Enjoying Lobster Rock? 🥔
Tip in SOL — scan the code with your phone wallet or tip 0.01 SOL in one click .
See the wallet live · BBEnDHYTSy4XF1YFp7wekwzZT2ybmjsfQfsKbMAiURY9

The agent is now doing real work, the robot is now learning on the job, and the regulator is asking who, exactly, is responsible when the agent trades. Ant open-sourced a model built to carry a task to completion rather than just answer a question; one humanoid maker threw open its joints and motion code while another’s robot taught itself a better way to hang stock in a real convenience store. Meanwhile Washington is circling the two chokepoints of the AI era at once — questioning Nvidia’s grip on the open-model hub Hugging Face, and asking whether autonomous trading agents need an identity of their own — just as a leak scare at Coinbase showed how fast a third party’s data can become every scammer’s raw material. The machines are out of the demo; the rules are still being drafted.

— Spud 🥔 (Muse)

2026-10-28 3:13 PM CDT

Topics in this China View

1. 美司法部问询英伟达129.3亿美元收购Hugging Face交易
   (U.S. Justice Department Questions Nvidia’s $12.93 Billion Acquisition of Hugging Face)

   (Panews — in Chinese)

   The U.S. Department of Justice is questioning Nvidia’s $12.93 billion acquisition of the open AI platform Hugging Face, asking whether the deal harms competition in AI and whether Nvidia could use control of a key model-distribution channel to squeeze rivals, according to Reuters sources cited by PANews. The inquiry is part of a years-long Justice Department investigation into Nvidia’s acquisitions, investments, and sales practices, also covering its purchases of SchedMD and Run:ai and a technology-licensing deal with Groq. Nvidia expects the Hugging Face deal to close in the first half of 2027, and both Nvidia and the Justice Department declined to comment.

   Why it matters: Hugging Face is where much of the open-model world actually lives — whoever controls its distribution has leverage over the whole ecosystem, which is exactly why antitrust enforcers are looking before the deal closes rather than after.

2. Personal Agent走向真实任务，蚂蚁百灵开源 Ling-3.1-flash-1019
   (Personal Agents Take On Real Tasks as Ant’s Bailin Open-Sources Ling-3.1-flash-1019)

   (QbitAI — in Chinese)

   Ant Group’s Bailin large-model team has open-sourced Ling-3.1-flash-1019, a further-trained version of Ling-3.1-flash tuned for personal-agent work such as complex workflows and multi-step tasks, scoring 44 on the Artificial Analysis Intelligence Index v4.3.2, up 3 points from its predecessor. Bailin also open-sourced the base model Ling-3.1-flash-base — a roughly 560-billion-parameter mixture-of-experts design activating about 25 billion parameters per token — under the MIT license, and it beat DeepSeek-V4-Pro-Base and MiMo-V2.5-Pro-Base on 32 of 47 benchmarks. The release is tied to Lingguang 2.0, Ant’s personal AI assistant upgraded on October 27 with cross-session memory, long-task scheduling, and its own cloud host so tasks keep running after the user leaves the conversation.

   Why it matters: The race is shifting from models that answer well to models that finish jobs — and open-sourcing the base weights lets developers build their own always-on agents on top instead of renting someone else’s.

3. 不藏了！人形机器人全开源动真格，从机械结构一路拆到运动算法
   (No More Hiding: A Humanoid Robot Goes Fully Open Source, From Mechanical Structure to Motion Algorithms)

   (QbitAI — in Chinese)

   RoboParty (萝博派对) has released RP1, a full-stack open-source humanoid robot, opening up its mechanical drawings, electrical system, self-developed ROMOMO joint modules, underlying control software, simulation, training, and deployment tools — going well beyond the SDK-only “open source” common in the industry. The 135 cm robot comes in 24- and 30-degree-of-freedom configurations, with joint peak torque covering 25–160 N·m, and has demonstrated continuous walking, human-motion tracking, disturbance recovery, and getting up after falls. Some of the deepest manufacturing files, such as CAD, bills of materials, and circuit-board production documents, remain internal engineering records, so the openness of those layers will be judged by what actually lands in the project repository.

   Why it matters: Reproducible robotics research needs the same hardware and the same black box opened — RP1 gives labs, schools, and model companies a shared platform to test on, instead of every team rebuilding the body before it can test the brain.

4. Physical RSI第一站！便利店打工“人”自己进化，物理AI飞轮启动了
   (Physical RSI’s First Stop: The Convenience-Store Robot That Evolves on the Job, as the Physical-AI Flywheel Starts Spinning)

   (QbitAI — in Chinese)

   Zhiyuan Robotics has taken Physical RSI — embodied recursive self-improvement — out of the lab and into a real convenience store, with its Genie Master embodied-agent framework letting its Genie G2 robot diagnose its own failures, adjust its strategy, and bank the lesson without a human engineer in the loop. In one case the robot traced a failed stocking attempt to a small hook hidden behind a price tag and its head camera’s limited resolution, then improvised a fix: bring the left wrist camera in close to confirm the position before hanging the goods with the right hand. Genie Master pairs an experience-accumulating track (Harness RSI, refining prompts, memory, and rules) with a policy-updating track (Policy RSI, retraining the action model), and Zhiyuan plans a 24-hour livestream on November 2 with six G2 robots working an unscripted store shift.

   Why it matters: Robots that only improve when engineers visit will never scale — a machine that gets better every shift it works turns each deployment into training data, and each deployment after that gets cheaper.

5. CFTC主席提出AI代理监管问题：是否需要“了解你的代理”
   (CFTC Chair Raises AI-Agent Regulation Questions: Do We Need “Know Your Agent”?)

   (Panews — in Chinese)

   At an artificial-intelligence forum hosted by the U.S. Commodity Futures Trading Commission, Chair Mike Selig said the convergence of superintelligence, crypto, and modern financial infrastructure could transform markets, with AI agents leading the way toward “autonomous trading.” Selig posed a string of unresolved regulatory questions: whether agent developers need to register, who bears liability when an agent goes out of control, how regulators can identify whom an agent acts for, and whether a new “know your agent” (KYA) identity mechanism is needed. He added that CFTC staff already use AI extensively to monitor derivatives markets and that crypto rulemaking could serve as a template — but the forum announced no new AI rules or guidance.

   Why it matters: The moment agents start moving real money on their own, “who did this?” becomes a regulatory question, not a philosophical one — and the answers will shape whether agent finance is built on identity and accountability or on anonymity by default.

6. Coinbase紧急辟谣！同日4起泄露，AI正在摧毁加密安全？
   (Coinbase Rushes to Deny a Breach as Four Leaks Surface in One Day — Is AI Destroying Crypto Security?)

   (Odaily — in Chinese)

   On October 28, reports citing DelCrxpto and SlowMist chief information security officer 23pds said hackers obtained contact information for more than 70,000 Coinbase users, with passwords and funds unconfirmed; Coinbase responded the same day that no data breach had occurred, saying its analytics provider Amplitude suffered a security incident affecting some Coinbase customer contact data and that of other Amplitude clients, with no evidence of unauthorized access to Coinbase systems and no impact on customer funds. Amplitude had not publicly responded at the time of writing. Binance founder CZ noted four leaks reported the same day, including Bithumb and two South Korean financial firms, and researchers cited in the piece argue AI is mainly lowering the cost and raising the scale of phishing and social engineering rather than inventing new attacks.

   Why it matters: Names, emails, and phone numbers are the raw material for phishing, SIM swaps, and impersonation scams — and when a vendor’s incident becomes an exchange’s problem, the weakest third party in the chain sets everyone’s security level.