Trillium's Nathan Lambert says RSI hype oversells AI and distillation bans backfire
In an a16z podcast conversation with partner Lisha Li, Trillium founder Nathan Lambert pushed back on industry hype around recursive self-improvement (RSI), arguing the truth sits between marketing and fear and that his real worry is frontier labs' own unhardened infrastructure. He frames model progress as a function of compute, data and talent, saying Chinese labs are not short on talent but are constrained by compute, which makes open-weight releases a pragmatic route to visibility and market access, citing Moonshot AI and Zhipu as examples. He also says existing public research does not convince him that distillation is the decisive factor behind fast followers, doubts that banning it would work and expects such moves to do more harm than good, while promising Trillium will publish its open SFT and mid-training stack, task sets and experiment trajectories so outsiders can verify what RSI can actually do.
As debate over recursive self-improvement and the open-versus-closed gap intensifies, a practitioner-level assessment helps ground the discussion in compute, data and engineering rather than takeoff narratives.