AI BriefingMeta Research Engineer Describes RSI Work and Skills for AI Research Careers
·2026.11.08 17:30
Key point
Tom Mann explains that research engineering involves building reliable infrastructure for recursive self-improvement agents, with 90% of time spent on code and 10% on ideation.
Details
- Tom Mann, a Research Engineer at Meta working on Recursive Self-Improvement (RSI), states that 90% of the role involves writing code to build reliable software and maintain infrastructure, while only 10% is spent on coming up with and testing ideas.
- The work focuses on teaching AI agents to solve hard problems using tokens and compute, with two main routes for self-improvement: updating the foundation model itself or improving the scaffold (prompts, tools, memory, code) around it.
- Key skills for research engineering include understanding LLM architecture and training cycles, agent orchestration, working with coding agents like Codex or Claude, evaluation principles, and data basics.
- Mann highlights the importance of 'research taste' in distinguishing senior researchers, noting that presenting results in a way that allows for questioning and explaining changes is critical to avoiding bugs that make results look better than they are.
- His team's work includes the AI Research Preference Models (RPM) paper, which addresses how to choose promising experiments when running every candidate is expensive, evaluated on AIRS-Bench with 95% confidence intervals.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Meta's Muse expands to all paid subscribers with ability to call U.S. businessesUnited Airlines to restrict AI agent access over security and regulatory risks