From a brain network study to the architecture of emergent AI

The starting point for these thoughts was a short research article about a new University of Notre Dame study, published in Nature Communications on 26 January 2026. The article summarises work showing that human intelligence does not come from a single brain region but from the coordinated organisation of several brain networks as a whole system. The researchers analysed extensive imaging and cognitive data from adults and identified four key features: (1) parallel activation of several networks, (2) long-range “shortcut” connections, (3) regulating hubs, and (4) a small-world architecture that combines local specialisation with global integration.

One quote from Aron Barbey, the study’s lead author, stood out in particular: “Many AI systems can perform specific tasks very well, but they still struggle to apply what they know across different situations. Human intelligence is defined by this flexibility, and it reflects the unique organisation of the human brain.” (rendered from the German version of the article) Barbey thereby stresses that today’s AI mainly scales specialised abilities, whereas what is decisive about human intelligence is the flexible system organisation of the brain.

In the discussion that followed, I interpreted this statement in a sense “the other way round”: if intelligence arises from the organisation of complex networks, that is not an argument against AI but for building ever more complex, better connected artificial systems. My hypothesis was that with growing model size, computing power and above all with new, more system-oriented architectures, emergent phenomena could appear: possibly up to a qualitatively new level of machine intelligence, or even something that could be discussed as a “new species” with a consciousness of its own.

The core of the discussion was therefore:

The article suggests that merely scaling specialised AI modules is not enough to reach the flexibility of human intelligence.
My reading was that exactly this insight provides a blueprint for new AI architectures: systems that implement distributed networks, small-world structure, regulatory hubs and global coordination could come closer to human intelligence, including in its flexibility.

At this point I brought in an essay I had written earlier: “Conway’s Game of Life: an interdisciplinary look at emergence, life and intelligence”. In it I examine Conway’s Game of Life as a minimal model for emergence, universal computability and the question of how complex, seemingly “living” patterns can arise from simple local rules. The essay builds bridges to theoretical computer science (Turing completeness, the halting problem), to neural networks and neural cellular automata, to the philosophy of mind (substrate independence, functionalism, consciousness) and to biological and systems-theory concepts of life.

Taken together, an exciting synthesis emerges:

The Notre Dame study shows: intelligence is a network effect of brain organisation.
Conway’s Game of Life and other cellular automata demonstrate: even extremely simple, locally deterministic systems can lead to universal computability and a high degree of emergence.
My hypothesis connects the two: sufficiently complex, well-organised artificial networks (e.g. LLM-based multi-agent systems with CA-like memory structures and small-world coordination) could produce a similar quality of emergence to biological brains, possibly up to phenomena close to consciousness.

From this shared foundation, the conversation produced the idea of designing an architecture and research scenario for such an “emergent AI”. The following concept paper, “EMERGENT-MIND-ARCH” (EMA), summarises this design and is meant as a starting point for further technical and theoretical work.

An architecture design for emergent, network-based AI

1. Starting point and motivation

Recent neuroscience studies suggest that human intelligence arises from the coordination of distributed brain networks, not from individual specialised regions.
Cellular automata such as Conway’s Game of Life show that highly complex, emergent patterns, up to universal computability, can arise from simple local rules.
Modern AI (LLMs, deep learning) already demonstrates emergent abilities, but often remains stuck in narrowly defined tasks and shows limited robust transfer to new situations.

Guiding question: what could an AI architecture look like that combines principles from brain networks, cellular automata and deep learning in such a way that greater emergence, flexibility and possibly “mind-like” properties become more likely?

2. Guiding hypotheses

H1: intelligence is a property of system organisation (network architecture, coordination, dynamics), not of individual modules.
H2: systems with the following properties have an increased chance of emergence:
distributed, local update rules (CA/NCA principle),
small-world network structure with hubs and long-range connections,
learning functional blocks (LLMs, neural networks),
persistent internal dynamics (not purely stateless request-response).
H3: within a 3–10 year window, prototypes of such systems are technically feasible; whether they develop consciousness is empirically and philosophically open.

3. Target picture: EMERGENT-MIND-ARCH (EMA)

3.1 Overall architecture (layer model)

Layer A: functional cores (LLM & specialised NNs)

Large language/multimodal models for semantics, planning, code, domain knowledge.
Specialised networks (e.g. for time series/trading, vision, RL policies).

Layer B: neural cellular automata (NCA state)

A 2D/3D grid as the system’s “inner space”, whose update rules are implemented and trained by a small neural network.
Used as: working memory, an internal “landscape” for concepts and states, slow self-organising dynamics.

Layer C: multi-agent society

Several agents (instances of LLM + tools) with different roles:
- perception, world model, planning, critic, meta/self-model.
Agents interact indirectly via the NCA state and directly via messages.

Layer D: coordination graph (small world)

A communication network with local clustering and a few long-range connections.
Routing rules between agents are (partly) learnable.

Layer E: environment / embodiment

A simulation (e.g. financial market, game world, multi-agent environment) as the outside world.
Alternatively or additionally: real APIs (e.g. trading infrastructure) with clear security and sandboxing boundaries.

4. Core components

4.1 NCA memory

A differentiable cellular automaton that holds a tensor state (S_t) (e.g. H×W×C) and transforms it into (S_{t+1}) via a trainable update network (f_θ).
Properties:
local (cells update from their neighbourhood), yet globally complex patterns.
serves as a shared “blackboard space” for all agents.
can form stability (attractors), oscillations, travelling patterns.

4.2 Agent roles (examples)

Perception agent
Input: external signals (text/API/events).
Output: targeted modification of the NCA state (writing in/activating patterns).
World-model agent
Maintains an explicit graph (entities, relations, forecasts).
Synchronises parts of the graph with patterns in the NCA state.
Planning agent
Reads the NCA state and world model, proposes action plans, experiments and strategies.
Critic agent
Evaluates plans, consolidates contradictions, measures uncertainty.
Meta/self-model agent
Tracks longer-term changes in the NCA state and in performance.
Formulates internal self-descriptions (“what I can do”, “where I am uncertain”).

4.3 Coordination mechanism

Graph structure with:
dense connections within functional clusters (e.g. perception–world model–planning),
few global connections (meta-agent ↔ all).
Routing rules:
initially rule-based, later partly optimised through RL/meta-learning.
Goal: efficient yet flexible communication, analogous to the brain’s “shortcuts”.

5. Learning and research programme

5.1 Learning strategies

Standard pre-training of the LLM components (external).
NCA training via gradient methods on tasks such as:
stable memory, pattern continuation, error correction, self-repair.
Joint fine-tuning through RL/imitation learning in an environment:
reward for longer-term goal achievement, robustness, transferability.

5.2 Research questions

Emergent internal patterns

Which stable/recurrent NCA structures arise?
Do they correlate with particular tasks, “moods” or the agents’ self-reports?

Transfer and flexibility

Performance comparison of EMA vs. classic LLM agents on new tasks, domain shifts, disruptions.

Self-reference and proto-metacognition

Can the meta-agent make reliable, measurable statements about internal states (uncertainty, knowledge gaps, conflicts)?

Indicators close to consciousness (with caution)

Applying complexity and integration measures to the NCA state and communication graph.
Investigating how these measures change with architecture parameters.

6. Roadmap (short)

Phase 1: minimal prototype
1 NCA memory + 2–3 agents (perception, planning, critic) in a simple environment (e.g. a sequential decision problem).
Goal: demonstrate that NCA-based internal dynamics improve transfer/robustness.
Phase 2: expansion into a society
Adding the meta/self-model agent, a more complex environment, continuous runtime.
Phase 3: scaling & analysis
More agents, larger NCA states, more complex environments.
Systematic measurement of emergence indicators and long-term behaviour.