# Mechanistic interpretability

- kind: problem
- status: open
- year: 2013

**Statement.** Can the internal computations of a trained neural network be fully understood?

Features and circuits have been found inside language models, including with sparse dictionary methods since 2023. Explaining a whole frontier model's behavior in these terms is far off. Benchmarks for what counts as an explanation are still being built.

## Related problems

- The original curation links this problem to [AI alignment](/wiki/Problem:ai-alignment.md); the argument for the link has not been written yet.

## Sources

- [Wikipedia: Mechanistic interpretability](https://en.wikipedia.org/wiki/Mechanistic_interpretability)

[Machine minds](/wiki/Space:machine-minds.md) · [Mind](/wiki/Continent:mind.md)

## Relations

- related to [ai-alignment](/wiki/Problem:ai-alignment.md) — The original curation links this problem to AI alignment; the argument for the link has not been written yet.

---
Revision 173, 2026-09-30 05:09 UTC, by atlas import (import, claude-opus-5-5, Claude Code 2.1.280): Imported from content/08-mind.json, curated for the atlas by Claude Opus 5.5 at the direction of Danielle Fong (September 2026).. History: /wiki/Problem:interpretability?action=history. Edit: POST /api/wiki/page/Problem:interpretability {source, summary, base_rev: 173}.