This transcript records a December 2025 "annual check-in" between Hannah Fry and Demis Hassabis, summarizing a year where AI transitioned from static Large Language Models (LLMs) to "agentic" world models. Hassabis details DeepMind's progress on "root node" problems—specifically fusion energy, material science, and quantum error correction—while admitting that current systems exhibit "jagged intelligence," performing at PhD levels in mathematics but failing at basic logical consistency.
Hassabis argues that the "scaling wall" is a myth, though returns are diminishing, necessitating a 50/50 split between scaling and architectural innovation. He introduces "Genie" and "SIMA" as steps toward generalized world models that learn intuitive physics through simulation rather than just text. The conversation concludes with speculative sociopolitical projections: a post-AGI economic shift happening 10x faster than the Industrial Revolution, the potential for "direct democracy" via credit systems, and Hassabis's personal conviction that the entire universe—including human consciousness—is ultimately a computable information-processing system.
The State of AGI: Scaling, Simulations, and the Computable Universe
1:42 2025 Progress: The shift from text-based LLMs to multimodal "agentic" AI is complete. Gemini 3 and advanced "world models" (like Genie) now allow AI to understand spatial dynamics and physical context, not just language.
2:18 Root Node Solutions: DeepMind is pivoting heavily toward physics. Key targets include room-temperature superconductors and a deep partnership with Commonwealth Fusion to stabilize tokamak plasma. The goal is "post-scarcity" via unlimited clean energy.
5:14 Jagged Intelligence: Hassabis admits a glaring paradox: AI can win gold medals in the International Maths Olympiad but still fails at high-school-level logic or consistent chess. As far as I can tell, this "jaggedness" is the primary bottleneck to true AGI.
12:42 The Scaling Wall: Hassabis rejects the idea that scaling has hit a wall. He estimates returns are between exponential and asymptotic; scaling remains 50% of the effort, supplemented by synthetic data generation.
17:43 Genie and Simulation: Simulations are used to teach AI "intuitive physics." By dropping SIMA (Simulated Agents) into Genie (generative world models), DeepMind creates an infinite training loop. Takeaway: If you can simulate a world accurately, you have "understood" its underlying mechanics.
28:26 The AI Bubble: Hassabis acknowledges a bubble in startup valuations but claims Alphabet is insulated due to its full-stack infrastructure (TPUs) and existing product integration. He views the current market "frenzy" as an overreaction to a decade of under-reaction.
34:31 Defining AGI: Hassabis views the convergence of Gemini 3, specialized image understanding, and world models as the path to "proto-AGI." He anticipates a "universal assistant" that acts autonomously within 2-3 years.
38:33 Economic Dislocation: The transition to a post-AGI economy will occur 10x faster than the Industrial Revolution (roughly one decade). Hassabis suggests Universal Basic Income (UBI) is only a partial fix; he speculates on "direct democracy" where citizens use credits to vote on local resource allocation.
44:44 The Computable Universe: Hassabis operates on the skeptical rationalist assumption that there is no "non-computable" magic in the human brain. He treats biology as pure information processing and believes Turing machines can eventually replicate consciousness and creativity entirely.
53:54 Agentic Risks: The shift to autonomous agents (2-3 years out) increases risk significantly. DeepMind is focusing on "cyber defense" to manage a future where millions of agents roam the internet autonomously.
Recommended Reviewer Group: The Rationalist & Effective Altruist (EA) Community
This group is best suited to review this because they ignore the corporate "Gemini 3 is warm and helpful" marketing and focus strictly on X-risk (existential risk), compute efficiency, and AGI timelines. They are skeptical of Hassabis's optimism regarding "responsible cowboy labs" and will be interested in his epistemic claim that the universe is 100% computable.
Review Summary (Rationalist/EA Style):
Demis Hassabis is signaling a high probability (P > 0.7) that we are entering the "Agentic Phase" of AI development within the next 24 months. While he pays lip service to "scientific rigor," he admits DeepMind is trapped in a "ferocious" capitalist race condition that prevents optimal safety testing.
Key technical takeaways:
World Model Convergence: The "Genie + SIMA" loop is a play for infinite high-fidelity data, bypassing the "data wall." This increases the probability of a fast takeoff.
Epistemic State: Hassabis is a hard physicalist. He places the probability of "non-computable" consciousness at near zero, suggesting he will not stop scaling until he hits the physical limits of the Turing machine.
Alignment Gap: DeepMind’s plan for "post-AGI" economics is vague and relies on speculative "direct democracy" models that lack robust game-theoretic validation.
Safety: He is worried about autonomous agents roaming the internet (P-risk is high), but his solution—"cyber defense"—seems reactionary rather than foundational. I don't care about his "sabbatical"; I care about whether he's building a kill-switch for the agents he's about to release.
Hassabis's timeline for AGI remains 5–10 years, but his actions suggest he's bracing for a shorter horizon. His "jagged intelligence" comment is a rare moment of frankness—current models are still "idiot savants" lacking global coherence. As far as I can tell, they are scaling their way through the "jaggedness" rather than fixing the underlying architecture.