Specialist Models for Swarm Intelligence

The future of AI is not one giant model that does everything. It is thousands of specialists — a designer, a researcher, a planner, a modeller — each trained to be the best in the world at one job, working together as a swarm.

Today the frontier is monolithic. General models are built to be deployed everywhere at once, and every additional point of capability costs more than the last: bigger runs, more data, more compute, for gains that flatten just as workloads turn agentic and start running for hours. Generality bought this way is real. It is also expensive, and only a handful of labs can afford it.

We take the opposite path. Hold the job fixed — interior design, floor-plan reasoning, furniture selection, shopping research, literature review — and a focused model trained on rich human-generated data, sharpened with reinforcement learning against that one domain, and evaluated continuously will beat a generalist many times its size at that job, at a fraction of the serving cost. Human data is the raw material: how people actually describe, judge, and want things is what a specialist is made of.

Specialists compose. Each agent in a swarm is independently optimised and replicates to meet demand, and the swarm collectively exhibits capabilities no member has — the same way civilisation outperforms any individual human. Generality is a property of the system, not of any single model. Read why AGI is swarm intelligence.

The system should also improve itself. Our training loops generate hypotheses; modify their own prompts, code, and procedures; evaluate the outcome; and keep what works. Researchers set the objectives and constraints. The system explores the search space.

This is what we build at Logarithms: the research behind frontier specialist models, and the products they power in our studio — DecorLM for interior design, Scaffold for 3D creation, Cobots for dependable agentic work.