Sina Habibian Podcast

Sina Habibian × Jeffrey Quesnelle

Nous Research’s bet is that open AI needs the capacity to train foundation models independently, with Psyche coordinating distributed GPU participation.Quesnelle says DisTrO cuts training communication requirements by roughly 1,000:1, though each node still needs substantial fast memory and local interconnect.The approach could make frontier research less dependent on a few model publishers, while verification bottlenecks and unresolved safety risks remain.