DeepMind argues against the singularity

The lab that popularised the term AGI has now written down what it thinks comes after it — and the picture is deflationary. In a 57-page position paper, fourteen Google DeepMind researchers, including AGI-namer Shane Legg and AIXI theorist Marcus Hutter, define superintelligence as a system 'more intelligent and cognitively capable than large organisations of humans.' Not a deity. A bar a coordinated swarm of human-level agents could clear without any single model exceeding a person.

"Runaway acceleration cannot be ruled out." — the paper's own hedge, after 57 pages arguing a hard takeoff is unlikely

They lay out four ways to get there: keep scaling today's models, hit a new algorithmic paradigm, let AI improve its own design, or have superintelligence emerge from a large enough collective of ordinary agents. The last is the freshest framing — it implies ASI may already be latent in scale rather than waiting on a breakthrough.

The argument that matters is what the paper pushes back on. The field's default intuition is a single 'transformative step change' — a hard takeoff where a self-improving system rockets past us overnight. DeepMind says that framing is likely wrong, and that even a superintelligence stays bound by the speed of light, the thermodynamics of computation, complexity theory, and Gödel's limits. Capable beyond any organisation, not omnipotent.

This is a stance, not a result — the paper hedges that runaway acceleration 'cannot be ruled out.' But it is a notable one to come from this roster, and it cuts against the recursive-self-improvement narrative that rival labs have leaned into. When the people who named the destination start arguing the trip is a long series of ordinary-looking steps, the timeline debate moves.