The singularity has a specific shape in the popular imagination: one machine, exponentially self-improving, leaving humanity behind. A new essay affiliated with Google DeepMind wants to replace that image with something stranger — intelligence as a sprawling social system in which humans, agents and institutions shape each other.
The piece, "Artificial Symbiotic Intelligence," published on the DeepMind Institute's essay site and reported by The Decoder on October 3, is authored by Benjamin Bratton, Blaise Aguera y Arcas and James Manyika. Its core claim: AGI will not arrive as a lone superintelligence but as an ecosystem of people and machines coexisting, adapting and making decisions together. The central research problem shifts accordingly — from building a bigger brain to coordinating and governing a complex network of agents, humans and the connective tissue between them.
Two academic preprints underpin the argument. "Agentic AI and the Next Intelligence Explosion," by James Evans, Bratton and Aguera y Arcas, supplies the social and institutional framing. More striking is the empirical companion, "Reasoning Models Generate Societies of Thought," in which Junsol Kim, Shiyang Lai, Nino Scherrer, Aguera y Arcas and Evans analyze reasoning traces from models including DeepSeek-R1 and QwQ-32B. They find the models spontaneously produce patterns resembling internal debate — shifting perspectives, raising objections, reconciling conflicting approaches — without anyone programming that behavior. Reinforcement learning that rewards only reasoning accuracy appears to make multi-perspective, conversational cognition emerge on its own.
The essay then goes to work on the concept of the agent itself. What users experience as a coherent persona, the authors argue, is a temporary assemblage of models, roles, memories, tools and ethical orientations — a collage that is reassembled with every request, anchored by nothing like the brain's physical continuity. Treating agents as digital twins with fixed human identities, they warn, misrepresents the technology and understates what coordinated agent swarms could actually do. Humans directing such swarms become plural themselves; the authors dub these responsive mirrors "parasocial mirrors."
Their historical framing is a "cognitive tipping point." As urbanization, specialization and falling birth rates shrink human populations in industrialized countries while AI agent instances multiply, the balance between biological and synthetic thinkers could tip so fast that, on the scale of history, it looks like a leap — the way industrialization compressed centuries of muscle-driven labor into decades of machine work.
On governance, the essay is pointedly skeptical that better models or markets alone will suffice. The authors point to the courtroom — fixed roles, procedures, precedents, feedback loops — as the template for institutions that can coordinate humans and machines in defined roles. They also reframe alignment: imposing a fixed value set from above is "a dead end"; values instead take shape through ongoing negotiation among people, agents and institutions, differing across fields as adoption speeds vary.
It is easy to file this under technologist philosophy. But the source matters: Aguera y Arcas leads a research organization inside Google, Manyika is the company's chief technology officer for AI, and Bratton directs the Antikythera think tank. When the people building frontier systems argue that institutions and interfaces — not raw model scale — are the binding constraint, that is a statement about where Google thinks the next returns will come from. And it is a quiet, academic rebuttal to the doomsday framing that has dominated the past two weeks of resignations, essays and exponential-risk headlines.
Comments (0)
Log in to join the discussion
Log InNo comments yet