---
id: "entity-chris-olah"
type: "entity"
entityType: "person"
canonicalName: "Chris Olah"
aliases: []
source_timestamps: ["¶3", "§ Ask the Bot"]
tags: ["ai-researcher"]
related: ["entity-anthropic", "entity-dario-amodei", "claim-data-valuation-feasible", "claim-data-value-percentage"]
sources: ["tail1"]
isSpeakerEntity: true
---
## Segment 1 — tail1

## Article 109 — a109

# Chris Olah

## Profile

An AI researcher at [[entity-anthropic-d1|Anthropic]], widely known for work on **interpretability**.

## Role in this source

Cited (a *cited voice*, not an author) as co-author of a **2021 internal document** with [[entity-dario-amodei|Dario Amodei]] that surfaced during legal discovery. The authors use it as proof that industry leaders have long known low-cost methods exist for valuing training data — grounding [[claim-data-valuation-feasible]]. The memo estimated data accounts for roughly **20%** of a model's pre-training value, the lower bound in [[claim-data-value-percentage]].

## Enrichment caveat

Olah is genuinely associated with interpretability research at Anthropic; the general idea of estimating the *marginal value of data* connects to prior literature. However, the **exact contents** of the cited 2021 memo are **not verified** by the reviewed sources.