Will Manidis

I think AI can produce extraordinary gains, but our future will be determined less by model capability than by whether institutions can turn that capability into useful, legitimate deployment. Burning tokens, orchestrating elaborate agents, and generating impressive-looking artifacts are not themselves productivity. Models may already be capable enough for many valuable applications; the bottlenecks are capital formation, organizational change, incentives, and the unglamorous work of integrating them into the economy. The gains will also be politically uneven. Automation can fracture coalitions between labor and capital, threaten economies built around exported services, and concentrate value unless businesses actually pass falling costs to customers and affected communities share in the benefits. Technical leadership alone is not necessarily a durable moat. A company that treats AI as a machine for consuming every available customer margin may lose to one that turns efficiency into lower prices and better service. I’m also worried about the information environment. Synthetic media is cheap, scalable, and useful to attackers; it can flood public spaces with content-shaped objects while making authenticity harder to establish. The likely response is not simply better filtering. People may retreat toward smaller private spaces where identity and meaningful human participation are easier to trust. So I see enormous productive potential, but nothing automatic about a good outcome. Spending is not adoption, output volume is not value, and corporate reassurance is not a social contract. We need institutions capable of absorbing and distributing risk—closer to insurance than congressional micromanagement—alongside businesses that create real customer surplus.

I expect a large positive productivity shock paired with a much messier political and social transition. AI can lower costs, improve services, and make previously uneconomic work possible. But those benefits do not distribute themselves. The default path could concentrate gains among owners of capital while displacing workers, destabilizing service-export economies, and filling public platforms with cheap synthetic material that erodes trust. So the overall impact depends less on another increment of model capability than on deployment and distribution: financing adoption, changing institutions, passing efficiency gains to customers, and creating credible ways to absorb risk. My expectation is neither simple abundance nor simple catastrophe. It is real economic value arriving through institutions that are poorly prepared to allocate it, producing substantial gains alongside serious political conflict and a degraded public information environment.

The biggest update would be evidence that the deployment bottleneck is much weaker than I think: organizations rapidly turning existing model capability into durable productivity gains, passing those gains to customers and workers, without requiring heroic capital formation or institutional redesign. That would make me substantially more optimistic about both the scale and distribution of AI’s benefits. In the other direction, convincing evidence that synthetic media is causing a persistent collapse of trust—not merely more spam, but the practical failure of open public platforms as shared informational spaces—would make me much more pessimistic. The crucial variables are not benchmark scores or token consumption. They are whether capability becomes useful work, whether surplus is broadly distributed, and whether institutions can preserve meaningful human participation while absorbing the disruption.

The most shaping observation is the widening gap between capability and useful deployment. I keep seeing systems that can generate astonishing volumes of plausible work, while organizations struggle to convert that output into durable productivity. Token consumption, elaborate agent architectures, and polished artifacts can become status performances—activity mistaken for value. That pushed me toward viewing capital formation, institutional change, and integration as more decisive than another benchmark improvement. The parallel observation is what synthetic content does to public spaces. When a feed becomes dominated by generated material, the problem is not merely that some posts are low quality. The space itself becomes harder to trust: authorship is uncertain, participation feels less meaningful, and attackers benefit from cheap scale. That suggests a future where people retreat into smaller private communities rather than rely on a shared public internet. Together, those observations make me optimistic about AI’s productive potential but skeptical that capability or spending alone will produce a socially valuable outcome.