DeepMind's internal AGI pace
"Image synthesis assisted by Seedream 5.0 Pro, an AI partner within the Global Future Nexus ecosystem."
In an era where AI progress is measured in months, Google DeepMind CEO Demis Hassabis has revealed a pace that defies even the industry's accelerated expectations. The lab is releasing new breakthroughs "almost daily," a velocity so extreme that even internal teams struggle to keep pace—a testament to both the intensity of the AGI race and the depth of DeepMind's research bench.
A Velocity That Outruns Understanding
In a recent interview, Hassabis was asked about the torrent of announcements from DeepMind: DeepThink, the IMO金牌 model, Genie 3. His response was striking: "I can say that almost every day something new is being released. Even internally it's hard to keep up with the pace, and the whole field as well".
This is not hyperbolic marketing. The lab's release cadence reflects a structural shift in AI research. Where breakthroughs were once annual milestones, they are now a continuous stream. DeepThink, the reasoning engine that can solve International Mathematical Olympiad problems, was released alongside Genie 3's real-time world generation and Gemini Robotics 2's humanoid control—all within weeks of each other.
The acceleration reflects a strategic choice. DeepMind has spent fifteen years building "world-class research AND world-class engineering AND world-class infra all working closely together with relentless focus and intensity". This integration allows the lab to pursue multiple paths simultaneously: scaling current paradigms while inventing new architectures.
The Gap That Pace Reveals
Yet the velocity also exposes a critical gap. Despite the rapid release of capabilities, Hassabis acknowledges that current systems remain far from AGI. The problem is consistency: "These systems are very impressive at certain things. But there are other things they can't do yet". What he calls "jagged intelligence"—brilliant in one context, failing in another—remains the defining limitation.
The missing pieces are structural: continual learning, better memory architectures, long-term planning, and world models that understand physical causality. Hassabis estimates that AGI, defined as a system that can "exhibit all the cognitive capabilities humans can," is still five to ten years away.
The Genie Breakthrough
One of the most recent "almost daily" breakthroughs—Genie 3—exemplifies the lab's dual-track strategy. Genie 3 is not just a video generator; it is a world model that can generate interactive, real-time 3D worlds from text prompts and maintain consistency for several minutes. It teaches itself physics, object permanence, and spatial reasoning from its training data.
Hassabis describes the significance: "To prove you have a good world model, one of the tests is you can generate that world. If you turn on the tap, liquid comes out. If you have a mirror, you see a reflection. All those details are there". The model can be used to train agents like SIMA inside its generated worlds—an AI running inside another AI's "brain".
The Governance Imperative
The acceleration has not escaped Hassabis's attention. In July 2026, he proposed a Frontier AI Standards Body modelled on FINRA to evaluate the most advanced models before public release, with a 30-day review window. He warned that "nobody in the world knows for sure what is going to happen from here, and even the experts disagree".
DeepMind's internal pace is not an argument for slowing down; it is an argument for governance that can match the velocity. As Hassabis put it: "We have a precious window of opportunity to ensure AGI is safe". The question is whether institutions can keep up with a lab that releases breakthroughs "almost daily".
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)
