# Chris Olah

> ~1990– · AI Researcher
>
> **Recorded contribution:** Neural network interpretability; distill.pub; Anthropic researcher; feature visualization

## How to use this dossier

Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory.

## 1. Historical orientation

AI researcher Chris Olah is known for neural-network interpretability, unusually clear visual explanations, and institution-building around model transparency. He co-founded the online research journal Distill, worked at Google Brain and OpenAI, and co-founded Anthropic. His work includes feature visualization, circuits-style analysis, and later sparse-autoencoder investigations of internal model features. Olah's interpretability work treats explanation as reverse engineering: optimize inputs that activate a feature, inspect how features compose across layers, and build diagrams whose claims can be tested against model behavior. Distill joined this method to interactive publishing so exposition itself became part of the research artifact.

## 2. The problem inherited

Neural networks achieved high performance while their internal representations remained difficult to inspect, making failures, abstractions, and safety claims hard to understand mechanistically.

## 3. The central contribution

Olah and collaborators developed visual and circuits-oriented methods for interpreting learned features and made interactive explanation itself a serious research medium.

## 4. Reconstruct the mechanism

1. Select an internal activation, neuron, direction, or circuit whose behavior needs explanation.
2. Optimize or search for inputs that strongly activate it and inspect natural examples.
3. Trace how upstream features combine and how downstream computation uses the activation.
4. Intervene by ablating or modifying the feature and test whether the predicted behavior changes.

## 5. What changed downstream

- Feature visualization and circuits work influenced mechanistic interpretability research.
- Distill demonstrated that interactive, open exposition can communicate mechanisms more effectively than static prose alone.
- Feature visualization and circuits research supplied a vocabulary for mechanistic interpretability and inspired laboratories to study internal representations rather than evaluate models only through outputs.

## 6. Attribution, limits, and uncertainty

- Interpretability projects are collaborative, and vivid visualizations can invite overinterpretation or cherry-picking.
- A human-readable feature is not necessarily causal, complete, monosemantic, stable, or sufficient to certify safety.
- A compelling visualization can be selected or anthropomorphized, neurons may be polysemantic, discovered circuits may not be complete or causal, and analysis of small models may not scale straightforwardly.

## 7. Reconstruction lab

Train a small image network, optimize an input for one channel, find natural activating examples, ablate the channel, and compare the causal change with your visual interpretation. Intervene on the proposed feature or connection and predict a specific output change; reject the explanation if the causal test does not match the visual story.

## 8. Evidence trail

- [Feature Visualization](https://distill.pub/2017/feature-visualization/) — Distill
- [Circuits thread](https://distill.pub/2020/circuits/) — Distill
- [Chris Olah](https://colah.github.io/) — Chris Olah

---

*Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*