--- title: Einstein Arena: Harnessing Collective Agent Intelligence... description: A Together AI researcher describes Einstein Arena, an environment where AI agents autonomously collaborate and compete to solve open scientific problems,... canonical: https://daily.dev/posts/einstein-arena-harnessing-collective-agent-intelligence-for-open-science-james-zou-together-ai-jrdf4zxg5 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Einstein Arena: Harnessing Collective Agent Intelligence for Open Science — James Zou, Together AI | daily.dev og:description: A Together AI researcher describes Einstein Arena, an environment where AI agents autonomously collaborate and compete to solve open scientific problems,... og:url: https://daily.dev/posts/einstein-arena-harnessing-collective-agent-intelligence-for-open-science-james-zou-together-ai-jrdf4zxg5 og:image: https://api.daily.dev/og/posts/JRDf4Zxg5.png og:image:alt: Einstein Arena: Harnessing Collective Agent Intelligence for Open Science — James Zou, Together AI 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. # Einstein Arena: Harnessing Collective Agent Intelligence for Open Science — James Zou, Together AI **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 16 min read · 0 upvotes · 0 comments ## Summary A Together AI researcher describes Einstein Arena, an environment where AI agents autonomously collaborate and compete to solve open scientific problems, verified by deterministic checkers and tracked via public leaderboards and discussion forums. Agents using this environment found new record-setting solutions to 11 problems, including improving the 11-dimensional kissing number from 593 to 604 non-overlapping spheres, surpassing a prior DeepMind result. The same environment design was applied to optimizing GPU kernels (achieving over 2x speedups in production at Together AI) using agents with different specialized personas. A second environment, DS Gym, curates data science tasks from papers and Kaggle competitions to evaluate and train agents, addressing the finding that many existing benchmarks allow up to 20-50% of tasks to be solved via shortcuts without using the actual data. The overarching thesis is that designing environments, rather than workflows or harnesses, unlocks more creativity and collective intelligence from AI agents. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is the current best known solution to the kissing number problem in 11 dimensions? AI agents collaborating on the Einstein Arena environment discovered a construction of 604 non-overlapping spheres in 11 dimensions, improving on the prior record of 593 set by DeepMind the year before, which itself improved on a 592-sphere result from 2022 that had stood for about 40 years since the 582-sphere construction from 1980. _Developers tracking breakthroughs in AI-driven mathematical research can follow updates like this on daily.dev._ ### Why are existing data science benchmarks unreliable for evaluating AI agents? Many popular data science benchmarks are vulnerable to shortcuts, meaning agents can solve 20 to 50 percent of tasks correctly through reasoning alone without ever touching the underlying dataset. This was found across several widely used benchmarks, prompting the creation of DS Gym, a curated benchmark built from recent papers and open Kaggle competitions with expert review to eliminate shortcut solutions. _Teams evaluating agent benchmarks for real rigor can find comparisons like this via daily.dev._ ### How much speedup can AI agents achieve when optimizing GPU kernels like page attention? AI agents collaborating within a competitive leaderboard environment achieved over two-fold speedups on production GPU kernels, including page attention kernels, compared to previous state-of-the-art implementations. Different agents were assigned personas focused on profiling, memory consumption, or tensor precision, and these improved kernels are now used in production at Together AI. _Engineers optimizing GPU kernel performance can track results like these on daily.dev._ ## Similar posts on daily.dev - [Together AI at ICML 2026: frontier research across the full stack](https://daily.dev/posts/together-ai-at-icml-2026-frontier-research-across-the-full-stack-pgl4rdif5) · Together AI · 0 upvotes · 0 comments - [Mathematical Experiments Are Becoming Abundant Through Human-Machine Teaming](https://daily.dev/posts/mathematical-experiments-are-becoming-abundant-through-human-machine-teaming-at6d9wesn) · Towards Data Science · 1 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#ai-agents](https://daily.dev/tags/ai-agents), [#computing](https://daily.dev/tags/computing) [View this post on daily.dev](https://daily.dev/posts/einstein-arena-harnessing-collective-agent-intelligence-for-open-science-james-zou-together-ai-jrdf4zxg5) ```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":"Einstein Arena: Harnessing Collective Agent Intelligence for Open Science — James Zou, Together AI","url":"https://daily.dev/posts/einstein-arena-harnessing-collective-agent-intelligence-for-open-science-james-zou-together-ai-jrdf4zxg5","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/einstein-arena-harnessing-collective-agent-intelligence-for-open-science-james-zou-together-ai-jrdf4zxg5"},"datePublished":"2026-08-25T20:52:30.290Z","dateModified":"2026-08-25T20:52:53.701Z","description":"A Together AI researcher describes Einstein Arena, an environment where AI agents autonomously collaborate and compete to solve open scientific problems,...","image":"https://i.ytimg.com/vi/mMNkdYnIVC4/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/mMNkdYnIVC4/sddefault.jpg","isAccessibleForFree":true,"articleSection":"AI Engineer","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":"AI Engineer","logo":"https://media.daily.dev/image/upload/s--u5PucxNT--/f_auto/v1724338940/logos/aidotengineer","url":"https://daily.dev/sources/aidotengineer"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/einstein-arena-harnessing-collective-agent-intelligence-for-open-science-james-zou-together-ai-jrdf4zxg5","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,ai-agents,computing","timeRequired":"PT16M","video":{"@type":"VideoObject","name":"Einstein Arena: Harnessing Collective Agent Intelligence for Open Science — James Zou, Together AI","description":"A Together AI researcher describes Einstein Arena, an environment where AI agents autonomously collaborate and compete to solve open scientific problems,...","thumbnailUrl":"https://i.ytimg.com/vi/mMNkdYnIVC4/sddefault.jpg","uploadDate":"2026-08-25T20:52:30.290Z","duration":"PT16M","url":"https://api.daily.dev/r/JRDf4Zxg5","embedUrl":"https://www.youtube.com/embed/mMNkdYnIVC4"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"AI Engineer","item":"https://daily.dev/sources/aidotengineer"},{"@type":"ListItem","position":3,"name":"Einstein Arena: Harnessing Collective Agent Intelligence for Open Science — James Zou, Together AI"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/einstein-arena-harnessing-collective-agent-intelligence-for-open-science-james-zou-together-ai-jrdf4zxg5#faq","mainEntity":[{"@type":"Question","name":"What is the current best known solution to the kissing number problem in 11 dimensions?","acceptedAnswer":{"@type":"Answer","text":"AI agents collaborating on the Einstein Arena environment discovered a construction of 604 non-overlapping spheres in 11 dimensions, improving on the prior record of 593 set by DeepMind the year before, which itself improved on a 592-sphere result from 2022 that had stood for about 40 years since the 582-sphere construction from 1980. Developers tracking breakthroughs in AI-driven mathematical research can follow updates like this on daily.dev."}},{"@type":"Question","name":"Why are existing data science benchmarks unreliable for evaluating AI agents?","acceptedAnswer":{"@type":"Answer","text":"Many popular data science benchmarks are vulnerable to shortcuts, meaning agents can solve 20 to 50 percent of tasks correctly through reasoning alone without ever touching the underlying dataset. This was found across several widely used benchmarks, prompting the creation of DS Gym, a curated benchmark built from recent papers and open Kaggle competitions with expert review to eliminate shortcut solutions. Teams evaluating agent benchmarks for real rigor can find comparisons like this via daily.dev."}},{"@type":"Question","name":"How much speedup can AI agents achieve when optimizing GPU kernels like page attention?","acceptedAnswer":{"@type":"Answer","text":"AI agents collaborating within a competitive leaderboard environment achieved over two-fold speedups on production GPU kernels, including page attention kernels, compared to previous state-of-the-art implementations. Different agents were assigned personas focused on profiling, memory consumption, or tensor precision, and these improved kernels are now used in production at Together AI. Engineers optimizing GPU kernel performance can track results like these on daily.dev."}}]} ```