Collective intelligence is the best route to safe, inclusive superintelligence, and to sharing the returns from creating it.

The claim

Collective intelligence is the best route to safe, inclusive superintelligence, and to sharing the returns from creating it. A superintelligence built by a handful of labs carries their blind spots into everything it touches, and its returns go to those few. I rest the case for the collective on correctability, not on raw smarts: Vaccaro et al. (2024) found that thrown-together human-AI teams do worse on average (g = -0.23) and beat both the human and the AI only when the human was the stronger party (g = 0.46), so "best" has to mean easier to catch and fix, and the collective has to be designed, not assembled. We build it in the open, where anyone can see what it is taught, add what it is missing, and fix what it gets wrong.