#332 — Can We Contain Artificial Intelligence?

Episode Summary

Executive Summary: Sam Harris and Mustafa Suleiman discuss Suleiman’s path from philosophy and conflict resolution into AI, DeepMind’s major breakthroughs, and his thesis in The Coming Wave: intelligence and synthetic biology will proliferate rapidly, creating major gains but also serious risks. Suleiman argues the central challenge is containment—keeping powerful technologies accountable to humans—while warning of labor disruption, misinformation, and geopolitical competition.

Main Topics: Suleiman’s background and path into AI (Priority: 4/5): He describes moving from philosophy and public service into entrepreneurship, conflict resolution, and eventually DeepMind, driven by a desire to do practical good at scale. DeepMind’s technical breakthroughs (Priority: 5/5): The conversation reviews DeepMind’s early bet on deep learning, reinforcement learning, Atari DQN, AlphaGo, AlphaZero, and AlphaFold as landmark demonstrations that AI can discover new strategies and knowledge. The coming wave thesis (Priority: 5/5): Suleiman defines the coming wave as general-purpose technologies—especially AI and synthetic biology—that spread rapidly, get cheaper, and reshape society much like past waves such as fire and electricity. Risk: containment and misuse (Priority: 5/5): He argues the key issue is not runaway superintelligence alone, but the loss of human control over increasingly capable systems and their use for misinformation, surveillance, military applications, and asymmetric threats. Labor disruption and economic change (Priority: 4/5): The discussion challenges the idea that AI will only create new jobs; Suleiman and Harris focus on AI as a true substitute for human cognition, especially in white-collar and knowledge work. Open source, scale, and the proliferation of power (Priority: 4/5): Suleiman warns that frontier capabilities tend to diffuse quickly, making containment harder; he emphasizes how falling costs and scaling compute will spread advanced models widely. Geopolitics and governance (Priority: 3/5): They touch on regulatory capture, the need for friction and oversight, and tensions around competition, cooperation, and supply chains—especially in relation to China.

Key Arguments: Suleiman says his concern about AI is longstanding; he built DeepMind around “safely and ethically” developing AGI and insisted on safety oversight as part of the Google acquisition. He argues that superintelligence talk can distract from more immediate and practical harms: misinformation, destabilization, easier abuse of power, and uncontrolled deployment. He defines intelligence as a power-enhancing capability that can increasingly perform actions, not merely generate text, making modern AI an “artificial capable intelligence” problem as much as a language problem. He contends that general-purpose technologies inevitably proliferate and become cheaper, so AI’s spread is not optional; the real task is governance and containment. He rejects the idea that AI labor effects will mirror past industrial transitions, because AI directly targets cognitive work and can substitute for judgment, creativity, and emotional labor. He emphasizes that open-source diffusion and rapidly falling serving costs mean frontier capabilities will become broadly available, raising the risk of abuse and reducing barriers to entry. He argues that governments and companies need practical mechanisms of control, accountability, and friction, rather than relying on a fantasy of keeping systems permanently boxed up. He believes the upside is enormous—massive productivity gains, better education, medicine, and personal assistance—but only if society proactively manages the downsides.

Data Points: DeepMind founding year: 2010 - Suleiman says DeepMind was founded in the summer of 2010. Atari DQN compute: 2 petaflops - He says Atari DQN used two petaflops of computation over several weeks. Go board configurations: 10^170 possible configurations - Used to illustrate the complexity of Go and why new methods were needed. AlphaFold structures open-sourced: 200 million protein structures - He says AlphaFold 2 open sourced roughly 200 million protein structures in one go. Protein structures previously experimented on: 190,000 proteins - He contrasts the prior scale of experiments with AlphaFold’s leap. GPT-3 size: 175 billion parameters - He cites GPT-3 as a frontier model released in 2020. Open-source training/serving scale reduction: 2 billion parameters - He says GPT-3-like capabilities are now regularly being trained at about 2 billion parameters. Compute increase over a decade: 5 billion times - He says Inflection’s Pi uses about five billion times the compute of Atari DQN. Compute growth rate: 10X per year for 10 years - He summarizes the decade-long scaling trend as roughly one order of magnitude per year. Top smartphone access: Top 2 billion people - He argues smartphones are a highly meritocratic technology available across income levels in the developed world.

Pivotal Quotes: "building artificial general intelligence safely and ethically for the benefit of everyone" — Mustafa Suleiman: He describes the founding mission of DeepMind and the safety-first framing from the company’s early days. "the focus on this, you know, sort of intelligence explosion ... has consumed way more time than the idea justifies" — Mustafa Suleiman: He explains why he thinks superintelligence has been overemphasized relative to near-term risks. "if we don't have mechanisms to shape that and restrict its capabilities, then it potentially leads us into some quite catastrophic outcomes" — Mustafa Suleiman: He defines the containment problem and why human control over technology matters.

Implications: The conversation argues that AI’s benefits are real and likely enormous, but society must prepare for rapid diffusion, labor disruption, misinformation, and misuse. The key challenge is governance: building mechanisms that keep powerful systems accountable before they become too widespread to control.