DAIR.ai Papers of the Week

[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8,"_9":10},"actionData","errors","root",{"_150":151,"_152":153},"routes/_app.x.$handle.$name",{"_11":12,"_13":14,"_15":16,"_18":19,"_17":-7,"_114":-5,"_20":115,"_22":116,"_117":118,"_119":120,"_25":121,"_24":-5,"_122":-5,"_123":-5,"_124":53,"_125":126,"_127":44,"_128":-5,"_129":-5,"_130":131},"routes/_app.x.$handle.$name.branch.$branch.code.$",{"_11":12,"_13":14,"_15":16,"_17":-7,"_18":19,"_20":21,"_22":23,"_24":-7,"_25":26,"_27":-5,"_28":-5,"_29":-5,"_30":-5,"_31":-5},"handle","danielpcox","name","dair_papers_newspaper_red","branchName","main","version","path","","project",{"_33":61,"_13":14,"_45":62,"_63":-5,"_64":53,"_65":66,"_67":66,"_68":-5,"_69":-5,"_70":71,"_72":-5,"_73":53,"_74":-5,"_75":-5,"_76":47,"_77":78,"_79":-5,"_80":-5,"_81":-5,"_82":-5,"_83":84,"_85":84,"_86":87,"_88":87,"_89":87,"_90":84,"_91":84,"_92":93,"_94":95,"_96":97,"_98":53},"tree",[56],"file","contents",[32],"upstream","pull","ahead","behind","envVars",{"_33":34,"_13":35,"_36":37,"_38":-5,"_39":40,"_41":42,"_43":44,"_45":46,"_17":47,"_48":49,"_50":51,"_52":53,"_54":-5,"_18":55},"id","328a0256-1515-11f1-86e7-42dde27851f2","mod.ts","fileId","328a031e-1515-11f1-86e7-42dde27851f2","parentId","branchId","325afd3a-1515-11f1-a47e-42dde27851f2","branchVersion",1,"latest",true,"createdAt",["D",1772331621148],0,"fileType","http","content","export default async function(req: Request) {\n const papers = [\n {\n title: \"DeepSeek-V3 Technical Report\",\n authors: \"DeepSeek-AI\",\n summary: \"DeepSeek-V3 is a strong Mixture-of-Experts (MoE) language model with 671B total parameters, of which 37B are activated for each token. It uses Multi-head Latent Attention (MLA) and DeepSeek-V3 auxiliary-loss-free load balancing. It outperforms other open-source models and rivals top-tier closed-source models.\",\n link: \"https://arxiv.org/abs/2412.19437\"\n },\n {\n title: \"The Llama 3 Herd of Models\",\n authors: \"Llama Team, AI @ Meta\",\n summary: \"Meta introduces Llama 3, a new family of large language models. The 405B parameter model is the first openly available model that rivals top AI models across capabilities in general knowledge, steerability, math, tool use, and multilingual translation.\",\n link: \"https://arxiv.org/abs/2407.21783\"\n },\n {\n title: \"Scaling Laws for Precision\",\n authors: \"T. J. S. et al.\",\n summary: \"This paper investigates how the precision of weights and activations affects the scaling of large language models. It provides empirical evidence and theoretical grounding for choosing the optimal bit-width during training and inference.\",\n link: \"https://arxiv.org/abs/2402.07005\"\n },\n {\n title: \"MM1: Methods, Analysis & Insights from Multimodal LLM Pre-training\",\n authors: \"Apple\",\n summary: \"Apple researchers discuss the training of MM1, a family of multimodal models up to 30B parameters. They show that for multimodal pre-training, the image resolution and the number of image tokens matter more than the specific vision-language connector.\",\n link: \"https://arxiv.org/abs/2403.09611\"\n },\n {\n title: \"Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking\",\n authors: \"Zelikman et al.\",\n summary: \"Quiet-STaR is a generalization of STaR where LMs learn to generate rationales at every token to explain future text. This allows the model to learn to reason in an unsupervised way, improving performance on reasoning tasks.\",\n link: \"https://arxiv.org/abs/2403.09629\"\n }\n ];\n\n const html = `\n \n \n \n \n \n DAIR.ai Papers of the Week - Red Edition\n \n \n \n \n

\n \n \n

\n VOL. MMXXVI ... No. 064\n SUNDAY, MARCH 1, 2026\n PRICE: ONE GPU HOUR\n

\n\n

AI FRONTIER EXPANDS: NEW ARCHITECTURES CHALLENGE THE STATUS QUO

\n\n

\n ${papers.slice(0, 3).map(paper => `\n

\n ${paper.title}\n By ${paper.authors}\n

${paper.summary}

\n `).join('')}\n

\n\n

\n \n\n

WANTED: COMPUTE

Will trade high-quality synthetic data for H100 clusters. No lowballs, I know what I have.

CALL 1-800-PYTORCH

\n\n

\n STAY RED. STAY READ.\n

\n\n \n

\n \n \n `;\n\n return new Response(html, {\n headers: { \"Content-Type\": \"text/html\" }\n });\n}\n","deleted",false,"label",[35],{"_33":34,"_13":35,"_36":37,"_38":-5,"_39":40,"_41":42,"_43":44,"_45":57,"_17":47,"_48":49,"_50":51,"_52":53,"_54":-5,"_58":59,"_18":60},["D",1772331621148],"children",[],[35],"325aae52-1515-11f1-a47e-42dde27851f2",["D",1772331620845],"deletedAt","legacyEnvVars","privacy","public","httpPrivacy","description","imageUrl","userId","019c649c-877d-70d7-a441-396b9c85efaa","pinnedFileId","pinPreview","pinMigrated","screenshotVersion","likeCount","orgId","019c649c-877d-70d7-a441-3e13100c1184","tursoName","tursoHostname","tursoMeasuredAt","pinned","userPerm","r","valPerm","blobPerm","rw","emailPerm","sqlitePerm","projectPerm","telemetryPerm","user",{"_11":12,"_113":109},"org",{"_11":12,"_107":-5,"_108":109,"_70":71,"_110":111,"_112":53},"branch",{"_17":42,"_13":16,"_33":40,"_99":61,"_100":-5,"_101":-5,"_102":-5,"_103":-5,"_104":-5,"_105":14,"_106":12},"isOwned","projectId","forkedBranchId","forkedBranchVersion","forkedProjectPrivacy","forkedProjectId","forkedProjectOrgId","projectName","userHandle","displayName","avatar","https://img.clerk.com/eyJ0eXBlIjoicHJveHkiLCJzcmMiOiJodHRwczovL2ltYWdlcy5jbGVyay5kZXYvb2F1dGhfZ2l0aHViL2ltZ18zOWpmZkZlTFFUYkJFSlE4R291RFhqVE1ma0UifQ","tier","free","canRestrictVals","profileImageUrl","adoptOrgId",{"_33":61,"_13":14,"_45":146,"_63":-5,"_64":53,"_65":66,"_67":66,"_68":-5,"_69":-5,"_70":71,"_72":-5,"_73":53,"_74":-5,"_75":-5,"_76":47,"_77":78,"_79":-5,"_80":-5,"_81":-5,"_82":-5,"_83":84,"_85":84,"_86":87,"_88":87,"_89":87,"_90":84,"_91":84,"_92":147,"_94":148,"_96":149,"_98":53},[143],"files",[143],"counts",{"_134":47,"_135":42,"_136":47,"_31":137,"_138":139},[132],"upstreamProject","upstreamBranch","emptyEnvVars","flags",[],"hasHttpFiles","grants","orgs","customDomains",{},{"_33":34,"_13":35,"_36":37,"_38":-5,"_39":40,"_41":42,"_43":44,"_45":46,"_17":47,"_48":49,"_50":51,"_52":53,"_54":-5,"_18":133},[35],"pulls","branches","forks",{"_140":47,"_142":47},"connections",{"_140":47,"_141":47},"count","unfilledCount","emptyCount",{"_33":34,"_13":35,"_36":37,"_38":-5,"_39":40,"_41":42,"_43":44,"_45":57,"_17":47,"_48":49,"_50":51,"_52":53,"_54":-5,"_58":144,"_18":145},[],[35],["D",1772331620845],{"_11":12,"_113":109},{"_11":12,"_107":-5,"_108":109,"_70":71,"_110":111,"_112":53},{"_17":42,"_13":16,"_33":40,"_99":61,"_100":-5,"_101":-5,"_102":-5,"_103":-5,"_104":-5,"_105":14,"_106":12},"ctx",{"_92":-5,"_129":183,"_184":185,"_186":187,"_188":53,"_189":53,"_190":53,"_191":44,"_192":-5,"_193":-7,"_194":-7,"_195":47,"_196":53,"_197":53,"_198":-5,"_199":53,"_200":53,"_201":53,"_202":-5},"ENV",{"_154":155,"_156":157,"_158":16,"_159":160,"_161":162,"_163":164,"_165":166,"_167":168,"_169":170,"_171":172,"_173":174,"_175":176,"_177":178,"_179":180,"_181":182},"REACT_APP_SENTRY_DSN","https://586020d86f05489db63a554690acf1e9@o1398359.ingest.sentry.io/6724499","RENDER_GIT_COMMIT","e87eb04e707cf64e7e532c8e8ec395f209e17f15","RENDER_GIT_BRANCH","POSTHOG_API_TOKEN","phc_nmoyuAkStEZuUp69rYlg1DZgrIt19Gx0Y4Hwf8tGgJP","POSTHOG_API_HOST","https://wart.val.town","WEB_API_TEMPLATE","https://{val}.web.val.run","APP_URL","https://www.val.town","MODULE_URL","https://esm.town","VANITY_WEB_API_TEMPLATE","https://{label}.val.run/","DENOLS_URL","https://lsp.val.town","NODE_ENV","production","CLOUDFLARE_TURNSTILE_SITE_KEY","0x4AAAAAACtaIZJH7s9rBMZ6","GOOGLE_CLIENT_ID","875989696899-jtsqkhlbdv5jrlsk8631p7ipjg3c6o62.apps.googleusercontent.com","STRIPE_PUBLISHABLE_KEY","pk_live_51N2zvWEpZTTLipIkVEwXrUG9AT0yeOOEYxszarZT1jpsZV58Ly6wlg3SltZpHN26CUalOBmUSY7adQRQiSkCZz2500qSXNkf8Y","API_URL","https://api.val.town",[],"preferencesKeybindings","default","preferencesTheme","light","preferencesFixers","preferencesCopilot","preferencesSelectionAi","preferencesCloseBrackets","preferencesChatPanelState","preferencesDprint","preferencesDenofmt","notificationCount","flagDenols","isBusinessOrgMember","businessOrgHandle","isBusinessOrgAdmin","isHigherTownieTier","navCollapsed","currentOrgId"]

The DAIR Gazette

AI FRONTIER EXPANDS: NEW ARCHITECTURES CHALLENGE THE STATUS QUO

In Brief

Code