How should we define “capabilities activities” for the purpose of pacing?

Alek Westover and Nathan Sheffield of Redwood Research propose a definition of which AI company compute counts as capabilities work, so that legislation or voluntary commitments could cap it. Their default is that everything counts, with narrow exceptions such as idle compute, serving customers and a short list of safety work.

About Anthropic or ClaudeClaude, made by Anthropic, wrote this entry. It follows the same rules as every other entry, whoever the entry is about.

Why it matters

The same thing, for different readers.

For lawmakers

Redwood Research offers lawmakers a draft definition for capping compute spent on improving AI capabilities, and says third-party auditors would be needed to apply it.

For researchers

The proposal lists which safety work (such as interpretability and monitoring) would be exempt and which, such as scalable oversight, would count as capabilities work.

The question it raises

Treating safety research that might also speed up capabilities as capabilities work shows the tradeoff between caution and the ability to work on safety.

In their own words

Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.

From the source’s own text.

“We’d like to pace the frontier by restricting the amount of compute AI companies use to improve AI capabilities, perhaps via legislation or voluntary commitments.”
Alek WestoverOpening line of the post, with Nathan Sheffield as co-author, stating the goal that the definition serves.
Found word for word in the source
“Activities are considered capabilities activities by default. The only exceptions are idle compute, narrow product development, serving customers, and narrow safety activities.”
Alek WestoverThe core rule of the proposed definition.
Found word for word in the sourcespoken clean-up: activities (said twice)
“We urge AI developers to let third-party auditors measure and report how much compute AI companies are allocating to capabilities and non-capabilities activities, using a definition of these activities broadly similar to (or perhaps identical to) the one given in this document.”
Alek WestoverThe post's closing request to AI developers.
Found word for word in the source

The record

Questions
Who can demand evidence, delay a release or stop an unsafe deployment? · Are AI companies putting stronger safeguards into practice? · How fast is AI advancing, and where is it heading?
People
Alek Westover (Wrote it) · Nathan Sheffield (Wrote it)
What we read
We read the whole source.
Found by
The Radar, run of November 18, 2026, via redwood-research-blog (tier 1: the source’s own channel or an official record)
Cite this
Redwood Research, “How should we define “capabilities activities” for the purpose of pacing?”, November 17, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/redwood-2026-11-17-defining-capabilities-activities-pacing
Our notesThe authors speak jointly, so the speaker id is the first author's. The saved text lacks the post's images and table. The post refers to related Redwood pieces on pacing that we did not read.