A new preprint on arXiv takes up the question of Collective AI: the idea that many independently developed models could work together as a kind of shared intelligence. The authors note that such models do accumulate diverse and complementary capabilities, so there is raw material for collaboration.
But the abstract warns that a growing model collection does not automatically realize Collective AI. Simply having more models is not enough; something must actively coordinate them.
The paper argues that the gap must be addressed by understanding model capabilities — the available text cuts off there, so the proposed characterization and coordination methods are not described in the excerpt. The preprint is listed as new on arXiv under number 2610.24739.