Working at Meta's Superintelligence Labs after OpenAI

A researcher who moved from OpenAI to Meta Superintelligence Labs says daily work shifts with project stage and immediate deliverables. Milestones such as a major training or reinforcement‑learning run come on a roughly 10‑month cadence, and work intensifies as deadlines approach.

Solutions are judged against the current model version, so missing a deadline can mean the issue no longer exists in the next iteration. Frontier labs differ from typical Big Tech settings because compute, not headcount, is the limiting factor. Adding people divides available compute, so teams stay small and require high‑bandwidth communication.

Team boundaries are fluid: individuals own projects but frequently collaborate across groups, and the presence of many senior staff gives people broader scope. Clear communication and deep engagement with code are essential.

meta superintelligence, openai, reinforcement learning, training run, compute bottleneck, small teams, model version, deadlines, high bandwidth, code engagement