Source
Dreyer et al. (2024), LIT-075, “Comparing cooperative geometric puzzle solving in ants versus humans”, PNAS, read in full in NOTE-040 with the supplement.
What was actually shown
The task. A T-shaped load moved through three chambers and two slits, built at five scales so that body size to load size stays roughly constant. Humans in groups pull only on assigned handles. Performance is path length in configuration space, normalised by the corridor width d_cor; for humans, also attempted transitions on a hand-built state graph.
Ants (Fig. 2A). Large groups (n = 28) reach about 90–100% success by about 400 d_cor, small groups (n = 48) about 50%. Single ants are three composite multi-day runs. After a collision large groups keep more of their speed and turn less (Fig. S5a,b). In simulation, alignment persists 5.9 s and 4.8 d_cor for large groups against 0.7 s and 1.9 d_cor for small ones (Table S4). An existing transport model with size-independent rules reproduces the gap.
Humans (Fig. 2B; SI Note 7). Individuals reach 100% within about 50 d_cor; 91 ± 1% of their transitions are consistent with depth-first search. Restricted groups (n = 48) are 32% worse than the individuals’ median, talking groups (n = 43) 7% better. Restricted groups start moving within about a second and take the greedy option more often. The talking-group gain comes from leaving a dead-end state: 91% of groups return after trying it, against 38% of individuals.
What could have come out otherwise. Restricted human groups could have matched or beaten individuals, as ant groups beat smaller ones. They did worse.
What this does not say
- That the differences are significant. The paper gives mean ± 1 SE and CDF bands and no test. “Significantly” in the text has no test behind it, and nothing is controlled for age, composition or repeat exposure.
- That human groups fail by majority vote. A one-shot vote with greediness P = 0.2, set “by adjusting only parameter P”, fits “reasonably good … in nine out of ten” forks. No alternative model is compared, and at three of five forks the vote is among a random quarter of the group.
- That talking groups follow a random leader. That comes from “personal observations”, since voices were not recorded (SI Note 6). The claim that leader-following is “strictly superior” to majority vote when most choices are wrong is a 200-run stylised simulation.
- That ant groups have collective memory. The “short-term collective memory” is persistence of heading relabelled. The order parameter is measured only in simulation, whose large-group carrier count (50.8) is well below experiment (79.9).
- That simple minds scale and complex ones need communication. That is the abstract’s generalisation from one ant species (P. longicornis), one puzzle and one population recruited in Israel.
Connections
- Collective intelligence as resources against coordination costs (THEORY-402) takes this as its clearest empirical instance: more pullers help ants, whose individual cognition contributes little, and aggregation through weak channels costs humans the plan each brings.
- Synchronised functional units (LIT-933). Nowak et al. count any binding of dynamics as a functional unit. Here a group bound by forces alone performs worse than one member, which is a case their scenario does not distinguish.