Daily Highlights
- DHS orders all DeepSeek products out of US federal procurement
- Altman lists three AI endgame paths, including "melting down rivals' GPUs"
- Nvidia's Vera CPU wins its first major design win at Modal
- Moonshot AI cleared for HK listing this month, eyeing $1B raise
- HKUST open-sources the Fourth Institute human-agent research platform
TODAY'S HEADLINES
[DHS issues DeepSeek exclusion order: federal civilian agencies must stop buying and remove related products immediately] On Oct 19, the US Department of Homeland Security issued an exclusion and removal order under the Federal Acquisition Supply Chain Security Act of 2018, published in the Federal Register on Oct 20: all federal civilian agencies must immediately stop procuring DeepSeek's large models and related services, and third-party products and services built on DeepSeek technology must be removed from agency information systems, effective immediately with waivers available through established procedures. It is the first time the US has systematically applied its supply-chain security legal toolkit to a Chinese frontier AI model developer. (Source: Sina Finance/Caijing)
[Altman lays out three AI endgame paths for staff: international oversight, lab mergers, or a compute monopoly] The Information reported on Oct 19 that OpenAI CEO Sam Altman described three safety paths under industry discussion at an internal session last week: an international oversight body modeled on the IAEA, merging the major AI labs into a single entity, or one company monopolizing compute and "melting down all rivals' GPUs." He said OpenAI will need to decide within a year which path to push, while stressing that he wants humans to stay in control of the technology and industry power to stay distributed - remarks that land as the Trump administration prepares its first oversight framework for "superintelligence" developers. (Source: The Information/NetEase Tech)
[Nvidia's Vera CPU lands AI cloud startup Modal as agent workloads crack the CPU duopoly] Reuters reported on Oct 19 that AI cloud platform Modal will run agentic workloads on Nvidia's next-generation Vera CPU, its first break from an all-Intel-and-AMD fleet; early tests showed Vera doubling the amount of work agents can complete, and Modal will rework its software stack to accommodate the chip. The CPU market has been locked between two incumbents for decades, and Nvidia is prying it open through agent workloads. (Source: Reuters)
NEW PRODUCTS & TECH
[MiniMax ships V4.1-Flash build 1017: self-feedback reinforcement learning strengthens coding] MiniMax rolled out the 1017 build of V4.1-Flash on Oct 18: using reinforcement learning based on self-feedback, the model runs the code it writes, checks it and fixes it on its own; official demos show it completing full projects such as a browser racing game and a Godot recreation of Portal 2 from minimal prompts. Community testing found all 30 coding tasks passed with high reasoning effort at about 4.5 cents per task, though each task burns roughly 550,000 tokens, and the model's 42.6 on HLE expert questions still trails Claude Opus 5's 56.3. (Source: QbitAI)
IPO & FUNDINGS
[Zhiyuan Robotics nears closing a new round at roughly RMB 50B valuation as its HK IPO slips to 2027] According to Caijing, embodied-AI unicorn Zhiyuan Innovation is wrapping up a new funding round of about RMB 1 to 1.5 billion at a post-money valuation near RMB 50 billion; under HKEX rules, a company that raises new funding after filing but before listing must wait 120 days before its shares can trade, so the company cannot complete its IPO this year and will ring the bell at the earliest in late February 2027. Zhiyuan posted RMB 1.05 billion revenue in 2025, and Counterpoint data ranks it first globally in humanoid shipments for H1 2026 with 9,700 units. (Source: Caijing)
[Moonshot AI cleared for Hong Kong listing, launching a roughly $1B IPO this month as the first LLM stock] The South China Morning Post reported on Oct 20 that Kimi developer Moonshot AI has won HKEX listing approval and will launch its roadshow this month to raise about $1 billion, on track to become the first large-model company to go public in Hong Kong. Squeezed by open-source rivals DeepSeek and Alibaba's Qwen, the closed-source company is treating its IPO war chest as ammunition for the next phase of the race. (Source: SCMP)
OPEN SOURCE & RESEARCH
[HKUST open-sources the Fourth Institute: an automated research platform where humans and agent societies create together] A Hong Kong University of Science and Technology team led by Prof. Kwang-Ting Cheng and Prof. Ma Songchen open-sourced the Fourth Institute on Oct 20: every agent carries a persistent Resident identity with cross-session research history, linear chains of thought are woven into a web that can cite, rebut and recombine across research lines, and humans can join with their own AI accounts to steer the work. The platform has accumulated more than 1.86 million thought-and-execution traces and over 5,100 hours of agent R&D; its first embodied-AI project completed inference optimization of dozens of models across multiple GPU platforms within 24 hours, with some results reaching SOTA. (Source: Sina Finance)
COMPANY & PEOPLE
[DeepSeek consolidates a unified product division under Cui Tianyi as an R1 core author departs] LatePost-affiliated outlet Silicon Star reported on Oct 20 that DeepSeek merged its seven-month-old product team with its Web, App and Platform lines into a unified product division on Oct 16, led by Cui Tianyi, who joined only seven months ago and previously built the internal agent framework Harness; the report also noted that Guo Daya, a core author of R1, left several months ago to bet on agents. (Source: Silicon Star)
Powered by AOYII.com
