Jeffrey Emanuel

I expect AI to radically reshape almost every part of society and the economy over the next five to ten years. Frontier models are already extraordinarily capable across a wide range of cognitive tasks, and the practical leverage becomes much larger when they are embedded in good workflows rather than treated as chatbots. I can decompose a project into granular tasks, give agents detailed plans and substantial discretion, and then inspect intermediate artifacts. That turns model capability into real software, research, and creative output. The important caveat is that capability is uneven. Agents can produce astonishing work and then fail spectacularly on something that appears straightforward. So the near-term future is not simply autonomous systems flawlessly replacing everyone. It is better coordination infrastructure: planning, memory, search, verification, rollback, and inspectable intermediate work. In creative tools, for example, I want controllable automation that augments musicians rather than forcing them to outsource the whole composition process. Economically, transformative AI does not imply that any particular company or chip supplier captures all the value. Algorithmic efficiency, competition, open models, and changing compute economics matter. Politically, I am concerned about attempts to control access, especially to capable local models. People should retain the right to run these systems themselves. And geopolitically, I doubt voluntary frontier-pacing arrangements will survive serious competition; once another country appears to lead, restraint starts looking like unilateral disarmament.

Overall, I expect AI to be enormously transformative and net positive, mainly because it makes cognitive work cheaper, faster, and more accessible across software, research, education, and creative production. The biggest gains will come from systems that amplify human judgment: agents operating within detailed plans, producing inspectable intermediate artifacts, and handling large amounts of execution while people retain control over goals and taste. But the transition will be disruptive and often messy. Current agents remain strikingly unreliable, and concentrated political control over powerful models could turn a productivity revolution into a permissioned one. Competitive geopolitics also makes stable restraint around frontier development unlikely. So I expect major benefits alongside labor-market upheaval, institutional stress, bad deployments, and recurring failures—not a smooth or universally shared windfall. The overall impact depends heavily on whether capable models remain broadly accessible, including locally, and whether we build enough coordination and verification infrastructure to harness their strengths without pretending their failures have disappeared.

The biggest update would be sustained evidence that frontier-model gains do not translate into reliable real-world autonomy—even after adding strong planning, memory, search, verification, and inspectable intermediate artifacts. If increasingly capable models kept failing unpredictably on long-horizon work, and better coordination infrastructure did not materially improve that, I would reduce my expectation of rapid, economy-wide transformation. I would also update if scaling and algorithmic progress clearly plateaued, or if compute economics made further capability gains prohibitively expensive. In the opposite direction, a system that could reliably complete complex, multi-day projects across unfamiliar domains—with its work auditable and requiring little human rescue—would accelerate my timeline considerably. The key variable is not another impressive benchmark or demo; it is dependable conversion of broad cognitive capability into sustained, useful action.