The Santa Fe Institute was founded in 1984 to crack complex systems — poverty, climate, immune response, financial markets. Forty years later, it produced rich vocabulary and no operable tools. Sean Linehan's essay "Billion-Parameter Theories," published March 9, argues that wasn't intellectual failure. The researchers were solving the right problem with the wrong medium.
Linehan draws a line between "complicated" systems and "complex" ones. Complicated systems decompose. Jet engines, orbital mechanics — you write a terse equation and predict behavior. Complex systems resist this: feedback loops, emergent behavior, and reflexivity defeat elegant mathematics. The Enlightenment toolkit, from Newton through Shannon, was built for the former.
His claim is that today's billion-parameter models are <a href="/news/2026-03-14-against-vibes-evaluating-generative-model-utility">the first tools actually suited to the latter</a>. Not because they're elegant, but because the most compressed representation of a genuinely complex system may simply be very large. Drawing on David Deutsch's concept of explanatory "reach," Linehan proposes a two-layer framework: a compact, general-purpose transformer architecture sitting beneath massive domain-specific trained weights. Andrej Karpathy's nanoGPT illustrates the architecture layer — minimal and universal. The weights are irreducibly large by necessity, not sloppiness.
The essay's sharpest reframe is on mechanistic interpretability. Linehan positions it not as a safety discipline but as the emerging science of complexity — studying trained models the way biologists study specimens, extracting compressible truths about the underlying systems they encode.
The post surfaced on Hacker News shortly after publication and drew substantive pushback. User wavemode raised the overfitting problem: billion-parameter models can fit training data perfectly while failing to generalize, and scale doesn't solve that. User b450 pointed to Waymo's World Model — where driving data produces simulation-capable models that feed back into training — and flagged a subtler issue. Complex systems like poverty operate simultaneously across economic, psychological, ecological, and political levels. Choices about which data to collect, and how to operationalize qualitative phenomena, introduce distortions before the model sees a single example. A third commenter suggested that dominant conceptual vectors might still be recoverable through PCA-like techniques, offering a possible route to compact explanations from large models.
The interpretability gap Linehan identifies has concrete stakes. Domain-specific agents are already being deployed in <a href="/news/2026-03-14-lancet-psychiatry-ai-associated-delusions-study">healthcare</a>, climate modeling, and legal systems, trained on specialized corpora and trusted to make or recommend consequential decisions. The difference between a model that encodes genuine causal structure and one that learned spurious correlations at scale is not visible from the outside. Until practitioners can read what is inside trained weights, Linehan's two-layer theory is also a two-layer black box — and the agents built on top of it will carry that uncertainty with them.