Seconds

I expect AI to create enormous practical value, especially by making software cheap enough for tiny audiences—or even one person. Things too niche to justify a conventional development team can now be built, tested and revised quickly. I’ve seen repeated attempts at personally useful tools cross from “cute demo” into genuinely useful territory as models improved. Predictions that capability had already plateaued have aged badly. But this does not mean instant abundance. Intelligence can produce plans, code, translations and designs at extraordinary speed; it cannot immediately manufacture more housing, food, energy infrastructure or physical commodities. The digital world can move much faster than the physical one, so strong AI and stubborn scarcity can coexist. The other constraint is human judgment. Automation still works best when someone states the intent precisely, reads the output, understands the model’s assumptions and catches errors that are locally plausible but contextually wrong. Faster generation moves more of the bottleneck toward taste and review—it does not abolish the need to know what you asked for. Security boundaries also need to be concrete rather than ceremonial. An “air gap” is not magical protection if a system can observe exploitable digital inputs or use side channels. Agents may be better than humans at finding leverage in those interfaces. So my default future is powerful, widely accessible AI producing huge consumer benefits, alongside very ordinary bottlenecks: physical constraints, access controls, ambiguous instructions and people failing to inspect what the machine actually did.