{"id":503,"date":"2026-09-19T16:13:15","date_gmt":"2026-09-19T16:13:15","guid":{"rendered":"https:\/\/halobates.de\/blog\/?p=503"},"modified":"2026-09-19T16:13:15","modified_gmt":"2026-09-19T16:13:15","slug":"dark-code","status":"publish","type":"post","link":"http:\/\/halobates.de\/blog\/p\/503","title":{"rendered":"Dark code"},"content":{"rendered":"\n<p>The <a href=\"https:\/\/huggingface.co\/deepseek-ai\/DeepSeek-V4.1-Flash\/blob\/main\/DeepSeek_V41_Tech_Report.pdf\">DeepSeek V4.1 technical report<\/a> has this interesting tidbit:<\/p>\n\n\n\n<blockquote class=\"wp-block-quote\">\n<p>\nCoding agent training environments are built from &#8230;\npublic GitHub repositories that meet a star-count threshold. Environment construction is carried\nout collaboratively by multiple specialized agents. First, an agent determines whether the project\ncan be built and fully run inside a container and whether it can be automatically verified; if so, it\nselects a specific turn or commit as the task starting point, designs several sufficiently complex\nimplementation directions, and produces concrete evaluation points, including both fail-to-pass\nand pass-to-pass points, along with a construction report, fetching external resources from the\nweb as needed. Next, a separate agent sets up dependencies, the initial working directory, test\ncode, and task descriptions in an isolated container, performs self-testing, removes any traces\nthat could leak the task solution, and packages the environment as a new image layer. Then,\nmultiple distinct agents attempt the task, and an independent quality-inspection agent reviews\nthe environment together with the solving agents\u2019 trajectories, &#8230;\n<\/p>\n<\/blockquote>\n\n\n<p>So essentially as part of model training the AI models solve random problems in github repositories. If you&#8217;re working on a public github project with enough stars, and looking at a problem, some AI agent might have already solved it somewhere in the dark. Maybe the solution was bad, or maybe it was better than yours.<\/p>","protected":false},"excerpt":{"rendered":"<p>The DeepSeek V4.1 technical report has this interesting tidbit: Coding agent training environments are built from &#8230; public GitHub repositories that meet a star-count threshold. Environment construction is carried out collaboratively by multiple specialized agents. First, an agent determines whether the project can be built and fully run inside a container and whether it can [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[1],"tags":[21,22],"_links":{"self":[{"href":"http:\/\/halobates.de\/blog\/wp-json\/wp\/v2\/posts\/503"}],"collection":[{"href":"http:\/\/halobates.de\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/halobates.de\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/halobates.de\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/halobates.de\/blog\/wp-json\/wp\/v2\/comments?post=503"}],"version-history":[{"count":2,"href":"http:\/\/halobates.de\/blog\/wp-json\/wp\/v2\/posts\/503\/revisions"}],"predecessor-version":[{"id":505,"href":"http:\/\/halobates.de\/blog\/wp-json\/wp\/v2\/posts\/503\/revisions\/505"}],"wp:attachment":[{"href":"http:\/\/halobates.de\/blog\/wp-json\/wp\/v2\/media?parent=503"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/halobates.de\/blog\/wp-json\/wp\/v2\/categories?post=503"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/halobates.de\/blog\/wp-json\/wp\/v2\/tags?post=503"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}