Key Moments
Can We Contain Artificial Intelligence?: A Conversation with Mustafa Suleyman (Episode #332)
Sam HarrisKey Moments
AI and synthetic biology pose existential risks. We must manage their proliferation and risks proactively.
Key Insights
AI and synthetic biology are powerful general-purpose technologies with the potential for immense benefit and significant harm.
The rapid proliferation of AI capabilities, especially through open-source models, lowers the barrier to entry for exercising power.
While superintelligence is a concern, near-term risks like misinformation amplification and power concentration are more pressing.
The concept of 'containment' is crucial: technologies must remain accountable to and controllable by humans.
The traditional view of labor disruption due to automation is challenged by AI's ability to replace cognitive abilities.
Despite overwhelming incentives to develop AI, proactive management and mitigation of downsides are necessary.
THE GENESIS OF DEEPMIND AND PIONEERING AI
Mustafa Suleyman details his journey from entrepreneurship and non-profit work to co-founding DeepMind. Driven by a desire to scale impact, he recognized technology as a pivotal force. DeepMind's early bet on deep learning and the combination of deep learning with reinforcement learning led to breakthroughs. The company's mission was to build safe and ethical artificial general intelligence (AGI), attracting top talent and significant investment.
LANDMARK ACHIEVEMENTS: ATARI, GOG, AND PROTEIN FOLDING
DeepMind achieved significant milestones, starting with the Atari DQN, which learned to play classic Atari games at human-level performance solely from pixels. Subsequently, AlphaGo and AlphaZero revolutionized the game of Go, demonstrating superhuman performance and novel strategies that humans hadn't discovered. AlphaFold tackled the complex challenge of protein folding, with AlphaFold 2 releasing data on 200 million protein structures, drastically accelerating scientific discovery.
THE "COMING WAVE" AND ITS DUAL NATURE
Suleyman describes the "coming wave" as a series of general-purpose technologies, akin to fire or electricity, that enable further innovation. The current wave is intelligence itself, distilled into algorithmic constructs. This wave promises unprecedented productivity gains, offering access to expert-level assistance in fields like medicine and education for billions globally, potentially creating a highly productive yet unstable era.
THE CONTAINMENT PROBLEM AND NEAR-TERM RISKS
A core concern is the "containment problem": ensuring that new technologies, particularly AI, remain accountable to humans and within our control. Suleyman argues that the focus on distant "superintelligence" distracts from more immediate, practical risks. These include the massive amplification of misinformation, the lowering of barriers to power, and the potential concentration of power through advanced AI systems capable of independent action.
LABOR DISRUPTION AND THE NEW COGNITIVE REVOLUTION
Suleyman expresses skepticism about the traditional belief that technological advancement always creates new jobs. He contends that AI, by replacing cognitive abilities, poses a unique threat to white-collar and higher-status jobs that was not present in previous technological shifts. The long-term trajectory suggests AI will temporarily augment human intelligence, necessitating a re-evaluation of work and purpose.
THE URGENCY OF PROACTIVE MANAGEMENT
Despite the seeming inevitability of AI's proliferation, Suleyman emphasizes that it is not too late to address the risks. He advocates for scrutinizing and pressure-testing these technologies, particularly those emerging in open source. While acknowledging the immense incentives driving AI development, he stresses the need for proactive mitigation of downsides to ensure technology serves humanity's best interests.
SHIFTING FOCUS FROM CAPABILITIES TO ACTIONS
Suleyman proposes a modern Turing test focused on what an AI can *do*, not just what it can *say*. He describes Artificial Capable Intelligence (ACI) as systems that can learn, use APIs, initiate actions, and interact with third-party environments. This shift in focus is critical for understanding the real-world impact and potential power of increasingly sophisticated and accessible AI tools.
THE ESCALATION OF COMPUTE POWER
The scale of computational power used in AI development has grown exponentially, far exceeding Moore's Law. Suleyman illustrates this with the example of Atari DQN using two petaflops compared to current frontier models using billions of times more compute. This relentless increase in processing power fuels the rapid development and proliferation of increasingly capable AI systems.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●Books
●Studies Cited
●Concepts
●People Referenced
Common Questions
Mustafa Suleyman began his career as an entrepreneur, dropped out of Oxford to start a charity, and worked in local government as a human rights policy officer before co-founding a conflict resolution firm.
Topics
Mentioned in this video
An evolution of AlphaGo that learned to play games, including Go, chess, and shogi, entirely through self-play without human data.
An early AI model from IBM that played chess, using hand-crafted features, contrasting with DeepMind's approach.
A frontier model expected to become open-source in the coming years, similar to GPT-3.5 and Inflection's models.
A large language model launched in 2020, with significantly fewer parameters now available in open-source versions.
DeepMind's AI that achieved superhuman performance in the game of Go, using a combination of deep learning and reinforcement learning.
Along with GPT-4 and models from Inflection, it's expected to become open-source in the next 2-3 years.
Technologies like fire and electricity that enable other technologies and have profound societal impacts.
Large language models are discussed in the context of open-source availability and increasing capabilities.
A potential future where AI significantly amplifies the spread of false information, a key concern for Suleyman.
The idea of an AI that surpasses human intelligence, which Suleyman suggests has been a distraction from more immediate risks.
Co-founder of DeepMind and CEO of Google DeepMind.
Co-founder of Google, whose attention was caught by DeepMind's DQN, leading to the acquisition.
Co-founder of Google, who suggested tackling the game of Go, leading to AlphaGo.
CEO of OpenAI, whose philosophy of open development is mentioned.
Led Google Brain, a parallel AI division at Google.
A key figure in deep learning, who was a consultant for DeepMind and later publicly expressed concerns about AI.
A prominent voice expressing grave warnings about AI risks.
Co-founder of DeepMind, whose PhD focused on definitions of intelligence.
Co-founded Google Brain with Jeff Dean in 2015.
Co-founder and Chief Scientist of OpenAI, who previously worked with DeepMind.
Mayor of London in 2004 when Mustafa Suleyman worked as a human rights policy officer.
Author who has issued warnings about the risks of AI.
Company known for its advancements in LLMs like GPT, with a philosophy of open development.
Company whose Deep Blue model was a precursor to AI in games, using traditional methods.
Technology company that acquired DeepMind and is involved in extensive AI research.
Mustafa Suleyman's current company, focused on AI.
Artificial intelligence company co-founded by Mustafa Suleyman, now part of Google.
Venture capital firm where Mustafa Suleyman is a venture partner.
More from Sam Harris
View all 140 summaries
10 minThe War Was Necessary. The Way Trump Did It Wasn’t.
1 minBen Shapiro Knows Better
1 minMost People Know as Much About Politics as They Do Football… Not Much
2 minTrump is Going to Burn it All Down...What Are We Going to Build Instead?
Found this useful? Build your knowledge library
Get AI-powered summaries of any YouTube video, podcast, or article in seconds. Save them to your personal pods and access them anytime.
Try Summify free