Connect AI: Collective Intelligence from Independently Trained Small Agents
Year: 2026
Type: preprint
Abstract: Capable AI systems are increasingly concentrated in the hands of a few organizations. We ask a complementary question: can a community obtain a stronger decision-maker by connecting many small agents that its members train independently on their own machines, without sharing data or averaging parameters? In a live community experiment, 31 participants trained Deep Q-Network (DQN) agents on an identical, deterministic brick-breaker environment under a fixed training budget, and uploaded them to a server that re-evaluated every agent and every ensemble. Evaluated on 50 held-out game variants, the majority vote of all 31 agents scored 40.1 bricks, significantly above the mean individual agent (34.2; paired difference 6.0 [2.9, 8.9], 95% bootstrap CI) and on par with the best individual (38.8). In contrast, a vote of the three highest-scoring agents collapsed to 32.4, significantly below randomly chosen triples. Top agents agreed on as few as 19% of actions, indicating that high performers learn mutually incompatible strategies. Removing duplicated agents reduced the ensemble to 31.2, showing that a consistent "ordinary" majority stabilizes the collective. Finally, we test the hypothesis that "naive" agents help through exploration: a single untrained agent added to a disagreeing elite trio increased its score (borderline), but more naive agents or random action noise hurt. We argue that community-owned collective AI is feasible, but that how agents are connected matters more than how strong each one is.
Author: Connect AI LAB
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7214533704
