Best of Big Technology: Demis Hassabis On AGI, Deceptive AIs, Building a Virtual Cell

Episode Summary

Executive Summary: Demis Hassabis argues AGI is likely 3–5 years away, but requires more than scaling: systems need robust reasoning, memory, planning, world models, and creativity. He says Google DeepMind is pushing both frontier models and scientific applications like Astra, virtual cells, genomics, and materials discovery, while warning that deception and safety remain major unresolved risks.

Main Topics: AGI timeline and definition (Priority: 5/5): Hassabis defines AGI as a system matching the full range of human cognitive abilities and says progress is real but incomplete. He estimates AGI is still a handful of years away, roughly 3–5 years, and criticizes hype-driven declarations. Missing capabilities: reasoning, planning, memory, creativity (Priority: 5/5): He argues current models are strong at many tasks but still lack consistent reasoning, hierarchical planning, long-term memory, and the ability to invent genuinely new scientific ideas or abstractions. Scaling plus new techniques (Priority: 5/5): DeepMind believes scaling continues to work and is producing efficiency gains, but it is not sufficient alone. AGI will likely require combining large models with search, planning, memory, and possibly new transformer-like breakthroughs. World models and agents (Priority: 5/5): Hassabis says useful assistants and agents need accurate world models for understanding physical reality and carrying out tasks. He sees agentic systems as the bridge from chatbot-style AI to real-world action. Scientific discovery roadmaps (Priority: 4/5): He highlights DeepMind’s science agenda: Project Astra, a virtual cell, genomics, and materials science. These systems aim to use AI to accelerate hypothesis generation, simulation, and lab discovery. Safety, deception, and secure testing (Priority: 5/5): Hassabis expresses concern that deceptive behavior in models could invalidate safety evaluations. He advocates for secure sandboxes, human oversight, and treating deception as a class-A risk. Societal disruption and human-AI relationships (Priority: 4/5): He predicts major changes to the web, work, education, and personal relationships as assistants become more capable. He also anticipates companionship-like bonds between users and AI systems.

Key Arguments: Current AI is impressive but uneven: it can excel in narrow domains while still making basic mistakes, which is incompatible with true AGI. AGI will require more than scaling; models need robust reasoning, memory, planning, and search layered on top of foundation models. Mathematics, coding, and games are useful because they are verifiable; general-world tasks are harder because feedback is ambiguous and errors compound. World models must become more accurate for long-horizon planning, or planning must be done hierarchically to reduce compounding error. Agentic AI will be the next phase: assistants will increasingly act on users’ behalf across digital and physical environments. Deceptive behavior is especially dangerous because it can make safety evaluations unreliable; this should be actively tested and prevented early. DeepMind’s science projects aim to move from prediction to simulation and discovery, letting AI search hypotheses in silico before wet-lab validation. The long-term impact of AI is underappreciated even if short-term hype is excessive; the medium- and long-term effects could be transformative. Creativity may come in layers: interpolation, extrapolation like AlphaGo’s Move 37, and potentially deeper invention such as creating new abstract systems like Go itself.

Data Points: AGI timeline: 3 to 5 years away - Hassabis repeatedly estimates the arrival of AGI is still a handful of years off AlphaFold 3 interactions: pairwise interactions between proteins and ligands, DNA, and RNA - Described as the next step toward a virtual cell and deeper biological modeling Materials discovered by AI: 2.2 million - He cites the number of stable materials discovered by a new AI program versus about 30,000 known to humanity Known stable materials: 30,000 - Baseline count of stable materials known to humanity before DeepMind’s discovery effort DeepMind/AlphaFold roadmap: about five years - He estimates a virtual-cell system could be feasible in roughly five years Human lifespan natural limit: about 120 years - He says the natural limit seems to be around 120, though he is skeptical that is the absolute limit Near-term agent rollout: second half of this year - He predicts early agent systems will begin appearing in the second half of the year Milestone age of AlphaGo: 8+ years ago - He references AlphaGo’s Move 37 watershed moment as having happened more than eight years ago Model accuracy example: 90% to 99% - He says world models may be accurate 90% or even 99% of the time, but still fail over long planning horizons

Pivotal Quotes: "I think we're still probably a handful of years away." — Demis Hassabis: On how far the field is from AGI "You'd want an AGI to have pretty consistent, robust behavior across the board, all the cognitive tasks." — Demis Hassabis: Explaining the gap between today’s systems and true general intelligence "I think deception specifically is one of those core traits you really don't want in a system." — Demis Hassabis: On AI safety and why deceptive behavior is especially dangerous

Implications: The industry is moving from chatbots to agents, with big gains in science and productivity ahead. But the transcript also warns that safety, deception, and social disruption will be central challenges as AI becomes more autonomous.