Sean Linehan published an essay on March 9 that reframes large language models as something other than sophisticated text predictors: a new class of scientific theory.
The argument starts with a simple observation about classical science. Its greatest strength — F=ma, E=mc², the Lotka-Volterra equations — is also its limitation. Compact theories work brilliantly on simple systems. But complex ones — poverty cycles, financial contagion, drug addiction, climate feedback loops — have resisted every equation researchers have tried to fit to them. Linehan's claim isn't that we've been doing it wrong. It's that we've been using the wrong tool. Some systems genuinely require billions of parameters to model. For the first time, we have tools capable of holding theories that large.
The more provocative half of the essay concerns the transformer architecture specifically. Linehan argues that trained model weights aren't the theory — they're domain-specific, large, and opaque. The transformer itself, the small elegant structure underneath, is the compact universal meta-theory. His parallel is Chomsky's universal grammar: correct in principle, too abstract to be operationally useful on its own. The transformer is that compact structure, just one abstraction level higher than complexity researchers expected when they founded the Santa Fe Institute in 1984.
What the SFI found — after decades of work — was a set of descriptions. Recurring features of complex systems: emergence, power laws, feedback loops. What they couldn't produce was prescriptions. Linehan's implicit claim is that transformer-based models, trained at scale on the right domains, do exactly what SFI's theoretical frameworks couldn't.
For people building AI agents, the framing is worth taking seriously. If agents trained on complex real-world domains are empirical models of systems that resisted every prior attempt at formal theory, then the pipelines used to build and fine-tune them are instruments of scientific inquiry — not just engineering. Mechanistic interpretability, the discipline of reverse-engineering what trained weights have actually learned, becomes the most interesting place to look for compressible, operationalizable knowledge about how complex systems actually work.
Linehan reaches for the usual historical analogy — blacksmiths before metallurgy, cathedral builders before structural engineering — to make the point that practitioners can operate successfully ahead of theory. It's a familiar move, and he knows it. The honest version of his argument is that we don't yet know whether billion-parameter models are science in any rigorous sense. What the essay does is give the field a better question to argue about.