NextFin News - Google DeepMind CEO Demis Hassabis spoke with Axios co‑founder Mike Allen at the Axios AI+ SF summit in San Francisco on 2025-12-04. The conversation, held in a historic bank venue, covered DeepMind’s technical priorities, the new Gemini 3 model, the path to artificial general intelligence and the risks society must manage.
The onstage discussion was hosted by Mike Allen and drew on Hassabis’s experience as a scientist and product‑minded leader. Over the course of the interview, Hassabis emphasized the role of the scientific method in DeepMind’s work, outlined near‑term milestones such as multimodal convergence and world models, and described both the utopian possibilities and the concrete threats posed by advanced AI.
Nobel laureate platform and public engagement
Hassabis described how his recent Nobel Prize has changed the way audiences receive his messages. He said the award is a sort of shortcut to almost anyone to to know that you're...expert in your field
and that it gives him a broader platform to speak about issues he cares about, particularly AI safety and responsible deployment.
Science first: method, rigor and product thinking
Throughout the interview Hassabis returned to the primacy of science: I'm a scientist first. The scientific method...is the most important maybe idea humanity's ever had.
He said DeepMind’s edge comes from combining rigorous research, engineering and infrastructure: we blend world-class research with world-class engineering with world-class infrastructure
, and that gives the organization its advantage in a ferocious competitive landscape.
Near‑term technical priorities: multimodality and world models
Hassabis outlined the short list of technical frontiers he expects to move significantly in the next 12 months. He emphasized the convergence of modalities in Gemini — images, video, text and audio — and the cross‑pollination those capabilities enable. He highlighted recent image work (NO Banana Pro) and interactive video experiments (Genie 3), describing Genie 3 as an interactive video model...you can sort of generate a video but then you can start walking around it like you're in a game or simulation and it stays coherent for a minute.
Agents and the universal assistant
On agent systems Hassabis described DeepMind’s ambition for a universal assistant that is present across devices — phone, laptop, glasses — and becomes part of everyday life. He observed that agents today cannot reliably complete delegated tasks end‑to‑end, but predicted that a year from now we'll start having agents that are close to doing that
. He stressed the need for guardrails as agents gain autonomy and continual learning capabilities.
Gemini 3: nuance, personality and surprising capabilities
Hassabis characterized Gemini 3 as a step change in both capability and style: it answers succinctly. It pushes back a little bit...it has a bit of a step change in its kind of intelligence and therefore usefulness.
He recounted being surprised by one‑shot game generation and front‑end creative work: Gemini 3 can oneshot games
and produce high‑quality websites and aesthetic outputs more quickly than prior systems. He also described multimodal understanding in video as particularly underappreciated: point your phone at something and the model can help as a mechanic — ultimately ideally through glasses so your hands are free.
AGI: timeline and what remains to be solved
Hassabis was explicit about his estimate for AGI: We're definitely not there now...quite close. I think we're like five to 10 years away
. He defined AGI as a system exhibiting the full range of human cognitive capabilities, including inventive and creative thought. While scaling current systems will be essential — and could in principle be enough — he said his best guess is that one or two additional breakthroughs, at the level of a transformer or AlphaGo innovation, will likely be required in addition to aggressive scaling.
Risks now and catastrophic scenarios
Hassabis listed concrete near‑term dangers and longer‑term concerns. On misuse he warned about pathogens designed with AI and attacks on critical infrastructure: pathogens created by an evil actor using AI...energy or water cyber terror using AI by a foreign actor
. He described the probability of catastrophic outcomes as non‑zero
and said that warrants putting significant resources into mitigation. Regarding autonomous systems acting outside human control, he acknowledged the risk as agents become more agentic and emphasized research on staying within guardrails.
Competition with China and global leadership
On the U.S. versus China question Hassabis said the West still leads algorithmically, though China is closing fast: we're still in the lead...but China is not far behind...the lead is only a matter of months as opposed to years at this point
. He singled out chips as a separate area where dynamics differ, but repeated that algorithmic innovation in the West remains a key advantage.
Industry dynamics: bubble risks and recruiting
Asked about an AI investment bubble, Hassabis said some parts of the industry look frothy — $50 billion seed rounds...seems a little bit unsustainable
— while reiterating his conviction that AI will be one of the most transformative technologies ever. On talent, he described DeepMind’s recruiting posture: mission matters, and top researchers want to work on the most impactful, cutting‑edge problems.
Games, human adaptation and sport
Hassabis reflected on gaming as both training and inspiration. He said games are microcosms of something in the real world
and provide abundant practice for critical decision moments. Asked about sports, including the World Cup, he noted teams already approach DeepMind for analytics and suggested AI will continue to optimize elite performance, down to precise positioning on set pieces.
Closing note: preparation and responsibility
Across topics Hassabis returned to two themes: prepare society for transformative change and deploy AI responsibly today. He described the Nobel platform as an opportunity to open doors to leaders and governments on those questions and emphasized that capitalist incentives and enterprise demands for guarantees will also push providers toward more responsible behavior.
References and further reading
Axios coverage and event reporting: Exclusive: "Transformative" AGI is on the horizon, DeepMind's Hassabis says and Exclusive: Some AI dangers are already real, DeepMind's Hassabis says. For DeepMind background: DeepMind.
Explore more exclusive insights at nextfin.ai.
