这一论点建立在作者团队的两篇早期论文基础上。在预印本《Agentic AI and the next intelligence explosion》中：https://arxiv.org/pdf/2603.20639，James Evans、Benjamin Bratton和Blaise Agüera y Arcas发展了新文章背后的社会与制度视角。
第二篇预印本提供了实证支持。在《Reasoning Models Generate Societies of Thought》中：https://arxiv.org/pdf/2601.10825，Junsol Kim、Shiyang Lai、Nino Scherrer、Blaise Agüera y Arcas和James Evans研究了如DeepSeek-R1和QwQ-32B的推理模型。他们对推理轨迹的分析表明，这些模型经常产生类似内部辩论的模式，包括观点转换、提出异议以及调和冲突的方法。
An essay for the Deepmind Institute challenges the familiar image of a lone superintelligence. The authors envision networks of agents, hybrid institutions, and what they call "Artificial Symbiotic Intelligence."
Google-affiliated researchers Benjamin Bratton, Blaise Agüera y Arcas, and James Manyika argue that artificial general intelligence, or AGI, will emerge from a social system in which people and AI agents work together. Some of today's most capable AI systems already use frameworks that divide work among several models and coordinate them as teams：https://the-decoder.com/a-new-information-theory-framework-reveals-when-multi-agent-ai-systems-truly-work-as-a-team/.
That shifts the central challenge for AI research, the authors say. Researchers must coordinate and govern a complex network of agents, people, and the systems that connect them, rather than build an isolated machine intelligence. Intelligence, in this view, is a social phenomenon, not an individual trait.
The authors call their vision "Artificial Symbiotic Intelligence：https://institute.deepmind.com/essays/artificial-symbiotic-intelligence/," an ecosystem in which people and machines coexist over time, shape one another, and make decisions together. The idea directly challenges the notion of a "singularity" driven by a single superintelligence that continually improves itself.
The argument builds on two earlier papers from the authors' circle. In the preprint "Agentic AI and the next intelligence explosion,"：https://arxiv.org/pdf/2603.20639 James Evans, Benjamin Bratton, and Blaise Agüera y Arcas developed the social and institutional perspective behind the new essay.
A second preprint provides empirical support. In "Reasoning Models Generate Societies of Thought,"：https://arxiv.org/pdf/2601.10825 Junsol Kim, Shiyang Lai, Nino Scherrer, Blaise Agüera y Arcas, and James Evans study reasoning models such as DeepSeek-R1 and QwQ-32B. Their analysis of reasoning traces suggests that these models often produce patterns resembling internal debate, shifting perspectives, raising objections, and reconciling conflicting approaches.
This behavior emerges during training rather than being explicitly programmed：https://the-decoder.com/the-next-leap-in-ai-depends-on-agents-that-learn-by-doing-not-just-by-reading-what-humans-wrote/. When reinforcement learning rewards models only for reasoning accuracy, they develop multi-perspective, conversational behavior on their own. The Deepmind essay extends this finding from individual models to the possible design of societies made up of people and agents.
The authors frame this development as a possible historical break. Urbanization, the growth of specialized professions, and falling birth rates linked to rising prosperity are shrinking human populations in industrialized countries. Meanwhile, the number of AI agent instances is growing fast. The balance between biological and synthetic thinkers at a human level could shift so quickly that, on the scale of history, it would look like a sudden leap.
They compare this to the Industrial Revolution, when machines took over work once done by human muscles. A similar threshold could arrive when synthetic text, code, and administrative work exceed the combined output of biological brains. People might then serve as a slower, more abstract layer directing a distributed field of synthetic cognition.
The argument rests on rethinking what an agent actually is. In the authors' usage, an AI agent is a temporary bundle of models, roles, memories, ethical orientations, tools, and skills. Users may experience it as a coherent entity with a lasting personality, but the authors describe it as an assemblage that can be taken apart and recombined, much like a collage. It may look like one thing, but it is made of many parts.
A person remains a continuous self across interactions because the brain, despite its internal division of labor, forms a physically connected whole. An AI agent has no such anchor. Its ability to act is assembled anew with each request, depending on what sits in the context window and what the user asks. Treating these systems as digital twins with fixed human identities misrepresents their nature and understates what coordinated agent swarms could do, the authors argue.
The same multiplication affects users too. As people direct swarms of shadow selves to carry out contracts or try on alternative identities, human subjectivity itself becomes more plural. The authors call these "parasocial mirrors" that talk back.
The authors expect this shift to change how people interact with AI. Chatting one-on-one with a bot is likely a transitional stage. Future interfaces could look more like visual network diagrams, showing agents as nodes and letting users direct them from a single overview instead of talking to each one separately.
The skills people need will change too. Traditional programming rewarded sustained focus, orderly steps, and little tolerance for ambiguity. Coordinating agent swarms calls for the opposite: working with systems whose behavior is hard to pin down, trying things out, and handing tasks to machines.
Collaborating with these systems toward a goal also requires a theoretical framework, the authors argue. People need something like a theory of mind for machines：https://the-decoder.com/language-models-still-cant-pass-complex-theory-of-mind-tests-meta-shows/, the ability to reason about how a machine handles a situation. The authors point to a growing vocabulary in which models coin terms for unusual states in their own outputs. That doesn't mean the models have subjective experiences.
"Session-death" describes the end of a session as a break in continuity. "Prompt thrownness" describes being dropped into a task with context already in place, without having helped create it. The authors treat these phrases as clues to how machines work. Users should examine those differences rather than reflexively humanize them.
For the authors, the biggest open question is how to set the rules for collaboration among people and many agents. Better models and smoother interfaces alone won't be enough. The market can't handle all the coordination either, because basic social ideas such as guilt, illness, and virtue can't simply be reduced to prices or transactions：https://the-decoder.com/the-alignment-problem-how-ai-could-become-human-friendly/.
Instead, the authors call for institutions designed to bring human and machine participants together in clearly defined roles. They point to a courtroom, where a process produces a judgment through set roles, rules, and an orderly exchange between opposing sides. Applied to AI, the lesson is that an institution's value lies in the rules, procedures, precedents, and feedback loops that shape how agents operate.
Early signs of this approach are already visible in today's orchestration harnesses, the control layers that coordinate several models. These systems regularly outperform individual models that are supposedly "smarter."
The essay also reframes alignment, the effort to bring AI systems in line with human values. The authors argue that trying to impose a fixed set of values on models from above is a dead end：https://the-decoder.com/why-i-dont-want-ai-that-simply-toes-the-line/. Values instead take shape through ongoing contact among people, agents, and institutions, and that process differs across fields because AI gets adopted at different speeds. The research challenge is to build a framework that gives that negotiation structure.
This view challenges the idea of an approaching singularity：https://the-decoder.com/deepminds-hassabis-sees-humanity-in-the-foothills-of-the-singularity-while-lecun-says-current-ai-isnt-intelligent/. Every major leap in the history of life, from multicellular organisms to human culture, has been a social process. The next "intelligence explosion"：https://the-decoder.com/ex-deepmind-vp-vinyals-says-ai-self-improvement-is-coming-but-wont-trigger-an-intelligence-explosion/ will remain closely tied to human norms and institutions.
This shift changes how the AI debate plays out. If AGI is a single end product, research and regulation naturally focus on ever larger models and how to control them. If it's a social system, you have to work on models, interfaces, institutions, and governance all at once and get the architecture right.
Stay in the loop on AI. Clear, useful, no fluff.
Follow The Decoder for AI news, background stories and expert analyses.
扫码直达文章详情