Wes Roth · 2026-09-17 · notable
Wes Roth — 'Google is SO back...'
Wes Roth's 17 September 2026 video walks through Dream-RSI, a project from Google and DeepMind researchers that lets an AI agent replay its own past experiments as a simulator and test better research strategies before spending compute.
Wes Roth explains Dream-RSI, a Google method that lets an agent rehearse experiments against its own history.
What is it?
'Google is SO back...' went up on Wes Roth's channel on 17 September 2026. The episode covers Dream-RSI, a recursive self-improvement method from researchers at Google, Google DeepMind, the University of Maryland and the University of Virginia. Roth's framing question is whether learning to choose the right experiment could become a building block for recursive self-improvement.
How does it work?
Dream-RSI, as the video's sources describe it, records the tree of discoveries an agent produces while it explores. That finished tree then becomes a replay simulator over the search space already visited, so alternative exploration policies can be scored against it without running a single new experiment. The best policy is redeployed to gather fresh data and the loop repeats. Roth builds the episode from the project page, the arXiv paper and DeepMind's earlier AlphaEvolve announcement rather than reporting independently.
Why does it matter?
The expensive part of automated research is evaluation, not idea generation. The Dream-RSI project page claims 2.43x fewer generations on VGG16 at comparable performance and 162x fewer discovery calls than SimpleTES on a Lasso task — the kind of saving that decides whether an agent-driven search is affordable at all. Roth's video is commentary on a technical report, so those numbers are the authors' claims, not independent measurements.
Who is it for?
people following automated AI research