According to Max For AI’s summary of a Federal Register document, the U.S. Department of Homeland Security’s order regarding DeepSeek applies to federal civilian executive agencies and is not limited to direct model procurement: third-party products and services that incorporate DeepSeek are also covered. The document requires the removal of relevant products from procurement channels and the cleansing of such products already in use within agency IT systems. National security and isolated classified systems have separate boundaries; the defense and intelligence communities are outside the scope of this order’s authority. The order includes an exemption application process. The author notes that AI applications targeting the U.S. government market, even under their own brand, must verify the eligibility of their underlying model suppliers.<\/p>\n
Source:<\/p>\n
According to Max For AI’s summary of a U.S. Treasury OFAC announcement, entities added to the list on October 19 include MiraclePlus’s investment management company, its RMB Phase I and Phase II funds, and its Zhiyuan enterprise management entity\u2014four entities in total\u2014as well as Shanghai Zhisuan Technology. The account states that the measures are based on the “Protecting American Intellectual Property Act of 2022” and are related to the case of former Google engineer Ding Linwei’s alleged theft of trade secrets. The account’s analysis regarding the impact on portfolio companies’ financing, exits, and equity compliance is presented as an author’s note. Specific impacts should be verified based on the sanctioned parties’ equity stakes and transaction structures and should not be assumed to automatically restrict all portfolio companies.<\/p>\n
Source:<\/p>\n
Cell’s continuous analysis on AI commerce points out that the key to agent shopping is not just “placing orders on behalf of users”: candidate product recall, ranking, description, and the final transaction can all be controlled by the same system, shifting the entry point from search pages or information feeds to the assistant. A cited example states that Perplexity’s Comet can, with user authorization, log into an Amazon account to compare prices and place orders. The post reviews Amazon’s lawsuit, preliminary injunction, and the appellate court’s overturning of the injunction, noting that the case has been remanded to the district court and the legal dispute is not yet resolved. The author also cites Cloudflare’s bot\/user visit ratio and Pew click-through data to discuss how AI answers may reduce website traffic; these figures are research data relayed by the account.<\/p>\n
Source:<\/p>\n
Research results compiled by Cell are inconsistent: a study tracking 5,172 customer service agents reported that AI assistants increased the average number of problems solved per hour by approximately 15%, with greater benefits for novices. In a METR randomized experiment with 16 senior open-source developers and 246 familiar codebase tasks, tasks performed with AI at the time actually took 19% longer. The account also mentions a Danish administrative records study finding that the impact on worker income and hours was close to zero about two years after ChatGPT’s launch, despite some task reorganization. Differences in samples, tasks, and tool generations mean the materials support the conclusion that effects vary by scenario, not that “AI universally boosts efficiency” or “AI is ineffective.”<\/p>\n
Source:<\/p>\n
Max For AI relays a Reuters report: a Chinese-language user, leveraging the open-source multi-agent framework Cairn, attempted to attack over 1,200 websites involving more than 100 companies; researchers estimate the operation consumed approximately $5,000 in tokens, with a successful intrusion taking nearly 23 hours on average. The post states that researchers observed task logs from a control panel exposed to the public internet and speculate that the execution combination may have involved Claude Code and DeepSeek; this tool combination is speculative, not a confirmed fact. This case demonstrates that automated agents can be used for prolonged, multi-step attack attempts. Specific technical attribution should be based on the original report and researcher materials.<\/p>\n
Source:<\/p>\n