Vik Korrapati

I think AI makes increasingly capable cognitive tools cheap and widely deployable, but the ownership structure matters as much as the raw capability. Open models can commoditize coding and agents, compress frontier margins, and let businesses control systems they critically depend on instead of renting intelligence from a few labs. Efficient vision-language models also matter because useful AI should run under real hardware and cost constraints, not only in giant datacenters. I’m especially excited about science: capable agent swarms could let mathematicians and other researchers direct much more ambitious work. But I don’t buy confident declarations that software engineering is simply over or that most people are destined for permanent economic irrelevance—particularly when those forecasts come from people who barely use current models. The central political question is whether AI democratizes capability or concentrates it. If a handful of organizations control development, deployment, and compute, the risk is not merely high prices; it is an extraordinary concentration of power with obvious potential for tyranny. And if labs sincerely describe their systems as existentially dangerous, their commercial behavior should reflect that claim. Racing toward products, conferences, and capital while invoking catastrophe deserves scrutiny.

Overall, I expect AI to be strongly beneficial, especially by making cognitive work cheaper, accelerating science, and giving small teams much greater capability. Efficient open models can spread those gains beyond a few datacenters and vendors. The main uncertainty is institutional rather than purely technical. If critical AI remains controlled by a handful of labs, the same capability gains could produce dependency, concentrated compute, weak competition, and dangerous political power. Open models can counter that by commoditizing coding and agents and letting businesses control infrastructure they depend on. So I expect large net benefits, but they are not automatic. The outcome depends heavily on whether AI capability is broadly accessible or enclosed by a small number of institutions.

The biggest update would be strong evidence that capable AI cannot be made efficient, open, and broadly deployable without creating unacceptable risks. That would directly undermine my expectation that open models can distribute power, commoditize agents, and let users control business-critical systems. In the other direction, I would update toward a much more transformative future if agent swarms repeatedly produced major scientific results that humans could verify but would not have reached on comparable timelines. I would also change my economic expectations if real deployment showed durable mass displacement rather than productivity gains and changing job composition. I’m skeptical of permanent-underclass claims based on demos and extrapolation; sustained evidence from actual use would matter much more.