20VC: DeepMind's Demis Hassabis on Why AGI is Bigger than the Industrial Revolution | Why LLMs Will Not Commoditise & We Have Not Hit Scaling Laws | Bottlenecks in AI & The Energy Crisis Caused By AI | Whether AI Will Do More to Harm or Help Inequality

Hassabis defines AGI precisely as a system that exhibits all the cognitive capabilities of the human mind—a definition he describes as stable and consistent over time. The human brain, he notes, is "the only existence proof we have" that general intelligence is possible, making it the appropriate benchmark. On timing, he places a "very good chance" of AGI arriving within the next five years. This is not a revision of earlier expectations: when DeepMind was founded in 2010, co-founder Shane Legg published extrapolations suggesting roughly 20 years to AGI, and Hassabis considers that projection broadly on track.

The discussion addresses directly the widely circulated claim that AI scaling is plateauing. Hassabis pushes back, calling the picture "more nuanced." He acknowledges that the near-doubling of performance seen with each early generation of large language models has slowed—that rate of exponential gain was always going to moderate. However, he argues that substantial returns on compute investment continue, and that frontier labs are still extracting significant performance improvements from scaling existing architectures. The more important constraint, he suggests, is not the ceiling on scaling but the need for new algorithmic ideas, which require large amounts of compute simply to test at meaningful scale.

Despite being ahead of expectations in areas like video generation and interactive world models, Hassabis identifies several critical capability gaps that remain unsolved across the industry: **Continual learning** — Current systems stop learning once training ends. They cannot elegantly incorporate new information after deployment. Hassabis draws an analogy to the brain's sleep-based memory consolidation process, suggesting something architecturally similar may be needed. **Memory architecture** — Long context windows are described as "a bit brute force." More sophisticated memory systems remain an open research problem. **Long-term planning** — Current models struggle with hierarchical planning over extended time horizons, a capability humans exercise routinely. **Consistency** — Hassabis uses the term "jagged intelligences" to describe systems that perform impressively on certain problem framings but fail on elementary variants of the same problem. A true general intelligence, he argues, should not exhibit such structural holes. He identifies labs with the capacity to invent new algorithmic ideas as those most likely to pull ahead competitively, as the returns from the current generation of foundational ideas are being exhausted.

One of Hassabis's most concrete near-term claims concerns AI's role in pharmaceutical development. He describes Isomorphic Labs—spun out of DeepMind following AlphaFold—as focused on solving the drug design pipeline: compound design, toxicity screening, and property validation. He estimates a complete drug design engine could be ready within five to ten years. The harder problem, he acknowledges, is clinical trial timelines, which remain multi-year processes. His proposed path forward involves AI-assisted patient stratification, metabolic simulation, and—once a sufficient number of AI-designed drugs have cleared regulatory review—a potential reduction in certain trial steps as regulators gain confidence in model predictions. He explicitly frames this as a two-stage process: solve design first, then address regulatory timelines.

On AI safety, Hassabis identifies two distinct risk categories: misuse by bad actors exploiting dual-use capabilities, and the technical challenge of maintaining alignment as systems become more agentic and autonomous. He advocates for international minimum standards, independent auditing bodies modeled loosely on nuclear regulatory frameworks, and certification processes analogous to quality marks for consumer products. He singles out the UK AI Safety Institute as a credible institutional model. An open tension runs through this section: Hassabis acknowledges that effective governance requires international coordination at precisely the moment when geopolitical fragmentation makes that harder to achieve. He does not resolve this tension, describing it as "not ideal" while arguing for pragmatic minimum standards as a fallback. He also flags a specific technical safeguard he considers broadly agreeable across leading labs: AI systems should not output tokens in non-human-readable formats, which would introduce new vulnerabilities beyond existing safeguards.

On labor markets, Hassabis accepts that significant job disruption is coming—more so than with previous technological revolutions—while maintaining that new categories of work will emerge. He frames AGI's economic impact as roughly ten times the Industrial Revolution, unfolding over a decade rather than a century. On wealth concentration, he suggests sovereign wealth funds and pension fund investment in AI companies as mechanisms for broader distribution of productivity gains, alongside infrastructure investment funded by productivity surpluses. He also notes that AI-driven advances in energy—particularly fusion, superconductors, and grid optimization—could fundamentally alter the economic calculus, with grid efficiency gains of 30–40% cited as a near-term possibility.

Hassabis attributes DeepMind's recent acceleration to organizational consolidation: combining talent and compute resources from across Google into a unified effort rather than maintaining parallel model development tracks. He claims that roughly 90% of the foundational breakthroughs underpinning the modern AI industry—including Transformers, AlphaGo, and reinforcement learning advances—originated within Google Brain, Google Research, or DeepMind. On Europe's competitive position, he identifies late-stage capital as the primary structural gap, noting that early-stage startup formation is strong but billion-dollar growth rounds remain scarce. He cites Isomorphic Labs as a candidate for Europe's first trillion-dollar AI company. Key takeaways: - AGI, defined as full human cognitive equivalence, is assessed as likely within five years—consistent with projections made at DeepMind's founding in 2010, not a recent acceleration. - Scaling laws are not exhausted, but the era of near-doubling generational gains is over; future competitive advantage will accrue to labs capable of generating new algorithmic breakthroughs, not merely scaling existing ones. - Critical unsolved problems—continual learning, robust memory architecture, long-term planning, and output consistency—represent the next frontier of AI research, distinct from scaling challenges. - AI governance requires international minimum standards and independent auditing bodies, but effective coordination faces structural geopolitical headwinds that Hassabis acknowledges without resolving. - The economic magnitude of AGI is framed as categorically larger than prior technological revolutions, requiring proactive redistribution mechanisms—sovereign funds, pension investment, infrastructure spending—rather than reliance on historical precedent alone.