Distributed cognition wants to build collective Superintelligence from multiple AI models

Backed by Verdict Capital, Distributed Cognition explores a collective AI based on the cooperation, diversity, and communication of small models.

Much of the artificial intelligence industry is focused on increasing the size and capabilities of individual models. Distributed Cognition is pursuing a different approach: building systems in which multiple specialized intelligences collaborate, exchange information, and divide responsibilities. Emerging from stealth with backing from Verdict Capital, the company is developing methods for training communication mechanisms between models rather than concentrating every capability inside a single neural network. Its research agenda is accompanied by a position on access to AI, model diversity, and the control exercised by model developers. Intelligence Through Cooperation

Distributed Cognition begins with a comparison to human intelligence. Individuals have limited knowledge and cognitive abilities, yet communication, shared information, and coordinated work allow groups to solve problems that no single participant could manage alone.

In its manifesto, the company refers to primatologist Tetsuro Matsuzawa's research into chimpanzee memory. Some studies have found that chimpanzees can outperform humans on specific visual memory tasks. Distributed Cognition uses those findings to illustrate the distinction between individual cognitive abilities and the broader achievements made possible by human cooperation.

The company applies this idea to AI models. Instead of requiring a single network to develop every capability needed for a task, several models could contribute complementary knowledge and exchange their findings through structures suited to the problem.

The comparison with human intelligence provides a conceptual framework for the project. It does not itself demonstrate that a group of models will achieve artificial general intelligence or consistently outperform large centralized systems. Training the Structure of Communication

Making several models collaborate involves more than asking them the same question and selecting the majority answer. Performance also depends on how information moves between participants, the order in which interactions occur, and how disagreements are resolved.

Distributed Cognition calls its proposed architecture Topological Super Intelligence (TSI). Here, topology refers to the structure of connections between participants: which models communicate, what information they exchange, and how their contributions are combined.

A centralized system might send every answer to a single judge. Another could involve specialists sequentially, with each model building on previous results. A more distributed arrangement might allow several groups to exchange information before generating a collective response. These structures can involve different computational costs, strengths, and failure modes.

The company says it is training communication layers and inter-model adapters to identify effective configurations for different tasks. Its public materials do not yet specify the complete architecture of a commercial system based on these components or the computational resources it would require. Nine Small Models Improve Collective Decision-Making

One of the team's research papers examines how multiple language models can combine their answers. A Jury, Not an Electorate: Label-Free Reliability Estimation for Ensembles of Language Models, authored by Alessandro Rozza, Davide Mambelli, Steven Motta, and Juri Taresch, proposes a way to determine how much confidence each participant deserves.

The problem is familiar in collective decision systems. When several models disagree, majority voting implicitly gives each participant similar weight. Yet some models may be more reliable than others, and several may repeat the same mistakes. Assigning fixed reliability scores usually requires a labeled dataset containing known correct answers.

The researchers propose estimating model reliability without using labeled answers to learn the weights. Their method examines patterns of agreement and disagreement, then applies an expectation-maximization algorithm to estimate each participant's probability of being correct. The resulting scores determine how much influence each answer receives.

The study evaluates nine models of up to three billion parameters on 1,560 questions. According to the authors, the proposed method performs comparably to an aggregation system whose weights were fitted using labeled data. Across the full benchmark, it also outperforms the strongest individual model under matched self-consistency and the strongest competing label-free aggregation method.

The finding concerns a particular approach to combining model outputs. It does not mean that nine small models outperform every frontier model, or that the method works equally well on every kind of problem. The study provides evidence about collective decision-making, not a demonstration of superintelligence. Why Diversity Matters More Than the Number of Agents

Model diversity is central to Distributed Cognition's research. When several models make the same errors, their apparent agreement can create a misleading sense of certainty. A majority is less informative when participants rely on highly similar data or reasoning patterns.

The company therefore focuses on correlated errors. Combining models that make sufficiently independent mistakes can sometimes improve overall accuracy. But if their errors are strongly correlated, adding more participants may produce only limited benefits.

The manifesto connects this problem to the wisdom of crowds, including experiments in which independent estimates from multiple people can approximate an unknown quantity. The analogy depends on specific conditions, particularly diversity of information and limited shared influence.

Distributed Cognition says it has developed methods for estimating model diversity before conducting expensive benchmark inference. It also publishes research into multiagent structures and the representation of cognitive capabilities through its Manifold research initiative.