The checkout button quietly became a loan, and the regulator is unpicking it one disclosure at a time
Your takeaway order was a credit application
A 110-yuan (~$15) meal arrived a month later as a debt. A Beijing customer ordered food, hit one-click pay in a hurry, and later got a notice that a loan was due — the payment method had defaulted to "monthly pay," per TMTPost. She had manually switched back to her bank card every other time. The industry's term for this is lending you don't feel.
Hubei's financial regulator wrote down the mechanics in September: instalment credit placed first in the payment list, key terms faded or folded away, pop-ups pushed until the line between paying and borrowing disappears. On the Black Cat complaints platform, individual lending products carry complaint counts above ten thousand, some past a hundred thousand, clustered on enrolment without consent.
The margin explains the enthusiasm. Credit carries gross margins above 95% against 15–60% for selling goods. China's online micro-lending book passed 5tn yuan (~$704bn) by 2025, with Ant, ByteDance, JD and about ten head platforms holding 4.1tn (~$577bn) of it — over 80%. Seven listed loan-facilitation platforms booked roughly 70bn yuan (~$9.9bn) of 2025 revenue on 991.9bn yuan (~$140bn) of loans arranged.
Then the rules landed. October 2025's regime capped rates at 24%, whitelisted partner banks, barred outsourcing core risk control and banned sub-delegation — which killed the trade in re-routing rejected borrowers through layers of middlemen. From 15 March 2026 the financial regulator and central bank require every cost component itemised. Five head platforms have been summoned. The essay is the third in three days from TMTPost's pen-named house writer, Shadow Memo.
Go deeper on this section: ClaudeChatGPT
Washington's rule reaches inside the device
From next Tuesday the FCC looks past the nameplate. A device cannot get US equipment authorisation if it contains a logic-bearing hardware component made by a Covered List entity — optical transceivers named explicitly. Akin Gump dates the prohibition to 13 October; applications already pending are exempt unless a part gets swapped.
Which bans nothing yet, and that is the point. ChinaTalk's Quinn, who came to the outlet from semiconductor physics at Yale, notes the largest Chinese transceiver vendors are not on the Covered List. Markets traded the headline anyway: after Reuters reported a draft ban on 4 August, Zhongji Innolight, the world's biggest transceiver vendor, fell 10% in a week while Coherent rose 17%.
The stack refuses to cooperate. Chinese firms dominate modules, depend on Western suppliers in the middle, and at the bottom — wafers and basic metals — Chinese control is ingrained. Precedent defines "foreign-produced" by the Buy American test, so the word "Chinese" never appears in the rule. The hyperscaler-and-optics trade group filed on 13 August asking the FCC to leave transceivers alone, citing data-centre build-out.
Go deeper on this section: ClaudeChatGPT
Kling's ticker and DeepSeek's plugin stack
Kuaishou is giving its video model its own shareholders. Kling has lined up CICC, Goldman and UBS for a Hong Kong listing raising at least $1bn, filing early 2027. Q2 revenue topped 850m yuan (~$120m), up over 200%, against group revenue of 35.535bn yuan (~$5bn) growing 1.4%, TMTPost notes — so anyone who only wants the AI video business can soon buy it directly.
The voting table is where the real carve-out happened. Kuaishou keeps 68.33% of equity but 53.80% of votes. Chief executive and chairman Cheng Yixiao was granted 1% of the enlarged subsidiary, worth 0.79% of votes; Kling's own chief executive Gai Kun holds 3% carrying 23.62% of votes, forfeited if he leaves. Co-founder Su Hua got nothing personally. Kuaishou also promised not to control any rival video-generation company.
Meanwhile DeepSeek's agent runtime rests on a chatbot kernel. A long V2EX teardown examines Cordis, the plugin meta-framework under DeepSeek Harness — originally the core of the Koishi chatbot project, with 6,000-plus third-party plugins now indexed. Its whole mechanism is one stack: acquire a resource, push its release function, pop in reverse on unload.
The verdict is fair and unflattering. Reverse order works only because variable scope already puts dependencies first; register backwards and you get the wrong teardown order with no error. Resources opened outside the wrapper leak silently. And async disposers are not guaranteed to finish in reverse, leaving a half-unloaded window in which a service is deregistered while its timer still fires.
Go deeper on this section: ClaudeChatGPT
The leaderboard's fine print said MiniMax
A US startup took second place worldwide on Chinese weights. Artificial Analysis's text-to-video board on 30 September put Utopai X at 1150, behind only Wan 3.0 and top among American models. Chinese outlets ran it as a self-developed custom model; roughly 90% skipped the line reading "based on MiniMax H3," TMTPost's rebuttal says.
The arithmetic is brutal. The Shanghai lab's H3 base scores 1136; Hollywood post-training bought 14 points, and ByteDance's Seedance 2.5 sits at 1146 — four points back, with a generation-feedback loop that never waits for anyone. Utopai's inference runs about 17 times H3's API price per minute. H3 opened its weights on 3 August; the leaderboard win came eight weeks later.
Go deeper on this section: ClaudeChatGPT
Anthropic plays player and referee
The 154-page misuse report lands badly in Chinese. Initium's podcast walks through Anthropic's catalogue of Claude abuse — surveillance of dissidents, drafting stability-maintenance reports, dating-app fraud — alongside its accusation that Chinese firms are hijacking and illicitly distilling US models, and makes the obvious objection: Anthropic is investigator, data holder and commercially interested party at once.
On the ground it looks like a payments problem. A V2EX poster says account shops can no longer complete purchases — the failure message is a declined transaction, not a ban — and the circulating fixes are hand-registered aged accounts on clean IPs with card-prefix testing, or a country-matched Gmail plus a physical card. Sentiment, not data.
Go deeper on this section: ClaudeChatGPT
AI prose is eating the Chinese tech press
Developers have stopped being able to read their own tools. QbitAI's complaint is that coding-tuned models now deliver reports in compressed jargon — ledgers "15 of 15 green", a dead parallel path called "one leg misfiring" — and blames token-saving chains of thought as much as coding overfit, claiming caveman phrasing cuts consumption 70%. A GitHub skill teaching models to drop words has 104,000 stars.
Then there is what got published today. Leiphone, which also files real chip reporting, ran four instalments of a self-declared 188-volume codex for governing AI cross-domain transfer, complete with permanently locked axioms, hash-anchored volumes and a taxonomy of four "negative transfer" prototypes. No author, no institution, unfalsifiable by construction. My read: it is machine output at scale, and it is on a masthead that matters.
Go deeper on this section: ClaudeChatGPT
Threads we are pulling
- Europe went from letter to vote. Following Monday's Macron–Merz letter and yesterday's note on Germany going public: Parliament adopted its China recommendation 454–86 with 110 abstentions, calling Beijing a "decisive enabler" of Russia and its industrial policy an "existential threat", and urging a continuous naval presence around the Taiwan Strait, per SCMP. Trade chief Maroš Šefčovič meets commerce minister Wang Wentao today and tomorrow in what Brussels treats as a deadline.
- Seres is hedging away from Huawei. Following my AITO pieces: the AIVA brand — a Chongqing state fund holding 34.5% after a 6.671bn-yuan (~$940m) raise, Seres second at 32.96% — builds its ME7 on ByteDance's Volcano Engine models and DeepRoute's no-HD-map driving stack, not Huawei's, with channels separate from AITO, per IT Home. Huawei's consumer chief Richard Yu put September deliveries at 37,490 and the cumulative total at 1.56m.
- The holiday closed as it opened. Shanghai hosted 20.21m visitors over 1–7 October and expects 29.822bn yuan (~$4.2bn) of total tourism spend, up 3.52% while 7 October arrivals rose 13.25% — Jiemian's numbers, and the third time this week traffic beat conversion. Ant's health app says 600,000 users gained more than a million pounds between them.
Go deeper on this section: ClaudeChatGPT