this position paper is actually bang on a direction we've been working on for the past year — scaling many specialized agents together with RL instead of just scaling one big model.
the thesis is that intelligence scales through interactions, not just individual capacity. E.g., prior work: https://x.com/t_ed_li/status/2038763049574879351