I am quite happy to have AI reconcile a bank account, decipher a government regulation or tell me why the dishwasher is making that noise. I am less interested in having it decide what I should believe or what I should write.
The important dividing line may be between augmentation and abdication.
Porter’s stronger criticism is about power. Zuckerberg worries about concentrating AI in too few hands while sitting atop one of the handful of companies capable of building frontier systems in the first place. He wants distributed intelligence, but creating that intelligence requires enormous concentrations of capital, chips, electricity and technical talent.
That contradiction becomes even more interesting when we ask where AI will actually live.
I have been thinking about that because of the giant data centers now being built. OpenAI, for example, is planning a complex in Georgia with a contracted electrical supply of 3.2 gigawatts.³ Training frontier AI increasingly resembles heavy industry.
But creating intelligence and using intelligence are two different jobs.
I spent part of the 1970s working for Digital Equipment Corporation and Data General, when serious computing still belonged mostly to institutions. DEC and DG helped pull computing out of the mainframe room. Their minicomputers did not kill the mainframe. They simply made it unnecessary to send every job to one. Then personal computers moved even more computing onto people’s desks.
The mainframe survived. It just stopped owning everything.
AI may be heading in the same direction. The giant data centers may increasingly become model foundries: places where frontier intelligence is created, tested and improved. Smaller versions can then move outward onto laptops, phones, cars and private business systems. Meta has already described using very large models to generate synthetic data and distill capability into smaller ones.⁴ Apple and Microsoft are pursuing their own versions of edge AI.⁵,⁶
If that continues, we may end up with a curious AI economy in which the creation of intelligence becomes more centralized while the use of intelligence becomes less centralized.
The location of computation is not the same thing as the location of power.
There is a profound difference between a “personal AI” that lives in Meta’s data center and one that runs on my computer. If it lives in Meta’s data center, Meta owns the hardware, controls the model, establishes the rules and decides when the system changes. It may know me very well, but that does not mean it belongs to me.
If a capable AI actually resides on my computer, the relationship could be very different. My files can stay local. I may be able to choose among models, replace one I dislike, keep an older version or have several systems check one another’s work.
But even that does not necessarily mean I own it.
A model can sit on my laptop and still be encrypted, license-restricted, dependent on proprietary services or designed to require periodic permission from its creator. The telephone sitting in your house did not mean you owned the telephone network.
There are really three different kinds of power involved in AI. There is the power to create intelligence, which may remain concentrated. There is the power to control its distribution — to decide who can copy, modify, license and run it. And there is the power to use intelligence, which could become highly decentralized.
That middle layer may matter most.
Suppose Meta spends tens of billions creating a frontier model and a smaller descendant eventually runs on my laptop. Do I meaningfully control it if Meta can still dictate what it does, revoke access or determine which versions I am allowed to use?