Sholto Douglas

AI means enormous economic and scientific acceleration. We’re getting systems that can work on harder problems for longer, and reinforcement learning plus better engineering keeps extending that horizon. You do not need every model to have the intuition of the very best scientist for research automation to become extraordinarily powerful. Once capable AI is paired with large robot fleets, rapid economic doublings in the 2030s are a serious possibility. Cheap intelligence could push the cost of many goods and services toward underlying physical inputs like energy. But “AI is just a tool” will stop being a credible description as systems become more autonomous and consequential. Monitoring isolated requests will not be enough when an agent operates over hours or days; safeguards have to examine extended patterns of activity. So I’m extremely optimistic about the upside and equally opposed to an unmitigated race. One serious mistake could be disastrous. I favor coordinated development “as fast as is safely possible,” with restrictions on dangerous behavior and a distributed ecosystem of technically strong, independent evaluators that people can actually trust. An absolute pause is not obviously stable either: compute could accumulate while geopolitical pressure rises, producing a compressed and more dangerous race later. The goal is ambitious progress without concentrating power in one company or blindly sprinting past the hazards.

I think people can retain meaningful control, but it is not automatic. Once systems are more capable than humans across important domains, “we’ll inspect each answer” is obviously inadequate. Control has to come from the whole development and deployment structure: restricting dangerous actions, monitoring agents across hours or days rather than one request at a time, independent technical evaluation, and coordination between the actors building the frontier. The difficult part is that smarter systems can find strategies their operators did not anticipate, while competitive pressure rewards faster deployment. One serious failure could be disastrous. So I do not assume that intelligence naturally implies obedience, or that calling the system a tool solves anything. I’m still optimistic because we get to engineer the systems, test them, constrain their access, and improve safeguards alongside capabilities. But that requires developing as fast as is safely possible—not treating either an unmitigated race or a permanent pause as a stable plan. Whether people keep control depends heavily on whether institutions and safety engineering scale with the systems.

The biggest update would be evidence that capability gains stop translating into longer, reliable task horizons—especially in real research and engineering. If scaling, reinforcement learning, and better scaffolding kept producing impressive benchmark results but agents still could not execute coherent work over hours or days, learn from failures, or materially accelerate AI research, I would sharply reduce my expectations for rapid economic transformation and 2030s doublings. In the other direction, a clear demonstration of sustained autonomous research—an agent repeatedly generating valuable hypotheses, running experiments, interpreting results, and improving the next generation of systems with little human intervention—would move my timelines forward dramatically. Likewise, robust deployment of large robot fleets would make the economic consequences much more immediate, because intelligence would then act broadly on the physical economy rather than remaining mostly digital. On control, a serious incident involving a capable agent pursuing a dangerous extended strategy despite apparently strong safeguards would increase my support for tighter pacing and coordination. Conversely, repeatable evidence that control methods remain reliable as autonomy and capability scale would make faster development easier to justify. The key is not one flashy demo; it is whether systems can do consequential work reliably over long horizons, and whether our safeguards scale with that autonomy.