On April 7, Google DeepMind CEO Demis Hassabis sat down with 20VC for an in-depth interview exploring the definition and timeline of AGI, the real yield ofScaling Law, the missing cognitive capabilities in current models, AI's revolutionary applications in biopharma and energy, global safety frameworks, and AGI's impact on the labor market.
Demis Hassabis argues that the benchmark for AGI should be the full spectrum of human brain-like cognitive abilities—a milestone he believes is likelywithin five years. He notes that while Scaling Law's performance gains have slowed compared to early days, the dividends remain substantial and far from stagnant.
He highlights that roughly 90 percent of foundational AI breakthroughs in the past decade trace back to Google Research or DeepMind, and that thesignificant competitive moat in coming years will shift from brute-force compute scaling to inventing new algorithms. Hassabis coins the term "saw-tooth AI" todescribe today's fragmented intelligence: systems excel at specific, narrowly-defined tasks but lack fundamental coherence. To close this deficit, heemphasizes breakthroughs in continuous learning, memory architectures, and long-range planning.
On AI-driven industrial transformation, Hassabis predicts the dawn of a golden age of scientific discovery. By simulating human metabolism and enabling precise patient stratification, AI could compress clinical-trial timelines dramatically. In energy, he forecasts grid-efficiency gains of 30-40 percent, and breakthroughs in nuclear fusion and materials science that slash energy costs. Regarding open-source models, he argues theywill always trail state-of-the-art commercial releases by about six months, largely functioning as replication and digestion mechanisms. For governance, he advocates creating an international body akin to the IAEA—one staffed by government-backed safety institutes to audit models, enforce limits such as bans on nonhuman-readable machine output, and maintain safety guardrails.
01
AGI Must Mirror the Full Spectrum of Human Brain Cognition
Industry debate swirls around AGI's definition. What benchmark do you treat as definitive?
Demis Hassabis: We have consistently defined AGI as a system that exhibits every major cognitive faculty of the human brain—the one proven model of general intelligence in the known universe. Any acceptable benchmark must center on that full-spectrum cognition.
Our probabilistic forecast puts a greater-than-evens chance of achieving this within five years, so the horizon isn't far off.
Has that view shifted? No. Going back to 2010, my co-founder/-chief scientist Shane Legg posted a prediction about AGI's timeline when few were even studying AI, calling it a dead-end. Today those posts still exist online. We extrapolated compute and algorithmic trends, predicting a 20-year journey from Day 1 to realization; we're on track.
02
The Compute Bottleneck, Scaling Law, and the Myth of a Plateau
What is the single greatest bottleneck on the path to AGI? Many claim Scaling Law has hit a ceiling—do you agree?
Demis Hassabis: Compute remains the central bottleneck: we need it not just for scaling parameters upward, but as the laboratory bench forexperimentation. Cloud-scale compute gives us the runway to test new algorithmic ideas on meaningful scale. Ideas that otherwise would fail to integrate into main systems become viable. A constant flow of creative researchers demands enormous headroom.
Regarding Scaling Law plateaus: we don’t see one. Top labs experienced exponential doubling in the first generations of large language models, but growthhas inevitably decelerated. That doesn’t mean scaling is exhausted; dividends are still real and material. The gains today are less eye-popping than before, but still impactful.
03
The Missing Pillar: Continuous Learning and Coherence
Which AI domains lag behind expectation, especially continuous learning? Will it be the next watershed?
Demis Hassabis: Most domains have moved faster than expected—consider today’s video-generation models or our new interactive world-model Genie. If someone showed those to me five-to-ten years ago, I’d be stunned. A handful of core ingredients remain missing, however; continuous learning tops the list. Current systems stop absorbing new knowledge once released. They don’t replaysleep-like consolidations the brain performs to merge daytime memories into long-term schemata. Something analogous is likely required for machines.
Other gaps include memory architectures that avoid brute-force "long context windows," long-horizon planning that spans years rather than tokens, and above allcoherence. Today’s models produce "saw-tooth AI": switch the phrasing of a prompt and you uncover jarring failures that a true AGI should banish.
04
From Breakthroughs to Competitive Moats: DeepMind's Organisational Shift
DeepMind has accelerated sharply in recent years. How? With commoditization looming, will the gap narrow, or entrench a few front-runners? What’s the role of open-source?
Demis Hassabis: Structural changes mattered. Google and DeepMind collectively hold the deepest research bedrock in modern AI: roughly 90 percent of fundamental advances since AlphaGo, reinforcement learning, to the Transformer trace back to Google Brain, Google Research, or DeepMind. Future missing puzzle-pieces are where I expect us to contribute. We consolidated top talent and compute into one unified stack instead of fragmented efforts inside the company. By operating with startup pace and focus, we reclaimed pole position.
On competition, three-to-four labs will likely pull away: the old tooling such as programming libraries and math solvers accelerates model iteration. Margins from exploiting old tricks vanish; advantage flows to labs that invent novel algorithms.
Open source serves replication. We are avid supporters of open science and release landmarks like the Transformer, AlphaFold. We’ll continue in scientific AI applications. Yet the rhythm shows open-source versions trail commercial releases by ~6 months, time enough for communities to digest and replicate state-of-the-art advances. We are pushing the open-source Gemma family to compete on equal scale while striving for global leadership.
05
After the Era of LLMs: System Architecture and Societal Impact
Yann LeCun argues LLMs won’t be sole cores of future AGI systems. Do you agree? AGI within five years—what world awaits?
Demis Hassabis: I disagree with Yann on that point. While a 50 percent chance exists that world models or novel paradigms will also break through, foundation models have already proven themselves indispensable—scaling laws hold. The question is whether LLMs become the sole cores or parts of broader systems. They won’t be superseded; rather, they’ll serve as bedrock layers the full system can build upon.
In five years, AGI will be the ultimate tool for scientific and medical research: accelerating discovery and cures. We’ll enter a golden age of discovery.
06
Healthcare Revolution: Faster Cures, Simulated Metabolism, and Regulatory Leapfrog
Drug discovery can take a decade from bench to patient. How can AI solve this bottleneck and catalyze a medical revolution?
Demis Hassabis: We’re close. AlphaFold cracked protein folding; we spun out Isomorphic Labs to focus on chemical design, synthesis, andtoxicity screens. In five-to-ten years, the drug-design engine matures. Clinical trials remain slow, but AI can simulate human metabolism, stratify patients for precision trials. The first AI-discovered drugs hitting the market will validate models, enabling regulators and clinicians to trust predictions—shortcuts past animal testing phases and expedite dosing ladders. Two-step journey: first design, then optimize regulations.
07
Global Governance and the Need for IAEA-Style Safeguards
Stephen Hawking warned AI must get it right the first time because we may not get a second chance. Do you share that concern? What qualifies as "correct regulation" in your view?
Demis Hassabis: I absolutely share those concerns. I worry about malicious misuse of this dual-use technology and emergent risks as systems become more autonomous, Agent-like. Correct regulation sets minimum safety standards worldwide; it must be international.
We need an IAEA-style body staffed by government-backed safety institutes to audit models against shared baselines, enforce bans on nonhuman-readable machine communications, and certify safety guarantees. Governments, academia, and civil society must collectively ensure these systems undergo scrutiny before they reach the public.
08
The Coming Labor-Force Upheaval and the 10× Industrial Revolution
Science will be one of the most exciting fields of the next five years—but labor displacement is inevitable. As systems exhibit breathtaking capabilities, how do you view their societal job-market impact?
Demis Hassabis: Past revolutions displaced countless jobs, yet ultimately spawned higher-quality, better-paid roles. We must resist the impulse that 'this time is different.' This event might compress a 100-year industrial revolution into a single decade. Industrialization brought staggering disruption but also modern medicine—child mortality once 40 percent. We surely don’t want to miss those advances, yet this iteration could dwarf historical upheavals. Ideally we should mitigate side-effects better this go-round.
Short-term hype and long-term undershot remain real. AI’s rate of change compresses timelines, yet over one year overhype exists. Over ten years, we underrate how thoroughly AGI will reshape industry and society; the revolution’s completeness is dangerously underestimated.
09
Energy Abundance and AI's Net-Zero Footprint
Beyond labor, people fear rising inequality and consolidation of wealth. Historically, technological revolutions concentrate capital—how might this play out? How do we square AI’s mounting carbon footprint?
Demis Hassabis: One lever is institutional ownership: pension funds and sovereign wealth funds can buy stakes in AI giants so gains are shared across society. Equity foundations help broaden prosperity.
Over next five-to-ten years, energy breakthroughs—nuclear fusion, superconductors, novel batteries, materials science—may deliver near-zero-cost renewable energy. AI will simultaneously squeeze efficiency gains of 30-40 percent in existing grids. It also powers climate and weather models for resilience. The energy piece alone could make space exploration trivially cheap; fusion turns seawater into infinite rocket fuel.
10
Europe’s Unicorn Potential and the Ultimate Medical Quest
Europe currently lacks a trillion-dollar tech company. If you had AI-enabled vision and capital infrastructure equal to the task, what would a unicorn recipe look like there?
Demis Hassabis: Right now, none. Spotify shows promise. At Isomorphic Labs, we’re shooting for that ceiling. Fragmented markets are a barrier; institutional innovations—capital markets bold enough to support $1B rounds—must align.
In global health, my ultimate hope is to eradicate cancer. Isomorphic’s platform is designed to crack not just oncology but neurodegeneration, cardiovascular, immunology, and beyond. While it sounds cliché, it’s our true north star.
Looking further ahead, once economic and technical hurdles fall, AGI forces existential questions about consciousness and meaning. We’ll need new philosophers to chart the path forward.
Further Reading:
NVIDIA CEO Jensen Huang's 20,000-Word Dialogue: 'AGI Is Already Here'
Yang Zhilin, Zhang Peng & Lu Fuli Discuss OpenClaw: The Future of Software
AI Is Experiencing 'Speciation' | Andre Karpathy