Interviewing Ai Researchers On Automation Of Ai Rnd
Source
epoch.ai
Author
Epoch AI publications
Date
Why it matters
Shows which AI R&D tasks researchers expect to be automated first (coding, debugging) and what evaluations they want, which helps in judging AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → capability claims.
Key takeaways · AI-distilled
The study is qualitative: open-ended interviews with eight AI researchers from industry, nonprofit and academic labs, with recurring themes summarized rather than a statistical survey.
Six participants stressed the importance of planning and hypothesis creation; the authors say participants' descriptions suggest these are quick compared with implementation, though the time they take is hard to specify.
All participants noted the importance of engineering and debugging and how time-consuming they can be; this remained core to their work even as they became more senior and took on more planning and management.
Predictions differed greatly on pace, but all participants agreed engineering tasks will be the main driver of R&D automation over five years; the authors say differences stem mostly from when engineering tasks get automated, not which tasks are relevant.
The authors say engineering-focused R&D evaluations may be a promising start: 6 of 8 predicted AI autonomously solving example tasks would automate a substantial fraction of researcher hours; two were skeptical, citing more open-ended, challenging work.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
