In chess, there used to be a time when a human and an AI together could beat any AI alone. They called the pairs centaurs. Eventually, standalone AI outperformed centaurs. Work is in its centaur era right now: AIs do the work, and humans stay in the loop to verify it.
But human verification is a stopgap that gets weaker as AI capabilities scale. As the cost of AI autonomy goes down, the demand for human verification becomes unbearably high. Models will write increasingly complex software for themselves, and human verification will no longer be able to keep up. The only thing that will matter on the other side is whether we built airtight safety mechanisms.
That safety can only be built by whoever owns the frontier.
We are closing loops: building safety mechanisms that allow fully autonomous work to scale. The same mechanisms scale to every loop, including recursive self-improvement of code and chips.