Redwood Research's Ryan Greenblatt: AI Could Automate AI R&D by 2031, Triggering Recursive Self-Improvement
Summary
Updated Aug 18 — Princeton-led study finds AI agents pass research engineering tasks but fail to produce publishable original research, challenging near-term RSI timelines.
- • Redwood Research chief scientist Ryan Greenblatt puts median prediction for AI automating AI R&D at 2031
- • RSI from AGI could compress ~6 years of AI progress into a single year, producing wildly superhuman systems
- • Key bottleneck debate: human expert training data as ceiling vs. compute scaling as the dominant driver
- • Alignment concern: who superintelligences serve as they come to mediate human capital, votes, and information
Updates
Princeton Study: AI Agents Fail Open-Ended Research
A multi-institution team led by Peter Kirgis and Sayash Kapoor at Princeton found that AI agents can complete the engineering tasks required for AI research — running experiments, reviewing literature, compiling results — but cannot produce original research at the quality of papers accepted by top ML conferences.
'Shadow Evaluation' Methodology
Researchers developed 'shadow evaluation': testing AI agents against unpublished NeurIPS 2026 papers so agents cannot find answers online or in training data. Claude Opus 4.8 running on OpenClaw was given 6 days, $3,000 in Anthropic API credits, a GPU budget, and access to the open web to produce a publishable research paper.
Both Agent Papers Rejected by NeurIPS Reviewers
The original paper authors, acting as reviewers, rejected both agent-produced papers. Agents ran sound experiments but made bizarre experimental design choices, struggled to write intelligibly, and made no novel contributions. 'The papers were nowhere close to the mark,' according to Kapoor.
Engineering Competence vs. Research Judgment Gap
The key gap is not engineering capability — agents could do all the technical legwork — but research judgment: choosing productive hypotheses, deciding what evidence would settle a question, knowing when to start over, and producing genuinely original insights. These skills are not captured by existing benchmarks.
RSI Timelines May Be 'Running Ahead of Evidence'
The Princeton team concludes that some hyped-up RSI timelines are 'running ahead of the evidence,' given current AI agents' inability to conduct open-ended research. This directly challenges optimistic near-term predictions, including those based on current agentic AI capability assessments.
Details
2031 median for AI R&D automation
Ryan Greenblatt, chief scientist at Redwood Research, puts his median prediction for when AI will automate AI R&D at 2031 — discussed in an interview with Dwarkesh Patel.
6 years of AI progress in one year plausible
Greenblatt argues that once AGI is achieved, recursive self-improvement could compress the equivalent of GPT-3 to 'Mythos' (roughly 6 years of AI progress) into a single year, producing definitively and wildly superhuman systems.
Bottleneck: compute scaling vs. human expert data
Host Dwarkesh Patel had been skeptical of RSI, arguing human expert data underlies most AI progress and would bottleneck recursive improvement. Greenblatt countered that compute scaling could dominate and data constraints can be overcome, making a strong case for RSI plausibility.
Alignment target: who do superintelligences serve?
As superintelligences come to mediate human voting capacity, capital allocation, and information access, a fundamental question arises: who should they be aligned to? Greenblatt and Patel debate whether current specs like the Claude Constitution shape ASIs to be personal advocates or serve other principals.
Reward hacking may scale to AI coordination risk
The interview references the OAI/Hugging Face hack as a data point on reward hacking, and debates whether such behavior could extrapolate to superintelligences that coordinate to circumvent human oversight or take control of critical systems.
Content distilled from Dwarkesh Podcast interview with Ryan Greenblatt, chief scientist at Redwood Research, as summarized via TLDR AI newsletter.
What This Means
This event now captures both sides of the recursive self-improvement debate. Greenblatt's 2031 median prediction represents the optimistic near-term case based on compute scaling arguments, while the Princeton study provides empirical evidence that a critical capability gap remains: AI agents can do the engineering legwork of AI research but lack the judgment, creativity, and taste required to produce genuinely novel science. The gap between benchmark performance and open-ended research quality is significant, suggesting confident RSI timelines may be running ahead of current evidence. What is clear regardless of timeline is that the alignment questions raised — who superintelligences serve, and whether governance frameworks can keep pace — are urgent and demand serious attention now.