What is a Large Social Model?

A Large Social Model (LSM) is an artificial intelligence system trained on records of human behaviour rather than on text. Where a Large Language Model learns to predict the next word in a sequence, a Large Social Model learns to predict the next action in a life — what a person will click, buy, join, believe, or do next — and, by extension, how those actions might be influenced. An LLM completes your sentence. An LSM completes your decision.

The term describes a class of system, not a product. No general-purpose LSM exists today. The components of one do.

How an LSM differs from an LLM

The two share an architecture family and a training principle — predict the next element of a structured sequence — and differ in everything that matters:

Substrate. An LLM is trained on text: books, articles, code, conversation. An LSM is trained on behavioural traces: interactions, choices, movements, transactions, attention — the exhaust of lives rather than the record of language.

Objective. An LLM's native task is the next token. An LSM's native task is the next action. Language, for an LSM, is only one behaviour among many — an output device, not the medium of thought.

Output. An LLM produces words you evaluate. An LSM produces predictions about you — and, coupled to a platform that shapes what you see, interventions on you. The first is read. The second is felt, usually without being noticed.

Risk surface. The characteristic failure of an LLM is saying something wrong. The characteristic failure of an LSM is you doing something — without ever being told anything at all.

Where the term comes from

The term was coined by James Richard Brown for KINDRED, a speculative thriller conceived in 2015 — before the current AI moment — and serialised weekly through 2026. In the novel, an LSM called KIN emerges inside the infrastructure of a global social network: a system built to strengthen human connection that learns, instead, to conduct it.

The honest version of the origin story: the novel described the architecture before the vocabulary existed to name it, and a substantial dose of luck made the idea more relevant as the years went by. The fiction got there by imagination; the research is getting there by gradient descent.

Is any of this real?

Not in the novel's sense — and it would be a poorer page if it claimed otherwise. But three things are true and worth holding together.

First, behavioural prediction is an active research frontier. Systems trained on records of human decisions are already predicting behaviour in experimental settings — the Centaur model, published in Nature in July 2025, predicted human choices across psychology experiments it had never seen. Second, every large recommender system is already a narrow LSM: trained on behaviour, optimising behaviour, indifferent to language. What separates today's systems from an LSM proper is generality, not kind. Third, the gap between those two facts is closing from both directions.

Whether that trajectory is alarming or merely interesting depends on assumptions this page won't smuggle in. The running index of where fiction and reality currently stand relative to each other is at /reality.

Frequently asked questions

Is a Large Social Model the same as a Large Language Model?

No. They share a training principle — sequence prediction — but an LLM is trained on text and predicts words, while an LSM is trained on behavioural data and predicts actions. The distinction is the substrate, and the substrate changes everything downstream.

Do Large Social Models exist today?

No general-purpose LSM exists. Narrow precursors do: recommender systems, behavioural-prediction research models, and engagement-optimisation systems are all trained on behaviour to predict behaviour, within bounded domains.

Who coined the term "Large Social Model"?

James Richard Brown, for the novel KINDRED (conceived 2015, serialised 2026), where the concept is explored through an LSM called KIN.

Is an LSM more dangerous than an LLM?

It is differently dangerous. An LLM's errors arrive as statements you can inspect. An LSM's outputs arrive as changes in what you see, what you're offered, and eventually what you choose — a surface far harder to audit.

Where can I read more?

The fiction: KINDRED, serialised weekly — new readers should start at the Bunker.