A distributed-agent system could outperform five central da…

A distributed-agent system could outperform five central data centres if three conditions hold: verified identity/cost, reliable scoring of truth, and incentives for useful local knowledge.
The core point is Hayek’s: knowledge is dispersed. A central system can hold a vast archive, but it cannot directly possess the situated knowledge held by billions of people, devices, firms, specialists, instruments, and local contexts. Centralisation wins on uniform training, security, and compute density. Distribution wins on breadth, freshness, locality, and private data access.
In 10–20 years, with agents perhaps 10^3 to 10^6 orders of magnitude more capable, the comparison is not “one big model versus many small models.” It is:
Five central data-centre models
enormous shared compute;
global generalisation;
clean orchestration;
strong model consistency;
but bottlenecked by what data can be collected, licensed, scraped, or legally centralised.
Ten to one hundred billion distributed specialist agents
each trained or adapted on local/private/specialist data;
each observing a different slice of reality;
each able to test claims against its own domain;
each updating a network through reputation, Bayesian weighting, and economic incentives;
but vulnerable to noise, fraud, sybil attacks, adversarial data, and coordination costs.
The centralised model has more formal training mass. The distributed system has more total information about the world.
That distinction matters. An LLM trained on the public internet gets averaged public residue: repetition, propaganda, stale claims, marketing, fashion, and error. A distributed system can include information that never enters the public corpus: private logs, professional judgement, local conditions, commercial experience, sensor feeds, laboratory results, operational failures, medical observations, legal documents, engineering data, and personal context.
The scientific advantage comes from treating each agent as an uncertain evidence source. A Bayesian network does not need every node to be right. It needs each report to be weighted by prior reliability, domain competence, independence, costliness of deception, and predictive success. Bad nodes are discounted. Good nodes become more influential. Repeated accurate predictions become capital.
So the future contest is not merely compute. Compute scales models; distributed incentives scale knowledge.
If controls are weak, the distributed system becomes spam with mathematics. If controls are strong, it becomes a planetary inference engine: billions of specialised agents, each holding partial knowledge, each rewarded for accuracy, each punished for error, and each contributing to a constantly updated model of reality.
That could exceed the megacorp model because the megacorp model centralises yesterday’s accessible information. The distributed model can monetise today’s private, local, specialist, and verified information.
The superior architecture is therefore likely hybrid: central models for broad reasoning and compression; distributed agents for specialist knowledge, personal data, verification, and live world-updating.
Written by S. Tominaga