According to court ruling documents compiled by @MaxForAI, Anthropic knowingly downloaded over 7 million e-books from pirated sites during the training of Claude; co-founder Ben Mann personally downloaded multiple batches of unauthorized books from Books3, LibGen, and PiLiMi. The ruling also found that Anthropic spent millions of dollars purchasing physical books, hired people to remove spines, cut pages, scan them into PDFs, then destroyed the original physical copies, in an attempt to build a permanently preserved "universal central library". Senior leadership also proactively halted licensing negotiations with multiple publishers, citing that "negotiating copyright is too much trouble". Ultimately, the court ruled Anthropic liable for "knowingly downloading pirated content and permanently storing it", rather than ruling the use of books for training itself infringing. The incident triggered reactions from multiple independent accounts: @beffjezos, retweeted by @steipete, stated that Anthropic "is effectively burning books and still wants exclusive access to the truth"; another retweet by @steipete from @ZackKorman further notes that Dario’s position goes beyond opposing open weight models, and his goal is to not allow any unregulated models to exist. This creates a clear rift with Anthropic’s long-projected image of prioritizing "AI safety first".<\/p>\n
Sources:<\/p>\n
According to reports from Chinese automotive industry media compiled by @MaxForAI, Lu Siyuan, head of XPeng’s AI Infrastructure department, is leaving the company to join OpenAI’s robotics division. Lu Siyuan holds a PhD in physics from the University of Southern California, and has long overseen end-to-end infrastructure work including model training frameworks, GPU clusters, in-house chip compilers, model quantization, and in-vehicle inference deployment, directly managing a team of around 200 people at XPeng. XPeng has not yet found a direct replacement, and the original team may be split into more than three groups managed by different leaders; since October 2025, XPeng’s core technology lines including intelligent driving, robotics products, and AI infrastructure have all seen leadership-level personnel changes. OpenAI launched its robotics business last year, and is actively seeking talent that understands training, chips, compilers, and automotive-grade mass production at the same time, which is the core motivation for recruiting Lu Siyuan. This move is seen as a new signal of continued talent outflow from Chinese automakers to Silicon Valley embodied intelligence companies.<\/p>\n
Sources:<\/p>\n
According to reports from multiple Taiwanese media compiled by @MaxForAI, the Keelung District Prosecutors Office in Taiwan has detained Zhang Denglong, a senior business development manager at NVIDIA, during an investigation into an illegal smuggling case of high-end AI servers, and has barred visitors and communication. Prosecutors accuse him of providing false documents during customer qualification reviews, helping Supermicro servers equipped with NVIDIA B300 chips pass internal reviews for transshipment to sanctioned countries (excluding North Korea, Russia, Iran). A total of 7 people have been detained in the case so far. Supermicro has confirmed that 4 employees of its Taiwan branch are under investigation, 2 of whom have been detained and 2 released on bail; Supermicro states that the company itself is not the target of the investigation and is cooperating with prosecutors. In May this year, prosecutors seized around 50 servers, and in June the search scope expanded to Supermicro’s Taiwan office, distributor Albatron, and data center operator IsCom.<\/p>\n
Sources:<\/p>\n
Developer @migtissera has published logs showing that months after he canceled his Cursor subscription, Cursor continued uploading source code and project dependencies to its servers. Between July 18 and 27, a total of 63,106 files totaling around 736 MB were uploaded; this did not occur while using Agent or Composer, it happened simply when "opening a few folders with Cursor as a regular IDE". The logs also show that Cursor’s code indexing only checks if a user is logged in, not if they have an active paid subscription; even when AI features are not used, background indexing may still continue running. Cursor’s official documentation acknowledges that when codebase indexing is enabled, files are split into chunks and uploaded to compute embeddings, plaintext is deleted after the request, but metadata such as embeddings, file hashes, and filenames may be retained long-term. After the story was compiled and spread by @MaxForAI, it is seen as a major shift in default privacy boundaries for developers following Cursor’s acquisition by SpaceXAI.<\/p>\n
Sources:<\/p>\n
Kimi officially announced the full open source release of K3, with weights, technical reports, MoonEP MoE communication library, KDA high-performance operators, and Agent reinforcement learning sandbox released simultaneously; K3 is a 2.78T total parameter \/ 104.2B activated parameter MoE model with native visual understanding and 1M token context window. In the concurrently open sourced MoonEP, Supported Devices now adds Zhenwu PPU, the new generation inference chip Zhenwu M890 (PPU 1.5) released by Alibaba T-Head in May this year, with integration reaching deep down to the low-level expert parallel communication library layer. Multiple accounts have cross-confirmed that inference services from Together AI, DigitalOcean, Fireworks AI, OpenCode Zen and others went live on day zero of the open source release, making K3 the fastest rising model in Hugging Face trending history. While domestic rumors that "GLM, Kimi, MiniMax will adopt Alibaba PPU 1.5" have not been fully confirmed, the completed integration of MoonEP marks the official start of large-scale binding between "Chinese AI models \u00d7 Chinese chips".<\/p>\n
Sources:<\/p>\n
Elon Musk confirmed while citing a third-party cybersecurity benchmark that Grok 4.6 is scheduled to launch around August 7, with 1.5T parameters and significant improvements to SFT and RL; several weeks later Grok 4.7 will launch, with 2.1T parameters, outperforming 4.6 in most dimensions, only with slightly slower inference speed but further improved token efficiency. Background context: the benchmark was published by Vercel CEO Guillermo Rauch, which found Grok 4.5 is already the strongest cybersecurity AI model in terms of price-performance, costing 10x less than Sol, 5.7x less than Opus 5, 2.2x less than Kimi K3, with performance close to Kimi K3. Musk’s statement moves Grok beyond just simple version updates to a clear "two-step scaling roadmap", corresponding to xAI’s next round of investment in long context and RL training.<\/p>\n
Sources:<\/p>\n
NVIDIA has led the founding of the Open Secure AI Alliance, with founding members including multiple US and European AI infrastructure and ecosystem companies such as Microsoft, Dell, OpenAI, SpaceX, and Anthropic, with OpenClaw, Nous Research also included on the member list. The alliance’s position reflects a shared judgment: in cybersecurity scenarios, "defenders need an open ecosystem with both open source and closed source", echoing the recent contrast from the OpenAI-Hugging Face security incident where closed models hindered forensics while open weight models helped contain the threat. Building on this, Anthropic publicly stated that it "does not advocate a full ban on open source models", but added three caveats: no sales of advanced chips and manufacturing equipment to China, cracking down on industrial distillation, and all sufficiently powerful models must pass mandatory security testing. This position is viewed by multiple Chinese independent accounts as targeted terms rather than neutral security advocacy. @NousResearch and @openclaw have publicly joined the alliance, and @AndrewYNg has publicly seconded Jensen Huang’s open weight position.<\/p>\n
Sources:<\/p>\n
SSI, founded by Ilya Sutskever, has announced a long-term strategic partnership with NVIDIA, with compute power set to increase to 10x current levels over the next 12 months, and NVIDIA will co-shape the design direction of future computing platforms with SSI. SSI stated in its announcement that its core research has passed basic validation and entered the "scalable" stage. Multiple bloggers have expanded discussion around this milestone: @Gorden_Sun notes that SSI has already begun large-scale scaling of compute power, interpreting this as "a major research problem has been solved, ready for production"; @imjustnewatai, retweeted by @MaxForAI, combines Ilya’s podcast statements, the public background of SSI members (Daniel Levy, Yair Carmon, Hugh Zhang, etc.) and NVIDIA’s investment pace to speculate that SSI is researching a new mechanism that allows models to continuously learn after deployment, contrasting with the current pre-training paradigm. @oran_ge has concurrently confirmed that SSI and NVIDIA’s strategic partnership has entered the scaling preparation phase after basic validation.<\/p>\n
Sources:<\/p>\n
Statistics: Number of timeline posts scanned=360\/360, Number of bloggers matched=40\/40, Total matched tweets=212\/212, Weighted tweet score=168.2\/168.2, Original tweets=82\/82, Retweeted tweets=39\/39, Crawl attempts=2\/2, Boundary coverage status=tail_confidently_crossed_target_boundary\/tail_confidently_crossed_target_boundary<\/p>\n","protected":false},"excerpt":{"rendered":"
Today\u2019s top stories include data copyright disputes for AI companies, new talent movement trends between China and the US, and technology supply chain investigations amid geopolitical tensions.<\/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-1518","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\/1518","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=1518"}],"version-history":[{"count":0,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1518\/revisions"}],"wp:attachment":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1518"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1518"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1518"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}