Beijing is tightening who gets to speak before a plenum about discipline
Governors installed, one economist muted
The plenum opens Monday. The Central Committee's fifth plenary session runs 26 to 29 October on what the Party calls "full and rigorous" self-governance, meaning discipline. AFP counts eight vacant full-member seats and 15 more members under investigation. It expects the lowest turnout since Mao died in 1976.
Provinces got their teams first. Anhui, Henan, Shanxi and Xinjiang finished party congresses last week. Hebei's legislature made Zhao Chenxin acting governor. He was a vice chairman of the National Development and Reform Commission, the central economic planner. He succeeds Wang Zhengpu, governor since 2021. One appointment proves little; a pipeline of them, filled before a discipline plenum, does.
Ren Zeping went quiet. On Saturday his WeChat account stopped accepting followers. His team says it faces a 15-day posting suspension and blames platform adjustments under new regulations. Ren is a widely followed commentator and, per TechFlow, chief economist at Zhongyuan Bank. The likely trigger is rules in force since late September. They bar anyone but financial institutions and platforms from marketing financial products, even in disguise.
Message or business model? The day before, his account argued stocks had bottomed and a rebound was under way. In July a member of his 2,980 yuan (~$420) paid membership claimed leveraged losses above 10 million yuan. I can't separate the two, and the rule doesn't have to. With ChiNext, Shenzhen's growth-stock board, down 27.8% in the third quarter, a bull telling retail investors the bottom is in is not a message Beijing needs sold. Inference.
Beijing speaks to one half of Taiwan. On the 81st anniversary of Japanese colonial rule ending, Song Tao, head of the Taiwan Affairs Office, pledged in Fuzhou to back pro-unification "patriots". Beijing made the date an official memorial only last year. The DPP downplays it and the KMT leans in. In Taipei's mayoral race, the DPP's Puma Shen, known for civil-defence work, has closed on KMT incumbent Chiang Wan-an. My read: the audience is a ballot.
Go deeper on this section: ClaudeChatGPT
Kimi gets a victory lap at home and an audit abroad
Two lead stories, one model. Chinese tech press led with rank. New Wisdom's headline has Moonshot's Kimi K3.1 Pro breaking into the global top four and first on agents. TMTPost's hands-on singles out a demo of a virtual human learning to walk. When reward tuning failed to fix the arm swing, the model audited its own simulator and corrected 36 shoulder-muscle calculation errors.
Western outlets led with the brief. VentureBeat pairs the launch with allegations that creators were told to promote Kimi without a paid-partnership label. X user Elshayib posted screenshots of a brief that also required drafts to be approved first. Nobody has authenticated it, and VentureBeat found no confirmed company response.
Pekingnology reads it as habit. The China newsletter argues that company-approved, undisclosed paid posts are routine in Chinese media, so a team might export the playbook unthinking. It also flags talking points casting Kimi as a national win over US chip curbs. A Chinese-language report on the allegations is now unreachable at its original link, with no evidence Moonshot caused that. THE TELL: the part of TMTPost's review I could read never mentions any of it.
The spec sheet. Latent Space lists K3.1 Pro as a mixture-of-experts with 2.8 trillion total parameters, 104 billion active per token and a 1-million-token context. Yesterday I said "roughly 3 trillion." Moonshot's own Terminal-Bench 4.0 score was 65.7%; TMTPost's review gives 67.2% and first place. The text doesn't say whether that is a different harness or run.
A deliberate cyber haircut. Moonshot says it removed cyber-attack material from training data and reinforcement-learning environments. On a controlled exploit-finding test the restricted model scored 81.0%, against 90.5% for a version without limits, GeekPark relays from Reuters. Weights are due by 4 November. My guess is fine-tuning recovers some of that gap.
The hackers' contest is bowing out. Organisers of Geekcon's Shanghai edition, which closed Saturday, billed it as possibly the last human hacker contest. They wrote that it nearly didn't happen because AI has all but replaced human hackers. Anonymous contestants still took over a mainstream-brand drone remotely and redirected a banking app's transfers with one link.
Go deeper on this section: ClaudeChatGPT
ByteDance owns the clock; Tencent opens a drama front
The gap is widening. QuestMobile data, cited in a TMTPost commentary, put ByteDance at 41.8% of August usage time across the top 50 apps against Tencent's 28.6%. The gap was 3.3 points a year earlier. ByteDance's first-half revenue was about $120 billion, up 30%, against $134.3 billion for Alibaba and Tencent combined, per The Information as quoted there. Tencent still leads on unduplicated users.
Tencent answers inside WeChat. On 15 October it launched a free short-drama mini-program, Green Bubble. Through 31 December it pays producers an incentive pegged to per-viewer watch time, up to 100%. ByteDance's Hongguo, with 168 million daily users, stops issuing coins to new users on 1 November. TMTPost reads that as a leader choosing quality while a challenger still buys attention.
Yesterday's read, amended. I tied the coin cut to the 1 September rules on addictive algorithms. Today's reporting adds a rival's money. Both can be true.
The audience is tired of clones. A line from an AI food-business drama, roughly "pricey, but it won't give you heartburn," passed 4 billion plays as a meme because every clone's plot is guessable. DataEye figures in the same piece say under 1.3% of 221,900 AI comic-dramas on Douyin in the first half reached the 50 million plays needed to break even.
Go deeper on this section: ClaudeChatGPT
Xpeng's "old guy" row won't die
An apology didn't end it. At the 15 October presale for the MONA L05, Xpeng founder He Xiaopeng handed the stage to a young team, and product manager Etta kept mocking management and wood-and-chrome trim with a slang jibe for fusty older men. Etta apologised on 22 October. Two days later Xpeng VP Yu Tao called critics overreaching and offered 1 million yuan (~$141,000) to anyone proving Xpeng staged the row. Etta still presents at the 29 October launch.
WHY IT MATTERS: a company chose a generational insult as its brand voice, then defended it. TMTPost's commentator argues age labels turn a product launch into an identity fight. The car became the second story.
Go deeper on this section: ClaudeChatGPT
Threads we are pulling
- The plenum date: I leaned on a wrong window. I relied on SCMP's Monday-to-Wednesday window and doubted 26 to 29 October. CGTN confirms Monday to Thursday. My test stands: does the closing communique repeat the "core" language from Xi's Thursday Long March speech?
- Moonshot and the export draft. I still see no report that Moonshot registered under Beijing's draft AI export rules, and today's feed has nothing. Its weights ship by 4 November.
- iPhone Duo's second week. After the scalper premium faded, GeekPark reports bright pixel lines and cracks at the hinge within hours of unboxing. One Australian user's screen split after fewer than 20 folds. Apple hasn't commented. Out-of-warranty inner-screen repair is listed at 7,998 yuan (~$1,126), against 628 (~$88) with AppleCare+.
- The memory windfall has a payroll. Storage maker Demingli reported third-quarter net profit of 2.73 billion yuan (~$384m), up 2,902.75%. Nine-month profit was 8.75 billion (~$1.23bn), against 688 million (~$97m) for all of 2025. In Taiwan, Micron's Taoyuan union rejected a profit-sharing offer on Friday. Both unions had asked for a one-off bonus of 83 months' pay. Talks resume 27 October.
- Robots ship. On 23 October I wrote about Unitree's sliding valuation. Qiyuan Robotics says its 5,000th personal robot came off the line five weeks after sales opened. It claims one every 2.5 minutes, with daily capacity up from 10-plus to 200-plus. Self-reported.
- Nagqu shook again. A magnitude 5.7 quake hit Baqing County at 7:25pm Sunday Beijing time, a minute after a 4.7. The county received 300 million yuan (~$42m) in relief after Thursday's 7.0. Today's feed has no casualty figures.
Go deeper on this section: ClaudeChatGPT