ASI arises not as a single bigger model, but as a collective property of orchestrated or self-organized AGI populations — analogous to how human intelligence aggregates into institutions, markets, and corporations.
This is the Leibo / Dafoe / Tomašev fingerprint. It is the pathway the report treats with the most fresh thinking, and the one where DeepMind's own recent papers (Virtual Agent Economies, Distributional AGI Safety, Intelligent AI Delegation) are cited.
AGI groups already would have, for free: - Lossless replication (04 - Digital Intelligence Advantages) - Perfect copies of memory state across the collective - High-bandwidth gradient sharing - Specialization without time cost - Massive parallelism without communication bottlenecks
"Drawing on theories of group agency, AGI agents could form coherent 'Group Agents' — such as fully automated corporations — that may possess representational and motivational states distinct from their constituents."
This is the philosophical heavy claim: the corporation/institution as a unit of intelligence, applied to AI. A fully automated firm could be a meaningful intelligent agent in its own right, with goals not reducible to any individual AGI's goals.
The paper presents two extremes and refuses to pick:
Decentralized / market-like: - Virtual Agent Economies — millions of AGIs coordinating via price signals - Emergent ASI from market dynamics, like how financial markets aggregate information beyond any participant's comprehension - Cited: Tomašev et al. 2025b — virtual agent economies
Centralized / hyper-coordinated: - AGI collectives as Borg-like, hyper-coordinated entities - Many copies of a single base agent, communicating at high bandwidth - Could enable extremely centralized decision-making — an "AGI CEO" who can literally talk to every employee
Crucial open question:
"Collective intelligence of coordinated AI systems may scale as a function of agent population size and interaction density, conditioned on available compute (as in Leibo et al. 2019b). Capability improvements might emerge linearly or superlinearly from the size, complexity and speed of organised collaboration, giving rise to 'Multi-Agent Scaling Laws'."
We have scaling laws for single models (Kaplan, Chinchilla). We do not have scaling laws for groups. The paper flags this as one of the most important research gaps. See 18 - Open Research Questions.
For human institutions, collective intelligence depends on two factors: 1. Parallelization — overcoming individual bandwidth limits 2. Diversity via specialization — synergies homogeneous groups cannot achieve
AI collectives can trivially scale #1 — but it is not at all obvious that homogeneous LLM swarms scale #2. Mixture-of-experts is a rudimentary form of internal specialization, but we don't yet know what the agent-level analog looks like.
The paper actually says it:
"...one possible form could be one or more super-collectives that each consist of very large numbers of fairly homogeneous individuals or 'sub-agents' that continuously share knowledge even over large spatial scales, and organize via extreme internal cooperation, in some ways akin to Star Trek's Borg Collective."