--- title: Reproduction of MIT's weight-space programming results description: Alex Krentsel says he has reproduced the results from an MIT blog post on weight-space programming. He also says he has one additional result beyond those.... canonical: https://daily.dev/posts/reproduction-of-mit-s-weight-space-programming-results-09m5alp50 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Reproduction of MIT's weight-space programming results | daily.dev og:description: Alex Krentsel says he has reproduced the results from an MIT blog post on weight-space programming. He also says he has one additional result beyond those.... og:url: https://daily.dev/posts/reproduction-of-mit-s-weight-space-programming-results-09m5alp50 og:image: https://api.daily.dev/og/posts/09M5AlP50.png og:image:alt: Reproduction of MIT's weight-space programming results 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. # Reproduction of MIT's weight-space programming results **[Collections](https://daily.dev/sources/collections)** · 2 min read · 0 upvotes · 0 comments ## Summary Alex Krentsel says he has reproduced the results from an MIT blog post on weight-space programming. He also says he has one additional result beyond those. Thom Wolf and lateinteraction both reposted the announcement. The repost text is cut short, so the extra result itself isn't stated here. Krentsel's own account is the place to look for it. ## Content A new paper on weight-space programming reports that frontier LLMs can read and edit the raw weight tensors of smaller neural networks directly, with no gradient descent and no training loop. ## What the paper reports - Identifying a network's behavior from its raw weights. - Repairing a failed control policy by editing 3.5% of its weights. - Raising Qwen3.5-0.8B's MATH-500 score from 7.9% to 48.1% by editing just 0.016% of its weights. @omarsar0 called it a must-read report and said the finding that LLMs are surprisingly good at reading and writing weights matters for recursive self-improvement (RSI). The framing in some circles is that it cuts autoresearch out of the RSI loop. ## Replication and caveats AlexKrentsel says he reproduced the results from the MIT blog post and has one additional result of his own. @yoheinakajima, @Thom_Wolf and @lateinteraction all reshared that note. Critics point out that the MATH-500 jump probably doesn't reflect new math ability. The edit most likely suppresses a pathological reasoning-loop bug in the model. The same fix failed on Gemma 3 270M, which suggests it doesn't generalize across models. So I'd treat the 48.1% as a bug fix, not a skill transplant. Still, an LLM making a precise surgical edit to 0.016% of a network's weights is a striking thing to see work at all. ## Other items in the same news cycle - DeepSeek open-sourced a prototype "super-agent" harness. - A new agentic benchmark, GDPx, has a best model score of only 12%. - People are still arguing over whether quantized local models can rival Claude on coding tasks. ## Questions this post answers ### Can large language models read and write the weights of another neural network? Yes, an MIT blog post on weight-space programming reports that LLMs are surprisingly good at reading and writing neural network weights directly, treating them similarly to how they handle code or text. The finding has been raised as relevant to recursive self-improvement, since a system able to rewrite weights can in principle modify itself. Alex Krentsel independently reproduced the results and reported finding one additional result beyond what the original post described. _Anyone tracking self-modifying AI systems can follow reproductions and discussion of weight-space programming on daily.dev._ ## Similar posts on daily.dev - [LLMs are not superhuman at reading and writing neural network weights](https://daily.dev/posts/llms-are-not-superhuman-at-reading-and-writing-neural-network-weights-yrkwtj9xo) · sean goedecke · 0 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#mit](https://daily.dev/tags/mit) [View this post on daily.dev](https://daily.dev/posts/reproduction-of-mit-s-weight-space-programming-results-09m5alp50) ```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":"Reproduction of MIT's weight-space programming results","url":"https://daily.dev/posts/reproduction-of-mit-s-weight-space-programming-results-09m5alp50","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/reproduction-of-mit-s-weight-space-programming-results-09m5alp50"},"datePublished":"2026-11-02T22:17:02.468Z","dateModified":"2026-11-03T05:37:51.317Z","description":"Alex Krentsel says he has reproduced the results from an MIT blog post on weight-space programming. He also says he has one additional result beyond those....","isAccessibleForFree":true,"articleSection":"Collections","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":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/reproduction-of-mit-s-weight-space-programming-results-09m5alp50","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,mit","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Reproduction of MIT's weight-space programming results"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/reproduction-of-mit-s-weight-space-programming-results-09m5alp50#faq","mainEntity":[{"@type":"Question","name":"Can large language models read and write the weights of another neural network?","acceptedAnswer":{"@type":"Answer","text":"Yes, an MIT blog post on weight-space programming reports that LLMs are surprisingly good at reading and writing neural network weights directly, treating them similarly to how they handle code or text. The finding has been raised as relevant to recursive self-improvement, since a system able to rewrite weights can in principle modify itself. Alex Krentsel independently reproduced the results and reported finding one additional result beyond what the original post described. Anyone tracking self-modifying AI systems can follow reproductions and discussion of weight-space programming on daily.dev."}}]} ```