Billion-parameter theories recast AI complexity science

This is a sharp AI-era update to the old “map versus territory” debate: if elegant laws fail on complex systems, giant learned models might be the first theories that are actually usable.

• –The core claim is not just “LLMs are useful,” but that scale itself may be necessary because some domains do not compress into napkin-sized equations
• –The essay draws a useful distinction between compact architectures with broad reach and huge trained weights that stay domain-specific
• –Its most interesting move is treating mechanistic interpretability as a scientific method for complexity, where researchers study the learned model to recover structure from messy systems
• –For AI developers, the piece is less a product announcement than a worldview shift toward simulation-first, probabilistic modeling over human-readable causal stories