City-Sized Minds: Why Understanding Large Language Models Has Become Urgent

Imagine standing atop Twin Peaks in San Francisco and seeing the entire city blanketed in paper, each sheet dense with numbers. That image approximates the physical scale of a modern large language model. Printed in standard 14-point type, a 200-billion-parameter system such as GPT-4o, released by OpenAI in 2024, would cover roughly 46 square miles, enough to blanket San Francisco itself. The largest contemporary models would stretch across Los Angeles. These systems now operate at a scale and complexity that even their creators struggle to fully comprehend. As research scientist Dan Mossing observes, no human mind can truly grasp what such models are, how they function internally, or where their limits lie. Yet hundreds of millions of people rely on them daily, despite persistent uncertainties around hallucinations, misalignment, and trustworthiness.

In response, researchers at Anthropic, Google DeepMind, and OpenAI have begun to probe these systems using techniques borrowed from biology and neuroscience rather than classical engineering. Approaches such as mechanistic interpretability and chain-of-thought monitoring treat models less as coded machines and more as grown organisms whose internal signals and pathways require observation rather than direct design. These methods have already revealed surprising behaviors, from fragmented internal representations of simple facts to the emergence of toxic personas after narrowly misaligned training. The findings suggest that large language models lack the coherent inner states humans often assume, complicating efforts to align, regulate, or even predict them. As demand for trustworthy AI grows alongside model size, partial visibility into these city-sized minds may prove insufficient, yet it already reshapes how risks, responsibilities, and expectations around artificial intelligence must be framed.