OpenAI CEO Sam Altman stated that if frontier reinforcement learning training is expected to significantly enhance capabilities, OpenAI will now establish a clear safety case before training begins, elevating monitoring, runaway risk, and the collaborative development of safety standards to the level of industry cooperation. This “slowdown” is not a halt, but an acceptance of the additional costs brought by safety cases and monitoring; he also supports independent audits and a unified federal safety framework. Fran\u00e7ois Chollet, from another perspective, emphasizes that the excessive concentration of power in frontier AI itself is a risk, necessitating multiple independent providers, including open-source models.<\/p>\n
Sources:<\/p>\n
@MaxForAI relayed a Zhipu announcement stating the company completed approximately $5 billion in equity and debt financing, with about $2 billion from share placement and about $3 billion from convertible bonds; the disclosed use of funds includes the next-generation GLM foundational model, a fully self-training system, and large-scale training and inference infrastructure. More notably, Zhipu has written “Fully Self Training” and RSI (Recursive Self-Improvement) into its R&D roadmap, with directions including model-generated data, Agent-generated training environments, and model participation in optimizing operators, kernels, scheduling, caching, and the Serving Stack. The above content is from the blogger’s summary of the announcement; the public post did not provide independent material proving that full RSI has already been achieved.<\/p>\n
Sources:<\/p>\n
@Gorden_Sun introduced a roadmap jointly published by Shanghai Jiao Tong University, Tsinghua University, ByteDance, Xiaohongshu, and Shanghai AI Lab, dividing self-improvement into L1 to L5: first executing according to human rules and accumulating results, then autonomously selecting improvement strategies, generating training tasks for weaknesses, accumulating skills and tools from real environments, and ultimately modifying the underlying methods of research and self-training. It provides a hierarchical framework for judging an Agent’s capability boundaries: being able to complete tasks, choose methods, self-train, adapt on-site, and improve its own mechanisms are not at the same level.<\/p>\n
Sources:<\/p>\n
@xiaohu relayed an Anthropic demonstration: after the payment interface error rate exceeded 2%, Claude Tag automatically pulled monitoring data in Slack, compared recent deployment and feature toggle records, and located the issue in about 15 minutes to a retry program lacking a concurrency limit, proposing two fixes. After the engineer confirmed “add limit, keep feature on,” it changed only 4 lines of code and continued monitoring for 10 minutes after deployment. The key to this case is not full delegation, but that the Agent can investigate the site, provide reviewable repair suggestions, while keeping merging and deployment in human hands.<\/p>\n
Sources:<\/p>\n
@poteto previewed that they, along with @mattyp and @roshan_s, will use only Grok Bot in a Grok Bot Galaxy livestream to start a new company from scratch, with a 3-day time limit; they later stated about 70,000 people registered for the livestream. Another practical test shared by Elon Musk shows Grok Bot would daily read user X bookmarks, launch Cursor Agent to create a demo, verify with screenshots or videos, then create a code branch and deploy a preview via Cloudflare. Both signals point toward a more specific Agent closed loop: sourcing from personal interests, automatically generating runnable prototypes, then handing them over for human review.<\/p>\n
Sources:<\/p>\n
@derek_wall90176 believes the real difficulty in AI video is stably reproducing shots according to design, therefore suggesting first having an Agent call Blender to create a white model, determining character positions, action relationships, camera angles, and spatial structure, then handing it to MiniMax H3 for rendering, separating “structural determinism” and “visual imagination.” The Dunhuang group dance production workflow shared by @LufzzLiz also adopts a similar approach: using GPT-6 Astra for video understanding, image generation, and breakdown, then using H3 for omnidirectional reference generation of two video segments, requiring five adult characters to maintain identity consistency during occlusion, turns, and formation changes. The currently visible materials are creators’ practical tests and tutorials, illustrating workflow and effects, not that models have solved all deformation issues.<\/p>\n
Sources:<\/p>\n
The Obsidian e-book reader released by @vista8 supports EPUB, PDF, MOBI, AZW3, FB2, and other formats, and can call Codex, Claude Code, ZCode, Kimi, or DeepSeek APIs to assist reading; highlights and notes are written to Markdown files, the code is open-source, and a mobile adaptation tutorial was later added. The value of such tools is not just another reader, but integrating file rendering, cross-model assistance, and portable note accumulation into the same workflow, reducing dependence on the built-in models of a single reading platform.<\/p>\n
Sources:<\/p>\n
@steipete indicated that the next version or development channel will use folder cloning with APFS, Btrfs, XFS, and ReFS to make worktrees approximately 80% faster while saving disk space; after targeting teams, he also stated that the current environment runs about 80 sessions in parallel without significant pressure. On the other hand, @rauchg announced that Steren, the creator of Google Cloud Run, has joined Vercel to oversee Fluid computing products such as Functions, Containers, Sandbox, and Builds, and directly pointed out that Agents require new computing primitives. The visible change is that competition among Coding Agents is no longer just about model response quality, but also about code isolation, concurrent sessions, sandboxing, and build\/run environments.<\/p>\n
Sources:<\/p>\n
@lifesinger relayed an observation from a friend, stating that Silicon Valley AI application startups often scale their growth teams to a size similar to their product and R&D teams because, with AI, product development might just be a matter of time; what truly creates a gap is traffic and growth. @bbkirstry, meanwhile, categorized AI monetization into seven types: saving time, improving efficiency, selling features, selling judgment, selling attention, selling connections, and selling assets. Both are analyses from bloggers rather than industry statistics, but together they provide an actionable insight: being able to build a product is just the starting point; the methods of capturing value\u2014such as services, subscriptions, content, consulting, and channels\u2014along with the ability to consistently reach users, are what determine whether commercial results can be achieved.<\/p>\n
Sources:<\/p>\n
Statistics: Scanned timeline count=476 Number of bloggers matched=57 Total matched tweets=331 Weighted tweet score=245.8 Original tweet count=130 Retweet count=87 Crawl attempt count=3 Boundary coverage status=tail_confidently_crossed_target_boundary<\/p>\n","protected":false},"excerpt":{"rendered":"
Discussion focused on shifting AI safety strategies from post-training to pre-training, explicit planning for model recursive self-improvement, and Agents demonstrating closed-loop handling and creative capabilities in real work scenarios.<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[19],"class_list":["post-1737","post","type-post","status-publish","format-standard","hentry","category-brief","tag-x--ai-"],"_links":{"self":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1737","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/comments?post=1737"}],"version-history":[{"count":1,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1737\/revisions"}],"predecessor-version":[{"id":1738,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1737\/revisions\/1738"}],"wp:attachment":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1737"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1737"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1737"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}