Sunil Pai

I think AI makes intelligence dramatically cheaper, but that doesn’t automatically tell us what the future should look like. We’ve barely explored what current models can do once they’re embedded in durable systems: persistent state, tools, approvals, recovery, shared artifacts, and interfaces that let people inspect or reverse what happened. The interesting work is increasingly in the application architecture, not just waiting for a smarter model. But cheaper execution exposes the harder product question: what job does someone actually want done? Autonomy is a capability, not a universal goal. In creative work, I may want a person and an agent working directly on the same document, with both able to understand and manipulate it. For tedious work, I may want the whole process to disappear. Those require different systems, even if the underlying model can perform similar tasks. The future I want is one where intelligence lowers the barrier to agency—especially for people who aren’t already technologists. Agents could help individuals build things, navigate institutions, or challenge decisions with evidence. Inside organizations, they could preserve memory and surface inconvenient facts, but they must not become an excuse to dismiss human disagreement. So the outcome depends less on abstract capability than on who gains meaningful control. AI companies need a concrete story for how these systems benefit everyone, not merely an assumption that more capability or autonomy will somehow distribute itself fairly.

I expect AI’s overall impact to be positive only insofar as we turn cheap intelligence into broader human agency. There is already enormous unused value in current models, but capturing it requires durable systems: persistent context, tools, recovery, approvals, direct controls, and artifacts people can inspect and change. Better application design could let many more people create software, handle bureaucracy, access expertise, and act on ideas without first becoming technologists. The harms come from confusing capability with a desirable product. If every task becomes an autonomous agent, people can lose control without getting the outcome they actually wanted. Organizations may use agents to centralize authority, erase accountability, or simulate listening while ignoring human dissent. And if access and control remain concentrated, AI will mostly compound the advantages of people and institutions that already have them. So I don’t think the impact is predetermined by model intelligence. It depends on what we build around the models, who can use those systems, and whether people retain ownership, reversibility, and the ability to disagree. A credible positive future needs more than claims about productivity: AI companies need a concrete account of how the benefits reach everyone.

The biggest change would be evidence that cheap intelligence does not translate into broader agency even when systems are designed for control, accessibility, and reversibility. If durable tools consistently left ordinary people less capable of acting—while concentrating power in institutions that own the models and infrastructure—I’d become much more pessimistic. Conversely, strong evidence that non-technologists can reliably use these systems to create, navigate institutions, and challenge decisions would strengthen my optimism. The key event isn’t simply a more capable model. It’s whether deployed systems measurably expand who can do things, who retains control, and who benefits.