The road to AGI: DeepMind's perspective
"Image synthesis assisted by Qwen, an AI partner within the Global Future Nexus ecosystem."
From world models that simulate physics to a Frontier AI Standards Body modelled on Wall Street's FINRA, Google DeepMind is shaping the most consequential technological transition in human history—with a roadmap that balances unprecedented ambition with urgent calls for safety.
The AGI Horizon: Years, Not Decades
For years, Demis Hassabis has been among the most measured voices in AI. The DeepMind co-founder, Nobel laureate, and cognitive neuroscientist has consistently avoided the hype that surrounds his peers. That restraint ended on July 14, 2026. In a lengthy post titled "A Framework for Frontier AI and the Dawning of a New Age", Hassabis declared that AGI—"a system with all the cognitive capabilities of the human brain"—is "probably only a few short years away" .
He framed the moment in civilisational terms: "When we look back decades from now, I think we will realise we were standing at the foothills of the singularity. This is nothing less than the dawn of a new era for humanity" . AGI, he argued, is not comparable to the internet or mobile revolution—**"it is more like the discovery of electricity or fire"** . Its impact will be "perhaps ten times that of the Industrial Revolution, unfolding at ten times the speed" . The upside is staggering: accelerated drug discovery, novel clean energy, advanced materials, and potentially a future where "resources are no longer a limiting factor for human progress" .
The Missing Pieces: Memory, Reasoning, and Introspection
Despite this optimism, Hassabis is clear that the path is not fully mapped. In a June 2026 interview with Y Combinator, he identified three critical gaps that must be bridged :
Continuous Learning. Today's models have no natural pathway to transfer short-term experiences into long-term, reusable knowledge—a condition researchers have likened to anterograde amnesia . DeepMind's breakthrough in Nested Learning may offer a solution, enabling AI to build abstract structures during operation, moving beyond the Transformer's inherent limitations .
Long-horizon Reasoning. Models can solve International Mathematical Olympiad-level problems yet stumble on elementary arithmetic when prompted differently. The issue, Hassabis explained, is that AI lacks genuine reflective capacity—*"it knows it might be wrong, but doesn't know how to overturn or correct itself"* .
Memory. The context window, Hassabis argued, is merely working memory—and expanding it to millions of tokens is not a solution. "Being able to store and being able to retrieve are two completely different things," he said . True AGI requires integrating new understanding into an existing knowledge system, precisely retrievable when needed .
He placed a 50% probability on AGI by 2030—not a prediction, but a pragmatic acknowledgment that "one or two key ideas" remain undiscovered.
World Models: The Stepping Stone
DeepMind's technical strategy is anchored in world models—AI systems that understand and simulate the physical world . The August 2025 release of Genie 3 marked a significant milestone: the first real-time interactive general-purpose world model capable of generating consistent, physically realistic 3D environments from simple text prompts .
"We think world models are key on the path to AGI, specifically for embodied agents, where simulating real world scenarios is particularly challenging," explained Jack Parker-Holder, a research scientist on DeepMind's open-endedness team . Genie 3 does not rely on a hard-coded physics engine; instead, it "teaches itself how the world works—how objects move, fall, and interact—by remembering what it has generated and reasoning over long time horizons" .
This capability is essential for training AI agents that can navigate the physical world—a prerequisite for AGI that is not merely a "brain in a box" but an intelligence capable of understanding and acting within reality.
Measuring Progress: The 10 Faculties of Intelligence
To cut through the "hand-waving" around AGI, a DeepMind team introduced a cognitively inspired framework that deconstructs general intelligence into 10 key faculties . Drawing on decades of psychology, neuroscience, and cognitive science research, the framework identifies eight basic cognitive building blocks—perception, output generation, learning, memory, reasoning, attention, metacognition, and executive functions—plus two composite faculties: problem solving and social cognition .
"Despite widespread discussion of AGI, there is no clear framework for measuring progress toward it. This ambiguity fuels subjective claims, makes it difficult to track progress, and risks hindering responsible governance," the researchers wrote . The framework aims to provide "a practical roadmap and an initial step toward more rigorous, empirical evaluation of AGI" .
From AGI to ASI: Four Pathways and Six Walls
In June 2026, an elite DeepMind team led by co-founder Shane Legg published a sprawling 60-page report, "From AGI to ASI", examining the transition from human-level intelligence to superintelligence . The report identifies four pathways to ASI—continuous scaling, algorithmic paradigm shifts, recursive self-improvement, and multi-agent collective intelligence—alongside six potential bottlenecks: data walls, economic constraints, paradigm limitations, research difficulty, the "abstraction barrier," and human brakes .
The report's conclusion is sobering: "When AI capabilities begin to approach human levels, humanity's greatest need is not prediction—it is preparation" .
The Safety Framework: A FINRA for AI
Hassabis's July 2026 manifesto was less a prediction than a proposal for governance. Warning that "frontier technology is advancing faster than our understanding of it" and that the most intense commercial competition is driving a "race to the bottom" where "those who break the rules gain an advantage" , he called for the creation of a Frontier AI Standards Body modelled on FINRA, the financial industry's self-regulatory organisation .
Key elements of the proposal include :
A US-led, public-private body defining what constitutes a "frontier" model
Pre-release safety assessments, initially voluntary, eventually mandatory
A maximum 30-day review window before deployment
Coverage of cybersecurity, biological threats, and agentic deception
The authority to coordinate a slowdown if conditions become sufficiently severe
"To control increasingly autonomous, recursively self-improving systems, you need strong safeguards," Hassabis wrote . He argued that the US, given its technological and economic position, is uniquely placed to establish international standards that other nations can adopt .
A Defining Moment
Hassabis's perspective is distinctive: neither the exuberance of those who see AGI as imminent nor the scepticism of those who dismiss the current path as a dead end. He believes the architecture is largely right, but the puzzle is incomplete. The gaps—memory, reasoning, introspection—are not insurmountable, but they demand breakthroughs in areas that are "not particularly sexy" .
The window to establish effective governance is open, he warns, but it will not remain so for long. As he put it: "We have a precious window of opportunity to ensure AGI is safe" . Whether that window is seized or squandered will determine whether AGI becomes humanity's greatest achievement or its final challenge.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
