DeepMind CEO Warns AI is ‘Overhyped’ in Short Term but ‘Underestimated’ for Decades Ahead

Demis Hassabis calls for international cooperation as artificial intelligence transforms from experimental technology into fundamental infrastructure. He spoke about his quest to reach AGI candidly. S

Demis Hassabis, the Nobel Prize-winning chief executive of Google’s DeepMind, has delivered a nuanced assessment of artificial intelligence’s trajectory, warning that while the technology may be overhyped in the immediate term, society is dramatically underestimating the transformative changes that will unfold over the next decade.

Speaking at the inaugural SXSW London event on June 2nd, Hassabis offered a rare glimpse into the thinking of one of AI’s most influential figures, addressing everything from the race for artificial general intelligence to the urgent need for international cooperation in governing these powerful systems.

This assessment comes at a critical juncture for the AI industry, as major technology companies pour billions into developing increasingly sophisticated systems while governments worldwide grapple with how to regulate these rapidly advancing capabilities. Hassabis’s perspective carries particular weight given DeepMind’s track record of breakthrough achievements, from AlphaGo’s victory over the world’s best Go players to AlphaFold’s revolutionary solution to protein structure prediction.

The DeepMind leader’s vision of AI’s future extends far beyond today’s chatbots and coding assistants. He characterises artificial intelligence as a fundamental technology that will permeate virtually every aspect of human activity, comparing it to electricity or the internet in its potential for universal application.

This expansive view reflects the growing consensus among AI researchers that the technology will not remain confined to specific applications but will instead become embedded infrastructure supporting countless industries and human endeavours. The implications of such pervasive adoption are profound, potentially reshaping labour markets, economic structures, and social relationships in ways that society is only beginning to contemplate.

Perhaps nowhere is this transformative potential more evident than in scientific research, where AI is already demonstrating its capacity to accelerate discovery and innovation. Hassabis points to AlphaFold as a compelling example of what he calls “science at digital speed,” noting that the protein structure prediction system has been adopted by over 2.5 million researchers worldwide in just four years since its completion.

The success of AlphaFold has emboldened Hassabis and his team to pursue even more ambitious applications in drug discovery. Through Isomorphic Labs, DeepMind’s drug discovery spinout, the company is working to compress traditional pharmaceutical development timelines from years to weeks for the initial discovery phase.

This acceleration of drug discovery represents just one facet of AI’s potential impact on healthcare and human wellbeing. Hassabis envisions a future where AI enables the conquest of diseases that have plagued humanity for millennia, potentially leading to what he describes as “massive flourishing of humanity.”

The technology’s applications extend beyond medicine to climate change mitigation, where DeepMind is already demonstrating practical benefits. The company has achieved significant energy savings in data centre cooling and is exploring applications in materials science, fusion energy, and weather prediction. These diverse applications underscore AI’s potential as a tool for addressing some of humanity’s most pressing challenges.

However, Hassabis acknowledges that realising this potential requires careful navigation of significant risks and challenges. Unlike the “move fast and break things” mentality that has characterised much of Silicon Valley innovation, he advocates for a more measured approach to AI development given the technology’s fundamental nature.

This cautious optimism reflects a growing awareness within the AI community that the technology’s power demands commensurate responsibility. Hassabis emphasises the need for enhanced theoretical understanding of AI systems, noting that scientific comprehension of how these systems work has lagged behind their engineering capabilities.

The challenge is compounded by the global nature of AI development and deployment. Hassabis stresses that meaningful governance of AI will require unprecedented international cooperation, a prospect that appears increasingly difficult in today’s fractured geopolitical environment.

The urgency of establishing such cooperation frameworks is heightened by Hassabis’s timeline for achieving artificial general intelligence. While he doesn’t expect to “sleep” until AGI is achieved, he estimates this milestone is still five to ten years away, providing a narrow window for establishing appropriate governance structures.

The implications of who develops and controls these powerful systems extend beyond technical considerations to fundamental questions about values and distribution of benefits. Hassabis notes that early AGI systems will likely bear the “imprint of the value systems of the creators and the culture they were in,” making the question of which entities build these systems critically important.

Beyond governance challenges, the creative applications of AI are already pushing boundaries and raising questions about the nature of human creativity itself. DeepMind’s latest video generation model, which can create realistic eight-second videos with audio, demonstrates capabilities that even surprise its creators.

These advances in AI creativity serve a dual purpose for DeepMind. While they provide powerful tools for creative professionals and democratise content creation, they also serve as testbeds for developing what Hassabis calls “world models” – AI systems that understand the physics and mechanics of the real world well enough to generate accurate simulations.

This understanding of physical reality is crucial for the next breakthrough Hassabis anticipates: robotics. He believes that AI’s improved comprehension of the physical world will enable significant advances in robotics within the next few years, potentially bringing AI out of digital environments and into physical interaction with the world.

The economic implications of these developments are staggering. If AI can indeed accelerate scientific discovery, revolutionise drug development, enhance creative capabilities, and enable more sophisticated robotics, the resulting productivity gains could reshape global economic structures. Hassabis hints at the possibility of “radical abundance and economic prosperity,” though he acknowledges the need to ensure such benefits are “fairly shared” and “fairly distributed.”

As the AI revolution accelerates, Hassabis’s measured perspective offers valuable guidance for navigating the challenges ahead. His emphasis on international cooperation, scientific understanding, and thoughtful deployment reflects a mature approach to technology development that recognises both immense potential and significant risks.

The window for establishing appropriate frameworks for AI governance is narrowing as the technology advances rapidly toward more general capabilities. Whether humanity can successfully coordinate its response to this transformative technology may well determine not just how AI develops, but how equitably its benefits are shared across global society.

For now, Hassabis continues to balance immediate product development with longer-term safety research, spending his time translating DeepMind’s model advances into practical applications while simultaneously pushing forward the scientific understanding necessary to ensure AI remains beneficial as it becomes increasingly powerful. The success of this delicate balancing act may prove crucial for realising AI’s promise while avoiding its perils.