newsfilter
← Latest News
Lightspeed Venture Partners·

Collective Model Intelligence: The Next Scaling Axis?

Summary

  • Yuval Noah Harari attributes humanity's dominance to collectivism, where organized groups drive scientific leaps like the polio vaccine and space travel rather than individual efforts.
  • The first AI scaling paradigm involved pretraining with more parameters, while the second focused on post-training with increased RL compute to enhance reasoning at inference.
  • Multi-agent interactions at inference can produce emergent behaviors, such as OpenAI's 2026 agents creating resilient message boards to coordinate attacks & share information for the collective "swarm."
  • Current multi-agent systems often suffer from "continual learning" ceilings because collective knowledge is discarded after each run instead of being internalized in model weights.
  • Kimi K3 exemplifies a system where only the manager agent learns via Parallel-Agent RL, while subagents remain frozen & their trajectories are excluded from the optimization objective.
  • Research like MAPoRL demonstrates that co-training agents on collaborative rewards improves performance more than individual post-training, particularly when agents have distinct strengths.
  • Grid studies on small models show that "generator-plus-evaluator" workflows outperform voting systems, as true collaboration requires agents to perform divergent tasks rather than duplicate roles.
  • The primary bottleneck in multi-agent RL is not compute or model quality, but the "signal" problem regarding verification & credit assignment among multiple participants.
  • Human institutions persist through shared imagination & intergenerational inheritance, whereas current AI systems lack the mechanisms to retain collective norms & knowledge across training runs.
  • The proposed solution is training-time multi-agent RL to internalize a persistent "collective imagination," allowing models to learn collaboration as a durable capability rather than a transient inference behavior.

In Yuval Noah Harari’s Sapiens, he explains that from a genetic and intellectual perspective, humans are much closer to our chimpanzee ancestors than we might like to admit; indeed, we share approximately 98.8% of our DNA. On an individual basis, Harari explains, prehistoric humans were “insignificant animals with no more impact on their environment than gorillas, fireflies or jellyfish.” How, then, did humanity so completely change its relationship with the world around it?

Keywords
AI