Connections Rule: Recursive Intelligence from Neurons to Civilization
An evidence-led thought experiment about connectomes, artificial intelligence, and civilization-scale networks
Connection matters—but connection is not sufficient
“Connections rule” can easily collapse into a cult of scale: more synapses, more parameters, more agents, more messages, therefore more intelligence. The evidence demands a harder statement. A capable system needs not only edges, but the right topology, direction, weight, timing, plasticity, energy budget, feedback, and a body and environment that give those edges consequence. Randomly wiring a million components does not produce a brain. Letting a hundred agents talk does not guarantee collective reason.
The claim of this essay is narrower and stronger: structured connection is one of the necessary foundations on which complex intelligence can integrate and expand; capability depends on how connections are organized, revised, tested, and governed.
The deepest layer of today's AI competition is not model size alone. It is the ability to create high-quality links: between data and parameters, models and tools, agents and agents, perception and action, digital systems and the physical world, and human judgment and machine execution.
“Connection” names several different things
At the neural scale, a connection is a physical and biochemical relation with direction, dynamics, and energetic cost. Inside a trained model, it may be a numerical parameter participating in a computation. In a tool-using AI system, it can be an API contract or an authority-bearing action. In society, it may be a communication channel, an institution, a market dependency, or a relation of trust. Treating all of these as identical would be a category mistake.
What survives across scales is an abstract systems question: which state can affect which other state, through what channel, under what rule, with what delay, and with what feedback? That question is useful because it forces us to describe structure and consequence together. A node's isolated capability matters, but system behaviour emerges from permitted influence, blocked influence, and the mechanisms that revise those boundaries.
The neuron's compact grammar: reach, integrate, transmit, feed back
A neuron offers a powerful—and frequently abused—template for thinking about intelligence. Dendrites and cell bodies receive signals from other cells and sensory pathways. Membrane potentials, ion channels, and internal state integrate those influences. Under appropriate conditions, action potentials propagate along axons and alter downstream cells. Chemical synapses dominate the familiar diagram, but nervous systems also contain electrical synapses, neuromodulators, glial effects, and feedback at multiple timescales. Basic neuroscience therefore does not describe communication as a single input-threshold-output line.[1]
Nervous-system function depends on wiring, but also on synaptic strengths, cellular properties, neuromodulation, and network state. Related circuits can preserve useful output through compensation and modulation, so a wiring diagram is not a complete operating manual.[15]
Software APIs, buses, and agent protocols resemble axons and synapses only in one architectural sense: they determine who can influence whom. A software message is not an action potential, and parameter updating is not a literal copy of biological plasticity. The analogy should generate design questions, not erase mechanisms.
From wiring diagram to behavioural prediction: what the digital fly proves
In 2024, the FlyWire collaboration published a whole-brain wiring diagram of an adult female fruit fly: 139,255 neurons and roughly 50 million chemical synapses. The resource includes cell classes, projections, and predicted transmitter identities, and demonstrates how researchers can trace pathways from sensory input toward motor output.[2]
A companion line of work turned more than 125,000 neurons and approximately 50 million connections into a leaky integrate-and-fire model. Stimulation of sugar, water, and mechanosensory pathways predicted neurons involved in feeding initiation and antennal grooming. Researchers tested portions of those predictions with optogenetic activation and behavioural experiments.[3]
Synapse-level connectivity plus predicted transmitter identity was sufficient to generate testable predictions for selected sensorimotor circuits and describe several sensory-to-motor transformations. It did not show that loading the connectome automatically creates a complete autonomous fly with its full behavioural repertoire, needs, learning history, and physical context.
NeuroMechFly supplies another missing layer. It combines anatomy, joints, muscles, control, and a physical environment to test behaviours such as walking and grooming.[4] A connectome constrains where signals can flow; embodiment determines what those signals do in a world with gravity, friction, contact, and consequences.
Connectivity is a skeleton and a constraint on function, but it is not sufficient for complete intelligence. We also need weights, timing, plasticity, neuromodulation, cellular dynamics, a body, and an environment. The digital fly matters not because intelligence has been copied whole, but because structure can now produce circuit hypotheses at unprecedented scale—and those hypotheses can fail in experiments.
A connectome is closer to a circuit diagram than a recording
A circuit diagram tells an engineer which components are connected and which paths a signal may follow. By itself it does not reveal every capacitor's present charge, the history of every input, temperature drift, or component ageing. A connectome has a comparable boundary. Electron microscopy primarily provides a structural snapshot, while a living nervous system continuously changes synaptic efficacy and overall excitability. Hunger, sleep, stress, and experience alter responses. Individuals with broadly similar wiring can adopt different strategies, and one circuit can change its function under neuromodulation.
This limitation makes the digital fly a stronger scientific platform, not a weaker story. Once omissions are explicit, experiments can ask why a prediction failed. Was transmitter identity wrong? Were weights inaccurate? Was cellular dynamics too simple? Was sensory or bodily feedback missing? Failure turns “connection” from a slogan into revisable parameters and localized uncertainty.
Three ways of building relationship models of the world
If knowledge is partly a model of how things relate, revelation and tradition, philosophy and logic, and science and engineering can be read as different epistemic postures. They are not a simple ladder from primitive to advanced, nor consecutive ages in which one replaces another. All three remain active in modern life.
Revelation and tradition: networks of meaning and norm
Religious traditions, myths, and rituals connect people to the world, individuals to ultimate meaning, and actions to norms. Their central questions often concern who we are and how we ought to live. Authority comes through scripture, tradition, experience, and community rather than repeatable experiment. Calling this a “fixed neural circuit” would erase the diversity and historical change within religions.
Philosophy and logic: disciplined exploration of possible relations
Philosophy examines concepts, premises, and implications: what causation is, how knowledge is possible, and how subjects relate to a world. Aristotle's causes, Hume's challenge to induction, and dialectical traditions all reorganize allowable relationship models. Logic disciplines inference, but a valid inference cannot make a false premise true.
Science and engineering: letting the world answer back
Science subjects relational claims to controlled intervention, measurement, modelling, and replication. Engineering then asks whether those relationships work reliably under constraints. An experiment is not literally a stimulus applied to a “world brain”; it is an institutionally organized comparison between prediction and observation.
Describing all three traditions as relationship models is a philosophical lens, not a neuroscience result. It compares how claims receive justification; it does not rank cultures, faiths, or disciplines.
Science is powerful not because it is always right, but because links can be revised
Scientific models fail too. Their institutional advantage is that claims are connected to methods, data, instruments, criticism, and repeatable procedures, giving later investigators a route to locate error. An equation is valuable not only because it compresses observations but because it exposes predictions. An experiment matters not only because it produces a number but because it records conditions, uncertainty, and failure modes. A scientific community is therefore a network of evidence designed—imperfectly—to rewrite itself.
Engineering connects that evidence network to material constraints. A bridge must carry load. A remote radio must remain safe under noise and disconnection. An AI system must invoke tools within an authority boundary. Whether design documents, code, tests, telemetry, and incident reports can be traced to one another determines whether an organization learns or merely accumulates files. Here “connection” stops being only a philosophical metaphor: every claim should reach evidence, and every action should return an observable result.
Connection recursively rebuilt at civilization scale
Communication and computing repeatedly expand who can affect whom, what can be jointly computed, and how results return as feedback. Telegraphy, radio, the internet, and distributed accelerators are genuine increases in connective capacity. Calling the result a “civilization brain,” however, remains an analogy. The network has no single body, goal, or naturally unified consciousness.
| Connection layer | Demonstrated capability | Primary example | Unresolved boundary |
|---|---|---|---|
| Transmission | Information crosses distance; high-speed interconnects let many processors train and serve models together. | Internet protocols and accelerator fabrics | Bandwidth is not understanding. Central infrastructure creates concentration and single points of failure. |
| Representation | Deep models encode statistical relations. AlphaFold maps sequence and evolutionary information into structural predictions.[8] RAG links generation to an external retrievable corpus.[9] | AlphaFold; Retrieval-Augmented Generation | Correlation is not automatically causal explanation. Retrieval imports errors, bias, and stale knowledge. |
| Collaboration | Models coordinate with people or other agents through language, planners, and rules. CICERO combined language modelling and strategic reasoning to achieve competitive human-level Diplomacy play.[10] | CICERO | Game performance is not general social intelligence; additional agents can amplify coordination cost and shared error. |
| Causal and predictive | World-model research learns environmental change for prediction or planning. Genie explores generative interactive environments derived from video.[11] | Genie and environment models | Plausible video prediction is not demonstrated physical causal understanding. Long-horizon consistency and controllability remain open. |
| Action | Vision-language-action models connect web-scale representations to robot instructions; RT-2 demonstrates one transfer path.[12] | RT-2 | Real-world safety, touch, reliability, long tasks, and out-of-distribution conditions remain difficult. |
| Biological | Brain–computer interfaces map neural activity to external-device control. Brain-to-text handwriting research demonstrated high-performance communication.[13] The registered Neuralink PRIME study targets control of external devices through an implanted BCI.[14] | Brain-to-text; PRIME early feasibility study | Long-term safety, stable decoding, privacy, consent, and equitable access must precede narratives of fusion. |
Neuromorphic hardware: borrowing principles, not copying brains
Intel's Loihi 2 supports research in event-driven spiking neural networks, with asynchronous computation, local state, and programmable learning mechanisms.[5] IBM's NorthPole instead places memory and computation tightly together on-chip and reports efficiency, density, and latency advantages on specified inference benchmarks. The paper calls it a brain-inspired neural inference architecture; it is not a replica of Loihi's mechanism.[6] Tianjic was designed to support both neuroscience-oriented and computer-science-oriented models in a hybrid architecture.[7]
Chip interconnect can be compared with fast axonal pathways, external stores with memory resources, and agent protocols with group communication. Similar system roles do not imply identical materials, learning rules, or consciousness. The useful question is not “is it a brain?” but “which organizing principle was borrowed, and under which benchmark and constraint does it work?”
From connections inside a model to connections around it
Early machine-learning narratives located intelligence inside one model: increase parameters and data, then expect capability to rise. RAG redraws that boundary. A model need not compress every fact into its weights; it can retrieve external material at runtime. Tool use connects language output to databases, calculators, compilers, and operational services. Multi-agent systems distribute planning, criticism, and execution among roles. The engineering unit is becoming a system of models, memory, tools, protocols, people, and environments.
System-level connection also magnifies responsibility. Who selects retrieved evidence? What authority does a tool call possess? Can one agent's error contaminate shared memory? When can a person stop an automated process? These are not governance details to add after deployment. They are properties of the links. Every new edge creates a capability, a failure path, and a power relationship that may need audit.
Compute interconnect is infrastructure, not intelligence
Fast fabrics organize thousands of accelerators into a computational body capable of training very large models. That infrastructure is indispensable, but bandwidth answers whether gradients, activations, and parameters can move on time. It does not determine which objective is learned, whether the data is sound, or whether the outcome is safe. Calling the fabric a civilization-scale axon can illuminate a communication bottleneck; it cannot equate throughput with cognition.
Why “connections rule” is still not enough
Connection increases what a system can sense, compute, and do; it also increases how far failure can propagate. Financial networks distribute risk and transmit crises. Social networks connect knowledge and construct echo chambers. Multi-agent systems divide labour and inherit a common false premise. Network value cannot be inferred from node or edge count alone.
- Noise and error propagation. High connectivity spreads weak information quickly unless provenance, confidence, and correction travel with it.
- Fragility and concentration. Dependence on a few platforms, models, clouds, or communication hubs turns efficiency into systemic risk.
- Coordination cost. As agent count rises, communication, conflict resolution, permissions, and shared-state management consume gains.
- Power and governance. Who may connect, who is recommended, and who can write shared memory are political and architectural choices.
- Category error. Synapses, parameters, hyperlinks, social relations, and causal edges are not the same entity. Comparison requires explicit boundaries.
The scarce resource in the next stage is not connection count but connection quality: clear semantics, reliable protocols, rapid feedback, fault isolation, evidence provenance, authority boundaries, and correction. Without those properties, a “civilization-scale neural network” is merely a larger fragile network.
Good connections have engineering properties
A high-quality link needs at least six properties. First, clear semantics: sender and receiver share a contract for meaning. Second, traceable provenance: a conclusion can be followed to data, model version, and operating context. Third, closed feedback: the system observes whether an action worked instead of treating a command as state. Fourth, bounded authority: permissions to read, write, and act are explicit. Fifth, fault isolation: one failing node cannot automatically collapse the network. Sixth, reviewable objectives: people can understand, contest, and revise what is optimized.
These properties explain why protocols, standards, audits, scientific replication, and institutional governance matter as much as model algorithms. Civilization-scale intelligence will not consist only of smarter nodes. It will require shared rules that let unlike nodes cooperate under disagreement, uncertainty, and limited trust.
Harness, Loop, and SDD: making connection an engineering discipline
Three structures can govern connection in a working system. A Harness determines which nodes may participate, which tools they may use, what authority they hold, and which test and deployment gates must be passed. A Loop closes specification, implementation, testing, review, deployment, and observation so that consequences can correct the network. A living SDD preserves interfaces, decisions, evidence, and known issues, allowing knowledge to cross time and team boundaries without depending on one person's memory.
They also reveal three recurrent failures. Without a Harness, a system reaches uncontrolled tools and data. Without a Loop, errors receive no feedback and the network repeats an old judgment. Without an SDD, experience evaporates and each iteration starts from zero. “Connections rule” becomes an engineering proposition only when links can be tested, rolled back, audited, and improved.
The recursive loop: from being shaped by connections to governing them
Beyond established fact, we can now attempt a bounded thought experiment. Natural selection produced nervous systems that perceive, learn, and act. Those systems produced humans capable of studying themselves. Humans externalized communication, memory, and computation into civilization-scale infrastructure. Those external systems now help reconstruct connectomes, search design spaces, control experiments, and generate hypotheses that can be tested.
- Nature forms neural connections.
Biological systems evolve perception, integration, action, and learning inside bodies and environments. - Conscious minds study connection.
Science makes some neural relationships measurable, shareable, and falsifiable. - Civilization builds external networks.
Language, writing, radio, the internet, databases, and AI extend collective memory and coordination. - Networks help investigate and alter the world.
Computation reconstructs connectomes, searches designs, controls experiments, and proposes testable models. - Humans begin to govern the direction of connection.
We must choose authority, objectives, evidence standards, risk allocation, and who benefits from network capability.
“Connections rule” does not mean that more links are always better. It means that those who can build links that are more accurate, efficient, verifiable, and governable gain a greater ability to organize distributed perception, knowledge, and action into a new scale of intelligence.
From competitive metrics to design questions
If this thesis is useful, the way we evaluate AI competition must broaden. Parameter count, training FLOPS, and benchmark scores still matter, but none describes a complete intelligent system. We must also ask: which reliable knowledge can the model reach? Which tools may it invoke? Are permissions minimal? Can outcomes feed back and correct future action? How do agents handle disagreement? Where can a person intervene? Will an error be isolated, amplified, or written permanently into shared memory?
These questions move attention from the smartest node to the most reliable network. The most valuable system may not be the model with the most dazzling single answer, but the one that remains attributable, bounded, observable, recoverable, and capable of learning from field evidence over time. For laboratories, companies, and public institutions, protocols, data governance, evaluation, audit, and human–machine division of labour are not accessories around intelligence. They are constitutive parts of system capability.
Connection also prevents responsibility from disappearing. When a decision passes through a model, retriever, tool, agent, and platform, accountability cannot dissolve into the chain. Civilization-scale networks need a responsibility graph alongside a capability graph: who supplied the data, who selected the objective, who authorized the action, who observed the consequence, and who could stop it. Recursive improvement is sustainable only when the growth of capability links is matched by the growth of responsibility links.
References
- Purves D, et al. Neuroscience, 2nd ed. “Synaptic Transmission.” NCBI Bookshelf. NBK10799.
- Dorkenwald S, et al. “Neuronal wiring diagram of an adult brain.” Nature (2024). doi:10.1038/s41586-024-07558-y.
- Shiu PK, et al. “A Drosophila computational brain model reveals sensorimotor processing.” Nature (2024). doi:10.1038/s41586-024-07763-9.
- Lobato-Rios V, et al. “NeuroMechFly, a neuromechanical model of adult Drosophila melanogaster.” Nature Methods (2022). doi:10.1038/s41592-022-01466-7.
- Intel Labs. “Loihi 2: A Brief Introduction.” Official technology brief. Intel Neuromorphic Computing.
- Modha DS, et al. “Neural inference at the frontier of energy, space, and time.” Science (2023). doi:10.1126/science.adh1174.
- Pei J, et al. “Towards artificial general intelligence with hybrid Tianjic chip architecture.” Nature (2019). doi:10.1038/s41586-019-1424-8.
- Jumper J, et al. “Highly accurate protein structure prediction with AlphaFold.” Nature (2021). doi:10.1038/s41586-021-03819-2.
- Lewis P, et al. “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.” NeurIPS (2020). arXiv:2005.11401.
- Bakhtin A, et al. “Human-level play in the game of Diplomacy by combining language models with strategic reasoning.” Science (2022). doi:10.1126/science.ade9097.
- Bruce J, et al. “Genie: Generative Interactive Environments.” (2024). arXiv:2402.15391.
- Brohan A, et al. “RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control.” (2023). arXiv:2307.15818.
- Willett FR, et al. “High-performance brain-to-text communication via handwriting.” Nature (2021). doi:10.1038/s41586-021-03506-2.
- ClinicalTrials.gov. “PRIME: An Early Feasibility Study of a Precise Robotically Implanted Brain-Computer Interface for the Control of External Devices.” NCT06429735.
- Marder E, Goaillard JM. “Variability, compensation and homeostasis in neuron and network function.” Nature Reviews Neuroscience (2006). doi:10.1038/nrn1949.