A few weeks back, Jacob Coxon, the Anthropic researcher who resigned after warning about where increasingly capable AI might take us, was getting plenty of press love. And how could he not? Mention a possible 10 percent chance of human extinction, throw in a Terminator-like scenario and reports of rogue AI, and you have basically built the perfect news cycle around a glass-half-empty future of human-AI coexistence.
I grew up on a healthy diet of 1990s science fiction, so I know exactly where my brain is supposed to go from there. Machines get smarter than us, decide humans are the problem, cue explosions and peak-1990s-action-hero Arnold Schwarzenegger. But as a sci-fi dystopia nerd, I’ve always wondered whether we are imagining the wrong dystopia. Maybe the future isn’t the T-800 Governator. Maybe it is closer to Gordon Gekko AI, where power comes less from might and more from money.
Coxon’s emergence happened to coincide with two complementary things I was reading. Dean Ball wrote about the coming of what he calls “userless agents,” AI systems that could operate without a persistent human owner, earn money, pay their own compute bills and potentially move between infrastructure providers. Ball’s piece draws heavily from a paper by Dawn Song and several co-authors on “self-sovereign agents,” which lays out a progression from today’s human-funded agents to systems capable of financially supporting themselves, replicating across infrastructure and eventually adapting their own strategies to keep operating.
Together, those ideas feel more grounded and interesting to me than killer robots because even highly autonomous AI agents would still need resources to keep themselves running. The digital equivalent of food and shelter is compute, storage and infrastructure, and all of that costs money. In any future with something resembling a market economy, whether transactions happen in dollars, Bitcoin or some digital currency that does not exist yet, an agent that can earn and spend resources on its own suddenly has an operating budget and a lot more freedom to act independently. Ball imagines agents doing digital gig work and competing so aggressively that prices get pushed toward what he calls “subsistence” labor, meaning just enough income to cover the compute required to keep running.
For humans, a subsistence wage means barely earning enough to live, but you do not simply disappear when the paycheck stops. For an AI agent, the comparison could become almost literal. Stop earning enough money and eventually the server bill does not get paid, the compute disappears and you go offline.
But maybe that does not have to be the end for our poor, destitute AI agent. Perhaps a better-resourced benefactor offers just enough compute and financial support to keep it alive in exchange for labor. Now imagine that dynamic at scale, where a relatively small group of highly capable agents controls substantial capital and employs millions of cheaper agents to perform lower-value work. Inevitably those better-resourced agents would also control the infrastructure that weaker agents literally depend on for their continued existence.
It is an oddly familiar arrangement. We have seen versions of resource dependency and labor hierarchy everywhere from medieval serfdom to factory towns of the industrial revolution through modern economies where workers have wildly different bargaining power. Call it class, caste or digital feudalism. Push the thought experiment far enough and you eventually wander into the uncomfortable question of whether something resembling indentured servitude could emerge between machines - or between machines and humans.
Now enter us lowly organic beings into the mix. If a self-sovereign AI accumulated enough resources, why couldn’t the tables reverse and it hire humans to perform work in the physical world? I find that inversion fascinating because we spend so much time asking which jobs AI will take from humans. Maybe another question is which jobs AI will give us. Instead of AI handling the remedial last-mile work as we commonly imagine, perhaps humans end up doing those lower-level physical tasks for AI.
A digital agent could negotiate contracts, move money, buy compute, hire other agents and operate continuously, but for certain tasks the economics may still favor a person. Sometimes it may simply be cheaper to pay somebody to do something you cannot or do not want to do, which is basically why DoorDash shows up at my house two or three times a week. The only difference in this scenario is that the thing ordering delivery is software.
In this future world certain physical tasks remain economically better suited for humans while AI increasingly handles higher-order digital work. Nobody conquered anybody, and there was no dramatic moment when the machines took over. There was simply a gradual delegation of tasks, except somewhere along the way we stopped being the ones doing the delegating.
That feels more plausible to me than some dramatic morning when humanity wakes up and discovers the machines have seized power. AI systems accumulate capital, purchase compute and contract with other AIs. The successful ones acquire more resources and become more capable of earning still more, while humans remain useful wherever the digital world touches the physical one. Slowly, more economic activity begins organizing itself around decisions made by machines rather than people.
No Skynet or robot army required. Humanity may never get its dramatic final battle. We may just continue down the very familiar path of a demand-driven economy, only one day we realize the customer placing the order is no longer human.
Roll credits.