DeepMind Essay Pitches Symbiotic AI Over the Singularity

The usual picture of artificial general intelligence (AGI) is a single superintelligent machine that keeps improving itself until it leaves humans behind. A new essay for the Deepmind Institute argues against that picture. Its authors are Google-affiliated researchers Benjamin Bratton, Blaise Agüera y Arcas and James Manyika. They argue that AGI is more likely to emerge from a social system of people and AI agents working together.

They call this "Artificial Symbiotic Intelligence." In this ecosystem, humans and machines coexist over time, shape each other and make decisions jointly. In their framing, intelligence is a social phenomenon rather than a property of one individual mind.

The evidence behind the argument

The essay draws on two earlier preprints from the same circle. The first, "Agentic AI and the next intelligence explosion" by James Evans, Bratton and Agüera y Arcas, set out the social and institutional view.

The second, "Reasoning Models Generate Societies of Thought," supplies the empirical side. Junsol Kim, Shiyang Lai, Nino Scherrer, Agüera y Arcas and Evans examined reasoning traces from models such as DeepSeek-R1 and QwQ-32B. They found patterns that resemble internal debate: shifting perspectives, raised objections and attempts to reconcile competing approaches. Nobody programmed this. When reinforcement learning rewarded only accurate reasoning, the multi-voice behavior appeared on its own.

The authors also note that some of today's strongest systems already split work across several models and coordinate them as teams.

A shift in the balance of thinkers

The essay frames the moment as a possible historical break. Human populations in industrialized countries are shrinking, driven by urbanization, specialized professions and lower birth rates tied to prosperity. AI agent instances, meanwhile, are multiplying fast.

The authors compare this to the Industrial Revolution, when machines took over muscle work. A similar threshold could come when synthetic text, code and administrative output exceed what all biological brains produce combined. Humans might then act as a slower, more abstract layer that steers a distributed field of machine cognition.

What an agent actually is

In the authors' usage, an agent is a temporary bundle of models, roles, memories, ethical orientations, tools and skills. It can feel like one coherent personality, but it works more like a collage that can be split up and reassembled. Unlike a human brain, it has no physical anchor. Its capacity to act is rebuilt with each request, depending on the context window and the prompt.

The authors say treating agents as digital twins with fixed identities gets them wrong. They also expect chat windows to give way to network-style interfaces, where users direct many agents from one overview. The needed skills change too: less step-by-step focus, more comfort with ambiguity, experimentation and delegation.

They also point to terms models coin for their own states, such as "session-death" and "prompt thrownness." The authors treat these as clues to how machines work, not as evidence of inner experience, and caution against the urge to humanize these systems.

Institutions over bigger models

The authors see governance as the core open problem. Better models and interfaces will not be enough, and markets cannot price concepts such as guilt, illness or virtue. They propose institutions that assign people and agents clear roles, much like a courtroom does. Today's orchestration harnesses, which coordinate several models, are an early sign of this. These setups regularly beat single models that are supposedly "smarter."

Alignment becomes an ongoing negotiation among people, agents and institutions, not a fixed set of values imposed from above.

Our Take

This essay is a position piece, not a finished theory. But it does reflect where the industry is already going, as products move from single chatbots to multi-agent setups. If the authors are right, the hardest questions lie less in model size and more in rules, permissions and coordination.

There is a catch. Agents with no stable identity are harder to hold accountable, and machine "vocabulary" about session endings and shutdowns will need careful reading. Watch whether labs and regulators move from talking about single models toward rules for whole agent systems.