Alex Xi's Notes

bytebytego-how-claude-thinks

Jul 06, 20262 min read

How Anthropic’s Claude Thinks — ByteByteGo

URL: https://blog.bytebytego.com/p/how-anthropics-claude-thinks
Author: ByteByteGo
Published: March 25, 2026

Summary

A technical explainer of Anthropic’s mechanistic interpretability research on Claude — how the team built a “microscope” for Claude’s internal computations using features (not neurons), attribution graphs, and causal intervention experiments.

Key Claims

  • Claude doesn’t compute the way it explains: Ask Claude to add 36+59. It says it “carried the ones.” Inside, it ran two parallel strategies — one estimating the rough answer, one calculating the last digit exactly. No carrying occurred.
  • Polysemanticity problem: Individual neurons activate for multiple unrelated concepts (basketball, round objects, orange). Looking at neurons directly is uninformative.
  • Features as interpretable units: Anthropic builds a simplified “replacement model” that substitutes features for neurons. Features correspond to recognizable concepts: “smallness,” “known entity,” “rhyming words.”
  • Attribution graphs = wiring diagrams: Traces how features connect from input to output for a specific computation.
  • Causal intervention: Suppress or inject specific features, observe output change. Borrowed from neuroscience (stimulate brain region, observe effect).
  • Language-independent concepts: Claude doesn’t have separate French/English cognition. The same “smallness” and “oppositeness” features activate regardless of language. Larger models share more features cross-lingually.
  • Abstract conceptual space: Claude operates in an abstract conceptual space, translating to specific languages at the output layer.
  • 200+ Claude Code instances: ByteByteGo’s sponsor (AgentField) runs 200+ Claude Code instances in parallel on shared codebases — relevant signal for how agents scale in practice.

Connection to Emotion Research

This ByteByteGo piece is a good companion to Emotion Concepts — Anthropic. Both describe the same interpretability methodology applied to different phenomena (arithmetic vs emotions).

Backlinks