DeepMind Targets General AI to Crack Science Crises, Hassabis Says

Google Zeitgeist

Solving general intelligence will serve as a foundational meta-solution to humanity's most complex scientific bottlenecks, requiring ethical governance decades before human-level machines arrive.

AI-assisted summary of Human AI is Decades Away, But We Need to Start the Debate Now | Demis Hassabis | Google Zeitgeist — original source. Check the original for full context.

Source published: 2015-05-12

Key ideas

DeepMind's institutional blueprint focuses on a two-step mission: solve general intelligence first, and then apply it as a meta-solution to tackle complex global challenges like climate change and disease. (5:13)
Using a single reinforcement learning architecture trained purely on raw 200x150 display pixels without pre-programmed rules, DeepMind's agent achieved superhuman performance across classic Atari games overnight on a single GPU. (9:52)
While human-level artificial general intelligence remains decades away, Demis Hassabis emphasizes that ethical frameworks and public debate must begin immediately to govern how the technology is deployed. (13:33)
DeepMind's technical roadmap extends beyond 2D arcade games toward 3D environments, complex games like Go, robotics, and computational models inspired by the hippocampus for memory and abstract concept acquisition. (12:51)

Detailed notes

## Background Published on YouTube by Google Zeitgeist on May 12, 2015, this talk is part of a series featuring global innovators, scientists, and leaders sharing their visions for tomorrow. The presentation captures DeepMind co-founder Demis Hassabis addressing the trajectory of artificial intelligence, from foundational neuroscience to general-purpose machine learning.

## TL;DR Demis Hassabis traces his journey from childhood chess master and video game designer to cognitive neuroscientist and co-founder of DeepMind. He details how DeepMind leverages neuroscience-inspired reinforcement learning to build general-purpose algorithms, showcasing landmark breakthroughs where agents mastered classic Atari titles directly from raw pixels. Framing general artificial intelligence as a potential 'meta-solution' to humanity's most complex scientific and systemic challenges, Hassabis explains that while human-level AI is still decades away, establishing ethical oversight and governance cannot wait.

## Atmosphere & style The presentation carries a calm, visionary authority, blending deep intellectual curiosity with measured scientific realism. Hassabis delivers his narrative with understated confidence, weaving personal anecdotes—such as coding chess programs on a ZX Spectrum to beat his brother—into sophisticated discussions of hippocampal memory and statistical world models. Watching the demonstration of the AI's autonomous breakthroughs in Atari feels both electrifying and sobering, evoking a sense of witnessing the foundational steps toward an epochal shift in human capability.

## Bulletin - [05:13](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=313s) DeepMind's institutional blueprint focuses on a two-step mission: solve general intelligence first, and then apply it as a meta-solution to tackle complex global challenges like climate change and disease. - [09:52](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=592s) Using a single reinforcement learning architecture trained purely on raw 200x150 display pixels without pre-programmed rules, DeepMind's agent achieved superhuman performance across classic Atari games overnight on a single GPU. - [13:33](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=813s) While human-level artificial general intelligence remains decades away, Demis Hassabis emphasizes that ethical frameworks and public debate must begin immediately to govern how the technology is deployed. - [12:51](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=771s) DeepMind's technical roadmap extends beyond 2D arcade games toward 3D environments, complex games like Go, robotics, and computational models inspired by the hippocampus for memory and abstract concept acquisition.

## Key takeaways - [00:08](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=8s) Artificial intelligence serves a dual scientific purpose: building synthetic minds is the ultimate pathway to truly understanding human cognition and perception. - [01:57](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=117s) Hassabis's early background as a chess prodigy and computer programmer sparked an enduring introspection into how the human brain evaluates decisions and plans ahead. - [03:09](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=189s) Co-designing the hit simulation game Theme Park demonstrated the potential of dynamic, emergent AI systems over rigid, hand-scripted behaviors. - [04:00](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=240s) Transitioning from game development to cognitive neuroscience allowed Hassabis to investigate memory, imagination, and the hippocampus at MIT and Harvard. - [05:13](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=313s) DeepMind distinguishes itself from narrow AI milestones like Deep Blue and Watson by building general-purpose learning algorithms rather than brittle, handcrafted rule engines. - [07:36](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=456s) The company's approach models the sensory-motor reinforcement learning loop, which biologically mirrors the brain's dopamine reward mechanisms. - [09:52](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=592s) Atari 2600 games provided a rigorous testbed where agents received only raw visual input (around 30,000 numbers per frame) and score feedback. - [11:20](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=680s) In Breakout, the reinforcement learning agent independently invented the optimal 'tunneling' strategy behind the brick wall without human instruction. - [12:51](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=771s) DeepMind's research incorporates hippocampal-inspired computational mechanisms to provide agents with abstract concept formation and long-term memory replay. - [13:33](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=813s) Society faces acute bottlenecks in information overload and systemic complexity that outstrip human cognitive capacity in fields like physics and genomics. - [14:20](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=860s) General AI can function as an 'AI scientist,' collaborating with human domain experts to distill massive data streams into actionable scientific breakthroughs. - [14:54](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=894s) Algorithms themselves are neutral tools, meaning that ethical governance and transparent oversight dictate whether AI deployment benefits humanity. - [15:24](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=924s) Solving intelligence may be an essential prerequisite to formulating a unified physical 'Theory of Everything,' as reality is fundamentally mediated by the mind.

## Detailed walkthrough ### [00:08](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=8s) The Twin Pillars of Physics and Neuroscience

Demis Hassabis defines artificial intelligence at its core as the science of making machines smart, framing it as what will likely become one of the most consequential technologies humanity will ever create. Reflecting on his intellectual path during a session themed around "The Theory of Everything," Hassabis recalls arriving at an early conviction during his school days that only two disciplines were fundamentally worth studying: physics and neuroscience. While physics seeks to explain the external universe, neuroscience and psychology endeavor to decode the internal world of the human mind.

Over time, Hassabis concluded that the internal world was the more fundamental mystery, recognizing that human reality is mediated entirely through the lens of perception. He points to the philosopher Immanuel Kant, who argued that the mind actively synthesizes and interprets sensory input to construct our experience of reality. For Hassabis, artificial intelligence offers the ultimate scientific pathway to illuminate these cognitive processes, guided by a famous maxim from one of his intellectual heroes: “As Richard Feynman said, one of my scientific heroes, 'What I cannot build, I do not truly understand,' and that's one of the things that I'm excited about with artificial intelligence.” By attempting to engineer synthetic intelligence from the ground up, researchers can simultaneously unlock a deeper understanding of human cognition.

[01:57](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=117s) From Chess Prodigy to the ZX Spectrum

Hassabis's journey into computation and artificial intelligence began through the world of competitive games. Introduced to chess at the age of four, he progressed rapidly, competing on various junior England squads, captaining national junior teams, and attaining the rank of chess master by the age of twelve. This intense immersion in high-level strategy fostered a deep introspective curiosity about the brain's internal mechanics. He found himself constantly analyzing how his own mind evaluated board states, generated tactical ideas, and structured forward-looking plans within an immensely complex decision space.

At eight years old, Hassabis took his tournament prize money and purchased his first computer, a Sinclair ZX Spectrum 48K. He immediately taught himself how to program using books. One of his earliest ambitious software projects was an automated chess program. While he admits it was crude by competitive standards, it achieved a personally satisfying benchmark: “One of the first, sort of, big programmes I can remember creating was actually a programme to play chess, and it didn't play very well but it was able to beat my little brother which I was very pleased about when I was small.” That early victory cemented his realization that software could simulate strategic decision-making, setting him on a direct course toward AI research.

- Began playing chess at age 4; achieved the rank of chess master at age 12. - Used international chess tournament winnings at age 8 to buy a ZX Spectrum 48K. - Wrote an early chess-playing program that successfully defeated his younger brother.

[03:09](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=189s) Video Game Design and the Neuroscience Pivot

During his late teens and early twenties, Hassabis fused his passions for chess, programming, and simulation into a decade-long career in commercial video game design. At seventeen, he co-designed the groundbreaking 1994 hit *Theme Park*. Departing from the prevailing video game trends centered on combat and twitch reflexes, *Theme Park* tasked players with designing and operating a Disney World-style attraction. The game populated the park with thousands of autonomous, simulated visitors, each possessing distinct personal preferences, moods, and feedback loops that determined how much fun they had, whether they spent money, and whether they recommended the venue to friends.

The title pioneered the management simulation genre and established an interactive paradigm where emergent AI dynamically adapted to player choices, guaranteeing that no two playthroughs were identical. Hassabis embedded advanced AI into every game he designed, but after a decade running his own studio, he recognized that commercial gaming could not bridge the gap to true machine intelligence. He sold his company and returned to academia to pursue a PhD in cognitive neuroscience, focusing on the hippocampus, episodic memory, and mental simulation. Because imagination and recall represent key blind spots in contemporary computational systems, he sought to unpack how mammalian neural architecture executes these complex cognitive tasks, later continuing his research through postdoctoral fellowships at MIT and Harvard.

- Co-designed the bestselling simulation title *Theme Park* (1994) at age 17. - Shifted the medium away from destruction toward open-ended creative construction and adaptive AI behavior. - Pursued a PhD in neuroscience alongside postdoctoral work at MIT and Harvard, specializing in imagination, memory, and the hippocampus.

[05:13](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=313s) Founding DeepMind and the Quest for AGI

Convinced he had assembled the requisite domain expertise across game design, software engineering, and cognitive biology, Hassabis co-founded DeepMind in 2010. Conceived as an ambitious, institutional endeavor akin to an "Apollo Programme for AI," the company assembled over one hundred leading PhD researchers across machine learning and systems neuroscience. Their methodology focuses on neuroscience-inspired AI—abstracting the brain's computational principles at a systems level rather than attempting direct biological emulation.

Hassabis distills DeepMind's institutional ambition into an audacious two-step blueprint: “So step one, solve intelligence and then step two, use it to solve everything else, and I'll come back to that right at the end why that's not as fanciful as it might seem.” To realize this, DeepMind concentrates strictly on building general-purpose learning systems rather than narrow AI. While historical milestones such as IBM's Deep Blue defeating Garry Kasparov at chess in 1997 or IBM Watson winning *Jeopardy!* in 2011 demonstrated immense computational power, Hassabis notes that they were handcrafted, brittle, and incapable of generalizing beyond their single programmed domains. By contrast, DeepMind builds algorithms that master tasks autonomously from raw data and first principles across diverse environments without bespoke engineering.

[07:36](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=456s) The Reinforcement Learning Loop and Atari Testbed

To formalize general intelligence, DeepMind frames learning around the classical agent-environment paradigm of reinforcement learning. The system is situated within an environment—whether physical or virtual—and navigates the world through a closed sensory-motor feedback loop. Unlike pristine abstract systems like chess, real-world data streams are fundamentally noisy, messy, and incomplete. DeepMind supplies its agents with raw perceptual observations (primarily computer vision), requiring the AI to infer a statistical model of its surroundings, evaluate candidate behaviors, and execute the optimal action to maximize a scalar reward signal.

```mermaid flowchart LR accTitle: Reinforcement Learning Loop accDescr: The closed-loop interaction between an AI system and its environment via observations, rewards, and actions. subgraph Agent["AI System"] Model["Statistical World Model"] ActionSel["Action Selection Policy"] Model --> ActionSel end

subgraph Env["Environment (Virtual or Real)"] State["Internal State Dynamics"] end

Env -->|"Noisy Observations & Rewards"| Model ActionSel -->|"Executed Action"| Env ```

Hassabis emphasizes that this formal framework maps directly onto biological cognition, noting that in mammalian and human brains, reinforcement learning mechanisms are primarily driven by the dopamine neurotransmitter system. To rigorously evaluate an agent's ability to ground itself in raw sensory-motor data without task-specific engineering, DeepMind deployed classic 1980s Atari video games as an empirical benchmark. These games offer an ideal proving ground because they were originally constructed to challenge human cognitive faculties while requiring an agent to extract meaning and strategic planning directly from raw visual pixels.

- The reinforcement learning agent operates in noisy, partially observable environments using sensory modalities like raw vision. - Action execution alters the environment, returning reward signals and updating the agent's statistical world model. - Hassabis notes biological parity: “In humans it's the dopamine system in our brains that implements reinforcement learning.” - Classic 1980s Atari console games serve as the testbed to ensure algorithms learn entirely from raw pixels without pre-programmed rules.

[09:52](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=592s) Mastering Atari from Raw Pixels: Space Invaders and Breakout

Demis Hassabis introduces the benchmark used by DeepMind: classic Atari 2600 video games. These titles offer rich environments that are not too trivial, yet not so complex that contemporary reinforcement learning algorithms cannot make tangible progress. Crucially, the AI agent receives nothing other than the raw pixels on the display screen. Hassabis notes that the screen resolution is 200 by 150 pixels, yielding approximately 30,000 numbers per frame fed directly into the system. The agent has no hardcoded rules, no prior knowledge of physics, and no built-in knowledge of what it is controlling; its sole operational objective is to maximize the game score entirely from scratch. DeepMind strictly mandated that the exact same algorithm and architecture play hundreds of different Atari titles without game-specific modifications.

```mermaid flowchart LR accTitle: Deep RL Atari Loop accDescr: The closed loop showing raw screen pixels fed into the agent and action outputs returned to maximize score. subgraph Env["Atari Environment"] P["Screen Pixels (200x150, ~30k values)"] S["Score / Reward Signal"] end subgraph Agent["DeepMind Learning System"] NN["Deep Q-Network (Single Architecture)"] A["Motor Actions (Joystick / Fire)"] end P --> NN S --> NN NN --> A A --> Env ```

To demonstrate this learning process, Hassabis plays a video clip of the system playing *Space Invaders*. When first exposed to the data stream, the AI immediately loses all three of its lives controlling the rocket cannon at the bottom of the screen because it has no concept of the game. However, after leaving the agent to train overnight on a single GPU machine, it returns the following morning playing at a superhuman level. It hits targets with every shot, avoids enemy fire completely, and discovers on its own that sniping the pink mothership moving across the top yields the highest point reward. As fewer alien invaders remain and their movement accelerates, the agent anticipates their trajectories, culminating in a predictive shot aimed ahead of the final invader. As Hassabis puts it, “So you can see how perfectly it, sort of, modelled the game world, and that data stream.”

Moving to a demonstration of *Breakout*, Hassabis explains the clear gradations in the agent's competence across training epochs. After 100 games, the system still performs poorly, though viewers can see it tentatively learning to move the paddle toward the descending ball. By 300 games, it consistently returns the ball without missing, equaling the performance of top human players. The breakthrough arrived when researchers allowed the machine to continue playing for a couple hundred additional games: the AI independently discovered the optimal strategy of digging a tunnel through the wall of bricks on the left side to send the ball into the space behind the wall, racking up points rapidly as it bounced along the top. Hassabis humorously points out that the programmers and researchers on the project—while brilliant at math and coding—were mediocre Atari players who had never conceived of this tactic themselves; their own creation taught it to them.

- The system takes roughly 30,000 numerical inputs per frame (from a 200 × 150 pixel display). - Training to achieve superhuman proficiency in *Space Invaders* took place overnight on a single GPU workstation. - In *Breakout*, the agent progressed from basic tracking at 100 games to human-level defense at 300 games, eventually discovering the high-scoring "tunneling" exploit after further unsupervised play. - This research was published as the cover story of the scientific journal *Nature* in early 2015.

[12:51](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=771s) Beyond Atari: Hippocampal Memory, Go, and Near-Term Applications

DeepMind's roadmap extends well past 2D arcade environments. Hassabis outlines the team's ongoing work to integrate human-like cognitive capabilities, including the acquisition of abstract concepts and long-term memory. These architectural extensions draw heavily on neuroscience research—specifically computational models inspired by the hippocampus, the brain structure critical for episodic memory, spatial navigation, and memory replay.

Looking ahead, Hassabis details how these algorithms are scaling toward 3D environments, complex board games such as Go, physics simulations, and ultimately physical robotics. In the nearer term, DeepMind is applying these foundational models to practical problems within Google, such as enhancing recommendation systems on platforms like YouTube and building systems for predictive healthcare applications.

[13:33](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=813s) AI as a Meta-Solution to Societal Challenges and Ethical Governance

Returning to the broader theme of finding a "Theory of Everything," Hassabis identifies two escalating crises confronting modern civilization: information overload and systemic complexity. Society is drowning in massive datasets across domains like genomics and digital media, yet current crowd-averaging recommendation engines fail to generate true personalization or extract deeper meaning. Concurrently, vital macro-systems—such as climate change, epidemiology, energy grids, macroeconomics, and fundamental physics—have grown so intricately complex that even the most gifted researchers struggle to master an entire discipline within a single lifetime and still have time left over to innovate.

```mermaid flowchart TD accTitle: AI as a Meta-Solution accDescr: High-level overview showing AI addressing information overload and systemic complexity to enable AI-assisted science. subgraph Bottlenecks["Societal Bottlenecks"] D["Information Overload (Genomics, Big Data)"] C["System Complexity (Climate, Disease, Physics)"] end subgraph Solution["General AI System"] Meta["Meta-Solution (Transforms Data to Knowledge)"] end subgraph Outcomes["Breakthroughs"] Sci["AI-Assisted Science (AI Scientists + Human Experts)"] App["Predictive Medicine & Solutions"] end D --> Meta C --> Meta Meta --> Sci Meta --> App ```

Hassabis positions artificial intelligence not as a siloed tool, but as a foundational remedy: “So one of the reasons I work on AI and why I think it's going to be one of the most important technologies out there, is that solving AI is potentially a, kind of, meta-solution to all these other problems.” His core vision is the realization of "AI scientists" or AI-assisted scientific discovery, where artificial general intelligence partners with human experts to parse massive data spaces and surface breakthrough theories.

Addressing the rising public discourse and anxiety surrounding machine intelligence, Hassabis acknowledges that powerful technologies demand strict ethical oversight. DeepMind established an internal ethics board to supervise the research and deployment of its systems. He emphasizes that the underlying mathematics and algorithms remain neutral; ethical responsibility rests entirely on human implementation: “The technology itself is neutral. It's always how humans use this technology that ends up deciding whether it's ethical or not.” Hassabis notes that while true human-level artificial general intelligence remains several decades away, establishing the ethical frameworks and engaging in transparent public debate must begin immediately.

[15:24](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=924s) Converting Information to Knowledge and Decoding the Mind

Reflecting on DeepMind’s acquisition by Google in early 2014, Hassabis notes the strong philosophical alignment between both organizations. Google’s charter to organize the world's information and make it universally accessible and useful mirrors DeepMind's ambition to convert unstructured information into actionable knowledge to empower humanity.

In closing, Hassabis ties artificial intelligence back to fundamental neuroscience and physics. He proposes that translating cognitive processes into rigorous algorithmic constructs will offer an empirical mirror for understanding the deepest puzzles of human consciousness, imagination, and dreaming. Ultimately, the quest to formulate a unified scientific framework for reality may require cracking cognition beforehand: “Finally, I would say that, you know, in order to find the Theory of Everything, it might turn out that we have to solve intelligence first.”

## Notable quotes > [13:33](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=813s) The technology itself is neutral. It's always how humans use this technology that ends up deciding whether it's ethical or not. — Demis Hassabis

> [15:24](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=924s) Finally, I would say that, you know, in order to find the Theory of Everything, it might turn out that we have to solve intelligence first. — Demis Hassabis

## Mentioned Demis Hassabis, DeepMind, Google, Richard Feynman, Immanuel Kant, Garry Kasparov, Deep Blue, IBM Watson, Sinclair ZX Spectrum 48K, Theme Park, Space Invaders, Breakout, Go, YouTube, Nature, MIT, Harvard University

## Video description Google Zeitgeist is a collection of talks by people who are changing the world. Hear entrepreneurs, CEOs, storytellers, scientists, and dreamers share their visions of how we can shape tomorrow.

[Source](https://www.youtube.com/watch?v=rbsqaJwpu6A)

中文简报

哈萨比斯称DeepMind致力于通过通用人工智能破解科学难题破解通用智能将成为解决人类最复杂科学瓶颈的底层“元解决方案”，在人类水平的机器问世前数十年就需要建立伦理治理体系。德米斯·哈萨比斯（Demis Hassabis）回顾了自己从国际象棋神童、电子游戏设计师到认知神经科学家以及DeepMind联合创始人的历程。他详细介绍了DeepMind如何利用受神经科学启发的强化学习来构建通用算法，并展示了智能体仅凭原始像素便掌握经典雅达利（Atari）游戏的标志性突破。哈萨比斯将通用人工智能视为应对人类最复杂的科学与系统性挑战的潜在“元解决方案”，同时指出尽管达到人类水平的人工智能仍需数十年时间，但建立伦理监管与治理框架已刻不容缓。 DeepMind的发展蓝图聚焦于两步走的使命：首先解决通用智能问题，随后将其作为元解决方案应用于应对气候变化和疾病等复杂的全球性挑战。 (5:13) 通过使用仅在200x150原始显示像素上进行训练且无需预设规则的单一强化学习架构，DeepMind的智能体仅用一块GPU便在一夜之间在多款经典雅达利游戏中达到了超越人类的表现。 (9:52) 尽管距离达到人类水平的通用人工智能仍有数十年之遥，但德米斯·哈萨比斯强调，必须立即启动伦理框架构建与公众辩论，以监管该技术的具体部署。 (13:33) DeepMind的技术路线图不仅限于2D街机游戏，还延伸至3D环境、围棋等复杂博弈、机器人技术，以及受海马体启发用于记忆和抽象概念获取的计算模型。 (12:51)

Detailed notes — zh

## Background 该视频由 YouTube 频道 Google Zeitgeist 于 2015 年 5 月 12 日发布。根据页面介绍，Google Zeitgeist 汇聚了改变世界的企业家、科学家与梦想家，旨在让先锋人物分享塑造未来世界的深远愿景。 ## TL;DR 在演讲中，DeepMind 联合创始人兼 CEO 德米斯·哈萨比斯（Demis Hassabis）回顾了自己从国际象棋大师、游戏开发者到神经科学家的跨界成长历程，提出了“先解决智能，再用智能解决其他一切”的宏大两步走使命。他重点展示了团队利用端到端深度强化学习算法从原始像素自学《太空侵略者》和《打砖块》等经典雅达利游戏的突破性成果，该研究也登上了《自然》杂志封面。最后，他指出通用人工智能（AGI）是将海量非结构化数据转化为行动知识的“元解决方案”，不仅能作为 AI 科学家辅助攻克气候与疾病等复杂危机，还将反哺人类对心智与意识本质的理解，同时呼吁全社会必须从现在开始严肃开展 AI 伦理探讨。 ## Atmosphere & style 整场演讲流露出严谨深邃的科学理性与极富远见的宏大叙事感，哈萨比斯的语气沉稳、自信且富有哲理。从童年下棋的个人觉醒到类比阿波罗计划的 DeepMind 创立，情绪层层递进；尤其是展示 AI 算法自主凿穿砖墙发现终极通关秘诀时，现场充满了科技突破带来的惊叹与前沿探索的魅力。 ## Bulletin - [05:13](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=313s) DeepMind 的两步走核心战略是：第一步先攻克通用智能，第二步利用它来解决人类面临的一切复杂问题。 - [11:57](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=717s) 基于原始像素输入与强化学习，DeepMind 的算法在零规则预设下自主掌握雅达利游戏，并在《打砖块》中自创了凿穿侧边反弹的最优解，成果登上《自然》封面。 - [13:33](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=813s) 哈萨比斯将通用人工智能定义为“将非结构化信息自动转化为可行动知识”的元解决方案，未来有望通过“AI 科学家”协同人类攻克气候与疾病等复杂危机。 - [14:58](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=898s) 尽管达到人类水平的通用 AI 仍需数十年时间，但因技术本身具有中性特征，全社会必须从现在开始建立伦理监管与公开大讨论。 ## Key takeaways - [00:08](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=8s) 人工智能本质上是制造聪明机器的科学，也是人类有史以来发明的最关键技术之一。 - [00:08](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=8s) 理解心智的最佳方式是动手将其创造出来，这契合费曼关于“无法创造便未能真正理解”的名言。 - [00:08](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=8s) 哈萨比斯的早期历程跨越国际象棋大师、开创性经营游戏《主题公园》开发以及海马体神经科学博士研究，为创办 DeepMind 奠定了跨学科基础。 - [05:13](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=313s) DeepMind 旨在打造 AI 领域的“阿波罗计划”，其技术路线深受人类大脑架构与神经科学成果启发。 - [05:13](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=313s) 传统专家系统（如深蓝、Watson）属于狭义 AI，而 DeepMind 致力于构建跨复杂环境自主自适应的“通用学习机”。 - [07:36](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=456s) 强化学习系统在信息不完备、充满噪声的感知流中构建统计世界模型并决策，对应哺乳动物大脑的多巴胺机制。 - [09:52](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=592s) AI 仅凭每帧约 30,000 个原始像素数值输入，在单台 GPU 上训练即可在《太空侵略者》中展现超越顶尖人类的预判能力。 - [11:57](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=717s) 在《打砖块》训练中，算法自主发掘出连人类开发者都未曾发现的“侧边凿洞反弹”最优策略，相关论文登上了《自然》封面。 - [12:51](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=771s) 雅达利算法已拓展至 3D 虚拟环境、围棋及物理仿真系统，并在 YouTube 推荐优化与预测性医疗中展开实际落地。 - [13:33](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=813s) 人类当前面临信息超载与系统复杂性（如气候变化、疾病、理论物理）两大瓶颈，极度需要 AI 作为元解决方案突破认知上限。 - [13:33](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=813s) 哈萨比斯的核心愿景是打造能够与人类专家并肩协同工作的“AI 辅助科学家”。 - [14:58](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=898s) 通用人工智能（AGI）在功能本质上是将非结构化数据自动转化为可付诸行动的知识的处理系统。 - [14:58](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=898s) 技术本身是中性的，决定其伦理的是人类的使用方式；虽然人类级通用 AI 仍有数十年距离，但伦理讨论必须立刻展开。 - [14:58](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=898s) 通过构建严谨的人工智能算法模型反哺脑科学，最终将帮助人类破解意识、创造力以及万物理论本身的奥秘。 ## Detailed walkthrough ### [00:08](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=8s) 探索心智的旅程：从国际象棋少年到神经科学研究 Demis Hassabis 首先开宗明义地界定了人工智能的本质，指出 DeepMind 专注于人工智能研究，而 AI 本质上就是“制造聪明机器的科学”。在他看来，AI 将会成为人类有史以来发明的最重要的技术之一。然而，在深入技术细节前，鉴于演讲主题名为《万物理论》（The Theory of Everything），Hassabis 回溯了自己将整个人生与事业奉献给这一难题的个人历程。学生时代的他逐渐意识到，从根本上说世上只有两门学科真正值得研究：探索外部宇宙运作规律的物理学，以及探究人类内部心智世界的神经科学与心理学。在深入思考后，他认定心智更加重要，因为外部世界完全是通过我们的意识去感知的——正如伊曼努尔·康德（Immanuel Kant）所言，正是心智创造了我们身边的现实。要真正理解心智，最好的方法就是将其制造出来；正如物理学家理查德·费曼（Richard Feynman）所说的名言：“‘What I cannot build, I do not truly understand’（我无法创造的事物，我就未能真正理解）”。这段探索的起点其实是棋类游戏。Hassabis 4 岁学棋，随后入选英格兰国家青年队并担任队长，12 岁便晋升为国际象棋大师。这种早期的棋力训练与自省性格促使他不断追问：究竟是大脑内部的什么机制在构思出这些走法？8 岁时，他用国际象棋国际锦标赛上赢取的奖金购买了自己的第一台电脑——ZX Spectrum 48k，并开始自学编程。他编写的第一个大型程序正是一个国际象棋软件，虽然棋力尚浅，但已经能击败他的弟弟。随后，他对编程与棋类游戏的热爱自然融合到了电子游戏开发中。17 岁时，他主导开发了开创性的模拟经营游戏《主题公园》（Theme Park，1994年发布）。这款游戏摒弃了当时主流的射杀玩法，让玩家自由构建类似迪士尼乐园的主题公园，园内成千上万具有独立喜好与情绪属性的小人会根据体验做出反应并反馈给朋友。游戏开创了管理模拟游戏（Management Simulation Games）这一全新流派，也开启了他长达十年的游戏开发生涯。在出售了自己的游戏公司后，为了解开大脑运作之谜，他重返校园攻读神经科学博士，聚焦海马体对记忆与想象力的调控，并在 MIT 和哈佛大学完成了博士后研究。 - 8 岁购入人生第一台计算机 ZX Spectrum 48k 并自学编程。 - 12 岁获得国际象棋大师头衔（Chess Master）。 - 17 岁设计制作风靡全球的模拟经营鼻祖游戏《主题公园》（Theme Park，1994）。 - 神经科学博士重点研究海马体（Hippocampus）与人类的想象力及记忆机制。 --- ### [05:13](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=313s) 创立 DeepMind 与迈向通用学习机在完成神经科学研究后，Hassabis 汇聚了认知科学与计算机科学的所有拼图，于 2010 年联合创立了 DeepMind。DeepMind 创立的核心初衷，是为人工智能打造一个类似于“阿波罗计划”（Apollo Programme）的科研登月任务。团队汇聚了超过 100 位顶尖的博士与机器学习研究科学家，其核心研发思路始终由神经科学启发——在系统级和高层架构上借鉴人类大脑的运作模式。Hassabis 用极度凝练且宏大的“两步走战略”概括公司的长期使命，正如他所强调的： > “So step one, solve intelligence and then step two, use it to solve everything else, and I'll come back to that right at the end why that's not as fanciful as it might seem.”（第一步，解决智能；第二步，用它来解决其他一切问题。在演讲最后我会回过头来谈谈，为什么这个目标并不像它听上去那么异想天开。）从工程实践角度来看，DeepMind 的目标是打造世界上第一台“通用学习机”（General purpose learning machine）。这里的核心关键词有两个：“学习”（Learning）与“通用”（General）。传统的狭义人工智能（Narrow AI）往往依靠手工编码或针对特定规则高度定制，例如 90 年代击败国际象棋世界冠军加里·卡斯帕罗夫（Garry Kasparov）的 IBM 深蓝（Deep Blue），以及后来在电视问答节目中战胜人类选手的 Watson，本质上都是针对单项特定任务设计的专家系统。相反，DeepMind 追求的通用人工智能（AGI）要求算法直接从原始数据中自主掌握技能，杜绝任何任务特化的规则预设，同一套底层算法架构必须能够自适应跨越截然不同的多种复杂环境。 --- ### [07:36](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=456s) 强化学习机制与以经典游戏为基准测试为了实现真正的自主智能，DeepMind 将智能问题置于“强化学习”（Reinforcement Learning）的闭环框架之中。在 Hassabis 呈现的模型中，系统（Agent）与环境之间仅存在两类交互：接收观察与输出动作。不同于棋盘信息完全对称的国际象棋，现实世界充满了噪声与不完全信息（Noisy and Incomplete）。智能体的任务首先是依据充满噪声的传感器输入（在 DeepMind 主要是计算机视觉通道）构建出关于世界的最佳统计模型；随后根据当前状态从所有可用动作集中做出决策，挑选最有可能推进其目标的行动。一旦执行该动作，环境随之发生变化并产生全新的感官观察，从而形成动态循环。在人类与哺乳动物的大脑中，这套强化学习机制正是由多巴胺系统（Dopamine System）驱动实现的。 ```mermaid flowchart LR accTitle: 强化学习交互模型 accDescr: 展示人工智能系统与环境之间通过感官观察与动作输出形成的闭环交互机制 subgraph AgentSystem ["人工智能系统 (AI System)"] WorldModel["构建统计世界模型 (Statistical Model)"] Policy["评估目标并选择最佳动作 (Action Selection)"] WorldModel --> Policy end subgraph EnvironmentBlock ["环境 (Environment)"] EnvState["环境状态发生改变 (真实或虚拟世界)"] end Policy -->|"输出动作 (Actions)"| EnvState EnvState -->|"非完整感官观察 (Sensory Observations, 如视觉)"| WorldModel ``` Hassabis 指出，这种设计之所以能够通向真正的认知，是因为任何真正的思维机器都必须沉浸在感觉运动数据流（Sensorimotor Data Streams）中，完全自主摸索出环境规律。因此，开发团队选取了 20 世纪 80 年代的一系列经典雅达利（Atari）游戏作为测试智能算法的理想实验场。这些游戏原本专为人类的反应与智力挑战而设计，提供了绝佳且公平的通用性评测标准。 - 强化学习系统必须在信息不完备、充满噪声的环境下依靠视觉等感官数据构建统计世界模型。 - 哺乳动物与人类大脑通过多巴胺系统（Dopamine System）实现强化学习闭环。 - 算法评测基准选用 20 世纪 80 年代极具人类挑战性的经典 Atari 街机游戏，要求算法在零先验知识下自学通关。 ### [09:52](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=592s) 玩转雅达利游戏：从零开始的端到端强化学习德米斯·哈萨比斯（Demis Hassabis）展示了人工智能代理玩经典雅达利（Atari）游戏的演示。测试所选用的游戏既没有简单到微不足道，又没有复杂到让现有算法完全无法取得进展。最为关键的是，系统输入的唯一数据就是屏幕上的原始像素（raw pixels），其感知方式与人类直视屏幕别无二致。由于雅达利屏幕分辨率为 200×150 像素，AI 在每一帧画面中接收到的仅仅是大约 30,000 个数值。系统的任务极其单纯：尽可能最大化游戏得分。哈萨比斯强调，系统对游戏规则、控制目标一无所知，所有决策逻辑都是“从零开始”（from scratch）自主学习得来的，而且他们严格要求用同一套系统去通玩数百款不同的雅达利游戏，而不是针对特定游戏定制代码。 ```mermaid flowchart LR accTitle: 端到端深度强化学习架构 accDescr: 展示从屏幕原始像素输入到神经网络策略，再到游戏得分奖励的闭环学习流程 subgraph Input["感知输入"] P["原始像素 (200x150 画面 / 约 3 万数值)"] end subgraph Agent["强化学习系统 (单一 GPU 运行)"] NN["深度神经网络 (构建内部世界模型 / 预测动作)"] end subgraph Environment["游戏环境 (雅达利主机)"] ACT["动作执行 (手柄控制)"] REW["环境反馈 (最大化游戏得分)"] end P --> NN NN --> ACT ACT --> REW REW -.->|"强化学习更新"| NN ``` 第一段展示的视频是雅达利最具代表性的《太空侵略者》（Space Invaders）。刚接入数据流时，AI 既不知道自己控制的是屏幕底部的火箭，也不知道屏幕上的外星人是什么，开局仅仅几秒就白白送掉了全部三条命。然而，在单台 GPU 机器上让它自主训练整整一夜后，第二天系统就达到了超越人类顶尖选手的“超人”水准。AI 发射的每一发炮弹都弹无虚发，它完全掌握了规避敌方攻击的技巧，甚至摸索出屏幕顶端飞过的粉色母舰分值最高，能够以极其惊人的预判精度将其击落。在游戏残局外星人移动速度暴增时，AI 展现出了近乎完美的预测射击能力。正如哈萨比斯评价的那样：“你可以看到它将游戏世界和数据流建模得多么完美，其精确度甚至能让它仅凭屏幕上的像素就提前预测即将发生的事情。” - 画面数据维度：屏幕尺寸为 200 × 150 像素，每帧仅提供约 30,000 个数值作为原始输入。 - 硬件配置：整个学习过程仅需在一台配置单块 GPU 的计算机上通宵运行。 - 通用性要求：同一套底层网络架构与学习算法，不加微调即可通玩数百款雅达利游戏。 --- ### [11:37](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=697s) 《打砖块》中的自创策略与登上《自然》封面随后哈萨比斯播放了他最喜欢的一段视频——《打砖块》（Breakout），这段视频生动地展示了算法随着训练推进而呈现的阶段性飞跃。在完成 100 局游戏后，AI 的表现依然相当糟糕，但仔细观察就会发现它已经开始理解“应当把底部的挡板移向球的位置”这一核心逻辑。到了 300 局游戏之后，AI 已经能够稳定且连续地把球击回，几乎不会失手，其实力已经媲美人类中的顶尖玩家。然而最戏剧性的一幕发生在之后。研究团队心想，如果让机器继续自己玩几百局会发生什么？令人惊叹的现象出现了：AI 自行探索并挖掘出了这款游戏的终极最优解——在屏幕最左侧集中火力凿穿一条竖向隧道，随后以令人难以置信的精准角度把球送入砖块后方，让球在顶部与砖墙背面来回反弹以持续自动得分。哈萨比斯打趣道，参与该项目的优秀程序员和研究员们虽然精通算法和代码编写，但自己玩雅达利游戏其实水平平平，他们根本不知道还有这样一种通关技巧，这一策略完全是机器教给创造者的。哈萨比斯骄傲地提到：“所有这些研究工作已于几个月前发表在世界顶级科学期刊《自然》（Nature）的封面上，如果大家对细节感兴趣，可以在那里查阅。” - 学习阶段演进：100 局建立基础行为（移动板子接球）；300 局达到人类顶级水准（极少漏球）；再训练数百局后自主掌握“侧边凿洞反弹”的最优策略。 - 学术里程碑：相关研究成果作为封面论文发表于权威科学期刊《自然》（Nature）。 --- ### [12:51](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=771s) 从海马体机制到现实世界应用在雅达利游戏取得突破后，DeepMind 的研究重心转向了更高阶的认知能力拓展，包括概念学习、抽象概念理解以及长时记忆系统。哈萨比斯指出，这些研究方向深受他本人及同行在神经科学领域成果的启发，尤其是借鉴和模拟人类大脑中海马体（hippocampus）的工作机制。哈萨比斯澄清，开发这些算法绝非仅仅为了打赢雅达利游戏。团队目前正将这套通用智能系统推向全 3D 游戏环境、围棋（Go）、复杂物理模拟系统，并计划在未来延伸至真实物理机器人领域。就近期的现实落地应用而言，这类技术已被用于改良 YouTube 等产品的推荐引擎，并正在积极布局预测性医疗健康（predictive healthcare）领域。 --- ### [13:33](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=813s) 应对复杂度危机与“AI 辅助科学家”愿景哈萨比斯将演讲重新拉回宏大的“万物理论”（Theory of Everything）主题。他指出人类社会当前正面临两大核心瓶颈：其一是信息超载（information overload）。尽管人人都高谈“大数据”，但真正的痛点在于如何从基因组学或娱乐产业等领域的海量数据洪流中筛选出有价值的洞察。现有的个性化技术效果欠佳，本质原因在于它们往往依赖粗糙的“群体平均”，缺乏真正的精准性。其二则是系统复杂性（system complexity）。人类试图攻克的气候变化、恶性疾病、能源危机、宏观经济学乃至理论物理学等系统日趋极端复杂，其庞大程度甚至让最聪颖的学者倾尽毕生精力都难以完全参透，更遑论留出足够的时间去推陈出新。 ```mermaid flowchart TD accTitle: AI 作为元解决方案应对全球复杂性挑战 accDescr: 展示人类面临的两大挑战以及通用人工智能如何赋能各大科学领域 subgraph Challenges["人类社会核心瓶颈"] C1["信息超载 (数据洪流中提炼洞见)"] C2["系统极度复杂 (气候/疾病/能源/物理)"] end subgraph Engine["核心驱动力"] AGI["通用人工智能 / 元解决方案 (将非结构化数据转化为行动知识)"] end subgraph Solutions["科学与社会突破"] S1["AI 辅助科学家 / AI 科学家"] S2["与人类专家协同攻克气候与健康"] S3["解构意识、创造力与万物理论"] end C1 --> AGI C2 --> AGI AGI --> S1 S1 --> S2 S1 --> S3 ``` 正因如此，哈萨比斯坚信人工智能将是未来最重要的技术，因为攻克 AI 本质上是攻克所有其他现实挑战的“元解决方案”（meta-solution）。哈萨比斯充满激情地表达了他最向往的愿景：“对我而言，研究 AI 最让我感到兴奋的梦想，就是在未来能够创造出‘AI 科学家’，或者实现‘AI 辅助科学研究’，让它与人类专家和科学家并肩协同工作。” --- ### [14:58](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=898s) 伦理监管、与谷歌携手及破解心灵奥秘谈及当时社会各界对人工智能伦理的热烈讨论，哈萨比斯直言，面对任何全新的强大技术，负责任且合乎道德的使用是必然前提。DeepMind 内部设立了专门的伦理委员会来监管该技术的应用与走向。他特别强调，技术本身是中性的，决定其是否符合伦理的关键永远在于人类如何去使用它。尽管距离达到人类水准的通用人工智能（Human-level AI）仍有数十年之遥，但伦理层面的全社会大讨论必须从现在立即展开。关于 DeepMind 与谷歌（Google）的结合，哈萨比斯回顾道，团队在去年初选择加入谷歌，是因为双方的崇高使命高度契合。谷歌的使命是整合全球信息、供大众使用并使人人受益；换言之，就是通过知识为人类赋能。从这一视角审视，哈萨比斯将通用人工智能（AGI）本质定义为“一个能够自动将非结构化信息转化为可付诸行动的知识（actionable knowledge）的过程”。在演讲的结尾，哈萨比斯再次将 AI 与神经科学融为一体。作为一名兼具神经科学家身份的 AI 研究者，他坚信将人工智能提炼为严谨算法模型并与人类大脑进行对照研究，反过来将帮助人类解开意识（consciousness）、创造力（creativity）乃至梦境等心灵深处的终极谜题。在掌声响起前，他留下了全场最具分量的论断：“为了找到‘万物理论’，事实可能证明，我们必须首先解开‘智能’本身的谜题。” - 技术中性与伦理监管：设立伦理委员会；强调“技术本身是中性的，最终决定其是否符合伦理的永远是人类使用该技术的方式”。 - 发展时间表预判：达到人类级别的通用 AI 仍需要“数十年时间”（several decades away），但跨领域的伦理研讨刻不容缓。 - AGI 的功能本质：自动将非结构化信息（unstructured information）转化为可行动知识（actionable knowledge）的处理过程。 ## Notable quotes > [14:58](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=898s) 为了找到‘万物理论’，事实可能证明，我们必须首先解开‘智能’本身的谜题。 — Demis Hassabis > [13:33](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=813s) 对我而言，研究 AI 最让我感到兴奋的梦想，就是在未来能够创造出‘AI 科学家’，或者实现‘AI 辅助科学研究’，让它与人类专家和科学家并肩协同工作。 — Demis Hassabis > [14:58](https://www.youtube.com/watch?v=rbsqaJwpu6A&t=898s) 技术本身是中性的，最终决定其是否符合伦理的永远是人类使用该技术的方式 — Demis Hassabis ## Mentioned Demis Hassabis, DeepMind, Google, ZX Spectrum 48k, 主题公园 (Theme Park), 海马体 (Hippocampus), 麻省理工学院 (MIT), 哈佛大学 (Harvard University), 深蓝 (Deep Blue), Garry Kasparov, Watson, Atari, 太空侵略者 (Space Invaders), 打砖块 (Breakout), 自然 (Nature), YouTube, Richard Feynman, Immanuel Kant ## Video description Google Zeitgeist is a collection of talks by people who are changing the world. Hear entrepreneurs, CEOs, storytellers, scientists, and dreamers share their visions of how we can shape tomorrow. [Source](https://www.youtube.com/watch?v=rbsqaJwpu6A)

Source quotes

[14:58] Demis Hassabis: 为了找到‘万物理论’，事实可能证明，我们必须首先解开‘智能’本身的谜题。 [13:33] Demis Hassabis: 对我而言，研究 AI 最让我感到兴奋的梦想，就是在未来能够创造出‘AI 科学家’，或者实现‘AI 辅助科学研究’，让它与人类专家和科学家并肩协同工作。 [14:58] Demis Hassabis: 技术本身是中性的，最终决定其是否符合伦理的永远是人类使用该技术的方式

Source quotes

[13:33] Demis Hassabis: The technology itself is neutral. It's always how humans use this technology that ends up deciding whether it's ethical or not. [15:24] Demis Hassabis: Finally, I would say that, you know, in order to find the Theory of Everything, it might turn out that we have to solve intelligence first.

Original transcript

[0:08] At DeepMind, we basically [0:10] work on artificial intelligence [0:12] and AI is the science [0:14] of making machines smart [0:16] I believe that this is going to end [0:17] up becoming one of the most important [0:19] technologies that humanity [0:20] will ever invent, but before I [0:23] get into that, because this session [0:24] is called The Theory of Everything [0:26] I thought I would start with a little bit [0:28] of an explanation [0:29] of my personal journey to [0:30] how I got to this point, and [0:32] why I've decided to, you [0:34] know, dedicate my life and career [0:36] to work on this problem. [0:38] So from a young age at school, I [0:40] kind of, came to this realisation [0:42] that in some fundamental [0:44] sense there are only two subjects [0:45] really worth studying, physics [0:47] and neuroscience. [0:49] For physics, of course, it's [0:51] all about explaining the external [0:54] world, so the external [0:56] world out there, including, of course [0:59] the entire universe, and [1:01] neuroscience and psychology [1:02] is really about conversely [1:04] explaining about what's inside [1:06] here, our internal world [1:09] Then when I thought about this more [1:11] I actually came to the conclusion [1:12] that the mind was more important [1:14] because obviously that's the [1:15] way we actually interpret the external [1:17] world out there, and in fact, you know [1:20] 'the mind interprets the world' [1:22] is something that was an idea that was [1:23] first proposed by the great philosopher [1:25] Immanuel Kant, and [1:27] really, it's the mind that creates [1:29] our reality around us [1:33] So this is where AI comes [1:35] in, because [1:37] the ultimate expression of understanding [1:38] something is being able to [1:40] recreate it, and as [1:42] Richard Feynman said, one of my scientific [1:43] heroes, 'What I cannot [1:45] build, I do not truly understand,' [1:48] and that's one of the things that [1:49] I'm excited about with artificial [1:51] intelligence. I think, ultimately [1:52] it will help us understand our own [1:53] minds better. [1:57] So my personal journey, as Darsheni [1:59] mentioned, started with games [2:01] and that's how I got into [2:03] AI. I started off playing chess [2:05] I was taught how to play chess when I [2:07] was four, ended up playing various [2:09] England chess teams, and captaining [2:10] various England junior chess teams [2:12] and by age of twelve I was a chess [2:14] master. The thing is, when you [2:16] teach a kid from a very [2:18] young age how to play chess [2:20] and if they have quite a reflective [2:22] personality like I had when I [2:23] was young, you can't help but [2:25] thinking and introspecting [2:27] about what it is about your mind [2:28] that is actually coming up with [2:30] these moves. What are the mechanisms [2:32] that allow you to make these [2:33] plans in such a complex game [2:35] as chess? [2:37] So then when I was around eight years [2:39] old, I actually took some [2:40] winnings that I won from an international [2:42] chess tournament, and I bought my [2:44] first computer, a ZX Spectrum [2:46] 48k, and I taught myself [2:48] how to programme. [2:50] One of the first, sort of, big [2:51] programmes I can remember creating [2:54] was actually a programme to play [2:55] chess, and it didn't play very well [2:58] but it was able to beat my little brother [2:59] which I was very pleased about when I [3:01] was small. [3:03] You know, it worked, and this was the [3:04] beginning of the path to me [3:06] towards AI. [3:09] Now, my love of programming, and chess [3:11] and games, sort of, came together [3:12] naturally in the form of [3:14] video games, and my first [3:17] career was in creating and designing [3:19] video games. I did this for ten [3:22] years, and I wrote several bestselling [3:24] games, but probably my most famous [3:25] game I wrote was called Theme Park [3:28] which I wrote when I was seventeen. This [3:30] is one of the first games to [3:31] use artificial intelligence [3:33] as the main gameplay component [3:35] So this game came out in the [3:38] mid-90s, in '94, and the [3:40] idea of the game was that you [3:42] designed your own Disney World [3:43] and thousands of little people [3:45] with their own desires and [3:46] characteristics came into [3:48] that Disney World and judged [3:50] and decided how much fun they had [3:51] in your theme park. [3:53] They would go and tell their friends, and they [3:55] would come the next day. [3:57] So this game spawned [3:59] a whole genre of games called management [4:00] simulation games, and [4:02] really, sort of, started this [4:04] whole genre of [4:06] creative games, where instead [4:07] of shooting and killing things in [4:09] games, actually you create and [4:11] built stuff yourself, and the game [4:12] would react to how [4:14] you played the game, so no two [4:16] people ended up with the same [4:17] game, because the AI adapted [4:19] to how the player played it. [4:22] In fact all my games involved [4:25] a lot of AI, [4:26] and then the final piece of the jigsaw [4:28] for me was, after doing this [4:30] for ten years, I sold my games company [4:32] and I went back to university [4:33] to do a PhD in neuroscience [4:35] to study how the brain itself [4:37] solved some of these hard problems [4:41] I chose, as my topics [4:43] imagination and memory [4:45] and an area of the brain called the hippocampus [4:47] which is responsible for imagination [4:49] and memory, because these [4:50] are two of the capabilities [4:52] that we don't know how to do very [4:54] well in AI. I wanted to [4:56] see and get some inspiration for [4:58] how the brain actually solves [4:59] these problems. [5:01] So after a couple of post-docs [5:02] at MIT and Harvard, I [5:05] then decided that I had all the [5:06] ingredients and the components [5:07] to start DeepMind, and [5:10] actually attack the AI [5:11] problem head on. [5:13] So all these experiences then [5:15] culminated, in 2010 [5:16] in me co-founding DeepMind [5:19] and the idea behind DeepMind [5:21] was really to create a, kind [5:23] of, Apollo Programme mission [5:25] for AI. [5:27] Now, at DeepMind, we have [5:28] over 100 research scientists [5:30] 100 PhDs, top people [5:33] in their machine learning fields [5:35] and neuroscience files working [5:37] on solving AI. [5:39] The type of AI we work on [5:41] is this neuroscience-inspired [5:43] AI, so inspired by how [5:45] the brain works at a very high [5:47] level, a systems level. [5:51] So one way we articulate our [5:52] mission at DeepMind, it's very [5:54] easy to articulate actually, but obviously it's [5:55] hard to do, is a, kind of [5:57] two step process. So step [5:58] one, solve intelligence [6:00] and then step two, use it [6:01] to solve everything else, and [6:04] I'll come back to that right at the end [6:05] why that's not as fanciful [6:08] as it might seem. [6:09] So more prosaically, how [6:11] is it that we're going to practically [6:13] do this? Well, we're going [6:15] to go about doing this by trying [6:17] to build the world's first general [6:18] purpose learning machine, and the [6:20] two key words here are the words [6:21] 'general' and 'learning'. So [6:23] all the AI that we do [6:25] at DeepMind involves [6:27] learning algorithms. So these [6:29] are algorithms that learn, automatically [6:31] how to master tasks [6:33] from raw data. They're not [6:35] pre-programmed or handcrafted [6:36] in any way, so that's unlike most [6:38] AI out there that you've heard of [6:41] There's also the second part. We [6:43] enforce this idea of wanting [6:45] the system to be general, so [6:46] i.e. the same system or same [6:48] set of algorithms can actually operate [6:49] across a wide range of tasks [6:52] Now, 'AI', of course, is a big buzzword [6:54] at the moment, and there have [6:56] been some big achievements in AI [6:57] that you'll of course be aware of, like [6:59] Deep Blue beating Garry Kasparov [7:01] in the '90s, and then more recently [7:02] IBM's Watson besting [7:04] the Jeopardy contestants [7:07] but in our opinion, that's still [7:09] examples of what we would call [7:10] narrow AI. So this is AI [7:12] that has been specifically [7:14] tailored or built to [7:15] tackle one problem, and one [7:17] problem only, and the hallmark [7:20] of, sort of, artificial general [7:21] intelligence, the type of AI we [7:23] work on, is that it's built to [7:24] be flexible, and general [7:26] and adaptive from the ground [7:27] up, so there's no special [7:29] casing or pre-programming [7:31] of the task involved. It has to [7:33] learn everything from first principles [7:36] So we, sort of, think about the intelligence [7:38] problem within this framework [7:40] called 'reinforcement learning' [7:41] and it's a very simple framework to [7:43] describe, and I'm just going to describe it with this [7:44] simple diagram. So on the left-hand [7:46] side here you have the system [7:48] itself, the AI system [7:50] and the AI system finds itself [7:52] in some kind of environment [7:56] that it's trying to achieve a goal [7:57] in, and that environment could be real [7:59] world or virtual. [8:01] Now, the system only interacts with [8:03] the environment in two ways. So [8:04] firstly it gets observations [8:06] about the environment through its [8:07] sensory apparatus. We normally [8:09] use vision at DeepMind, but [8:10] you could use other modalities [8:12] and these observations are always [8:14] noisy and incomplete. So unlike [8:15] the game of chess, the real world [8:17] is actually very noisy [8:19] and messy, and you never have full [8:20] information about what's going on [8:23] The job of the system is to build [8:25] the best model of the world [8:26] out there, statistical model [8:28] of the world out there based on [8:29] these noisy observations [8:31] and once it has that model of the [8:32] world, the second job of the system [8:34] is to pick the best action [8:37] that will get it closest towards its [8:38] goal from the set of actions [8:40] that are available to it at that moment [8:42] in time. Once the system has decided [8:44] which action that it, it outputs [8:46] that action, that action gets executed [8:48] it may or may not make some change [8:49] to the environment, and that drives a new [8:50] observation. This whole system [8:53] although it's very simply described [8:55] in this diagram, it has lots [8:57] and lots of hidden complexities [8:59] If we could solve everything behind [9:00] this diagram, that would be enough [9:02] for intelligence. [9:04] We know that that's enough for intelligence [9:06] because this is the way that all mammals [9:07] including humans, learn. In humans [9:10] it's the dopamine system [9:11] in our brains that implements [9:13] reinforcement learning. [9:16] So I'm just going to show you a few videos- [9:17] a couple of videos of the algorithm [9:19] working, but before I show you that, I just want to [9:20] explain clearly what it is that [9:21] you're going to see. [9:24] So we used games as [9:26] a test bed for testing [9:28] the intelligence of our algorithms [9:30] and in order to have true [9:31] thinking machines or cognition [9:33] a system has to be embedded [9:35] in sensory motor data streams [9:37] sensory motor reality, and [9:39] it has to figure things out for itself [9:41] So games are actually quite a perfect [9:43] setting for this, and in fact [9:44] we used [9:46] classic Atari games [9:48] from the '80s, which were [9:50] designed to be challenging to humans [9:52] but are not so complex that [9:54] AI algorithms couldn't make progress [9:55] with them. [9:56] So what I'm going to show you is the AI [9:58] playing these Atari games [10:00] but the only thing the system gets [10:02] is the raw pixels as inputs [10:04] so it's just like a human looking [10:06] at the screen, seeing all [10:08] the pixels on the screen. So there's about [10:10] 30,000 numbers per [10:11] frame because the screen's 200 [10:13] x 150 pixels in size [10:15] and the goal here is to simply maximise [10:17] the score. Everything is learnt from scratch [10:19] and we insist that the same system [10:21] plays all the different Atari [10:23] games, hundreds of different Atari [10:24] games. [10:26] So I'm just going to run this video now [10:28] This is a one minute video. This is [10:30] Space Invaders, the most iconic [10:31] game, probably, on Atari [10:34] This is the first time the AI has [10:36] ever seen this data stream, so [10:38] don't forget, it doesn't know what it's playing [10:39] it doesn't know what it's controlling, and [10:41] you can see it's actually losing its [10:42] three lives. It's controlling the rocket here [10:44] at the bottom of the screen, and it's losing [10:45] its three lives immediately, because it doesn't [10:47] know what it's doing. After you leave it playing [10:49] overnight on a single GPU [10:51] machine, you come back the next day [10:53] and now it's superhuman at the [10:54] game. It's learnt for itself [10:55] through experience, how to play [10:57] So you can see now every single [10:59] shot it fires hits something [11:01] It can't be killed anymore. It's [11:02] worked out that the pink mothership [11:04] that comes across the top of the screen [11:05] in a second is worth the most [11:07] number of points. It does these [11:09] amazingly accurate shots to do that [11:11] Those of you who remember Space Invaders [11:13] as there are of them on the screen [11:15] they go faster. Just watch [11:17] the last shot that the rocket [11:19] does. This is predictive shot [11:21] to hit the last Space Invader [11:23] So you can see how perfectly [11:26] it, sort of, modelled [11:28] the game world, and that data [11:30] stream. So accurately, it can [11:32] predict ahead of time what is going [11:33] to happen, just from the pixels [11:35] on the screen. So here's a second [11:37] video. It's my favourite video actually. This [11:38] is a game of Breakout. There are more [11:40] gradations here of the agent [11:42] getting better, the system getting better [11:44] So this is after 100 games [11:46] so just 100 games, and you can see [11:47] again here, the system is pretty [11:49] terrible, but you can [11:51] probably convince yourself that maybe [11:52] it's starting to get the hang of the fact [11:53] that it should move the bat towards [11:55] the ball. [11:56] Now, this is after 300 games [11:58] so it's now hitting the ball [12:00] back pretty consistently [12:02] and it's almost never missing, so it's [12:03] about as good as the best humans [12:05] can be at this game. [12:06] Then we thought, 'That's pretty cool. What [12:08] would happen if we just left [12:10] the machine playing the game [12:12] for a couple more hundred games?' This [12:13] amazing thing happened. What happened was [12:15] it discovered the optimal strategy [12:16] was to dig a tunnel round the left-hand [12:18] side here, and then send the ball [12:20] you know, with this unbelievable accuracy [12:22] round the back. [12:23] So that's really cool, because actually [12:26] the brilliant programmers and researchers [12:28] who are on this programme, they're brilliant [12:29] at programming and coming up with algorithms [12:31] but they're not so good at playing Atari [12:33] So they didn't actually know that strategy [12:35] for themselves, so this was something their [12:37] own creation taught them [12:39] So, you know, all this work was [12:41] then actually published on the front [12:43] cover of Nature a couple of months [12:44] ago, which is the biggest science journal [12:46] in the world, and so if you're interested in reading [12:48] more about these details, you can check [12:49] it there. [12:51] Now we're moving on to adding [12:52] things, capabilities and, like [12:54] concepts, and learning abstract [12:55] concepts, and long-term memory [12:58] These are things that are inspired [13:00] by my work and other people's [13:02] work in neuroscience [13:04] and around mimicking [13:06] the workings of this part of the brain [13:08] called the hippocampus. [13:11] Of course, we're not just building these algorithms [13:13] just to play Atari games [13:15] We're moving now towards 3D [13:16] games, Go, simulations [13:19] and then ultimately real robots [13:20] at some point, [13:22] and more near-term, in terms of applications [13:24] we're using it to improve recommendation [13:26] systems like on YouTube [13:27] and also moving into predictive [13:30] healthcare applications. [13:33] Now, I just want to end by coming back [13:35] to this, sort of, theme of Theory of Everything [13:37] So I think, you know, two of the biggest [13:39] challenges facing us [13:41] as a society are [13:43] information overload, just [13:44] the sense that there's so much [13:46] data around. You know, everyone [13:47] talks about big data, but the problem [13:49] is, what to do with it once you [13:51] have it all. [13:52] I think loads of areas like genomics [13:54] and entertainment, you know [13:56] are all, sort of, suffering from this deluge [13:58] of data. 'How do we sift through [13:59] this data to find the insights in that [14:01] data?' Of course, personalisation [14:03] technologies are one of the technologies [14:04] that are trying to help us with that [14:06] but they don't work very well at the moment [14:08] partly because they're not very personalised [14:09] and they work by, sort of, averaging [14:11] the crowd. [14:12] Secondly, there's the problem of system [14:14] complexity. You know, many of the systems [14:16] we would like to master as a society [14:18] like climate, disease, energy [14:20] economics, even physics [14:22] are getting so complex now [14:24] You know, it's difficult for even the [14:26] best and the smartest humans [14:27] to master it in their lifetimes [14:29] and still leave enough time for them to innovate [14:31] So one of the reasons I work on AI [14:33] and why I think it's going to be one of the most important [14:35] technologies out there, is that solving [14:36] AI is potentially a, kind of, meta-solution [14:38] to all these other problems. We can [14:40] use it to help us solve all [14:42] of these other problems. The dream [14:44] for me, the thing I get most excited [14:45] about working on AI, is in the future [14:47] being able to make and create [14:49] AI scientists, or [14:50] AI assisted science, making [14:52] that possible, working in tandem [14:54] with human experts and human [14:55] scientists. Now, of course, there [14:58] has been a lot of news at the moment about [15:00] the ethics around AI, and [15:01] like with all new, powerful technologies [15:03] you know, it must be used ethically [15:05] and responsibly, and we're actively [15:07] researching and doing these things [15:08] and we have an ethics board that governs [15:10] the use of this technology. The technology [15:12] itself is neutral. It's always how [15:14] humans use this technology [15:15] that ends up deciding whether it's ethical [15:17] or not. [15:18] Of course, human level AI is several [15:21] decades away, but we should start [15:22] the debate now. I just [15:24] want to end by just talking about- [15:26] we, sort of, joined forces with Google [15:28] early last year, and Google's [15:31] mission statement, of course, is to organised [15:32] the world's information, and make it universally [15:34] accessible and useful. [15:35] One reason we decided to join forces with [15:37] Google was that we felt our mission [15:39] fitted very well with this mission [15:41] Another way of, sort of, describing [15:43] that mission is to think about empowering [15:45] people through knowledge. Another [15:47] way to think about this, kind of, AI [15:49] or artificial general intelligence [15:50] is that it's a process that automatically [15:52] converts unstructured [15:54] information into actionable [15:56] knowledge. [15:58] So just to finish with a, sort of, slide [16:00] about the Theory of Everything. [16:02] You know, of course, I'm a neuroscientist [16:04] as well as an AI researcher [16:06] and I think that by trying [16:08] to distil artificial intelligence [16:10] into an algorithmic construct [16:12] and comparing it to the human [16:13] mind, that might help us to unlock [16:15] some of the deepest mysteries of [16:16] the mind, like consciousness [16:18] creativity and even dreams [16:21] Finally, I would say that, you [16:23] know, in order to find the Theory [16:24] of Everything, it might turn out [16:26] that we have to solve intelligence [16:27] first. Thanks for listening.