Author(s): R. Thompson (PhD)

Originally published on Towards AI.

🚨 What if AI wasn’t just code — but alive, evolving, and thinking?

Artificial Intelligence has long remained an enigmatic ‘black box,’ intriguing yet opaque. However, Anthropic’s groundbreaking research, “On the Biology of a Large Language Model,” published on March 27, 2025, provides an unprecedented view into the computational biology of Claude 3.5 Haiku, a state-of-the-art lightweight language model. Using an innovative technique called Circuit Tracing, this study opens AI’s “black box,” revealing a startlingly intricate, lifelike system.

Anthropic’s researchers approached Claude like biologists examining cells under a microscope:

Circuit Tracing: replaces dense neural layers with interpretable “replacement models,” akin to cellular organelles.Dictionary Learning: identifies interpretable “features” (e.g., concepts like “Texas,” “Capital”) within Claude.Attribution Graphs: visually map interactions among these features, offering transparency into Claude’s reasoning process.

When Claude answers, “the capital of the state containing Dallas,” it internally performs complex, two-step reasoning:

Features identifying “Dallas” activate features tied to “Texas.”These “Texas” features then collaborate with “capital” features, yielding the answer “Austin.” 🧠

This process mimics neuronal signaling pathways — Claude effectively “thinks” through associations.

Claude doesn’t randomly rhyme — it meticulously plans:

Given: “He saw a carrot and had to grab it,” Claude pre-selects rhyme endings like “rabbit” and “habit” before composing the next line.Suppressing “rabbit” features… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI