Coordination Systems / Research

All Intelligence is Collective Intelligence

Intelligence has never been an individual achievement. It accumulates between people, models, programs and memories. Coordination is where the value is made.

Three programs follow. Each takes one part of the claim: what a group keeps, how specialists combine, and how learning travels without being collected.

Cultural Intelligence

Human knowledge compounds because a group evaluates it, normalizes it and passes it on. The mechanism is social, and it sits outside any individual head.

Culture starts as signals people notice and argue over. Taste points shared attention at what might matter. Trust compresses a history of validation so nobody has to check everything alone. Valence is the charge a signal carries: pull, fit, novelty, belonging, refusal. What a community keeps is whatever survives all three.

Consensus coordinates a group. It does not make the group right. Every act of normalization keeps some distinctions and erases others, and the shape of the network decides which ones: who observes whom, whose judgment travels, where dissent stays legible.

We treat cultural intelligence as cumulative social learning: a shared ability to judge experience, keep its context, and build on judgments already made. The hard part is making that legible without flattening it into a single number.

Research lineage

  1. Herrmann et al. · 2007Humans Have Evolved Specialized Skills of Social Cognition: The Cultural Intelligence HypothesisLocates human cognitive advantage in social knowledge exchange rather than general problem solving alone.
  2. Mesoudi & Thornton · 2018What Is Cumulative Cultural Evolution?Names the loop culture compounds through: innovate, transmit, improve, repeat.
  3. Muthukrishna et al. · 2018The Cultural Brain Hypothesis: How culture drives brain expansion, sociality, and life historyModels the feedback between social learning and how much cognitive capacity a group can sustain.
  4. Golub & Jackson · 2010Naïve Learning in Social Networks and the Wisdom of CrowdsShows how influence and network topology determine whether distributed belief formation approaches truth.

Compositional Intelligence

A model is a component. General capability shows up when specialists are wired together at runtime.

At runtime a model can call other models, symbolic programs, tools, stored memories and a person. Capability then lives in the wiring. No single part of the graph contains it.

So the thing to search over is the whole graph: routing, control flow, memory, evaluators, and the models themselves. Vary the routing. Vary the memory. Vary the models. Then test each assembly against the world instead of against an intuition about architecture.

Assemblies that work become artifacts a network can inherit instead of rediscovering. Some of what they do can be fused back into the models themselves. Coordination finds the structure; keeping the structure is what makes it compound.

Research lineage

  1. Hu, Lu & Clune · 2025Automated Design of Agentic SystemsTreats agents as code and searches over prompts, tools, workflows, and their combinations.
  2. Akiba et al. · 2025Evolutionary Optimization of Model Merging RecipesDemonstrates evolutionary discovery across both parameter and data-flow composition spaces.
  3. Romera-Paredes et al. · 2024Mathematical Discoveries from Program Search with Large Language ModelsPairs language-model variation with evaluators to evolve interpretable programs.
  4. Wang et al. · 2024Mixture-of-Agents Enhances Large Language Model CapabilitiesShows that heterogeneous model outputs can be coordinated into stronger runtime performance.
  5. Zhang et al. · 2026Darwin Gödel Machine: Open-Ended Evolution of Self-Improving AgentsUses empirical evaluation and an archive of discoveries for open-ended agent evolution.

Open Epistemic Intelligence

Learning that stays local and still compounds.

Learning starts on a machine somebody owns, keeps its author attached, and compounds across the network while the raw experience stays put.

The person inside a local loop is the oracle. They live the experience, judge the result, and decide what is worth keeping. Their files, memory and taste stay on hardware they own. Learning starts from that authority. Nothing gets scraped into a central corpus first.

A node emits signed learning events and keeps the experience that produced them. Provenance, accrued trust and network position decide which events travel and how far. Shared memory here is a propagation protocol, selective by design.

What compounds is knowledge that still carries who learned it, in what context, under whose judgment. There is no canonical state and no single view. Communities coordinate without dissolving into each other.

Research lineage

  1. McMahan et al. · 2017Communication-Efficient Learning of Deep Networks from Decentralized DataEstablishes shared learning from locally computed updates without centralizing raw data.
  2. Hegedűs, Danner & Jelasity · 2021Decentralized Learning WorksDemonstrates viable peer-to-peer learning without a central aggregation server.
  3. Kleppmann et al. · 2019Local-First Software: You Own Your Data, in Spite of the CloudTies collaboration to local operation and durable ownership.
  4. Tarr et al. · 2019Secure Scuttlebutt: An Identity-Centric Protocol for Subjective and Decentralized ApplicationsThe lineage for signed, append-only histories that replicate subjectively.
  5. Wang et al. · 2024Voyager: An Open-Ended Embodied Agent with Large Language ModelsShows cumulative agent learning through a reusable library of executable skill artifacts.
  6. Packer et al. · 2023MemGPT: Towards LLMs as Operating SystemsTreats agent memory as managed tiers past the context window.