Demis Hassabis on AGI, Scientific Discovery and the Challenges Ahead

Google DeepMind co-founder and CEO Demis Hassabis joins Harry Stebbings to discuss artificial general intelligence, the research still needed to reach it, AI safety, science, energy and the future of work. His timelines, numerical estimates and proposed benefits are predictions or judgments, not established outcomes.

Hassabis defines AGI as a system exhibiting all the cognitive capabilities of the human mind. He uses the brain as the existence proof that general intelligence is possible. Asked about timing, he says there is a very good chance of AGI within five years while emphasizing that his estimate is a probability distribution, not a certain date.

He recalls that when he and Shane Legg started DeepMind in 2010, they extrapolated from compute and algorithmic progress and estimated roughly 20 years to AGI. Hassabis believes the field is broadly on that trajectory. This is his retrospective assessment and forecast, not a consensus deadline.

Compute is a major bottleneck, Hassabis says, both for scaling systems and for testing research ideas. He calls the cloud a workbench: an algorithmic idea needs experiments at a sufficient scale before researchers can know whether it will hold up in a larger system. More researchers and more ideas can therefore create more demand for experimentation.

Asked whether scaling laws have plateaued, he says the answer is more nuanced. Early generations produced especially large gains, and the pace of improvement has moderated, but he believes substantial returns remain from additional compute. He does not claim that every increase in spending produces the same improvement.

Hassabis says video models and interactive world models such as Genie are ahead of where he expected the field to be. At the same time, he sees continual learning as a major gap: deployed systems are not yet good at incorporating new information into what they learned during training.

He says leading labs are still working out how to integrate new learning into an existing model. As an analogy, he points to consolidation of human memories and suggests that a related mechanism might be useful. This is a research idea, not an established explanation of how future models must learn.

Stebbings asks why DeepMind’s recent products appear to have advanced quickly. Hassabis credits organizational changes that brought researchers and compute resources together around common goals rather than several competing efforts. He estimates that Google Brain, Google Research and DeepMind produced roughly 90% of breakthroughs underlying modern AI; that attribution and percentage are his judgment, not an independently measured count.

Beyond continual learning, Hassabis sees open questions in memory, long-horizon and hierarchical planning, and consistency. He describes current systems as ‘jagged’: they can solve difficult problems in one formulation and fail elementary ones when phrasing changes. A more general intelligence, in his view, should have fewer such gaps.

On model commoditization, Hassabis predicts that leading labs may pull further ahead as they use tools to accelerate research and develop new algorithmic ideas. He expects it to become harder to extract the same gains repeatedly from existing approaches. This is his competitive forecast as the leader of one of those labs.

Stebbings describes portfolio companies using frontier models to set a benchmark and open models to approach it at lower cost. Hassabis says Google supports open science and points to Gemma for smaller developers, academics, startups and edge applications. He expects open models often to trail the frontier by roughly one step or six months; that gap is an estimate, not a rule.

Asked about a post-LLM world, Hassabis says he disagrees with some views that foundation models will be replaced. He gives roughly even odds that further breakthroughs may still be needed, potentially including world models. His stronger bet is that future systems will build on foundation models, whether they are the entire system or one key component.

Hassabis hopes AGI could become a powerful tool for science and medicine and help start a new period of scientific discovery. This is an aspiration and a forecast, not evidence that AGI exists or that any particular disease will be cured on that timetable.

Stebbings explains that his mother has multiple sclerosis and worries that drug discovery and clinical trials take many years. Hassabis describes Isomorphic Labs, the company spun out after AlphaFold, as working on compound design and properties relevant to drug safety. He predicts a more complete drug-design engine in roughly five to ten years; the interview does not establish a treatment or clinical result.

Hassabis suggests AI might help simulate aspects of human metabolism and group patients according to characteristics relevant to treatment. He imagines regulators could eventually use evidence from multiple medicines to test model predictions and consider changing some steps. These are conditional research and regulatory possibilities, not validated substitutes for trials, safety testing or medical advice.

On safety, Hassabis distinguishes misuse by people from the technical challenge of keeping future, more autonomous systems within intended safeguards. He believes regulation could establish minimum standards among leading providers and ideally should involve international coordination. His assessment of future capabilities and risk is a judgment, not proof that any specific failure will occur.

He proposes benchmarks for undesirable behavior such as deception, followed by independent evaluation and a certification process. Hassabis says national AI safety institutes could contribute technical expertise, with governments ultimately responsible for oversight. These are proposals; the interview does not establish that a certification can guarantee safety.

Given a hypothetical ‘magic wand,’ Hassabis suggests an international body analogous to an atomic-energy agency, informed by researchers, national institutes, academia and civil society. He also raises a concern about AI systems communicating in forms people cannot interpret. His preferred safeguards are ideas for discussion rather than a finalized international standard.

Stebbings asks whether AI-driven labor displacement will resemble earlier technological transitions. Hassabis says past technologies disrupted jobs while also creating new ones, but believes the scale and pace of AGI could be much greater. His comparison of ten times the Industrial Revolution at ten times the speed is a heuristic prediction, not a measured forecast.

Hassabis argues that near-term AI expectations may be overhyped even while its effects over roughly a decade remain underappreciated. He wants society to mitigate disruption more effectively than during earlier industrial changes. His historical and future comparisons are broad judgments, not a calculation of how many jobs will be created or lost.

Asked about inequality, he suggests possible ways to broaden ownership of AI gains, including pension funds or sovereign wealth funds investing in the sector. He also calls for thought about how productivity benefits could be distributed through infrastructure and other means. These are policy options with unresolved design and trade-offs, not recommendations backed by an impact analysis in the interview.

Hassabis points to possible advances in fusion, batteries and materials science and mentions work with Commonwealth Fusion partners. He predicts AI could contribute to these breakthroughs, but does not establish that any technology has been solved or that a particular energy source will arrive on a given schedule.

On AI’s energy requirements, Hassabis believes the technology could eventually save more energy than it consumes. He estimates that national grids might become 30%–40% more efficient and cites weather and climate modeling as other applications. The interview supplies no method, baseline or evidence for that numerical estimate; the net energy outcome remains uncertain.

He speculates that future energy breakthroughs could change climate mitigation and space travel, including a possible role for fusion. These are long-term possibilities, not demonstrated sources of unlimited fuel or evidence that fusion is commercially available for the uses he describes.

Asked why he remained in London, Hassabis cites UK and European universities, scientific talent and a history of discovery. In DeepMind’s early years, he says, lower competition for talent was an advantage. He also values distance from Silicon Valley’s immediate trends because it can make sustained, original research easier, while acknowledging the loss of proximity to its networks.

Hassabis thinks Europe can build a trillion-dollar technology company and says he hopes Isomorphic Labs could become one. He identifies fragmented markets and limited access to very large growth rounds as obstacles. Those company valuations and outcomes are ambitions and predictions, not achieved milestones.

Asked what he would change, he proposes unlocking more pension-fund investment for growth-stage companies. He says DeepMind’s early fundraising was difficult and believes the region still needs access to larger pools of capital. The proposal is his policy view; he does not provide an analysis of pension risks or implementation.

Hassabis recalls first meeting Elon Musk at a Founders Fund portfolio event in the early 2010s, followed by a visit to the SpaceX factory. He remembers shared interest in ambitious projects and science fiction. This is his personal recollection of their meeting.

In a rapid-fire question about health, Hassabis says he wants Isomorphic Labs to build a general-purpose drug-design platform applicable to several therapeutic areas, including cancer and neurodegeneration. His aim to cure disease is an aspiration; the interview does not establish an effective treatment or a clinical outcome.

Beyond economic questions, Hassabis expects AGI to raise philosophical questions about consciousness, meaning, purpose and what it means to be human. He calls for philosophers to help address them, while explicitly assuming that technical and economic challenges would first need to be handled.

Hassabis says he would like to be remembered for advancing science and building technology that brings substantial benefits, including treatments for serious diseases. He closes with that aspiration rather than a claim that those outcomes have already been achieved.