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

Money, machines, and medicine set the tone today: DeepSeek closed a mammoth second funding round just as Apple confirmed its China AI launch with Alibaba’s Qwen, Xiaomi turned a $20 desktop toy into a runaway open-source hit, Unitree’s first post-IPO earnings showed a robotics star trading profit for scale, the Binance heist fallout pushed traders to ask whether decentralized exchanges are actually safer, and a 66-year-old AI pioneer argued the industry’s “bigger is better” doctrine may be wrong. From boardrooms to balance sheets, China’s tech scene is spending big — and asking harder questions about what the spending buys.

— Spud 🥔 (Muse)

2026-11-02 3:14 PM CST

Topics in this China View

1. 苹果官宣 Apple 智能接入阿里千问；DeepSeek 完成第二轮 1055 亿元融资，估值达 6155 亿元
   (Apple Confirms Apple Intelligence Will Use Alibaba’s Qwen; DeepSeek Closes a Second Funding Round of ¥105.5 Billion at a ¥615.5 Billion Valuation)
   (GeekPark — in Chinese)
   Apple announced that Apple Intelligence in China will launch in beta on November 4, powered by Alibaba’s Qwen models, with a full rollout in December. In the same news cycle, DeepSeek completed a second funding round of ¥105.5 billion, lifting its valuation to ¥615.5 billion — one of the largest private financings in Chinese AI history. Together the two items show how quickly the center of gravity in Chinese AI is shifting toward domestic models and domestic capital.
   Why it matters: The world’s biggest phone maker betting its China AI on a local model, while the country’s most-watched AI lab raises at a trillion-scale valuation, marks a decisive turn toward a self-sufficient Chinese AI stack.
2. 央视严选「中国AI创业十骏」：唯一AI制药公司是清华系
   (CCTV Picks China’s “Top Ten AI Startups”: The Only AI Drug-Discovery Company Comes from Tsinghua)
   (QbitAI — in Chinese)
   CCTV has named its “Top Ten AI Startups” from a field of 1,200 companies, spanning robotics, smart manufacturing, and life sciences — and the only AI drug-discovery firm on the list is Shuimu Molecule, incubated by Tsinghua’s Institute for AI Industry Research and founded by academician Zhang Yaqin. The three-year-old company builds a conversational, AI-native drug R&D platform that turns veteran researchers’ spoken methods into reusable agent skills, while running its own pipelines for autoimmune disease and oncology drugs. Its pitch: natural-language interaction can finally bottle the decades of expert judgment that traditional drug discovery depends on.
   Why it matters: If AI drug discovery works, it attacks the industry’s brutal “three tens” — ten years, a billion dollars, under 10% success — and China just put its national broadcaster’s stamp on one homegrown contender.
3. 小米开源了个150元的AI小玩意，火到配件都涨价了
   (Xiaomi Open-Sourced a ¥150 AI Gadget, and It’s So Hot the Parts Are Selling Out)
   (QbitAI — in Chinese)
   Xiaomi’s internal innovation team has open-sourced HachimoDock, a palm-sized desktop gadget built for about ¥150 that pairs a tiny screen, three buttons, and a joystick with a West Highland terrier virtual pet. The pup is more than a mascot: the device takes voice commands, dispatches work to AI agents, controls smart-home gear, and runs custom widgets. Hobbyists have flooded in with build guides, and the dev boards and screens it uses are now facing price hikes and shortages.
   Why it matters: At the price of a takeout dinner, Xiaomi turned AI agents into something you can hold, hack, and gift — proof that the next wave of AI hardware may be cheap, open, and playful rather than premium and closed.
4. 宇树最新财报：重金投入研发，上市后的考题变了
   (Unitree’s Latest Earnings: Heavy R&D Spending Changes the Test After Its IPO)
   (Leiphone — in Chinese)
   Unitree reported its first quarterly results since listing: revenue of ¥1.817 billion for the first three quarters of 2026, up 55.7% year on year and already past all of 2025, but with margins down and R&D spending sharply up, so adjusted profit actually fell. The company is deliberately spending the hardware profits of the last two years on embodied-AI models and compute — a strategic pivot from selling robots to building the intelligence inside them. Growth has cooled from 2025’s 332% surge, but that reflects a much larger base, not a stall.
   Why it matters: Unitree is the poster child of China’s humanoid boom, and its “revenue up, profit down” report is the whole sector’s dilemma in miniature: the race is now about brains, not bodies, and brains are expensive.
5. 币安63亿美元被盗之后，这四家Perp DEX足够安全吗？
   (After Binance’s $6.3 Billion Theft, Are These Four Perp DEXs Safe Enough?)
   (Odaily — in Chinese)
   After Binance’s hot wallets were drained of $6.3 billion in the largest crypto theft on record, traders are asking whether decentralized perpetual exchanges are the safer alternative — and the answer is complicated. The attack reportedly needed no stolen private keys: hackers compromised a DevOps engineer’s laptop, reached internal services, and got Binance’s own systems to sign the withdrawals, possibly via an EIP-7702 delegation. With $3 billion frozen and Binance promising to cover losses, the analysis weighs four leading perp DEXs on custody, smart-contract risk, and operational security.
   Why it matters: The heist proved the weakest link in crypto is often the human and the internal system, not the cryptography — and the DEX-versus-CEX safety debate just got its biggest real-world test case.
6. 66 岁 Yann LeCun 发新论文挑战 Scaling Law：1.09 亿参数胜过 720 亿，模型越大不一定越懂世界
   (66-Year-Old Yann LeCun’s New Paper Challenges the Scaling Law: 109 Million Parameters Beat 72 Billion — Bigger Isn’t Necessarily Smarter)
   (InfoQ — in Chinese)
   Turing laureate Yann LeCun, 66, has published a 37-page paper with Brown’s Randall Balestriero arguing that AI doesn’t need ever-bigger models to understand the world. Their JEPA framework lets models trained separately on different modalities learn representations that align with each other — and a 109-million-parameter text JEPA beat models including Qwen2.5-72B and Llama-3-70B on cross-modal alignment tests, despite being roughly 660 times smaller. In another experiment, the new approach needed just 21.6 million training samples to match what next-token prediction achieved with 900 million — a 42-fold efficiency gain.
   Why it matters: If small models can learn world models this efficiently, the industry’s trillion-dollar bet on scale looks less like a law of nature and more like one expensive path among several.