Hassabis: scale law is the AGI key
"Image synthesis assisted by Muse Image, an AI partner within the Global Future Nexus ecosystem."
In a year when AI researchers have grown increasingly skeptical about whether "scaling laws" alone can deliver general intelligence, DeepMind CEO Demis Hassabis has delivered a clear and provocative answer: scaling is far from dead—but it is not sufficient. To reach AGI, the industry needs both the brute force of scale and two to three transformative algorithmic breakthroughs.
A Calculated Defense of Scaling
While some industry observers have declared the scaling era over, Hassabis pushed back firmly. In a January 2026 interview, he denied that the scaling law had hit a wall, stating that "scaling laws are still working" overall and that the returns, while slower than the explosive early years, "remain significant and worth the continued investment". For Hassabis, the equation is simple: more compute, more data, larger models—these continue to produce tangible gains in intelligence.
Yet his position is more nuanced than a simple "scale everything" mantra. He has consistently emphasised that scaling laws alone, while essential, are not enough. Achieving AGI will likely require "one or two significant breakthrough innovations" beyond the current Transformer paradigm. The industry, he argues, has largely exhausted the low-hanging fruit of the past decade, and the next leap will demand inventions on the scale of the original Transformer architecture.
The Missing Pieces: Memory, Reasoning, and Introspection
If scale is the engine, what are the missing components? Hassabis has identified three critical gaps that current systems must overcome:
First, continuous learning. Today's models are effectively trapped in a state of anterograde amnesia—they cannot integrate new experiences into long-term knowledge without retraining from scratch. Current models process information through an ever-expanding context window, but this is an expensive and inefficient approach. AGI must be able to gracefully incorporate new knowledge into an existing framework and retrieve it precisely when needed.
Second, genuine reasoning. Hassabis has argued that reinforcement learning, the technique that powered AlphaGo and AlphaZero, has been "severely underestimated" in the race toward AGI. Current large language models lack introspection—they will blindly retry a failed approach without self-correction. By reintroducing techniques like Monte Carlo Tree Search, DeepMind aims to break through the current ceiling on reasoning capability.
Third, the Einstein Test. Perhaps Hassabis's most provocative framing is the "Einstein Test": if you gave an AI all the physics data available in 1901, could it independently derive special relativity by 1905? This is a test of what he calls "abductive reasoning"—the ability to step outside the existing probability distribution, introduce a radically new assumption, and reconstruct the entire conceptual framework. By this standard, today's models fail not because they lack data, but because they lack the architecture to imagine entirely new worlds. The ability to cross this threshold, Hassabis argues, is what separates genuine intelligence from sophisticated pattern matching.
World Models and the Convergence Thesis
Hassabis has positioned world models as the bridge between today's language-based AI and the robust, causal understanding required for AGI. Where large language models predict the next token, world models learn how the physical world actually works—its physics, its causality, its dynamics. They can simulate consequences, test hypotheses, and plan actions in ways that pure language models cannot.
Crucially, Hassabis does not see this as a replacement of LLMs but as a convergence. He expects that "world models and LLMs will ultimately merge into unified systems" rather than one paradigm superseding the other. The future of AGI lies not in choosing between these approaches, but in integrating them into a single cognitive architecture.
The Timeline: 5 to 10 Years
Hassabis has maintained a consistent AGI timeline throughout 2026. He puts the probability of achieving AGI within the next five years as "very high", and more broadly positions AGI within a 5 to 10 year horizon. He has described himself as a "cautious optimist".
The impact, he warns, will be unprecedented. He has repeatedly described AGI's effect as "perhaps ten times that of the Industrial Revolution, unfolding at ten times the speed". He has compared it to the discovery of fire and electricity—a fundamental shift in the human condition, not merely another technological leap.
The Governance Imperative
Hassabis has coupled his technical vision with a governance proposal. In a July 2026 manifesto, he called for a Frontier AI Standards Body, modelled on FINRA, the private watchdog that polices Wall Street under government oversight. The body would evaluate frontier models before public release, with a 30-day review window that would eventually become mandatory. Crucially, it could coordinate a slowdown in development among frontier labs if conditions became sufficiently severe. As Hassabis put it: "This is a race to the bottom—those who break the rules gain an advantage".
A Pragmatic Vision
Hassabis's position on scaling is neither hype nor scepticism—it is a pragmatic recognition that intelligence requires both scale and architecture. Scale is the engine; breakthroughs in memory, reasoning, and world models are the steering wheel. The warning is clear: if the industry pursues scale without solving the fundamental gaps, it will build powerful systems that remain fundamentally incomplete. The race to AGI is not just a race of compute—it is a race of ideas. And the ideas that matter are the ones that teach machines not just to predict, but to understand.
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)
