--- title: Did Google actually build RSI? Dream-RSI first look description: A new Google DeepMind and University of Maryland paper called 'Dream RSI' proposes letting an AI agent test thousands of new exploration policies against... canonical: https://daily.dev/posts/did-google-actually-build-rsi-dream-rsi-first-look-hzlh4b6qq twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Did Google actually build RSI? Dream-RSI first look | daily.dev og:description: A new Google DeepMind and University of Maryland paper called 'Dream RSI' proposes letting an AI agent test thousands of new exploration policies against... og:url: https://daily.dev/posts/did-google-actually-build-rsi-dream-rsi-first-look-hzlh4b6qq og:image: https://api.daily.dev/og/posts/Hzlh4B6Qq.png og:image:alt: Did Google actually build RSI? Dream-RSI first look og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Did Google actually build RSI? Dream-RSI first look **[Fireship](https://daily.dev/sources/fireship)** · 4 min read · 5 upvotes · 0 comments ## Summary A new Google DeepMind and University of Maryland paper called 'Dream RSI' proposes letting an AI agent test thousands of new exploration policies against cached past attempts (code, scores, crash logs) rather than touching the underlying model, then deploying whichever policy reaches the best result fastest. In tests across eight algorithm and math problems using Gemini, the dreaming approach found a lasso solver beating a standard Python ML library in about 300 tries, versus 550 for a static policy and roughly 51,000 for the previous record. The video argues this isn't true recursive self-improvement by I.J. Good's original definition, since the underlying model never gets smarter, it just explores faster and more efficiently, similar to how other recent AI math breakthroughs (Jacobian conjecture, Navier-Stokes, Riemann hypothesis progress) relied on static models wrapped in custom orchestration rather than models that improve their own weights. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is Dream RSI and how does it work in Google DeepMind's paper? Dream RSI is a technique where an AI agent's past discovery attempts (code, scores, crash outcomes) are cached and used as a simulator, letting the agent test thousands of new exploration policies against that cached history without retraining or touching the underlying model. The best-performing policy, judged by which reaches the target result in fewest attempts, is deployed on the next real run, and the process repeats. It was tested using Gemini across eight algorithm and math problems. _Developers evaluating AI-driven discovery loops can follow deeper breakdowns like this one on daily.dev._ ### Does Dream RSI count as true recursive self-improvement (RSI)? No, not by I.J. Good's original 1965 definition, because the underlying model doing the improving does not itself get smarter each round. The same Gemini model writes every new exploration policy, so it can only find solutions it was already capable of producing, just faster and with fewer wasted attempts, rather than gaining genuinely new capability the way true RSI would require. _Track how definitions of AI self-improvement evolve alongside real research on daily.dev._ ### How much faster did Dream RSI find a lasso solver compared to other methods? The dreaming policy discovered a lasso solver that beat a standard Python machine learning library in about 300 attempts, compared to 550 attempts needed by a fixed exploration policy and roughly 51,000 attempts needed by the previous record holder using older methods. This was one of eight test problems spanning algorithm design and mathematics used to evaluate the approach. _Engineers benchmarking agentic search efficiency can keep tabs on results like this via daily.dev._ --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#google-gemini](https://daily.dev/tags/google-gemini), [#google-deepmind](https://daily.dev/tags/google-deepmind) [View this post on daily.dev](https://daily.dev/posts/did-google-actually-build-rsi-dream-rsi-first-look-hzlh4b6qq) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Did Google actually build RSI? Dream-RSI first look","url":"https://daily.dev/posts/did-google-actually-build-rsi-dream-rsi-first-look-hzlh4b6qq","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/did-google-actually-build-rsi-dream-rsi-first-look-hzlh4b6qq"},"datePublished":"2026-09-17T20:38:58.020Z","dateModified":"2026-09-17T21:05:09.343Z","description":"A new Google DeepMind and University of Maryland paper called 'Dream RSI' proposes letting an AI agent test thousands of new exploration policies against...","image":"https://i.ytimg.com/vi/LoLYw--s-5w/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/LoLYw--s-5w/sddefault.jpg","isAccessibleForFree":true,"articleSection":"Fireship","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Fireship","logo":"https://media.daily.dev/image/upload/s--75ndqrkr--/f_auto,t_logo/v1702882094/logos/fireship.jpg","url":"https://daily.dev/sources/fireship"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/did-google-actually-build-rsi-dream-rsi-first-look-hzlh4b6qq","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":5},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,google-gemini,google-deepmind","timeRequired":"PT4M","video":{"@type":"VideoObject","name":"Did Google actually build RSI? Dream-RSI first look","description":"A new Google DeepMind and University of Maryland paper called 'Dream RSI' proposes letting an AI agent test thousands of new exploration policies against...","thumbnailUrl":"https://i.ytimg.com/vi/LoLYw--s-5w/sddefault.jpg","uploadDate":"2026-09-17T20:38:58.020Z","duration":"PT4M","url":"https://api.daily.dev/r/Hzlh4B6Qq","embedUrl":"https://www.youtube.com/embed/LoLYw--s-5w"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Fireship","item":"https://daily.dev/sources/fireship"},{"@type":"ListItem","position":3,"name":"Did Google actually build RSI? Dream-RSI first look"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/did-google-actually-build-rsi-dream-rsi-first-look-hzlh4b6qq#faq","mainEntity":[{"@type":"Question","name":"What is Dream RSI and how does it work in Google DeepMind's paper?","acceptedAnswer":{"@type":"Answer","text":"Dream RSI is a technique where an AI agent's past discovery attempts (code, scores, crash outcomes) are cached and used as a simulator, letting the agent test thousands of new exploration policies against that cached history without retraining or touching the underlying model. The best-performing policy, judged by which reaches the target result in fewest attempts, is deployed on the next real run, and the process repeats. It was tested using Gemini across eight algorithm and math problems. Developers evaluating AI-driven discovery loops can follow deeper breakdowns like this one on daily.dev."}},{"@type":"Question","name":"Does Dream RSI count as true recursive self-improvement (RSI)?","acceptedAnswer":{"@type":"Answer","text":"No, not by I.J. Good's original 1965 definition, because the underlying model doing the improving does not itself get smarter each round. The same Gemini model writes every new exploration policy, so it can only find solutions it was already capable of producing, just faster and with fewer wasted attempts, rather than gaining genuinely new capability the way true RSI would require. Track how definitions of AI self-improvement evolve alongside real research on daily.dev."}},{"@type":"Question","name":"How much faster did Dream RSI find a lasso solver compared to other methods?","acceptedAnswer":{"@type":"Answer","text":"The dreaming policy discovered a lasso solver that beat a standard Python machine learning library in about 300 attempts, compared to 550 attempts needed by a fixed exploration policy and roughly 51,000 attempts needed by the previous record holder using older methods. This was one of eight test problems spanning algorithm design and mathematics used to evaluate the approach. Engineers benchmarking agentic search efficiency can keep tabs on results like this via daily.dev."}}]} ```