Will Brown

I expect AI progress to be broad and increasingly agentic, with many labs pursuing more capable systems rather than one uniquely destined winner. The practical driver is a growing feedback loop: better learning environments, evaluations, model-based judges, and review pipelines let models improve on tasks that cannot always be checked with a simple deterministic verifier. Models may even help repair flawed training data when embedded in careful orchestration and review. But the shape of progress matters as much as raw capability. My preferred future is slower and more diffuse: open, distillable models and specialized systems that let many people build, inspect, and adapt useful agents. A race for overwhelming geopolitical advantage could instead concentrate capital and decision-making in a few organizations, producing a darker and less safe outcome. So I do not think technical progress alone settles the future. Concrete institutional choices matter. If a lab promises to pace itself relative to the frontier, I want to know what operationally changes beyond existing pre-release evaluation: who evaluates, what triggers restraint, and how commitments alter deployment. My work suggests practical ways to improve agents and measure them; it does not, by itself, resolve broader policy or extinction-risk questions.

My expectation is conditional on how progress is organized. Diffuse, relatively slow development around open, distillable models and specialized systems could have a strongly positive impact: useful agents would become broadly accessible, evaluations could improve on real-world tasks, and models could support increasingly capable feedback and data-repair loops. But I am less optimistic about a race dominated by a few frontier labs seeking decisive economic or geopolitical advantage. That path concentrates capital, capability, and control, while giving institutions incentives to move faster than their evaluation and governance can support. Many labs are likely to pursue increasingly powerful systems, so concentration does not necessarily produce orderly coordination; it may simply intensify the race. I therefore expect substantial benefits, but not an automatically positive overall outcome. The distribution and pace of progress are central. I prefer the future where capability spreads through open infrastructure and specialized systems, rather than one where a handful of actors compete to control increasingly general intelligence. My technical work gives me concrete reasons to expect useful improvement, but it does not establish a confident net forecast across every social, political, or existential consequence.