Epoch AI releases a taxonomy of 60+ R&D tasks to track progress toward automating AI development

Our taxonomy is a first draft, so we're very open to feedback and collaboration. If you're a researcher who thinks we've missed something in our classification, or would use it in your own work, we'd be excited to talk to you.

We propose an O*NET for AI R&D that includes 60+ tasks (with examples) across the research cycle. The taxonomy covers six categories, from deciding what to work on to communicating results.

Each task is rated 0–5 for how much we think AI automates it today.

Economists often study labor markets using the O*NET database, which breaks ~1000 occupations into tasks. But these tasks are too coarse-grained to track automation in AI R&D specifically, even in occupations closest to “AI researcher”.

With this breakdown, we can better understand what current forecasts and benchmarks are failing to track and motivate additional work on what’s being missed. It also helps anchor researcher surveys and AI usage studies on a shared vocabulary.

This week’s Gradient Update was written by @datagenproc, @joemkwon, and @ansonwhho.

All Gradient Updates are informal, opinionated analyses that represent the views of individual authors, not Epoch AI as a whole.

Read the full essay here: https://epochai.substack.com/p/toward-an-onet-for-ai-r-and-d

@EpochAIResearch @datagenproc @joemkwon @ansonwhho Decomposing frontier R&D into 60+ granular tasks is the right first step for measuring exposure. The real gap is mapping which of those tasks show declining wage premia or rising substitutability over time.