

8:48 这些系统能够自我学习，基于第一性原理直接从经验或数据中学习，真正受神经科学思想启发。

8:58 显然，这些系统今天的承诺是，它们有可能超越我们——程序员或系统设计师——已知的解决知识。

9:08 当然，这在科学发现等领域极其有价值。

9:15 所以我们自然从2010年代初期的游戏开始。

9:19 我一生中多次利用游戏，首先是训练我的大脑。

9:23 然后我在AI4里制作电脑游戏，最终第三种方式是训练我们的AI系统。

9:30 游戏是检验AI系统的完美试验场。

9:34 你可以从简单的游戏开始，比如70年代的雅达利游戏。

9:38 其实，这个系统DQN是首次有人建立的可以直接从原始数据中学习的端到端学习系统。

9:47 在这里，原始数据就是屏幕上的像素。

9:49 系统对游戏或所控制内容一无所知。

9:53 它只被告知要基于视频流输入，即像素数据，最大化分数。

10:01 我们大约在2013年左右掌握了各种雅达利游戏。

10:07 接着我们把系统扩展到游戏AI的重大挑战：是否能打造出能达到甚至超过世界冠军水平的围棋系统。

10:22 围棋无疑是人类发明的最复杂游戏。

10:28 已有数千年历史。

10:29 它也是最古老且最优雅的游戏之一。

10:33 衡量围棋复杂性的方法之一是：围棋有10的170次方种可能棋局。

10:40 这远远超过了可观测宇宙中的原子数。

10:44 关键是，这说明无法用蛮力方法制定围棋策略。

10:51 这是不可能的，完全不可行。

10:53 所以必须用更聪明的方法。

10:56 我们在2016年著名地赢得了100万美元挑战赛，击败了十次世界冠军李世石。

11:04 李世石是围棋传奇，是韩国特级大师，比赛被全球两亿人观看。

11:11 AlphaGo不仅赢得了比赛，重要的是，它还创造出了新的原创围棋策略。

11:19 虽然我们已经下了几千年围棋，职业下了几百年，它仍能找到未曾见过的策略。

11:28 最著名的是第二局红色标记的37手，如果你看相关纪录片——可在YouTube观看——你会看到世界顶尖选手有多惊讶。

11:42 他们在解说这个棋局时，都在讨论这步棋。

11:44 这简直是不可思议的一步。

11:46 但它决定用这步棋，在100步后将局势带向AlphaGo有利。

11:54 这再次让我看到此类系统发明和发现新知识的潜能。

12:01 这里说的，当然仅是游戏知识。

12:03 但我的梦想是把它推广到所有科学发现领域。

12:09 那么这些系统是如何工作的？

12:14 我们基本上是通过自我对弈来训练这些神经网络。

12:19 这实际上是AlphaGo，以及后续的系统AlphaGo Zero和AlphaZero，它们将我们在围棋上的方法推广到可以从零开始玩任何双人游戏。

12:31 你从一个版本1的系统开始，它并不真正了解游戏，只知道规则。

12:36 它随机下棋。

12:38 你让这个系统自我对弈，比如100,000局。

12:43 这就产生了一个由这100,000局游戏组成的新棋局数据库。

12:49 然后你用这些数据训练第二个版本，一个稍微更好的模型，版本2，它被训练来预测任何局面下可能的走法，以及哪个阵营（黑方或白方）更可能赢。

13:06 以及他们赢的概率有多大？

13:10 然后你可以用版本2去对战版本1，进行100局对决。

13:17 如果版本2赢得明显领先——比如胜率55%——你就用版本2替换版本1，产生一批更高质量的对局数据，再训练版本3。

13:33 如果你反复进行17、18次这样的循环，早上还是随机下棋，24小时内的版本17或18已经强过世界冠军水平。

13:46 看到这种自我提升的过程在极短时间内实现，真是非常不可思议。

13:56 如果我们考虑这些神经网络在做什么，就是将10的170次方级别的庞大搜索空间缩减到几分钟的计算时间内可处理的范围。

14:11 它通过神经网络来高效指导搜索机制，从而缩小范围。

14:19 如果你把可能的选择看成一棵树，每个节点是一个围棋局面，那么你不需要查看所有可能，只用神经网络引导，沿着最有趣、最有用的路线深入分析。

14:39 在这里，那些蓝色的节点就是。

14:41 用尽思考时间后，你选择迄今为止最好的、最有潜力的路线。

14:48 在这里，就是这个紫色的路线。

14:54 这就促成了我们不仅研究围棋，还研究任何双人完全信息游戏。

15:00 它甚至能发现象棋中新策略和新风格，这很了不起，尤其是象棋电脑已经非常强大。

15:11 例如像Stockfish这样的程序。

15:13 AlphaZero当时能击败Stockfish，在象棋领域几乎是难以想象的。

15:20 不仅击败了Stockfish，这里AlphaZero是白方，Stockfish是黑方。

15:27 在一个最著名的对局中，AlphaZero展示了所谓的“不朽逼迫”局。

15:34 白方获胜，因为它重视机动性胜过物质优势。

15:38 大多数象棋电脑重视物质价值。

15:40 下棋的你会看到黑方虽然物质更多，却几乎无法移动任何棋子。

15:47 它们全被困在角落里。

15:49 AlphaZero为了机动性主动牺牲了物质。

15:53 这对人类国际象棋大师和顶尖选手来说，不仅是一种有效策略，也是非常优美的棋风。

16:03 AlphaZero能够发现这种动态的新棋风，真是令人惊叹。

16:10 事实上，世界顶尖棋手对此也有评价。

16:13 加里·卡斯帕罗夫，我最喜爱的历史名将，曾说程序通常反映编程者的偏见和优先级。

16:21 但AlphaZero自己学习，我认为它的棋风反映了真理。

16:25 当时的世界冠军马格努斯·卡尔森说，看过这些对局和研究AlphaZero相关书籍后，

16:34 他表示自己最近受到了英雄之一AlphaZero的影响。

16:39 他实际上将很多这些理念融入了自己的棋局，近十年主导了象棋界。

16:49 这就是我们在游戏AI领域，在DeepMind最初十年里取得的一系列突破。

16:57 但这当然只是我们最终目标的练兵场，是达成其他目标的手段。

17:03 玩这些游戏并不是目的本身，尽管我很喜欢游戏。

17:07 目的是创造这些算法，能够普遍应用于解决现实世界的问题。

17:14 所以我们关注的是真实世界中的问题，不仅是科学问题，还有工业问题。

17:22 我们寻找三种不同的标准，来判断问题是否适合由我们为玩游戏开发的这类AI系统、理念和算法来解决。

17:34 第一，我们寻找的问题可以描述为庞大的组合搜索空间。

17:40 通常过于复杂，组合数太多，无法通过暴力破解找到解。

17:47 但也许我们可以用神经网络学到某种结构，从而高效地引导搜索。

17:55 第二，我们寻找的问题有明确的目标函数或某种可优化的指标。

18:03 在游戏中，这是非常容易的。

18:04 比如最大化得分或赢得比赛。

18:08 但实际上，很多现实世界的问题都可以归结为几个你想最大化的指标或目标函数。

18:16 最后，当然你需要大量的数据或经验来学习，理想情况下还需要一种准确且高效的模拟器，以生成更多合成数据来补充真实数据。

18:32 事实证明，如果从这个角度看问题，许多问题都符合这些条件，包括许多重要的科学问题。

18:45 我一直铭记在心的一个问题，是我作为剑桥本科生时首次接触到的蛋白质折叠问题，接下来我会简单介绍一下，给不了解生物学和蛋白质的人。

18:58 蛋白质极其重要。

19:00 它们是生命的基石，几乎身体内的所有功能都依赖蛋白质，从神经元的放电到肌肉纤维的收缩。

19:12 所以可以说，蛋白质是生命存在的基础。

19:17 蛋白质折叠问题其实很好描述。

19:20 蛋白质由基因序列决定，基因序列定义了氨基酸序列，自然界中氨基酸序列会自发折叠成通常非常美丽的蛋白质结构。

19:38 也就是说，从基因序列到蛋白质结构的转变。

19:42 蛋白质的三维结构非常重要，因为它在很大程度上决定了蛋白质的功能，即它在体内扮演的角色。

19:51 虽然结构不完全决定功能，但它对蛋白质的实际功能起着重要作用。

19:58 蛋白质折叠问题就是：你能否仅凭一维的氨基酸序列来预测蛋白质的三维结构？

20:07 能否用计算的方法，从序列预测出那令人惊叹的三维结构？

20:15 为什么这是一个难题？

20:17 60年代著名的蛋白质研究者Levental提出了一个猜想，被称为Levental悖论，他计算出平均蛋白质可能有大约10的300次方种形态。

20:37 但是，这些蛋白质在自然界和体内却能在几毫秒内自发折叠完成。

20:44 这就是悖论所在。

20:46 如果可能性如此之多，自然是如何做到的？

20:50 物理学是如何实现这一点的？

20:53 这让人有希望，表明此问题在计算上或许是可行的，因为物理学每秒在体内解决这个问题数十亿次。

21:06 吸引我关注这个问题的另一个原因，是有一个两年一度的CASP竞赛，可以称为蛋白质折叠领域的奥运会。

21:16 竞赛每两年举办一次。

21:17 由马里兰大学的John Malt教授带领的优秀团队负责组织。

21:23 比赛自1994年开始举办。

21:26 这是一场很棒的比赛，他们与实验科学家合作，后者用电子显微镜等昂贵而先进的设备费力发现这些结构。

21:39 并将尚未发表的新结构带入竞赛。

21:43 竞赛组织者知道真实结构。

21:45 但参赛的计算团队每两年有数百支队伍参与。

21:50 但每年有数百支计算团队参加每场比赛，每过几年都会有比赛。

21:56 他们用计算方法来预测这些结构。

22:00 参加比赛的蛋白质通常有大约100个。

22:03 然后在夏末，他们揭晓了真实的结构。

22:07 你可以将预测的结构和真实结构进行对比，查看预测误差。

22:17 我们实际上在2018年第一次参加了Alpha Fold 1。

22:22 我们2016年开始Alpha Fold项目，几乎是刚从韩国首尔的Alpha Go比赛回来之后的第二天。

22:32 我觉得我们已经准备好了。

22:33 我们的技术已足够成熟，可以应用在游戏之外，解决真正有意义的问题。

22:43 我们称这些为根本性问题，因为一旦解决，就会开启全新的研究分支和方向。

22:51 蛋白质折叠就是一个典型例子。

22:54 所以我们从2016年开始工作。

22:57 Alpha Fold 1 几年后准备就绪，我们将它投入了CASP13竞赛。

23:01 这图表显示了过去十年中最难蛋白质预测类别的获胜队伍得分。

23:13 你可以把它看作是准确率，预测的氨基酸有多少位置正确，误差在原子大小范围内。

23:26 可以看到，十年内几乎没有进展，停留在60分左右，大家都在等待。

23:35 - 实际上，如果能达到90分，就意味着误差在一个原子宽度内。

23:45 那就是原子级的精度。

23:46 实验人员告诉我们，要达到与实验方法竞争的精度，必须达到这个水平。

23:53 这样实验人员就能依赖预测结果，而不必费力费时地去找结构。

24:04 作为经验之谈，我的生物学朋友经常说，要用一整位博士生的四五年时间才能找出一个蛋白质的结构。

24:16 已知科学的蛋白质有2亿种，人类蛋白组有2万种。

24:27 Alpha Fold 1赢得了比赛，比第二名系统提高了近50％。

24:36 Alpha Fold 1首次将机器学习技术作为系统的主要组成部分。

24:44 但还不足以达到原子级精度。

24:47 我们不得不回炉重造，利用Alpha Fold 1的经验，从零设计了Alpha Fold 2，最终达到了原子级精度。

24:59 这让组织者在2020年底宣布问题已被解决。

25:04 这就是Alpha Fold的视觉示例。

25:09 你可以看到左侧是一个非常复杂的蛋白质。

25:13 绿色是真实结构，蓝色是预测结构，你可以看到蓝色与绿色高度重合。

25:21 右边显示了Alpha Fold 2的工作方式。

25:24 它通过迭代过程构建结构。

25:27 实际上它通过192步不断循环自我优化。

25:31 从一个蛋白质氨基酸的团状物开始，逐渐构建出更合理的结构。

25:39 最后精细调整结构，直到完成预测。

25:44 我们很快意识到，Alpha Fold不仅准确，而且极其快速。

25:54 它能在几秒钟内折叠一个普通蛋白质。

25:58 我们很快认识到能折叠所有已知的2亿种蛋白质。

26:05 一年内，我们利用谷歌云的大量计算资源完成了所有蛋白质的折叠，并与埃比科学研究所的同事合作免费发布于数据库。

26:18 地点就在剑桥附近。

26:20 我们向全球任何人提供免费、不受限制的访问权限。

26:24 如果你想想实验上要花五年时间完成的两亿个蛋白质结构，相当于一年的博士研究时间有十亿年。

26:35 所以，想想科学能够加速这么多，真的很令人惊叹。

26:41 它开启了全新的探索途径，尤其是对于那些研究较少的生物，比如某些植物类型的结构，它们对科学和农业研究非常重要。

26:57 但如果没有它们，几乎没有任何这些结构会被发现和公开。

27:03 现在它们全部都可用了。

27:05 此外，这两亿个蛋白质还可以进行整体观察，比较物种间的结构，研究宏观结构，找到进化中的共性。

27:19 因此，结构生物学领域出现了许多有趣的全新分支正在被探索。

27:26 当然，我们从一开始就考虑了安全问题，并且非常认真地承担作为人工智能先锋的责任。

27:32 在这方面，我们咨询了超过30位生物安全和生物伦理专家，确保发布的成果利大于弊。

27:46 现在我非常自豪地说，全球已有超过200万研究人员在使用它。

27:53 它的引用次数已超过3万次，现已成为生物学研究的标准工具。

28:02 希望在座的许多博士生们也都在使用它，充分发挥其作用。

28:06 它已经成为生物研究中标准的常用方法之一。

28:13 看到其他研究人员利用这些技术和结构取得的成果，确实令人惊叹。

28:20 我只挑了我最喜欢的六个例子。

28:25 朴茨茅斯大学的团队利用它研究环境中的塑料污染，设计可以降解塑料的新型酶类蛋白质。

28:37 我们与弗莱明中心合作，研究抗生素耐药性和热带地区贫困群体的被忽视疾病。

28:46 我们与被忽视疾病药物研究所合作。

28:50 这是一个很好例子，通过加速疟疾、利什曼病、寨卡病毒等许多未知结构疾病的研究。

29:01 现在，由于这些病毒和细菌的结构信息丰富，能直接进入药物研发阶段。

29:11 还有许多基础研究，例如解析纳米孔复合体结构，这是一种重要蛋白，控制细胞核孔的营养物质进出。

29:24 布罗德研究所完成了令人惊叹的工作，设计分子注射器，重塑蛋白质实现药物靶向输送。

29:35 这项技术甚至被用于研究生育机制。

29:40 现在，几乎所有生物和医学研究领域都在利用AlphaFold。

29:51 近几年我们不断推动系统改进，取得了更多进展。

29:58 今年早些时候我们发布了AlphaFold 3供学术界使用。

30:04 我们还扩展了AlphaFold 3，以处理分子间的相互作用。

30:08 你可以把AlphaFold 2看作是静态蛋白质结构的图像，但生物学其实是动态过程。

30:17 你需要理解各种生物元素如何相互作用。

30:22 当然是蛋白质与蛋白质之间，还有蛋白质与DNA、RNA及配体等生命重要分子之间的作用。

30:31 配体是指小分子，比如药物化合物，蛋白质如何与之结合？

30:41 我们还有一套叫Alpha-Proteo的工作，是AlphaFold的逆过程，利用AlphaFold技术设计新蛋白，可能自然界中不存在，具有特定功能。

30:59 什么样的氨基酸序列和基因序列能产生这种结构？

31:02 就像倒着运行设计新结构，这些新结构可以执行新功能，非常有助于药物、抗生素和抗体设计。

31:16 回过头来，如果看过去15年的所有工作，对科学和机器学习有什么意义？

31:30 无论是早期的游戏研究，还是现在的科学项目，AlphaFold是最佳范例——关键在于使搜索问题变得可控。

31:42 你面对的是极其复杂的问题。

31:44 有无数可能的解决方案，你必须找到最优解，就像在巨大组合空间中找针一样。

31:55 无法用蛮力解决，所以必须训练神经网络模型，让它了解问题的拓扑结构，从而有效引导搜索，最大化或找到目标的最优解。

32:14 我认为这是一种极具通用性的解决方案，一种极其通用的方法来处理众多问题。

32:22 所以我们回想一下围棋的例子。

32:25 我们试图用这些系统找到最佳的围棋着法，但你也可以把这些节点换成化学化合物，现在你要在化学空间内找到最佳分子，这就是药物设计的起点，找到能特异性结合目标的最佳分子，

32:46 而不会结合其他东西。

32:47 这样可以减少该化合物的副作用和毒性。

32:52 我们现在用来设计这些分子的技术，非常非常类似，这也是我们向药物发现不断迈进的下一步。

33:03 所以我觉得至少在生物学领域，我们正进入我称之为数字生物学的新时代。

33:10 我认为生物学在最根本的层面上是一个信息处理系统，试图抵抗周围的熵增。

33:18 我觉得这基本上就是生命的本质。

33:21 当然，它是一个极其复杂且涌现的信息处理系统。

33:26 我认为这正是人工智能的作用所在，就像我在这个房间学到的数学，是物理和物理现象的完美描述语言。

33:37 我认为人工智能有可能成为生物学的完美描述语言。

33:41 它非常适合处理生物学这类动态系统中涌现行为和互动的复杂性。

33:50 我觉得AlphaFold就是一个证明。

33:53 我希望十年后回头看，这不仅是一次孤立的突破，而是宣告了数字生物学这个黄金时代的到来。

34:04 我们也在努力推动这一进展。

34:06 我们成立了一家新衍生公司Isomorphic Labs，基于AlphaFold技术，向我刚才提到的化学领域深入。

34:14 尝试用人工智能从基本原理重新构想药物发现。

34:18 目前开发一款药物平均需要10年时间。

34:24 成本极其高昂，达到数十亿美元。

34:27 所以我想，为什么我们不能用这些技术，将时间从多年缩短到几个月，甚至有一天缩短到几周，就像我们把蛋白质结构发现从可能的数年缩短到现在的几分钟和几秒那样。

34:44 我们把这看作是以数字速度进行科学研究。

34:48 我试图将我们在技术领域的最佳成果带入自然科学。

34:56 我梦想有一天能创建一种虚拟细胞，一个计算细胞，也许是非常简单的，比如酵母细胞，并且可以在计算机内进行实验。

35:10 虚拟细胞得出的预测能够指导并形成你在实验室中的实际实验。

35:17 这样可以大大减少湿实验室里的搜索工作，更主要将湿实验室用于验证步骤，而非昂贵且缓慢的搜索流程。

35:29 当然，人工智能的应用不仅限于生物学，还可以用于科学、数学以及更广泛的医学领域。

35:39 我们取得了多项突破，不仅限于生物科学，还涵盖健康领域，例如通过视网膜扫描识别眼病、发现新材料、帮助等离子体聚变反应堆，

35:52 更快的算法。

35:54 人工智能还能为自身发现更优算法，比如更快的矩阵乘法，进行天气预报，甚至协助量子计算机及其误差修正。

36:05 这只是过去两三年我们所做工作的一小部分示例。

36:10 我认为人工智能几乎可以应用于所有领域。

36:14 我总是鼓励大学认真考虑多学科工作，将人工智能应用于特定专业领域的关键问题。

36:25 我相信未来五到十年内，通过这种方式会有许多重要进展。

36:31 最后我想稍微谈谈更为广泛的观点，不仅是人工智能在科学中的应用，还有通用人工智能的发展进程，我们离它有多近，以及我们在通用人工智能原始使命上的更广泛工作。

36:47 我们在理解世界的各个方面都取得了很大进展。

36:53 我们有时称之为世界模型。

36:55 我们特别自豪的是去年年底发布的新视频模型VO2。

37:00 它是最先进的视频生成模型。

37:03 能够仅凭文字描述生成视频，对吧？

37:09 或者仅凭单张静态图片。

37:11 所以，你知道，实际上，虽然有些视频看起来并不那么令人印象深刻，但如果你想想这个切番茄的视频，这就像是视频模型的图灵测试，因为通常你会看到番茄神奇地重新合在一起

37:24 或者你切到了手指，或者刀子飞到别处去了。

37:27 实际上，如果你思考这些系统为了真正理解世界的物理规律或者这里蓝莓周围的气泡所必须做的事情，对吧？

37:36 它仅仅是从文本生成这些东西，你知道吗？

37:38 蓝莓掉入一杯水中。

37:41 它正确地模拟了这些物理现象，或者这些小卡通人物的运动，比如这里的蜜蜂。

37:49 这真的有点令人震惊。

37:50 我想，即使是在五年前，如果你告诉我这有可能做到，而不需要内置某种特殊的物理理解，我也会觉得这不太可能。

38:02 然而，不知怎的，这些学习系统仅通过观看海量的YouTube视频，居然能学会现实世界的物理规律。

38:10 这非常疯狂，竟然有这个可能。

38:13 所以我们已经做到了。

38:15 我们用Genie 2更进一步，当然，我又戴上了游戏制作人的帽子。

38:21 这把那些视频模型推向了新的层次。

38:24 现在，通过文本指令，你可以生成整个游戏。

38:27 比如，我们在底部这里写了，“生成一个机器人在未来城市中可玩游戏世界”。

38:32 然后它就生成了这个。

38:34 你可以用QWE键和方向键控制它。

38:39 目前，这个世界只在几秒钟内保持一致，但我们正努力让游戏世界的连贯性持续更长时间，达到几分钟。

38:48 那样你就真正得到了我所说的世界模型，真正理解现实世界，理解现实世界中的交互以及物理规律。

39:01 当然，我们从一开始的2010年就非常重视安全问题，尽管那时几乎没人关注AI。

39:15 我们设想这将是一项为期20年的使命，而实际上，令人惊讶的是，十五年过去了，我们还算在正确的轨道上。

39:20 我们一直在规划成功。

39:24 如果我们要构建这些变革性的系统和技术，责任重大，我们必须确保它们以安全和负责任的方式部署。

39:35 我们开发的技术之一叫做Synth-ID，它使用AI对抗系统隐形水印，稍微调整像素、文字或音频，人眼或耳朵几乎察觉不到。

39:53 但检测系统能识别这些是合成生成的图像，无论是音频、图片还是视频。

40:02 随着这些技术的广泛应用，能够轻松区分合成图像和真实图像将变得越来越重要。

40:15 AI拥有令人难以置信的潜力，可以帮助解决我们最大的挑战，从气候到健康。

40:23 但显然，这将影响所有人。

40:26 我认为非常重要的是，我们不仅仅让技术人员作出决定，而是要让社会各界广泛参与其中。

40:35 过去几年我非常高兴看到，AI成为主流的一个结果是许多政府和社会各界开始关注它。

40:45 我觉得看到这些国际峰会非常好。

40:47 实际上，英国几年前在布莱切利公园主办了首次会议，汇集了政府首脑、学术界和民间社会，共同讨论这些技术，如何设立合适的防护措施，如何确保抓住机遇，

41:04 同时减少未来可能带来的风险。

41:07 鉴于这些技术呈指数级提升，这将变得越来越重要。

41:14 我对此的简要理解是，硅谷流行的格言是“快速行动，打破常规”。

41:23 这无疑推动了许多技术发展，也带来了我们每天使用的技术。

41:29 但我认为，这种变革性技术不适合这种做法。

41:34 我们应该采用科学方法，带着谦逊和尊重去对待这种技术。

41:44 很多事情我们还不了解。

41:46 关于这项技术未来如何发展，还有许多未知。

41:49 这是非常新的。

41:51 我认为只要我们以非凡的关怀和远见来对待，就能获得所有好处并将弊端降到最低。

41:59 但我认为只有现在开始研究和辩论才行。

42:06 最后，我想说的是，我们现在正在构建自己的大型多模态模型，试图将我展示的所有不同模型的最佳部分融合进一个系统。

42:15 我们称其为双子座系列。

42:16 我们的最新版本是双子座2.0，你们中有些人可能已经试过，它在许多领先基准测试中处于最先进水平。

42:23 我们用它来推进发展，我对下一代助手感到非常兴奋。

42:27 我称之为“通用助手”。

42:29 我们称它为Projastra，它实际上可以装在你的手机或其他设备上，也许是眼镜。

42:35 它开始成为一个可以随身携带的现实世界助手。

42:39 它帮助你丰富日常生活，提高效率。

42:47 AI的下一步，是将我所展示的与AlphaGo这类基于代理的模型结合，这些模型能高效搜索并在有限域内找到问题的良好解决方案，这里的例子是游戏。

43:03 但我们真正想做的是在更通用的模型之上构建这些搜索和规划系统，比如双子座这样理解真实世界运作的世界模型，进而规划并实现现实中的目标。

43:18 这当然对机器人技术至关重要，我认为未来两三年内这一领域将取得巨大进展。

43:30 那么我最后稍作推测，回顾图灵和他为奠定计算机科学基础所做的所有工作，这意味着什么。

43:43 我认为如果你看看我们所做的工作，我某种意义上把自己看作是图灵的继承者。

43:50 图灵机和传统计算的理念能走多远？

43:55 我记得在这里听过的一堂课是我最喜欢思考的话题之一——P=NP问题，这是计算机科学中的著名问题，探讨哪些问题在经典系统上是可解的。

44:10 显然，量子计算系统方面也有很多重要工作在进行。

44:14 剑桥和谷歌都有世界顶尖的量子计算团队。

44:21 很多问题被认为必须依赖量子计算才能解决。

44:25 还有很多现实世界的系统我们希望理解和建模。

44:29 我的推测是，实际上基于经典图灵机的系统，也就是构建这些AI系统的经典机器，它们能做的远比我们以前认为的多得多。

44:40 以AlphaFold和蛋白质折叠为例，蛋白质是量子系统。

44:45 它们在原子尺度上运作。

44:47 有人可能认为要用量子模拟才能找到蛋白质的结构。

44:54 然而，我们用神经网络已经能近似这些解决方案。

45:00 所以我认为，一个潜在的观点是，任何自然界中能被生成或发现的模式——即具有真实物理结构的实体——都可以被这些经典学习算法高效地发现和建模，比如AlphaFold。

45:16 如果这个结论成立，那将对量子力学乃至基础物理带来深远影响，这也是我和许多同事希望探索的领域，可能还需依靠这些经典系统的帮助。

45:31 帮助我们揭示现实的真相。

45:36 这让我回想起多年前我投身人工智能的初衷：我一直相信，这样构建的通用人工智能可能成为理解宇宙及我们自身位置的终极通用工具。

45:50 谢谢大家。

45:51 [掌声] 太棒了。

46:06 如果大家有问题，我们还有时间答疑。

46:09 先来一位，请举手，就这里。

46:11 您好。

46:12 谢谢。

46:13 谢谢。

46:14 感谢您的精彩演讲。

46:16 因为您有神经科学背景，且喜欢从根本问题入手思考。

46:22 在神经科学中，你是否曾遇到过一个根本性的问题，认为值得去解决，并且依然值得去解决，以更好地理解生物智能和人工智能？

46:34 是的，有很多。

46:35 我的博士研究的主题实际上就是记忆，还有想象力。

46:41 也就是未来思考，某种程度上的计划。

46:43 所以我真的很想了解大脑是如何做到的。

46:47 事实证明，海马体在两者中都有作用，所以我们或许可以用某些算法来模拟它。

46:54 所以我认为其中有很多关键点。

46:56 当然，还有关于创造力、做梦、意识等所有重大问题，我认为构建人工智能并与人类大脑对比，是解决这些根本性问题的最佳途径之一。

47:14 比如，意识的本质是什么？

47:16 大脑的物质基础是否有其特殊性，而这与用硅基算法模拟它有什么不同？

47:25 好问题，就问到这里吧。

47:29 你好，我其实有两个问题。

47:33 既然DeepMind成立于深度学习革命之前，我想知道如果深度学习没有兴起，你当时的心态是怎样的，计划如何推进？

47:47 这是第一个问题。

47:48 第二个问题是，既然你亲身经历过这么多高维复杂的问题，而我们知道梯度下降及其变种通常只会收敛到局部最优解，

48:07 你是否觉得这些系统居然能起作用很令人惊讶？

48:14 你觉得大部分自然界本质上是否具有次优性？

48:20 那么我们是否有可能构建出更优的“自然”？

48:23 这两个都是很好的问题。

48:26 第一个问题，之所以叫DeepMind，部分原因是“Deep”指的是深度学习，对吧？

48:33 深度学习的早期阶段，那时它还没被叫作深度学习，但已经逐渐普及。

48:40 那是Jeffrey Hinton发明的玻尔兹曼机等技术。

48:44 在2005、2006年前后，这些层级神经网络出现了，在学术界我们就觉得这是一个非常有前途的想法。

48:56 另外我们也押注了强化学习，以及两者的结合，对吧？

49:02 强化学习虽然如今又流行起来，但对我们解决AlphaGo这样的问题也非常关键。

49:08 所以你需要两方面的技术。

49:09 你需要深度学习来建模环境和世界，再需要强化学习来制定计划、解决方案并在现实世界中行动。

49:18 我们之所以押注这两条路径，即使当时还在起步，是因为我们知道传统方法那些专家系统无法扩展。

49:29 实际上，这也是我在这里以及MIT博士后时学到的，就是这些传统专家系统就像教派一样坚守某套方法，

49:41 你们大学课程里能学到的，不仅是做什么，还要知道不该做什么，以及为什么不行。我认为它们永远无法扩展到我要用AI解决的那类问题。

49:54 而学习系统看起来潜力无限，虽然刚开始很难让它们做成什么有意义的事。

50:02 问题就在于它们当时还没被足够规模化。

50:04 我们2010年创办DeepMind的另一个原因是发现计算范式在硬件方面发生了转变，比如GPU和其他硬件，当然是为游戏开发的，结果发现一切本质上都是矩阵乘法。

50:19 智能和游戏，以及计算机图形学。

50:22 所有这些因素汇聚起来，再加上神经科学的理解、功能性磁共振成像（fMRI）和过去十年神经科学的进展。

50:34 所以我觉得2010年是汇聚这些要素的绝佳时机。

50:39 我们下注，并非因为确定它会成功，而是相当确定其他方法不会成功。

50:46 这就是人工智能寒冬的根源，试图推动那些专家系统。

50:51 至于第二个问题，我不会否认确实很惊讶某些东西能够收敛，我们当时也不确定。

51:01 比如我展示过的Atari游戏那些案例。

51:04 所以，最初的几年，什么都没成功，是吧？

51:06 所以，我们甚至连一分都拿不到，如果你们有人还记得Pong，那是最早的电脑游戏之一，一个简单的网球球拍游戏，你能想象的最简单的游戏，我们都玩不过它。

51:17 所以，我们在想，我们是不是提前了10年、20年？

51:21 就像巴贝奇和他的差分机一样，是吧？

51:24 想法非常棒，机器也能工作，但他最终还是提前了50年或100年。

51:29 我常说，你最好领先时代五年，而不是领先五十年。

51:33 否则，你会经历很多痛苦，就像巴贝奇那样。

51:37 所以，我们很担心，但后来事情得以收敛，这让我们有信心去解决更难的问题。

51:45 我想你问的问题最后部分是关于自然界的事物，嗯，我的想法是，它们不是次优的，实际上可能相当优化，

51:53 因为它们经历了进化过程，不仅是生物学上的生命，还有地质和物理过程，比如小行星和物理现象的结合，它们能存活一段时间，因为它们在时间上是稳定的，是吧？

52:07 如果它们能持续稳定存在，那可能存在某种可学习的结构。

52:11 这就是我的推测。

52:13 - 很好，这里有人提问。

52:17 - 你怎么看待构建高带宽脑机接口和可植入的记忆及推理模块，使人类能够自主发现，而不仅仅是通过云端与AI对话？

52:33 - 我很喜欢这个领域，也一直在关注，并帮助别人制作脑电图帽之类的东西。

52:43 当然，问题在于读取脑信号的分辨率，以及理想情况下还能进行读写操作。

52:50 但我对类似新奥尔良项目或脑中芯片这类项目非常感兴趣。

52:54 显然，目前这些技术主要是为了帮助退伍军人和患者恢复身体功能。

53:00 我认为这方面会有惊人的进展，比如让脊椎受伤的人重新行走等。

53:07 我相信医学领域会有一些令人惊叹的突破。

53:11 但在那之外，或许如果这类技术普及且手术安全，有安全的操作方式，我可以想象这会成为我们跟上技术进步的一种方式。

53:23 某种意义上，这和我们现在身边的技术没什么区别。

53:28 我们随身携带手机、电脑等设备，24小时不离身。

53:32 我们已经和技术几乎共生了。

53:35 当然，有了这种接口，将是更进一步的融合。

53:38 但我不确定，也许在场的哲学家们可以回答，技术附着在你身上和你随身携带之间是否存在边界。

53:52 - 很好，那边有人提问。

53:54 这就是了，好的。

53:57 - 你好。

53:59 - 你确定吗？

54:04 - 嗨，你好？

54:06 啊，来了。

54:08 你怎么看人工智能的发展速度及其对经济发展的影响？

54:16 现在有很多人正在决策职业方向，在这种快速变化的环境下，很难预测应该选择什么行业。

54:27 - 这是个很复杂的问题，正如你说的，变化快得像闪电一样。

54:34 我们刚才还在和阿利斯泰尔讨论这个问题。

54:36 即便设计一个三年的计算机科学课程都很难，因为基础材料在三年内就发生了变化。

54:44 我唯一可以确定的是将会有大量变革，但这也带来了破坏和机遇。

54:52 举个编码的例子，我不知道你是不是计算机科学专业，但我仍然建议你学好编码和数学，因为你能更深入地运用这些新工具，

55:06 如果你了解它们是怎么构建的。

55:08 但另一方面，我认为编码将对更多类型的人更加普及。

55:15 因为借助AI，你可能可以用类似自然语言的方式编程，而不是使用复杂的计算机语言。

55:24 所以这将为创意人士打开新的领域，比如制作游戏、拍电影、开发应用。

55:32 这更多侧重于创意方面，而非工程技术方面的平衡。

55:39 但我也认为，这将增强工程师，让他们的工作效率提升十倍。

55:44 所以我仍然觉得，这很难确定。

55:47 但我要说的是，重点是利用业余时间拥抱这些工具，训练自己快速学习新信息的能力。

56:01 因为我认为未来十年基本上会是这样发展。

56:05 - 好的，这边有个问题，穿黄黑上衣的朋友。

56:12 你认为有没有哪些生物过程或行为模式，现有的深度规划技术无法建模？

56:22 我不是说无限增加模型规模直到可用那种。

56:27 你觉得有哪些过程在物理上无法用我们现有架构建模？

56:31 - 当然现在有很多过程还无法建模，不过这又回到了我演讲结尾提到的观点。

56:39 我不确定从极限上讲是否存在这样的情况。

56:41 我认为，如果物理能够解决，且存在可学习的结构，凭借足够的样本，或许可以反向工程出模型。

56:53 所以我看不出有什么理论理由阻止经典复杂系统对生物系统做出预测或模拟。

57:06 因此我实际上不认为存在什么界限。

57:12 当然，有些抽象问题，比如大数分解、密码学，是人类创造的系统，可能没有规律可循。

57:21 当然，自然数中可能存在某种结构。

57:24 很多人都猜测确实存在。

57:26 如果存在，那也可以被学习。

57:28 如果不存在，且分布均匀，那破解密码等问题可能需要量子计算机。

57:37 所以这些都是开放的猜想，但我认为大多数自然事物是在地质或生物物理时间中进化而成的。

57:46 这就意味着存在可供学习的结构。

57:50 这让搜索和预测变得可能可行。

57:54 - 好的，最后一个问题，请粉红衫的朋友提问。

58:01 - 这个问题代表剑桥大学游戏开发社团提问。

58:07 你提到了G&E2模型，目前它只能稳定运行几秒钟，你希望未来能稳定几分钟。

58:18 但我们社团的问题是，游戏世界的一致性是无限的。

58:25 玩《Minecraft》时，你希望转身时村庄依然存在，对吧？

58:29 - 是的。

58:30 - 那你觉得你现在的模型会被整合到开发流程中吗？

58:34 或者说，你怎么看AI、你的模型以及你正在做的工作？

58:39 你觉得它们将如何融入未来几十年的游戏开发？

58:43 - 好的。

58:44 我认为AI将以多种方式介入游戏开发。

58:48 其中之一是作为构建游戏所需资源的工具。

58:53 比如3D模型和动画。

58:55 我认为这些会在未来几年内实现。

59:00 你也可以把AI看作游戏平衡的工具。

59:03 所以你设想你设计了一款游戏，一夜之间，它可以进行一百万次通关。

59:10 然后早上，你作为游戏设计师就能得到一份报告，说这些地方不平衡，对吧？

59:14 或者降低某个单位的能力，或者其他什么。

59:18 我觉得开放世界游戏的漏洞测试也是如此。

59:20 我曾经做过模拟游戏和开放世界游戏。

59:24 它们的漏洞测试简直是噩梦，因为游戏的核心就是玩家几乎可以做任何事情，游戏会对此作出反应。

59:31 那么你如何测试一千万玩家各自独特的游戏历程呢？

59:36 实际上在发布前让AI玩家试玩，这能帮助你解决许多漏洞。

59:42 最后，我觉得更令人兴奋的是，AI角色更加逼真，推动剧情发展。

59:49 你曾梦想过一个大型多人在线世界，AI角色智能且会根据玩家的行为更新他们的信念和故事线。

1:00:00 这样会感觉是一个更加生动、真实的世界。

1:00:04 我认为我们正处于创造这类游戏的边缘。

1:00:09 最后，我们正在构建的世界模型，更偏向于通用人工智能。

1:00:13 你能否真的模拟这种方式？

1:00:15 这表达了对世界理解的能力。

1:00:17 你的模型理解这个世界吗？

1:00:19 如果它能在一定时间内生成世界，那显然它在某种意义上理解了经验事实，理解了背后的物理规律。

1:00:30 这更多是关于通用智能的，而我想我们也许有一天能拥有一个全息甲板，只需想象，世界就在你周围。

1:00:40 可能一旦我们有了通用人工智能，我们就能实现这个。

1:00:43 但我觉得那还需要一段时间。

1:00:46 太棒了。

1:00:47 看来以游戏为话题结束很合适，又回到了游戏上。

1:00:50 非常感谢大家的到来。

1:00:52 特别感谢Damage今天来和我们交流。

1:00:56 谢谢你们。

1:00:57 (观众鼓掌)

0:00 [APPLAUSE]

0:03 Thanks, Alastair, for that lovely introduction.

0:05 And it's so great to be back at Cambridge.

0:08 I always have a warm feeling when I'm sort of my homecoming

0:12 back to Cambridge.

0:13 And specifically, this lecture hall is Alastair reminded me.

0:15 I think it is the first lecture hall I was in.

0:17 It's always been my favourite lecture hall.

0:19 I remember telling, and I see a lot of my old friends here,

0:22 from my Cambridge days, Aaron, I think, about that one day,

0:27 maybe I come back to give a lecture in here

0:30 and talk about announcing AGI, and maybe a robot would walk on,

0:33 and sound everyone.

0:35 I'm not going to do that today to disappoint you.

0:39 But maybe in a few years' time, I'll come back again,

0:41 and I'll give that lecture.

0:43 But this is an amazing place.

0:45 It's such an inspiring place, and I'm

0:47 going to talk a little bit about how Cambridge has inspired

0:49 my whole career, actually, and hopefully is going

0:51 to do the same for many of you and the students in the room.

0:56 So for me, it started getting my journey on AI,

1:00 started with games, and specifically chess,

1:02 as Alastair mentioned.

1:04 So I was playing chess from the age of four years old,

1:06 and very seriously for the England junior teams

1:09 and things like that.

1:10 And it got me thinking about thinking itself.

1:13 How does our mind come up with these plans, with these ideas?

1:17 How do we problem solve?

1:18 And how can we improve?

1:19 Obviously, when you're playing chess at a young age

1:21 and you're trying to play competitively,

1:22 you're trying to improve that process.

1:25 And it was fascinating to me, perhaps more fascinating

1:28 than even the games I was playing,

1:29 was the actual mental processes behind it.

1:33 And in fact, AI and computers, I came across computers

1:37 in AI for the first time in the context of chess

1:40 and trying to use very early chess computers,

1:42 like the one on the right here.

1:44 I think this was my first ever chess computer.

1:46 There were physical boards where you actually

1:48 had to press the squares down to move the pieces.

1:51 And of course, we were supposed to be using these chess

1:53 computers to train opening theory and learn more about chess.

1:58 But I remember being fascinated by the fact

2:00 that someone had programmed this lump of inanimate plastic

2:03 to actually play chess really well against you.

2:07 And I was sort of really fascinated by how that was done

2:10 and how someone could program something like that.

2:13 And I ended up experimenting myself in my early teenage years

2:17 with the Amiga 500 computer, amazing home computer

2:20 back in the late '80s and early '90s.

2:23 And building those kinds of AI programs myself

2:26 to play games like Othello.

2:27 And really, that was my first taste of AI.

2:30 And I was hooked from then on.

2:32 And I decided from very early on that I

2:34 would spend my entire career trying

2:36 to push the frontiers of this technology.

2:41 So then that led me to Cambridge,

2:45 and which was really my three years here

2:47 were incredibly formative for me.

2:49 And I went to a school in-- a compensator school in North

2:53 London.

2:54 No one had ever gone to Oxbridge for in living memory.

2:59 But the reason I wanted to come to Cambridge

3:01 was all these inspiring stories that I'd heard about what

3:05 happened at Cambridge, all these amazing people

3:07 that I used to read about their biographies

3:09 and the work they'd done, especially people

3:12 like Crick and Watson in the top left there.

3:14 And I remember particularly a film,

3:17 The Race for the Double Helix.

3:18 It was an amazing film from the '80s,

3:20 if you haven't seen it, with Jeff Goldblum

3:22 at one of his early parts, with all the enthusiasm

3:25 that he plays, all of his parts, as he was Watson.

3:28 And they were having just such an amazing time

3:31 discovering, roaming around Cambridge,

3:34 working on things like DNA.

3:35 And I thought, that's what I want a piece of that.

3:38 And I want to feel what it's like

3:40 to be at the frontier of discovery

3:42 and what could be more exhilarating.

3:44 And that film actually really brought it to life

3:47 what that might be like.

3:50 And then, of course, all of my heroes, my scientific heroes,

3:54 a lot of them had gone through Cambridge.

3:56 People are Alan Turing and Charles Babbage, of course,

3:58 in the lecture hall that we now sit in.

4:02 And even places like the Eagle Pub,

4:04 where if you start at Queen's College,

4:06 one of the tours they give you on the first days

4:08 to go and show you the putered sort of table

4:12 that they were discussing and the DNA structure around.

4:17 And you can't help but be inspired by that

4:20 and walking down King's Parade.

4:22 And I almost felt like the intellectual giants of the past

4:26 were almost speaking to you from the stones.

4:29 And that's how I felt going for a late-night burger

4:33 at Gardeniers, that that was what was inspiring me around,

4:36 like all of these amazing people

4:37 that had walked those same steps over hundreds of years.

4:41 And that's the history that is sort of unrivaled here

4:45 at Cambridge that I think we can still draw on

4:47 and take inspiration from today.

4:49 And then there's a picture of me and Aaron there,

4:52 one of my best friends from Queen's,

4:54 obviously on the mathematical bridge there.

4:57 And then finally, when we, Alissa mentioned, obviously,

5:00 the Nobel Prize, and it was an honor of a lifetime

5:03 to go and collect that in Stockholm in December.

5:07 Amazing one week of activities.

5:10 But my favorite activity is when you get to sign

5:12 the Nobel book in the Nobel Foundation.

5:15 And that's the book there and one of my pictures of,

5:18 and I started sort of leafing through the book.

5:20 You sign your name and then you leaf back.

5:22 You wonder if it's crick in there and of course he is.

5:25 And then you go back further and then Einstein's signature

5:27 is there and it's just mind-blowing, really.

5:29 And I started to spend an hour just

5:31 photoing every page of the book.

5:35 So it's full circle for me of that picture.

5:39 And then really seeing that film in the late '80s.

5:45 So then in 2010, we started DeepMind in London.

5:50 As really at the time, it was a kind of an Apollo program

5:54 effort is the way we thought of it

5:56 for trying to build artificial general intelligence.

5:58 AI that was truly general and could

6:01 perform all the cognitive capabilities

6:04 that humans are capable of.

6:06 So it would be a truly general AI system.

6:09 In fact, the idea for that really comes from Turing

6:12 and Turing machines.

6:13 So something that's able to compute anything that

6:17 is computable as Turing showed with his Turing machines.

6:20 And really that's been the foundation for me

6:23 of one of the main things that I carried with me

6:26 from the lectures here at Cambridge

6:27 was all these theoretical underpinnings of computer science

6:31 and computation theory that people like Turing and Shannon

6:35 famously did in the '40s and the '50s.

6:39 So we started in 2010.

6:40 And it's amazing.

6:41 It's 15 years ago, which in some ways isn't that long ago.

6:44 But when we started out DeepMind, almost nobody

6:47 was working on AI, which is hard to believe today,

6:49 given that almost everyone seems to be working on AI today.

6:53 In just a matter of just over a decade,

6:55 things have gone accelerated incredibly.

6:57 And obviously, we've been part of that very exciting journey.

7:02 So our mission at DeepMind, from the beginning was,

7:06 we talk about building AI responsibly to benefit humanity.

7:09 But the way we used to articulate it when we started out

7:11 was in a two-step process, step one, solve intelligence,

7:15 step two, use it to solve everything else.

7:18 And it seemed very outlandish at the time in 2010.

7:21 And you could imagine trying to pitch

7:23 a venture capitalist on the basis of that mission.

7:26 It seemed pretty crazy.

7:27 But I still fundamentally believe in that today.

7:30 And I think more and more people are realizing

7:32 that AI, built in this general way,

7:35 could have these profound and transformative impact

7:38 on almost any field, which is obviously

7:41 the second part of that mission statement.

7:43 And I think, for me, that involves

7:46 accelerating scientific discovery itself, and medicine,

7:50 and advancing our understanding of the universe around us.

7:56 So back when we started out, there were basically two ways.

7:59 And in fact, when I was studying here in the '90s,

8:01 there are two ways to build AI, broadly speaking.

8:04 There's the expert system way, which

8:06 is you kind of pre-program an expert system directly

8:09 with the solution, things like deep blue

8:12 that beat Carrie Kasparoff at chess very famously

8:14 in the '90s, actually, while I was studying here.

8:17 That would be the pinnacle example,

8:19 probably, of an expert system.

8:21 But the problem with these expert systems was,

8:23 and why they never really scaled to full general intelligence,

8:26 is they can't deal with the unexpected.

8:30 If something unexpected happens that you didn't already

8:33 cater for, there's nothing in the system

8:35 that will allow it to deal with that.

8:38 And they're sort of inspired by logic systems,

8:41 and they were quite rigid and fragile and brittle

8:43 because of that.

8:45 Whereas the modern day approaches are built on learning systems.

8:48 So these are systems that are able to learn for themselves

8:51 and learn directly from experience or data

8:53 from first principles, and really inspired

8:56 by more neuroscience ideas.

8:58 And obviously, the promise of these systems that we have today

9:01 is that they can go beyond the knowledge, potentially,

9:04 that we, as the programmers or the system designers,

9:07 already know how to solve.

9:08 And of course, that's extremely valuable in areas

9:11 like scientific discovery.

9:15 So we started in the early 2010s with games, of course.

9:19 And I've used games many times in my life, first of all,

9:22 to train my own mind.

9:23 Then I used to build games in AI4, computer games,

9:27 and then finally, in a third way to train up our AI systems.

9:30 And games are the perfect proving ground for AI systems.

9:34 You can start with very simple games, like Atari games,

9:36 from the '70s.

9:38 And really, this system, DQN, was the first time

9:42 anyone had built an end-to-end learning system that

9:44 could learn directly from raw data.

9:47 So in this case, the raw pixels on the screen.

9:49 And it's not told anything about the games

9:52 or anything about what it's controlling.

9:53 It's just told to maximize the score based

9:56 on this video stream input, pixel stream input.

10:01 So we were able to master all different Atari games

10:05 in sort of around 2013.

10:07 Then we took these systems and we scaled them up

10:11 to really, I would say, the grand challenge of games AI,

10:14 which is can you create systems that

10:17 can play the game of Go at world champion level or beyond.

10:22 And Go, of course, is probably the most complex game

10:26 that humans have ever invented.

10:28 It's thousands of years old.

10:29 It's also the oldest game and one of the most elegant games.

10:33 But one of the ways you can just see the complexity of Go

10:35 is that there are 10 to the power 170 possible positions in Go.

10:40 So that's way more than there are atoms

10:42 in the observable universe.

10:44 And the important reason to point about that

10:46 is that you cannot come up with a strategy in Go

10:49 using brute force techniques.

10:51 It would be impossible, it would be totally intractable.

10:53 So you have to do something much smarter.

10:56 And we famously, in basic 2016, won a million dollar challenge

11:00 match against 10 times world champion, Lee Sedol.

11:04 And one of the legends of the game, a South Korean grandmaster,

11:08 and was watched by 200 million people around the world.

11:11 And not only did AlphaGo our system win that match,

11:14 importantly, it actually came up with new original Go

11:18 strategies.

11:19 Even though we've played Go for thousands of years

11:22 and professionally for hundreds of years,

11:25 it was still able to find never seen before strategies.

11:28 Most famously, this move 37 move here in red during game 2,

11:33 which if you watch the documentary on this,

11:36 you'll see-- which is on YouTube-- you'll

11:37 see that how surprised the best players in the world were.

11:42 They were commentating on the game about this move.

11:44 It was sort of an unthinkable move.

11:46 And yet, it decided to get this game

11:49 to in favor of AlphaGo 100 moves later.

11:54 So again, that told me about the potential

11:57 for these types of systems to invent and discover new knowledge.

12:01 So here, of course, we're just talking about game knowledge.

12:03 But obviously, my dream was to generalize

12:06 this to all areas of scientific discovery.

12:09 So how do these systems work?

12:14 We basically train up these neural networks

12:17 through a system of self-play.

12:19 So this is actually AlphaGo and also subsequent systems

12:23 that AlphaGo Zero and AlphaZero that

12:25 generalize what we've done for Go to play any two-player game

12:30 from scratch.

12:31 And you start off with a version 1 of the system

12:33 that doesn't really know anything about the game, just

12:36 the rules.

12:36 And it plays randomly.

12:38 And you play, say, 100,000 games against itself of this system.

12:43 And that creates a new database of game positions

12:47 from those 100,000 games.

12:49 And from that, you train a second version,

12:51 a slightly better version of the model, version 2,

12:55 that's trained to predict what the likely moves are

12:58 to be played in any one position,

13:01 and also who is more likely to win,

13:03 which side, black or white, is likely to win from that position.

13:06 And what percentage chance do they have of winning?

13:10 And then you can use that version 2 to play against version 1

13:15 and in a 100-game match-off.

13:17 And then if it wins by a significant margin-- so in this case,

13:21 55% win rate-- you replace the version 1 with version 2,

13:26 and you create a new database of games that

13:29 are slightly higher quality, and then you

13:31 learn a version 3 system.

13:33 And if you do this and you repeat it around 17, 18 times,

13:36 you go from playing randomly in the morning to 24 hours

13:40 or less later by the version 17 or 18,

13:44 you're stronger than the world champion level.

13:46 So it's quite an incredible thing process

13:49 to see this self-improvement process playing out

13:52 in a very, very short amount of time.

13:56 So if we think about what these neural networks are doing

13:59 is you're kind of reducing down this intractable search

14:03 space of 10 to the power 170 possibilities down

14:07 to something that's much more tractable

14:09 in a few minutes of compute time.

14:11 And it's doing this by narrowing down

14:15 using the neural network to efficiently guide

14:17 the search mechanism.

14:19 So if you think about this tree of possibilities

14:22 as-- and each of the nodes in this tree is a go position,

14:26 then instead of having to look at sort of every possibility,

14:30 you can actually use the neural network

14:32 to guide you just down the most interesting

14:36 and most useful lines to examine.

14:39 So in this case, the one's in blue.

14:41 And then after you've run out of thinking time,

14:44 you pick the best line, the most promising line

14:46 that you've seen thus far.

14:48 So in this case, this particular line in purple.

14:54 So this then leads to-- we then did not just go,

14:58 but any two-player perfect information game.

15:00 And it was even able to discover new strategies

15:03 and new styles of playing chess, which is kind of extraordinary,

15:07 but given that chess computers were so strong already.

15:11 So programs like Stockfish.

15:13 And AlphaZero was able to beat Stockfish at the time at chess,

15:18 which is almost sort of impossible to do.

15:20 And not only did it beat Stockfish,

15:23 so AlphaZero here is in white playing against Stockfish,

15:26 who's black.

15:27 But in this particular position, one of the most famous games

15:29 that AlphaZero played is called the Immortal Zugswang game.

15:34 White is winning here because it favors mobility over material.

15:38 So most chess computers favor material.

15:40 And you'll see that black-- those of you who play chess,

15:42 you'll see that black has more material,

15:45 but actually can't move any of its pieces.

15:47 They're all stuck in the corner.

15:49 And this is AlphaZero's sacrifice material for this mobility.

15:53 And actually, for human grandmasters and top chess players,

15:57 this is not only very effective styles,

15:58 but a very beautiful aesthetic style to play chess in.

16:03 So it's kind of amazing that AlphaZero was able to sort of discover

16:07 this new dynamic way of playing.

16:10 And in fact, some of the top chess players in the world

16:12 commentated about this.

16:13 So Gary Kasparov, my all-time sort of favorite chess player,

16:17 he said that programs usually reflect priorities

16:19 and prejudices of the programmer.

16:21 But because AlphaZero learns for itself,

16:23 I would say that its style reflects the truth.

16:25 And then the current world champion at the time,

16:27 Magnus Carlsen, said, read and looked at these games

16:32 and read about the books that were written about AlphaZero.

16:34 And I said, I've been influenced by one of my heroes recently,

16:38 one of which is AlphaZero.

16:39 So he actually incorporated a lot of these ideas into his own game

16:44 to dominate the chess scene for almost a decade now.

16:49 So we did all these landmark breakthroughs in Games AI

16:52 and over the first, I guess, sort of decade of DeepMind's existence.

16:57 But of course, these were just the training ground

16:59 for what we wanted to do and was just a means to an end.

17:03 It wasn't the end in itself to play these games,

17:05 much as I love games.

17:07 It was to create these algorithms that could be generally

17:10 useful for tackling real world problems.

17:14 So what we look for in real world problems,

17:17 not only scientific problems, but actually

17:20 industrial problems as well.

17:22 And we look for sort of three different criteria

17:24 that make it suitable, the problem suitable,

17:27 to be tackled by these types of AI systems and ideas

17:31 and algorithms that we developed for playing games.

17:34 Number one is we look for problems that can be described

17:37 as massive combinatorial search spaces.

17:40 So usually far too complex, far too many combinations

17:45 to brute force the solution.

17:47 But maybe there's some kind of structure

17:49 that we can learn about with our neural networks that

17:52 can guide that search very efficiently.

17:55 Secondly, we look for problems that

17:57 can be described with a clear objective function

17:59 or some sort of metric that you can optimize against.

18:03 So in games, that's very easy.

18:04 It's things like maximizing the score or winning the game.

18:08 But actually, there's a lot of real world problems

18:10 that you can boil down to a few metrics or a few objective

18:14 functions that you're trying to maximize.

18:16 And then finally, of course, you need

18:18 quite a lot of data or experience to learn from

18:21 and/or, ideally it's and a kind of accurate and efficient

18:25 simulator so you can generate more synthetic data

18:29 to augment the real data that you have.

18:32 And it turns out that there are a lot of problems

18:36 that can be couched in these terms

18:38 if you're looking at the problem from this angle

18:41 and including many, many important problems in science.

18:45 And the one that I always had in mind, actually,

18:47 from my days of first, coming across the problem

18:50 here at Cambridge as an undergrad,

18:52 is the protein folding problem, which I'll just quickly

18:54 describe to you for those that don't know about biology

18:57 and proteins.

18:58 So proteins are incredibly important.

19:00 They're the building blocks of life, pretty much every function

19:05 in the living body depends on proteins

19:09 from your neurons firing to your muscle fibers twitching.

19:12 So really, proteins are what makes life possible.

19:17 And the protein folding problem, then, is really easy

19:19 to describe.

19:20 It's basically a protein is defined by its gene sequence,

19:26 its genetic sequence, which then specifies an amino acid

19:30 sequence, which in nature then folds up spontaneously

19:35 into usually a very beautiful protein structure.

19:38 So you go from this genetic sequence to a protein structure.

19:42 And the reason the protein structure, the 3D structure,

19:44 is very important, is it goes a long way

19:46 to defining what function it has, what it does in the body.

19:51 So it doesn't totally describe the function,

19:54 but it has a big part to play in what it actually does in nature.

19:58 So the protein folding problem, then,

20:00 is this problem of can you predict the protein structure

20:04 directly from this one-dimensional amino acid sequence?

20:07 Can you predict that computationally, that incredible 3D

20:10 structure from that sequence?

20:15 So why is this such a hard problem?

20:17 Well, Levental is a famous protein researcher

20:21 in the '60s, described this conjecture

20:25 that became known as the Levental's paradox, which

20:29 is that he calculated there roughly 10 to the 300 possible

20:34 shapes that an average protein can take.

20:37 And yet somehow, in nature and in the body,

20:40 these proteins fold up spontaneously

20:43 in a matter of milliseconds.

20:44 So that's the paradox.

20:46 If there's so many possibilities, how does nature do this?

20:50 And basically, how does physics achieve this?

20:53 And that gives you hope that this

20:55 must be tractable computationally in some reasonable amount

20:57 of time, because physics does solve this problem billions

21:03 of times a second in the body.

21:06 And furthermore, what attracted me to this problem

21:09 was that there was a biannual competition called

21:11 the CASP competition, which is like-- you

21:13 can think of it as like the Olympics for protein folding.

21:16 And it happens every two years.

21:17 And it's run by some amazing people

21:19 led by Professor John Malt, University of Maryland.

21:23 And it's been running since 1994.

21:26 And it's a great competition, because they

21:28 work with experimentalists who painstakingly find

21:33 these structures using very exotic and expensive equipment

21:37 like electron microscopes.

21:39 And they newly discover structures

21:42 that haven't been published yet.

21:43 They put it into the competition.

21:45 So you actually, the competition organizers

21:48 know what the ground truth is.

21:50 But the computational teams, hundreds of teams

21:53 enter every year, every competition, every couple of years.

21:56 And they try with their computational methods

21:58 to predict those structures.

22:00 And it's usually around 100 proteins

22:02 that are in the competition.

22:03 And then at the end of the summer,

22:05 they reveal what the true structures are.

22:07 And you can compare the predicted ones and their distances,

22:10 the error in the predictions, to the real structures.

22:17 So we entered Alpha Fold 1, actually,

22:20 for the first time in 2018.

22:22 And we started this Alpha Fold project in 2016, actually,

22:25 pretty much the day after we got back

22:28 from the Alpha Go match in Seoul, in Korea.

22:32 And I felt that we were ready.

22:33 We had the techniques that were mature enough

22:36 and ready to now be applied outside of games

22:38 and to try and tackle really meaningful problems.

22:43 We called them kind of root-note problems,

22:44 because if they could be solved,

22:46 they open up whole new branches and avenues of discovery

22:49 that can be built on top.

22:51 And protein folding was a prime example of that.

22:54 So we started working in 2016.

22:57 Alpha Fold 1 was ready after a couple of years,

22:59 and we entered it into the Cas 13 competition.

23:01 And you can see, for the decade prior,

23:04 these bar charts are showing the winning score

23:07 of the winning team in the hardest category,

23:11 actually the hardest proteins that were being predicted.

23:13 And you can sort of think of it as a percentage accuracy

23:16 of how many of these amino acids

23:20 have you got in the right position

23:21 within a certain tolerance,

23:23 within the sort of width of an atom,

23:24 you need to predict it within.

23:26 And you can see there was sort of not much progress

23:29 for a decade, and we'll stuck at this 60 points level,

23:34 and you wait for the.

23:35 - Which is effectively, if you got to 90,

23:43 you would be within the width of an atom.

23:45 So you'd be at atomic accuracy.

23:46 And that's what we all told by experimentalists

23:49 was the accuracy you'd had to reach

23:50 so that it was competitive with experimental methods.

23:53 So the experimentalists could actually rely

23:56 on these predictions rather than having

23:58 to necessarily do the laborious painstaking work

24:01 to find that structure.

24:04 And just as a rule of thumb,

24:06 my biologist friends would always tell me

24:07 that, you know, take a PhD student,

24:10 their entire PhD, so four or five years,

24:13 to fold, to find the structure of just one protein.

24:16 And there are 200 million proteins known to science

24:21 and 20,000 proteins in the human proteome.

24:27 So we were at Alpha Fold 1,

24:29 we were able to win this competition

24:31 and sort of be better by almost 50% than the next best system.

24:36 And Alpha Fold 1, for the first time,

24:38 introduced machine learning techniques

24:40 as the main component of the system for the first time.

24:44 But it was not enough to reach this atomic accuracy.

24:47 We actually had to go back to the drawing board

24:49 with what we'd learned, and we architect it

24:51 for Alpha Fold 2 from scratch,

24:54 using all the learnings we had from Alpha Fold 1

24:56 to finally reach this atomic accuracy.

24:59 And that led the organizers to declare

25:01 that the problem had been solved in the end of 2020.

25:04 So this is an example of how Alpha Fold works visually.

25:09 So you can see on the left-hand side here,

25:11 it's a very complex protein.

25:13 The ground truth is in green,

25:14 the predicted structure is in blue,

25:17 and you can see how closely the blue overlaps the green.

25:21 And then on the right-hand side,

25:22 you can see how Alpha Fold 2 works.

25:24 It sort of builds up that structure in the iterative process.

25:27 It kind of recycles itself actually over 192 steps.

25:31 And then builds out starts as a scrunch ball

25:34 of protein matter, amino acids,

25:36 and then it builds out a more more plausible structure.

25:39 And at the end, it sort of refines the last parts of it

25:42 until it has the finished prediction.

25:44 We immediately then try to,

25:50 because Alpha Fold is so accurate,

25:52 but it's not just accurate, it's also extremely fast.

25:54 It's able to fold proteins in a matter of seconds,

25:57 an average protein.

25:58 We realized quite quickly

25:59 that we could actually fold all 200 million proteins

26:02 known to science.

26:05 And we, over the course of a year,

26:06 we used a lot of computers on the Google Cloud

26:08 to fold all of them,

26:10 and then put them out freely on a database

26:13 with our colleagues at Ember EBI,

26:14 just up the road at the Sanger Center,

26:18 just outside of Cambridge.

26:20 And we provided that for free, unrestricted access

26:22 to anyone in the world to use it.

26:24 And that 200 million proteins,

26:27 if you think about how long it takes

26:28 to do that experimentally for five years,

26:31 it's kind of like a billion years of PhD time

26:34 done in one year.

26:35 So, you know, it's kind of amazing

26:38 to think about how much science could be accelerated.

26:41 And it opened up whole new avenues of exploration,

26:44 because many of these structures,

26:47 especially for the less well-studied organisms,

26:50 like some certain types of plants,

26:52 they're very important for science,

26:54 and for agriculture, and agriculture research.

26:57 But that you wouldn't have any,

27:00 almost none of those structures

27:01 will have been found and be available.

27:03 So now those are all available.

27:05 And also 200 million,

27:06 you can look at them at an aggregate level,

27:09 and sort of look at structures across species,

27:13 and kind of meta structures,

27:16 and see what the commonalities are through evolution.

27:19 So there's really interesting new avenues

27:21 of branches of structural biology now

27:24 that are being explored.

27:26 And of course, we thought about safety from the beginning,

27:28 and we take our responsibility very seriously

27:31 at the forefront of AI.

27:32 And in this case, we, you know,

27:34 consulted with over 30 biosecurity

27:36 and bioethics experts to make sure

27:38 that what we're putting out into the world,

27:41 the benefits of that far outweighed

27:43 any risks associated with it.

27:46 And now I'm very proud to say that, you know,

27:48 over 2 million researchers are using it

27:51 from pretty much every country in the world.

27:53 And it's been sort of cited over 30,000 times now,

27:57 and it's become a standard tool in biology research now.

28:02 And many of you in the audience who are PhD students,

28:04 hopefully you're using it and making use of it.

28:06 And it's just part of the standard canon now

28:09 that is used for biology research.

28:13 It's been amazing to see what other researchers have done

28:16 with all of this technology and all these structures.

28:20 I've just called out six of my favorite examples.

28:25 People at the University of Portsmouth are using it,

28:27 research groups using it to tackle plastic pollution

28:30 in the environment, trying to design new enzymes,

28:33 which are types of proteins that can digest plastic.

28:37 We're working with the Fleming Center

28:39 on antibiotic resistance, neglected diseases,

28:44 like tropical diseases that affect the poorer parts of the world.

28:46 We work with Drugs for Neglector Diseases Institute.

28:50 And here's a good example of where we can accelerate research

28:53 in those areas where, you know,

28:55 whether it's malaria, leishmaniasis, Zika virus,

28:59 a lot of those structures are not known.

29:01 But now they can go straight to drug discovery

29:04 because they have a lot of the information

29:06 about the structures for those viruses and bacteria.

29:11 And then there's been a lot of fundamental research being done

29:13 on things like finding the structure of the nanopore complex,

29:17 which is a very important protein that lets nutrients in and out

29:22 of the nuclear pore of the cell.

29:24 It's an amazing work at the Broad Institute done

29:26 with drug delivery, sort of designing molecular syringes,

29:29 redesigning proteins that can deliver drugs targeted

29:33 to a particular part of the body.

29:35 And it's even been used in things like looking

29:38 at mechanisms of fertility.

29:40 So the amount of things that it's being used for is sort

29:43 of almost every area of biology and medical research now is making use

29:48 of alpha-fold.

29:51 We've continued in the last few years to develop more,

29:54 struck more developments in improving the systems.

29:58 We released alpha-fold 3 earlier this year for academics to use.

30:04 And we've extended now alpha-fold 3 to deal with interactions.

30:08 So you can think of alpha-fold 2 as a picture

30:11 of the static protein structure,

30:14 but really biology is a dynamic process.

30:17 So you need to understand how different biological elements interact

30:21 with each other.

30:22 So of course proteins with other proteins,

30:24 but also proteins with others, other molecules important to life,

30:28 things like DNA and RNA, and also ligands.

30:31 So small molecules, which, you know, things like drug compounds,

30:36 how does the protein bind with that compound?

30:41 And then we have a separate set of work, alpha-proteo,

30:44 which is doing the reverse of alpha-fold,

30:46 about making use of the alpha-fold techniques,

30:49 where if you want to design a novel protein,

30:52 maybe that doesn't exist in nature for a particular job,

30:55 a particular function, what is the amino acid sequence

30:59 and the genetic sequence that will give you that structure?

31:02 So it's kind of like running it in reverse

31:05 and trying to design new structures that will do novel things,

31:09 and again, could be extremely useful for designing drugs

31:13 and things like antibiotics and antibodies.

31:16 So taking a step back then, what is--

31:22 you know, if I look at all the work we've done in the last 15 years,

31:25 what's the implications for science and also

31:28 machine learning?

31:30 And if you think about what we've done with, first of all,

31:32 our games work, and then now with the scientific work

31:35 that we've been working on, of which alpha-fold is our best example,

31:39 it's all about making this search tractable.

31:42 You have this incredibly complex problem.

31:44 There's many, many possible solutions to the problem,

31:47 and you've got to find the optimal solution,

31:50 kind of needle in the haystack of that enormous combinatorial

31:54 search space.

31:55 And you can't do it by brute force,

31:57 so you have to learn this neural network model,

32:01 so it sort of learns about the topology of the problem

32:04 so that you can efficiently guide the search

32:07 to reach your-- to maximise or find the optimal solution

32:11 to the objective that you have in mind.

32:14 And I think this is an incredibly general way--

32:17 a general solution, an incredibly general way

32:19 to approach a whole myriad of problems.

32:22 And so we think about back to the Go example.

32:25 So we're trying to use these systems to find the best Go move,

32:28 but you could also change those nodes to be chemical compounds,

32:33 and now you're trying to find the best molecule in chemistry

32:37 space, in chemical space, and the best molecule,

32:40 and this is the beginning of drug design,

32:43 and that will bind specifically to the target you're interested in,

32:46 but nothing else.

32:47 So it reduces the side effects and the toxicity of that compound.

32:52 And it's a very, very similar techniques

32:54 that we're using in order to design these molecules now

32:58 as the next steps as we move more and more

33:00 into drug discovery.

33:03 So I think in biology, at least, I

33:05 feel like we're entering a new era now of what

33:08 I like to call digital biology.

33:10 So I think of biology in its most fundamental level

33:13 as an information processing system

33:15 that's trying to resist entropy around it.

33:18 And I think that's basically what life is.

33:21 Of course, it's a phenomenally complex and emergent

33:24 information processing system.

33:26 And I think that's where AI comes in, just like maths

33:29 and the maths that I learned in this room

33:31 was the perfect description language for physics

33:34 and physical phenomena.

33:37 I think that AI is potentially the perfect description language

33:40 for biology.

33:41 It's perfect for dealing with the complexities

33:43 of the emergent behaviors and interactions

33:46 that you get in a dynamic system like biology.

33:50 And I think alpha fold is a proof point of that.

33:53 And I hope when we look back in 10 years' time,

33:55 it won't be an isolated breakthrough,

33:57 but it actually has sort of heralded

34:00 in this new era, golden era of digital biology.

34:04 And we're trying to progress that ourselves.

34:06 We started a new spin-out company, Isomorphic Labs,

34:09 to build on our alpha fold technology

34:11 and move more into the chemistry space

34:12 that I was just talking about.

34:14 And actually try and reimagine drug discovery

34:16 from first principles with AI.

34:18 And right now, it takes an average of 10 years

34:22 for a drug to be developed.

34:24 And it's extraordinarily expensive,

34:25 it costs billions and billions of dollars.

34:27 And so I'm thinking, why can't we use these techniques

34:30 to reduce that down from years to months,

34:33 maybe even one day weeks,

34:35 just like we reduced down the discovery

34:37 of protein structures from potentially years

34:40 down to now minutes and seconds.

34:44 And so we think of this as a kind of doing science

34:47 at digital speed.

34:48 So I'm trying to bring the best of what we do

34:51 in the technology area to the natural sciences.

34:56 And my dream one day is to be able to create

34:59 maybe a kind of virtual cell, a computational cell,

35:03 perhaps of something very simple, like a yeast cell,

35:06 and that you can actually run experiments in silico on it.

35:10 And the predictions that you get out of the virtual cell

35:12 will actually tell you and form

35:14 your real world experiments in the lab.

35:17 And you can reduce down a lot of the search

35:19 that's done in the wet lab

35:20 and actually more use the wet lab for validation steps

35:24 rather than the very expensive and slow search process.

35:29 Of course, AI, and we've been using AI

35:33 not just in biology, but it can be used for science,

35:36 mathematics, medicine more generally.

35:39 And we've had a whole range of breakthroughs,

35:41 not just in the biological sciences,

35:43 but from things in health,

35:44 like identifying eye disease from retinal scans,

35:47 discovering new materials,

35:49 helping with plasma containing fusion reactors,

35:52 faster algorithms.

35:54 So AI discovering better algorithms for itself,

35:57 like faster matrix multiplication,

35:59 doing weather prediction,

36:01 and even helping with quantum computers

36:02 and error correction in quantum computing.

36:05 And that's just a small example

36:06 of some of the work we've been doing

36:07 in the last two, three years.

36:10 And I think AI will be almost be applicable

36:12 to pretty much every field.

36:14 And I always encourage actually universities

36:17 to start thinking very seriously

36:19 about multidisciplinary work

36:21 where you apply AI to the right questions

36:24 in a particular specialist field.

36:25 And I think there's many, many advances

36:28 to be made over the next five to 10 years by doing that.

36:31 So I'll just end then with a little bit

36:35 of a more general view about not just AI for science,

36:40 but the path to AGI and how close we are to that

36:42 and our more general work on the original mission of AGI.

36:47 And we've been making a lot of advances

36:49 in all areas of general understanding of the world.

36:53 We sometimes call them world models.

36:55 So we're particularly proud of our new video model

36:57 called VO2, which was just released

36:59 at the end of last year.

37:00 It's a state-of-the-art video generation.

37:03 And it's able to generate these videos

37:06 just from a text description, right?

37:09 Or a single static image.

37:11 So, you know, and actually,

37:13 although some of these videos may not seem that impressive,

37:15 if you think about this chopping the tomato one,

37:19 this is like the Turing test for video models

37:21 because usually you get, you know,

37:22 the tomato comes magically back together

37:24 or you're chopping through the fingers

37:26 or the knife moves off somewhere.

37:27 And it's actually, if you think about what the systems

37:29 had to do to really understand the physics of the world

37:33 or the bubbles around this blueberry here, right?

37:36 It's just generating that from text, you know?

37:38 Blueberries dropping into a glass of water.

37:41 And it's doing all the physics of this correctly

37:43 or the motion of, you know,

37:45 these little cartoon characters or the bee here.

37:49 It's kind of mind-blowing, really.

37:50 And I think, even if you told me five years ago

37:53 that this would be possible without sort of building

37:56 in some special understanding of physics or something,

37:59 I would have told you that that seems unlikely

38:01 that that would be possible.

38:02 But yet, somehow, these learning systems

38:04 are able to learn about real-world physics

38:07 just from watching, you know, many, many YouTube videos.

38:10 And it's pretty crazy that that's possible.

38:13 So we've done that.

38:15 We've gone a step further with Genie 2,

38:17 which, of course, bringing my games hat back in here.

38:21 And this is taking those VO models a step further.

38:24 So now, with the text instruction,

38:26 you can generate a whole game.

38:27 So, you know, here, at the bottom here,

38:29 we said, generate a playable world

38:31 as a robot in a futuristic city.

38:32 And then it just comes up with this.

38:34 And you can control it with, you know, QWE keys

38:37 and the arrow keys.

38:39 At the moment, it's only consistent for a few seconds,

38:42 but we're working to extend that to the consistency

38:46 of the game world last for many minutes.

38:48 And so then you've really got what I would call a world model,

38:51 a really an understanding of the real world

38:54 and how interactions in that real-world work

38:57 and the physics of the real-world work.

39:01 Of course, we've been working very hard

39:03 on the safety aspects of this

39:05 and from the very beginning in 2010,

39:08 we were working on planning for success,

39:11 even though almost nobody was working on AI back then.

39:15 We imagined that it would be a 20-year mission

39:17 and actually, amazingly, we're sort of on track 15 years in.

39:20 And we were sort of planning for success.

39:24 If we were to build these kinds

39:25 of transformative systems and technologies,

39:29 it would come with a lot of responsibility

39:31 as well to make sure they get deployed in a safe and responsible

39:35 way.

39:35 And one of the systems we technology we built

39:38 is called Synth-ID, which invisibly watermarks, actually

39:43 using an AI system, an adversarial AI system,

39:46 slightly adjusts the pixels or the text or the audio

39:50 imperceptibly to the human ear or eye.

39:53 But it can be detected by a detection system

39:56 that these were synthetically generated images,

39:59 whether that's audio, image, or video.

40:02 And of course, it's going to become increasingly

40:04 important as these technologies become widely deployed

40:09 that we were able to easily distinguish between synthetically

40:11 generated images and real images.

40:15 So AI has been-- it's incredible potential

40:18 to help with our greatest challenges from climate to health.

40:23 But obviously, this is going to affect everyone.

40:26 So I think it's really important that we engage--

40:29 not just the technologists deciding this,

40:32 but that we engage with a wide range of stakeholders

40:34 from society.

40:35 So I've been really pleased in the last couple of years

40:38 that one of the consequences of AI becoming mainstream

40:40 is that many governments have got interested in it

40:43 and many parts of all parts of society.

40:45 And I think it's been great to see these international summits.

40:47 Actually, the UK hosted the first one in Bletchley Park

40:50 a couple of years ago, bringing together

40:52 heads of government with academia and civil society

40:55 to discuss these technologies, how

40:58 to put the right guardrails on it,

41:00 how to make sure we embrace the opportunities,

41:04 but we mitigate the risks that are coming down the line.

41:07 And I think that's going to become increasingly important

41:10 given the exponential improvement

41:12 that we're seeing with these technologies.

41:14 So my shorthand for this is to say,

41:17 with a lot of Silicon Valley tries

41:19 to-- the kind of mantra in Silicon Valley

41:21 is like, move fast and break things.

41:23 And of course, that's created a lot of advances,

41:26 a lot of the technologies we all use every day to day.

41:29 But I think it's not appropriate, in my opinion,

41:32 for this type of transformative technology.

41:34 I think, instead, we should be trying

41:37 to use the scientific method and approach it

41:40 with a kind of humility and respect

41:42 that this kind of technology deserves.

41:44 And we don't know a lot of things.

41:46 There are a lot of unknowns around how this technology is

41:49 going to develop.

41:49 It's so new.

41:51 And I think with exceptional sort of care and foresight,

41:54 we can get all the benefits and minimize

41:58 the downsides of this.

41:59 But I think only if we start the research and the debate

42:03 about that now.

42:06 So just to end, then, we're now building our own big multimodal

42:09 models that try and take the best of all of these different

42:12 models I've shown you and put it into one system.

42:15 We call it the Gemini series.

42:16 Our latest one is Gemini 2.0, that some of you

42:18 may have tried, which is state of the art

42:20 across many leading benchmarks.

42:23 We're using it to further that I'm

42:25 very excited about the next generation of assistants.

42:27 I call it Universal Assistance.

42:29 We call it Projastra, where actually you have it

42:32 on your phone or some other devices, maybe glasses.

42:35 And it starts being an assistant that you can take around

42:37 with you in the real world.

42:39 And it helps you in everyday life to get to enrich your life

42:44 or to make you more productive.

42:47 And the next step, then, in AI is combining

42:49 what I've shown you with AlphaGo, these kinds of agent-based

42:53 models that are able to efficiently search through

42:58 and find a good solution to a problem in a limited domain,

43:02 in this case, in games.

43:03 But we want to actually build those types of search systems

43:06 and planning systems on top of much more general models,

43:10 like Gemini, these world models that

43:12 understand how the real world works,

43:14 and then can then plan and achieve things in the real world.

43:18 And of course, that's key to things

43:20 like robotics working, which I think

43:22 in the next two, three years is going to be a huge area

43:25 that's going to have massive advances in it.

43:30 So I'll just finish, then, by just having a slight conjecture

43:34 about what does this all mean if we think back to Turing

43:38 and all of the work that he did to lay down

43:40 the foundations of computer science.

43:43 And I think that if you see the work that we've done,

43:47 I see myself as a kind of Turing's champion in a way.

43:50 Like, how far can the Turing machines

43:52 and this idea of classical computing go?

43:55 And I feel like probably one of the lectures

43:57 that I took in this room was one of my favorite things

44:00 to think about is the P equals MP problem, which

44:03 is a famous problem in computer science of, you know,

44:06 what sorts of problems are tractable on classical systems.

44:10 And there's obviously a lot of great work

44:12 going on in quantum computing systems.

44:14 Many of that here in Cambridge, and also at Google,

44:17 we have one of the top quantum computing groups

44:20 in the world.

44:21 And there's a lot of things that are

44:22 thought to require quantum computing to solve.

44:25 A lot of real world systems that we

44:27 would like to understand and model.

44:29 And my conjecture is that actually classical Turing machines,

44:33 basically classical machines that these types of AI systems

44:36 are built on, can do a lot more than we previously

44:39 gave them credit for.

44:40 And if you think about alpha fold and protein folding,

44:43 proteins are quantum systems.

44:45 You know, they operate at the atomic scale.

44:47 And one might think you need quantum simulations

44:51 to actually be able to find the structures of proteins.

44:54 And yet, we were able to approximate those solutions

44:57 with our neural networks.

45:00 And so I think, you know, one potential idea here

45:03 is that any pattern that can be generated or found in nature--

45:07 so IE has some real structure, physical structure--

45:10 can be efficiently discovered and modeled

45:12 by one of these classical learning

45:14 algorithms, like alpha fold.

45:16 And if that turns out to be true,

45:18 I think that has all sorts of implications

45:20 for quantum mechanics and actually fundamental physics,

45:24 which, you know, is something that I hope to explore

45:26 and many of my colleagues hope to explore, maybe

45:29 with the help of these classical systems as well

45:31 to help us uncover what the true nature of reality might be.

45:36 And that leads me back to the whole reason why I started

45:39 my path on AI many, many years ago

45:41 is that I always believed that AGI, built in this way,

45:45 could be the ultimate general purpose tool

45:47 to understand the universe around us and our place in it.

45:50 Thank you.

45:51 [APPLAUSE]

45:54 Great.

46:06 We have time for some questions, if you will have questions.

46:09 First hand, a shot up, just here.

46:11 Hi.

46:12 Thank you.

46:13 Thank you.

46:14 Thank you for the great talk.

46:16 Because you have a background in neuroscience,

46:19 and you really like to think in terms of root node problems.

46:22 Was there ever a root node problem you came across

46:25 in neuroscience that you thought was worth tackling

46:28 and still worth tackling to understand

46:30 biological and artificial intelligence even better?

46:34 Yeah, there's many.

46:35 I mean, that's what I studied from my PhD, actually,

46:37 was memory, but also imagination.

46:41 So future thinking, kind of planning.

46:43 So I really wanted to understand how the brain did that.

46:47 And it turns out the hippocampus is involved in both

46:50 so that we could maybe mimic that with some of these algorithms.

46:54 So I think there are many key things in that.

46:56 Of course, there are all the big questions

46:59 around creativity, dreaming, consciousness,

47:03 all these big questions that I think building AI

47:07 and then comparing it to the human mind

47:09 is one of the best ways, I think,

47:11 will make progress with those sort of root node problems.

47:14 Like, you know, what is the nature of consciousness?

47:16 And is there something special about the instantiation

47:18 of the substrate of the brain,

47:20 versus the algorithmically mimicking that in silicon?

47:25 Great, good question just here.

47:29 Hi, I have two questions, actually.

47:33 So since DeepMind was founded before the deep learning

47:37 revolution, I wanted to know what your state of mind

47:41 was, had deep learning not picked up,

47:44 or how were you going to progress?

47:47 That's the first question.

47:48 And second question is, since you've had intimate experiences

47:53 with such challenging problems,

47:55 such high dimensional problems,

47:57 and we know that gradient descent and its variants

48:01 kind sort of converge to the optimum solution,

48:05 only locally optimum solutions,

48:07 were you surprised that anything works at all

48:10 in these systems at every point in time?

48:14 And do you think that most of the nature

48:18 is kind of suboptimal?

48:20 And so we can potentially build a more optimal nature?

48:23 So, look, I think they're both great questions.

48:26 So the first one, look, that's why we call it DeepMind,

48:30 partly because the deep refers to deep learning, right?

48:33 So deep learning had, well, the early parts of deep,

48:37 it wasn't called deep learning then,

48:38 but it was sort of just becoming common.

48:40 There was these Boltzmann machines

48:42 and things that Jeffrey Hinton had invented.

48:44 And in just a few years before, in 2006, 2005,

48:48 these hierarchical neural networks,

48:50 and it seemed like a super promising idea,

48:52 even back then, to us that come across it in academia.

48:56 And then the other thing we bet on

48:59 was reinforcement learning,

49:00 and the combination of that, right?

49:02 Which again is coming back into Vogue,

49:04 but it was also important for us

49:06 to solve something like AlphaGo.

49:08 So you need both parts of that.

49:09 You need the deep learning

49:10 to kind of model the environment and the world,

49:12 and then you need the reinforcement learning

49:14 to make the plans and the solutions

49:16 and take action in the world.

49:18 And so there was two reasons we bet on that,

49:21 even when it was just sort of the beginning of it,

49:23 is that we knew that the classical methods,

49:26 these expert systems, would not scale.

49:29 And actually, that's again one of the things I learned here,

49:31 and also my postdoc at MIT,

49:33 was that there were the kind of churches,

49:36 if you like, of the classical methods,

49:38 these expert systems,

49:39 and actually, that's something else you can learn here,

49:41 is like, you know, in your university courses,

49:43 is not just what to do, but also what not to do,

49:46 and why it may not, like I sort of thought about it,

49:48 and I felt it would never scale

49:50 to the kinds of problems that I wanted to solve with AI.

49:54 Whereas the learning system seemed to have unlimited potential,

49:57 though they were a lot harder to get anything to do,

50:00 significant to do at the beginning, right?

50:02 And this is the problem, 'cause they weren't scaled up enough.

50:04 And the other reason we started DeepMind 2010,

50:07 is we could also see the computing paradigm

50:10 was shifting on the hardware side,

50:12 GPUs and other things,

50:13 which of course also invented for gaming,

50:16 and turns out everything's a matrix multiplication, right?

50:19 Intelligence and gaming, and computer graphics.

50:22 So, and so all of those different influences came together,

50:27 and the understanding of neuroscience,

50:29 and fMRI machines, and neuroscience had advanced

50:32 a lot in that previous 10 years too.

50:34 So, I felt it was the perfect time

50:35 to sort of bring all of that together back in 2010.

50:39 And we were betting on that,

50:40 not necessarily 'cause we knew it would work,

50:42 but we knew that the other methods,

50:44 we were pretty confident, would not work.

50:46 And that's what the AI winters were about, basically,

50:49 with people trying to push those expert systems.

50:51 And then the second question, you know,

50:53 I think, I wouldn't say that, well, first of all,

50:58 it is surprising that some of these things converge,

51:00 and actually we weren't sure.

51:01 So, you know, that Atari stuff I showed you.

51:04 So, for the first couple of years, nothing worked, right?

51:06 So, we couldn't even get one point on,

51:08 if some of you remember Pong,

51:10 you know, one of the first computer games,

51:12 we couldn't, this sort of tennis, bat and ball game,

51:14 simplest game you could imagine,

51:16 we weren't able to get a point on it.

51:17 So, we were wondering whether,

51:18 were we like 10, 20 years too early?

51:21 Like, like, Babbage turned out

51:22 to be with his difference engine, right?

51:24 Amazing ideas, it worked,

51:25 but he was just in the end 50 years or 100 years too early.

51:29 And I always say, you wanna be like five years ahead

51:31 of your time, not 50 years ahead.

51:33 Otherwise, you know, you'll be in for a lot of pain,

51:36 like Babbage was.

51:37 And so, we were worried about that,

51:40 but then it did converge,

51:42 and then that gave us confidence

51:43 to tackle the harder problems.

51:45 And I think if you're asking,

51:46 I think the last part of your question was about

51:48 that the things in nature,

51:49 well, the things there, what I'm thinking is,

51:50 they're not sub-optimal,

51:51 they're actually probably pretty optimal,

51:53 because they've gone through an evolutionary process,

51:56 not just life with biology, but actually geologically

51:59 and physically, you know,

52:00 asteroids and physical phenomena combined together,

52:03 they survive some amount of time

52:05 because they're stable over time, right?

52:07 And if they're stable over time,

52:09 then there's probably some structure that's learnable.

52:11 That would be my conjecture.

52:13 - Great, there's a question down here.

52:17 - What do you think about building high bandwidth

52:23 brain-machine interfaces and implantable memory

52:26 and reasoning modules,

52:27 so that humans can be further empowered

52:29 to make discoveries autonomously,

52:30 as opposed to only talking to AI in the cloud?

52:33 - Yeah, I love that area,

52:35 and I've carefully followed it

52:38 and helped people building like EEG caps and things.

52:43 Of course, the problem is on resolution of these things

52:46 to get the readouts from the brain,

52:47 and also, ideally, when it read and write.

52:50 But I'm very fascinated by projects like New Orleans

52:53 or, you know, chips in the brain.

52:54 Obviously, right now, that's for, you know,

52:57 veterans and people to get functioned back in their body.

53:00 I think it's gonna be amazing things on that,

53:02 where I think people will be able to gain,

53:04 be able to walk again if they've broken their back

53:06 and things like that.

53:07 I think there's gonna be some credible advances

53:09 in medical sciences that will be amazing.

53:11 But then beyond that,

53:12 maybe if things, you know, if it becomes routine

53:15 and it's surgery safe and there's safe ways of doing this,

53:19 I could imagine that would be one way

53:21 for us to keep up with the technology.

53:23 And in some senses, it's no different

53:25 to what we already have today with our technologies

53:28 all around us.

53:28 We all have our phones all, you know,

53:30 24/7 with us and computers and other things.

53:32 So we're already symbiotic almost with our technology.

53:35 It would be one step further, of course, to have that.

53:38 But I'm not sure, maybe it's for the philosophers

53:40 in the room to answer what the difference,

53:42 whether there's a how boundary there,

53:44 if with the technology, if it's attached to you

53:47 versus, you know, it's just something you carry around

53:49 with you all the time.

53:52 - Great, some question there for here.

53:54 But that's one just, sure, man.

53:57 - Hi.

53:59 - Are you sure?

54:04 - Well, hello?

54:06 Ah, there we go.

54:08 What do you think about the speed

54:11 at which artificial intelligence is developing

54:13 and its effects on sort of economic developments?

54:16 There are a lot of people out there

54:17 who are deciding careers right now

54:19 who, given the sort of rapid change in the landscape,

54:23 make it really difficult to kind of predict

54:25 what they should go into?

54:27 - Yeah, so it's a very complicated one

54:30 because, as you say, things are changing

54:32 at kind of lightening speed.

54:34 And we were just discussing it with Alistair, actually.

54:36 I mean, even designing three-year computer science courses

54:39 is quite difficult, given that there are underlying

54:41 material changes in less than three years.

54:44 I think the only thing we can say for sure

54:47 is there's going to be a lot of change,

54:49 but I think that brings with it

54:50 disruption and opportunity.

54:52 So, I mean, I just give you an example on coding,

54:55 which I don't know if you're a computer scientist,

54:57 but, you know, so I still would recommend

55:00 that you get good at coding and math

55:02 because I think you'll be able to use these new tools

55:05 in a much deeper way

55:06 if you have an understanding of how they're built.

55:08 But on the other hand,

55:09 I think coding is going to be more available

55:12 to many, many more types of people.

55:15 Because of the way the AI, you'll be able to programme

55:17 sort of in natural language, probably,

55:19 rather than a quite complicated computer language.

55:24 And so that will open up fields for creative people

55:28 to, you know, build games, make films, make applications.

55:32 That maybe is more, the kind of balance of that

55:36 is more on the creative side than it is on the engineering side.

55:39 But I also think that will enhance engineers

55:42 to be able to do 10x of what they can do today.

55:44 So I still, I think it's difficult to know.

55:47 But what I would say is just to focus on

55:51 kind of embracing those tools in your spare time

55:55 and making, training yourself to be really good

55:58 at picking up new information extremely quickly.

56:01 Because I think that's basically

56:02 what's going to happen in the next 10 years.

56:05 - Okay, we've got one question

56:07 just in the here, the yellow and black top.

56:12 Do you think that there are any biological processes

56:15 or behaviours or patterns, which can't be modelled

56:19 with existing deep planning techniques?

56:22 I'm not saying like pro-modal computers until it works

56:25 and just make a bigger and bigger model.

56:27 Do you think there are some processes

56:28 which physically cannot be modelled with the architecture we have?

56:31 - I mean, there's certainly lots of processes

56:33 that can't be modelled today, but again,

56:37 this is sort of goes back to what I said

56:38 at the end of the talk.

56:39 I'm not sure in the limit that there are.

56:41 I think in the end, if physics can solve it

56:46 and there's some structure there to be learnt,

56:48 probably with enough examples,

56:50 one could reverse engineer a model of that.

56:53 And then I don't see any theoretical reason

56:56 why a classical system or be a very complex one

57:01 could not make predictions or simulations

57:04 of that biological system.

57:06 So I don't really see any in the limit

57:10 of what the limit of that could be.

57:12 I mean, there are lots of abstract things

57:14 like factorizing large numbers, cryptography,

57:17 where it's sort of human-made systems,

57:20 where there may not be any structure.

57:21 I mean, there may be structure in the natural numbers.

57:24 Lots of people conjecture there is.

57:26 If there is, then that will also be learnable.

57:28 If there isn't, and it's kind of uniform distribution,

57:31 then you would need a quantum computer

57:34 to crack cryptography, things like that.

57:37 So those are open conjectures,

57:39 but I think most things in nature have evolved

57:42 and over geological or biological physics time.

57:46 And so then that suggests to me

57:48 there is some structure to learn.

57:50 So that makes the search or the prediction

57:53 potentially tractable.

57:54 - Okay, and last question then,

57:58 we'll have the person in the pink shirt.

58:01 - This question comes from the behalf

58:04 of the Cambridge University Game Development Society.

58:07 So you mentioned about the G&E2 model

58:11 and how currently it's stable for a few seconds of consistency

58:15 and you hope to eventually have that a few minutes.

58:18 But I suppose the question that our society has is,

58:21 games that we actually play have consistency

58:24 that's indefinite.

58:25 When you're playing Minecraft,

58:26 you expect to turn around and the village is still there, right?

58:29 - Yes.

58:30 - So do you see your current model

58:32 being integrated into workflow?

58:34 Or what exactly, how do you see AI and your model

58:37 and what you're working on?

58:39 How do you see it integrating into game development

58:41 in the coming decades?

58:43 - Yeah.

58:44 Well, look, I think there's many ways AI is gonna come in.

58:48 So one is the tools to build the assets you need for games.

58:53 So 3D models, animations.

58:55 I think that's all gonna come in in the next couple of years.

59:00 I think you can think of AI as well for game balancing.

59:03 So you imagine you design a game and overnight,

59:07 it could play a million playthroughs of that game.

59:10 And then in the morning,

59:10 you could just get a report as a game designer

59:12 of like these things are unbalanced, right?

59:14 Or reduce the power of this unit or whatever it is.

59:18 I think also bug testing for open world games.

59:20 So I used to make simulation games, open world games.

59:24 They're a nightmare to bug test

59:26 because the whole point of them is that the player

59:28 can almost do almost anything

59:29 and then the game will react to you.

59:31 So how do you test 10 million people

59:33 having their own unique journey through your game?

59:36 Well, actually having AI players play it

59:38 before you release it could help you solve a lot of those bugs.

59:42 And then finally, I mean, I think excitingly as well

59:46 is like AI characters that are much more lifelike

59:48 that move the storyline on.

59:49 You used to dream about that massive multiplayer world

59:52 where actually the AI characters were intelligent

59:55 and actually updated their beliefs and their storylines

59:58 based on what the players were doing.

1:00:00 So it felt like a much more living, realistic world.

1:00:04 And I think we're on the cusp of having

1:00:06 or building those types of games.

1:00:09 And then finally, the world model we're building,

1:00:11 that's more about general AI.

1:00:13 And can you actually model the way?

1:00:15 It's an expression of being able to understand the world.

1:00:17 Does your model understand the world?

1:00:19 Well, if it can generate it for a sum amount of time,

1:00:22 then obviously it must be in some sense

1:00:24 understanding empirical evidence

1:00:27 that is understanding something about the underlying physics.

1:00:30 So that's more for the general intelligence

1:00:32 rather than I think we're going to maybe one day

1:00:35 we'll have this holodeck thing that you can just imagine

1:00:37 and it's just all there around you.

1:00:40 Probably we will be able to have that once we have AGI.

1:00:43 But I think that's still a ways off.

1:00:46 That's great.

1:00:47 Well, it seems like a nice place to finish a question

1:00:49 on games returning back to games.

1:00:50 But thank you all so much for coming.

1:00:52 And a particular special thank you

1:00:54 for the damage to coming in and talk to us today.

1:00:56 So thank you guys.

1:00:57 (audience applauds)

love

· 5d

特斯拉：我们将于 11 月 20 日登陆阿根廷。

原帖：https://x.com/Tesla_LatAm/status/2118674774440435905

love

· 3w

1Password 发布「1Password for Agents」平台。

AI Agent 登录网站时，会向你的密码库申请凭证；只有在你批准后，1Password 才会直接在目标网站自动填充。Agent 能完成任务，却看不到、也拿不到密码。

你始终掌握控制权：个人及家庭账户可免费选择使用。除非主动启用 Agent 连接器，否则任何 Agent 都无法访问你的密码库。

x.com/1Password

中文 English

0:00 您的AI代理需要访问权限，而不是密码。

0:03 一个密码，助代理安全登录，完成任务，让您轻松安排温馨夜晚或难忘外出。

0:12 预订下一个出城度假或出差的机票？

0:16 买生日、节日的完美礼物，不是所有场合的礼物。

0:22 或者清空日程，专注打造下一个项目。

0:25 一个平台保护您的秘密，让代理知道重要事项。

0:29 无需知道您的密码。

0:31 下次代理登录，只需一个密码。

0:35 看看可能发生什么！

0:00 Your AI agents need access, not your passwords.

0:03 One password for agents unlocks a safer way for your personal agents to log in

0:07 to complete tasks,

0:09 so you can organize a cozy night in or a memorable night out.

0:12 Book your next ticket out of town for vacation or for work?

0:16 Buy the perfect gift for birthdays, holidays, not everything in between.

0:22 Or clear your calendar so you can build your next project.

0:25 One platform to protect your secrets so your agents know what matters.

0:29 Without knowing your passwords.

0:31 So the next time your agents need to sign in, use one password.

0:35 See what's possible!

love

· 3w

Doug Leone 谈每个风险投资人都会犯的错误：过早卖出赢家。

红杉曾持有英伟达 20%、谷歌 10%、思科四分之一的股份，还是苹果的首位投资者，却都卖掉了。

真正伟大的公司在 IPO 后还能持续复利 20 年以上。如果一家公司面对的市场近乎没有上限，就不该卖出一股。持有，可能让 50 亿美元的收益最终变成 500 亿美元。

x.com/davidsenra

中文 English

0:00 每个风险投资者都犯过过早卖出赢家的错误，每一个风险投资者都是如此，因为如果你看看那些伟大的公司，他们在上市后积累了20年，比如Video、Google或Meta，如果你稍微有点敏感，知道这东西有

0:20 基本上是一个多年的无限市场，你绝不应该卖出股票，持有它们能赚得更多，因为如果你将50亿美元的收益复利增长，最终可能变成500亿美元，如果我们持有Google并且不卖出，或者Nvidia，这可能达到半万亿美元

0:37 美元，你知道，没有任何投资能给你这样的回报。你知道我们拥有Nvidia 20%的股份，想想看，我们拥有Google 10%的股份，还有很多其他，比如Apple，你开玩笑吧，我们是苹果的第一批投资者，还有Cisco，我们拥有Cisco系统四分之一的股份

0:56 你知道这些都是大股东，还是非常大的公司，但我们都会犯这个错误。有时候最好的财务决策根本不是财务上的。比如说，如果你考虑一下Sam Walton创造的财富，显然他把财富在积累到孩子之前就捐出去了，但

1:10 如果财富还集中在一个人手里，现在价值应该超过半万亿美元，无论现在数字是多少，你会明白他不是做财务决策，他只是喜欢沃尔玛。这又回到了Nick Sleep的观点，Nick Sleep有一句很棒的名言，他说最好的

1:22 投资者其实根本不是投资者，而是创业者，他们从不卖出。

0:00 every venture investor has made the mistake of selling to early their winners

0:04 every single

0:04 venture investors because if you look at the at the great companies they've

0:09 compounded 20 years

0:10 after the IPO and video or or Google or meta if you had a little sniffer that

0:18 this thing had

0:20 essentially an unbounded market for many years you should never sell share and

0:24 you can make way

0:25 more money holding those because if you're compounding a five billion dollar

0:28 gain that can turn into a

0:30 50 billion if we held Google and had sold gold that'd be worth or Nvidia that'd

0:36 be half a trillion

0:37 dollars you know and there's no investments that can give you that and you know

0:42 we own 20%

0:43 of Nvidia think about that I mean you know we own 10% of Google and there's

0:48 many others you know

0:50 Apple are you kidding me we're the first investor Apple in Apple Cisco we own a

0:54 quarter Cisco systems

0:56 you know those are big ownership and very big companies but we all made that

1:01 mistake sometimes

1:02 the best financial decisions are not financial at all so if you look at like if

1:05 you think about

1:05 the wealth created by Sam Walton obviously he gave it away before it

1:09 accumulated to his kids but it

1:10 comp you know if it was still in one person it'd be worth you know half a

1:13 trillion whatever the

1:14 number is now and you realize that he wasn't making a financial decision he

1:18 just liked Walmart

1:19 it goes back to what Nick sleeps at Nick sleep has this great quote where he's

1:21 like the best

1:22 investors are investors at all they're entrepreneurs you never sold

love

· 3w

读到新闻的那一刻，就能立即交易。

Jupiter 现已为 @DecryptMedia Money Accounts 内的交易功能提供支持。阅读新闻、兑换相关代币，全程无需离开 Decrypt。

每个应用都正在变成金融应用，而 Jupiter 是它们背后的引擎。

原始链接：https://x.com/JupiterExchange/status/2113269167784935921

中文 English

0:00 （欢快的音乐）- 木星的大事，本周重大公告。

0:08 交给你了，卡萝尔。

0:09 - 没错，鲍勃。

0:10 Decrypts、文章和币种页面都内置了购买框。

0:15 选择金额，点击购买。

0:17 木星就在页面上完成兑换。

0:21 从标题到持币，只需几步。

0:25 读它，换币，这就是新闻。

0:28 晚安。

0:29 （欢快的音乐）

0:00 (upbeat music)

0:02 - Big week for Jupiter, an exciting announcement.

0:08 Over to you, Carol.

0:09 - That's right, Bob.

0:10 Decrypts, articles, and coin pages

0:13 have a buy box built right in.

0:15 Pick an amount, tap buy.

0:17 And Jupiter handles the swap, right there on the page.

0:21 From headline to holding in a few taps.

0:25 Read it, swap it, and that's the news.

0:28 Good night.

0:29 (upbeat music)