AGI is Swarm Intelligence

Every AGI roadmap carries a quiet assumption: that general intelligence will arrive as a single artifact. One model, one context window, one mind that does everything. Scale it far enough, the thinking goes, and generality falls out the other side.

We think that assumption is wrong — and the reason is visible in the curve every lab is climbing.

You don’t beat a logarithm by pushing harder

Scaling laws are the most reliable result in modern AI, and they are also a warning: capability grows roughly logarithmically with compute. Each additional point of intelligence costs exponentially more than the last. The monolithic path to AGI is a race up an ever-steepening hill — possible, but punishing, and open only to the handful of organisations that can afford the climb.

That curve is why this lab is named Logarithms. We take it seriously enough to build around it. You do not beat a logarithm by pushing harder along it; you beat it by changing which curve you are on. A thousand small specialists, each riding the steep early section of its own learning curve, buy more capability per unit of compute than one giant model grinding along the flat end of its.

Generality is a property of systems

The most general intelligence we know of is not a brain. It is civilisation. No individual human can grow food, fabricate a chip, fly an aircraft, and argue a case in court — yet civilisation does all of these at once, continuously, at planetary scale. It manages this because it is made of billions of specialists coordinating through language, markets, and institutions. The generality lives in the system. No single member contains it.

Machine intelligence, we believe, will recapitulate this structure. A designer agent. A researcher agent. A planner agent. A modeller agent. A rendering agent. Each independently optimised for its task. Each able to replicate itself to meet demand. None of them general — and the swarm, collectively, exhibiting capabilities greater than any individual model.

Depth beats breadth at the frontier

General-purpose models are useful. Specialist models are exceptional. Hold the job fixed — floor-plan reasoning, furniture selection, scientific literature review — and a model trained on rich human-generated data, sharpened with reinforcement learning against that domain, and evaluated continuously on that objective will beat a generalist many times its size. Not at everything; at the job. That is the point. In a swarm, at the job is all any member needs.

This is also how human expertise works. Civilisation did not get better by making every person slightly more general. It got better by letting people go deep, and building the coordination that lets depth compose.

The loop that closes the argument

A swarm of specialists has one property a monolith cannot match: it can be improved piece by piece, and it can improve itself. Models should not only perform work — they should improve the systems that produce them. Our systems generate hypotheses; modify their own code, prompts, and training procedures; evaluate the outcome; retain what works; and repeat. Human researchers define objectives and constraints. The system explores the search space.

This is the loop our three research pillars form. Specialist models create better individual intelligences. The swarm composes them into something larger. Self-learning improves both — continuously, and increasingly without us. Run the loop long enough and research itself becomes something the system does, not something done to it.

What AGI will actually look like

AGI will not be a moment when one model wakes up. It will be an ecosystem quietly crossing a threshold: a population of specialist intelligences whose collective competence has become general, and whose improvement has become self-sustaining. No single neuron understands. No single human contains civilisation. No single model will contain AGI.

Generality is a property of the swarm. That is what we are building.

— Logarithms, London