Source
Pilgrim, Morford, Warren et al. (2025), LIT-046, “The Computational Foundations of Collective Intelligence”, read in full in NOTE-055. The empirical instance is Dreyer et al. (2024), LIT-075, read in NOTE-040, which LIT-046 does not cite. NOTE-055 says the paper “does not support or contradict any THEORY document on its own evidence”. This document states the account the paper puts forward, with that verdict attached.
What was actually shown
The inventory (Table 1). A group has the union of its members’ sensory information, the product of their state and action spaces with a collective state S_c on top, and their processes plus inter-individual ones. The caption says the operators denote totals, “not simple addition”, so no metric is given.
The constraints (§3). Coordination: synchronisation, communication cost, integrating the parts of a decomposed task, bandwidth. Cooperation: diverging preferences, social dilemmas, and “participatory computation”, where contributing and benefiting are one act.
The mapping (§4, Table 2). Aggregation cuts noise to σ_i²/n when errors are uncorrelated. Combined states can represent a distribution over hypotheses where a member holds one. Collective memory can outlast any member. The joint action space supports division of labour.
The trade-off (Fig. 2). The same budget can move a group further along a member’s resource–quality curve, put it on a different curve, or, with coordination costs, below the member.
Cases. Citing Sasaki et al. (2013), ant colonies beat single ants on hard discriminations and lose on easy ones, through early amplification of noisy assessments (§5.2). In LIT-075, ant groups improve with size on a transport puzzle, while human groups barred from speech do 32% worse than individuals and talking groups 7% better.
What could have come out otherwise. As stated, nothing. The framework names the outcomes after the fact. It cannot be wrong until its resources are measured.
What this does not say
- That the forms of collective intelligence derive from one principle. The link from resources to mechanisms is argued in prose. NOTE-055 rates the unification claim weak.
- When a group will beat its members. That needs the costs and gains in one unit. The framework lists constraints and models none.
- That collectives reason disjunctively or counterfactually. The paper flags this “Speculatively, we predict” and gives no example.
- That LIT-075 tests it. The ants-and-humans study is consistent with the trade-off, and was not designed or analysed against it.
- Anything new about averaging. The σ²/n result is the standard one for independent errors.
Connections
- The ants-and-humans finding (THEORY-608) is the evidence this account leans on, with its own caveats: no inferential statistics, and mechanisms fitted or observed informally.
- Minds of minded agents (THEORY-043) names this paper as a neighbour. The resource view is silent on whether a collective that computes is a subject.
- Variance pooling (THEORY-696) is another case of a group gaining from averaging over members, in growth rather than in error, with the same dependence on how correlated the members are.