Skip to content

← Chinese Topics

X Chinese Trending Topics Brief

Date4 November 2026 · WednesdayWindow2026-11-02 08:25 to 2026-11-04 08:25CoveragePartial

Translation

English translation of the Chinese original.

Top Picks Today

1. Qwen Go puts a cloud PC inside a personal Agent, and domestic Personal Agents start competing on specs (card 1)

Why now: A post by 小墨同学 at 08:25 on 11-03 drew 54K views, and a 09:58 five-product comparison drew 17K views and 102 bookmarks. Bookmarks outnumber likes, which suggests readers are saving it as "where do I go try this." Within 24 hours, someone else laid out the machine specs of Qwen Go and Grok Bot side by side. Overseas products Grok Bot and Muse were the hot topic a few days ago; this discussion is about the domestic equivalents.

Best angle: Take apart the phrase "free cloud PC." The product being from Qwen is confirmable; the cloud PC is part of an invite-only "Cloud Pro." The 4-core/4-thread CPU, 8GB RAM, 30GB disk, Hangzhou egress IP and Qwen3.8-Max are all user test claims; the company has published no specs, pricing model or model used, and there is no official statement about "free." Add a second layer of judgment: is this kind of product selling machines or tokens.

Format: A standalone original post, with a three-column table of "what the post claims / verified / unverified." Do not write it as a hands-on review.

Status: Ready to draft; cloud PC billing and long-term free policy await official word.

2. AI short drama "Armed Into Song": from 4 billion to 6 billion, and the team's background is misreported too (card 2)

Why now: A post with 136K views presents "4 billion plays, a team of stay-at-home moms and delivery riders from Xinxiang, Henan" as established fact; an interview transcript with 73K views gives a different set of details: one person makes a whole drama, an asset pipeline, Seedance 2.5 plus CapCut. Both versions spread within 24 hours, and the numbers keep climbing.

Best angle: Write it in three layers by verifiability. The 4 billion has a 10-08 source figure; the single-season peak popularity of 82.06 million is in a Paper interview; the 6 billion has no authoritative first-hand source; the team's location contradicts what the team itself said in a Red Star News interview on 11-03 — a six-person group spread across Guangdong, Anhui, Hubei, Jiangxi and Fujian, with the chief planner himself in Shijiazhuang, collaborating online. "Director He Xiushan" cannot be traced to a first-hand source; The Paper and Baidu Baike list Li Chaohua and Zhang Hansheng. Production with Seedance 2.5 and CapCut comes from the post's paraphrase.

Format: A fact-check post; or an episode breaking down "the cost structure of AI short dramas," clearly labeling any plays-to-revenue conversion as speculation.

Status: Ready to draft; the plays figure needs confirmation from the producer or platform.

3. Claude Code product lead Cat Wu leaves, and the Chinese-language circle follows the same day (card 3)

Why now: Her own post at 05:33 on 11-04 drew 467K views; 宝玉's paraphrase at 06:01 drew 38K views, and a second follow-up appeared 18 minutes later. Business Insider has confirmed the departure with Anthropic; her last day was Friday 10-30. Business Insider's headline frames it ahead of Anthropic's IPO.

Best angle: Don't just write "another one leaves." What actually carries information in her post is the change in user habits: two years ago every permission prompt had to be clicked by hand, now they are auto-approved by default; Claude Code went from an internal small tool to something that can take on large migrations. You can go on to discuss what a product lead leaving means, but state clearly that the company has not commented.

Format: A short commentary plus a timeline, citing her post and the BI report, without guessing where she goes.

Status: Ready to draft; verified.

Candidate Topics

1. Qwen Go: a personal Agent that comes with a cloud PC

  • Posts and engagement: 小墨同学 @xiaomovps's launch post (11-03 08:25, 54K views, 120 likes, 120 bookmarks) and a roundup of five domestic Personal Agents (11-03 09:58, 17K views, 71 likes, 102 bookmarks); @AlmostHuman07's specs post (11-02 16:53, 27K views, 58 likes, 58 bookmarks); @alin_zone's comparison with Grok Bot (11-03 15:15, 6,366 views, 14 likes, 8 bookmarks).
  • What it says: The post says Qwen Go is in the same category as Grok Bot and Muse, coming with a cloud PC configured as 4-core/4-thread, 8GB RAM, 30GB disk, no discrete GPU, egress IP in Hangzhou, running Qwen3.8-Max; alin_zone lays the two machines side by side (Grok Bot is 8-core, 16GB, 128GB) and judges it fine for everyday documents and scripts but tight when opening many tabs or running multiple Agents. 小墨同学 also lists MaxClaw, Coze 3.0, Kimi Work and TraeWork; Coze is marked as having a cloud PC and cloud phone, while Kimi Work runs locally.
  • Why it matters: Big domestic companies have started building personal Agents with cloud machines, and what readers most want to know is specs and pricing — exactly the two things with no official word.
  • Angles: The product's existence can be stated as fact; "comes with a free cloud PC" should be written as an invite-only feature plus user testing. A comparison paragraph could be added: a cloud machine suits long-running tasks, a local Agent suits people who don't want to send their files out.
  • Timing and verification: The launch post and AlmostHuman07's post are 24–48 hours old; the rest are within 24 hours. The product being from Qwen is confirmable (The Paper 10-13 report); the cloud PC is part of the invite-only "Cloud Pro," and the report's original wording is "supports advanced features such as the 'Cloud Pro' that can operate cloud phones and cloud PCs, all in invite-only testing" (Jimu News / 21 Caijing 10-23); the specs and Qwen3.8-Max are user test claims with nothing official published; no official statement of "free" was found.

2. AI short drama "Armed Into Song": after 4 billion, the numbers start climbing

  • Posts and engagement: 向阳乔木 @vista8's interview transcript (11-03 14:44, 73K views, 174 likes, 174 bookmarks); @xi_fen6510tui's plays post (11-03 18:58, 136K views, 230 likes, 148 bookmarks).
  • What it says: vista8 paraphrases an interview with the chief director: peak popularity of 82 million, over 6 billion plays, completion rate above 70% in later seasons; a team of about ten people using "one person makes a drama," relying on an asset middleware to keep characters consistent; a commitment to realism, no script reserve, adjustments based on comment feedback; war scenes use multi-image references in Seedance 2.5, voice is AI-generated, editing and scoring in CapCut. The xi_fen6510tui version is 4 billion plays, expected revenue of 4 to 10 million yuan, and a team of "stay-at-home moms and delivery riders from Xinxiang, Henan."
  • Why it matters: AI short dramas now have publicly citable cases with large play counts, but in Chinese-language spread the numbers and the team story are both being inflated, and readers need a version they can cite.
  • Angles: One is to break down the cost structure — where a core team of about ten puts its money and time; the other is to break down how the "ordinary people team" narrative got rewritten and explain why such stories are easy to embellish.
  • Timing and verification: Within 24 hours. The 4 billion is a 10-08 figure from data published by the publishing account, relayed by Newrank / Tencent News 10-28; the single-season peak popularity of 82.06 million and 11.41 million saves appear in a Paper interview; the 6 billion has no authoritative source; "stay-at-home moms and delivery riders from Xinxiang, Henan" contradicts the team's own account in a Red Star News interview on 11-03 (a six-person group living in several provinces, collaborating online); "director He Xiushan" has no first-hand source, and the Paper interview and encyclopedia entry list Li Chaohua and Zhang Hansheng; production with Seedance 2.5 and CapCut is the post's paraphrase. There are two versions of the launch date: Baidu Baike records a first season of 138 episodes launching 7-29, while a Xinbochang 9-11 breakdown records the AI-simulated-actor first season launching 8-22.

3. Cat Wu leaves Anthropic: product lead for Claude Code and Cowork

  • Posts and engagement: Cat Wu @_catwu's own post (11-04 05:33, 467K views, 5,054 likes, 731 bookmarks, 484 replies); 宝玉 @dotey's paraphrase (11-04 06:01, 38K views, 122 likes, 48 bookmarks); @JacksAISites's follow-up (11-04 06:18, 5,390 views, 13 likes, 3 bookmarks).
  • What it says: Her own post only says Friday is her last day, recalls two years ago getting Claude Code to build a reinforcement learning environment and watching it iterate for dozens of hours, notes that early on everyone clicked permission prompts by hand while now they are trusted and auto-approved by default, and thanks the Claude Code, Cowork and Tag teams; it gives no reason and no destination. She joined Anthropic in August 2024, previously an engineer at Scale AI and Dagster and also having worked at Index Ventures.
  • Why it matters: A departure at product-lead level, for exactly the tool Chinese developers use most; the information is clean and verifiable, well suited to a short commentary that doesn't speculate.
  • Angles: Write "the departure" as "two years of Claude Code going from internal tool to default trust," using her own retrospective to describe product change; you can also discuss how the team is arranged after one of two co-leads leaves, but note that the company has not commented.
  • Timing and verification: Within 24 hours, verified. A Business Insider report on 11-03 says Anthropic confirmed the departure, that her last day was Friday 10-30, and that she announced it on X on Tuesday afternoon; she was responsible for the Claude Code product and also oversaw Cowork.

4. Ten pieces of advice for fourth-year computer science students

  • Posts and engagement: Tw93 @HiTw93's original post (11-03 15:35, 77K views, 1,179 likes, 712 bookmarks).
  • What it says: Ten items running from picking a direction and a city and using AI to amplify your output, to putting half your salary into VOO and QQQ and staying out of A-shares and Hong Kong stocks, to finding a partner early, not borrowing money, building a personal brand, and putting your health first. The ratio of likes to bookmarks shows readers don't just agree — they're saving it to pass on to students.
  • Why it matters: Whether AI has killed the prospects of a computer science degree is a question that has recurred on the Chinese internet for the past six months, and this post's spread shows it hit fourth-year students' anxiety.
  • Angles: You could take just the first three items into a piece on "is a computer science degree still worth it in 2026"; item 4 is specific investment advice, so any write-up must say it is not investment advice and must not judge the market for readers.
  • Timing and verification: Within 24 hours. A personal-experience post with no data sources; the quote area is mostly agreement, and replies were not read one by one for this brief.

5. Follow-up: MIT weight-editing experiment challenged — "just turn off thinking and that's enough"

  • Posts and engagement: @oran_ge's follow-up post (11-03 22:34, 35K views, 253 likes, 53 bookmarks); his original paraphrase the day before (11-03 07:30, 19K views, 118 likes, 52 bookmarks); @xingyudang's Same recipe post (11-03 05:03, 83K views, 423 likes).
  • What it says: The original experiment modified 131,072 weights (0.016%), and after 50 forward passes raised Qwen3.5-0.8B's MATH-500 from 7.9% to 48.1%, GSM8K from 13.8% to 67.1% and IFEval from 21.0% to 62.3%. The follow-up post's claim is that "just turning off thinking gives a 6x improvement, no need to edit weights at all."
  • Why it matters: This is a public discussion of "whether model capability can be edited out of the weights." The original experiment was reported the day before, and now a citable third-party audit has appeared, which suits a follow-up rather than a repeat.
  • Angles: Tell three things in order. The author's original used the thinking-on setting; the third-party audit reran the unedited model with the settings recommended in the model card, got 44.0% on MATH-500 under non-thinking versus 48.1% after editing, and its preregistered conclusion is "mainly a settings effect"; the author has not responded. Then add the author's own Gemma 3 270M comparison, where a model that barely loops showed no established improvement from the same method, showing that weight editing can remove an obstacle but cannot add capability the model never had.
  • Timing and verification: The follow-up post is within 24 hours; the xingyudang post and the original paraphrase are 24–48 hours. The original is on the author's blog, the audit is Carlos Toledo's 11-03 report, and criticism appears in Honnibal, Sean Goedecke and an R&D World roundup. The audit also points out that the actual number of weights changed is 130,188, that "0.016%" was computed against a round 0.8B, and that the so-called deletion is in fact subtracting six times the projection along that direction.

6. Ant's Ming-3.0 speech models go open source, and the Chinese-language circle starts using them for voiceover

  • Posts and engagement: @Lonely__MH's hands-on post (11-03 22:19, 17K views, 296 likes, 422 bookmarks); @xueyu1125's workflow post (11-03 19:20, 5,695 views, 33 likes, 20 bookmarks).
  • What it says: Lonely__MH highlights voice cloning and timbre design, emotion and re-read tone control, and paralinguistic detail, saying it is MIT-licensed for free commercial use and supports one-click deployment; xueyu1125 gives the full chain: Claude or Codex generates storyboard prompts, Qwen Image 2.1 produces first and last frames and reference images, Minimax H3 generates video, and Ming-3.0-TTS-0.5B does the voiceover. The former's bookmarks are 1.4 times its likes, clearly "save it first and run it yourself."
  • Why it matters: A 0.5B model is small, and the post author says it runs locally on a single GPU; voiceover is the most easily overlooked step in AI short dramas and short videos, and this cross-references with card 2.
  • Angles: Write a fully open-source voiceover pipeline that actually runs, focusing on the compliance boundaries of voice cloning and on how to choose between the 0.5B tier and the 7.9B MoE tier.
  • Timing and verification: Within 24 hours. The release itself was Ant Bailing open-sourcing five Ming-3.0 speech models on 10-20; this window is a new round of spread in the Chinese-language circle. The MIT license is verified (Hugging Face model card, GitHub repo); the 0.5B is a dense model with a Qwen2.5-0.5B backbone, while the tiny tier is 7.9B total parameters with 1.3B active, so "0.5B" refers to only one of the tiers. The post's claim of "Ming-3.0-audio 126-billion-parameter MoE" was not confirmed in this check and must be flagged when writing.

7. MOSS-Live: full-duplex real-time speech that judges interruptions accurately

  • Posts and engagement: @hylarucoder's hands-on post (11-03 14:30, 5,868 views, 70 likes, 81 bookmarks); @ElliotChen's technical commentary (11-03 19:44, 16K views, 22 likes, 20 bookmarks).
  • What it says: hylarucoder puts it alongside GPT-Live-1, arguing that end-to-end speech sheds the stitched-together feel of VAD plus ASR plus LLM plus TTS; ElliotChen, who has worked two years in Voice AI, focuses on how interruption must not only stop fast but stop accurately: saying "mm" is agreeing, not telling it to shut up, and thinking mid-sentence doesn't mean it's the model's turn to jump in.
  • Why it matters: Chinese-language discussion of voice interaction often stops at "does it sound human," and these two posts break "when it should stop and when it should take a turn" into a discussable product problem.
  • Angles: Use concrete conversation scenarios to explain semantic turn-taking, distinguishing backchannels, thinking pauses and genuine interruptions; then cover its trade-offs in latency and cost against traditional cascaded setups.
  • Timing and verification: Within 24 hours. ModelBest released it on 11-02 and has opened an API and online demo (official model page, Machine Heart 11-02); it is not open source — the site marks only MOSS-TTS and MOSS-VL-Realtime as open source and there is no weights repository, so a more accurate phrasing than "beta" is "API and online demo open, free for a limited time"; figures like a "1.44% interruption rate" are official numbers relayed by the post and were not verified here.

8. Tencent enters the top three Seedance customers in China, Bilibili enters the top ten

  • Posts and engagement: @lxfater's original post (11-03 22:36, 16K views, 41 likes, 27 bookmarks); @yangyi's short comment (11-03 22:55, 5,241 views, 5 likes, 3 bookmarks).
  • What it says: lxfater relays a WeChat report: counting by token consumption and revenue contribution, Tencent ranks in the top three customers in China and Bilibili in the top ten; overseas revenue accounts for half, up nearly fivefold in the past six months; ByteDance signed a memorandum of understanding with the MPA, and TikTok negotiated a derivative-works deal with Disney. His judgment is that ByteDance has not kept its video models as an exclusive advantage.
  • Why it matters: Video models have reached the stage where peers buy from one another, which is directly relevant to teams building video products, and the half-overseas revenue share shows it isn't sold only domestically.
  • Angles: Write "how a model company turns rivals into customers," focusing on copyright and compliance as the precondition rather than on sales figures. The overseas revenue share can be cited; customer industry distribution should be handled cautiously.
  • Timing and verification: Within 24 hours. Based on the The Paper report on 11-03, the original being a 36Kr "Intelligent Emergence" exclusive on 11-03: overseas revenue at 50%, nearly fivefold growth in half a year, Tencent in the top three and Bilibili in the top ten all hold. The sources are people familiar with the matter and people close to Volcano Engine, not an official ByteDance announcement; "more than a quarter of customers from manufacturing, healthcare and automotive" does not appear in this report and is unverified.

9. CapCut 11.7.0 and Yichen's Skill: an attribution lacking evidence

  • Posts and engagement: @Astronaut_1216's original post (11-02 11:51, 21K views, 115 likes, 145 bookmarks).
  • What it says: The author says that after upgrading to CapCut 11.7.0, Yichen's Skill no longer works, so they rolled back to 11.5.0 and used the open-source tool LuLu to block CapCut and related updaters from going online; it ends with three suggestions: note which software versions a Skill supports, keep the installer and back up projects, and check auto-update.
  • Why it matters: It is really describing a general problem: automation that relies on reverse-engineered adaptation breaks when the software updates. Readers need to know how to protect themselves, not just remember that "CapCut blocked the Skill."
  • Angles: Turn version pinning into a method: how to back up the installer, how to block updates with a firewall, what to confirm before upgrading. Attribute the blocking as unproven rather than as a conclusion.
  • Timing and verification: 24–48 hours. 11.7.0 does exist — the Yingyongbao page shows CapCut desktop 11.7.0.14652, updated 11-02 to 11-03 (page); CapCut's official site, changelog and searchable official channels say nothing about "11.7.0 restricting third-party Skills"; Yichen's Skill repository was taken down on 10-25 after a lawyer's letter (issue #25), earlier than 11.7.0, so blaming the version upgrade lacks evidence.

10. An OpenAI team lead teaches resume writing: short, cut the irrelevant, write results, easy to scan

  • Posts and engagement: 宝玉 @dotey's paraphrase (11-03 23:24, 29K views, 276 likes, 371 bookmarks); Wulfie Bain @wulfie_bain_'s original post (11-03, metrics not collected for this brief).
  • What it says: The four points are: write short, cut irrelevant experience, write outcomes rather than only what you did, and let people scan to the key points quickly. Wulfie Bain's X bio is OpenAI's Applied AI Engineering lead for startups, and the post notes this is part three of a hiring-advice series (the resume installment), and that the team is hiring.
  • Why it matters: With 371 bookmarks and 276 likes, this is the classic "save it and fix my resume later" behavior; job hunting in AI is a high-frequency need in the Chinese-language circle right now, and this kind of first-hand advice is usually English-only.
  • Angles: Turn the four points into executable edits, each with one before and one after line; then add a fifth the original doesn't mention: when using AI to revise a resume, don't turn it into a generic version that could be sent to any company.
  • Timing and verification: Within 24 hours, verified. The four points correspond one-to-one with the original post, with nothing added or dropped; the X bio reads OpenAI Applied AI Engineering lead, Startups.

11. Can local models do real work now: a 125B model on an 8GB GPU, and Apex-35B hooked up to Claude Code

  • Posts and engagement: @NFT_Chen's Strata post (11-03 12:21, 9,327 views, 77 likes, 71 bookmarks); @lxfater's Apex-35B post (11-03 10:20, 7,583 views, 32 likes, 31 bookmarks).
  • What it says: Strata v0.2.10 claims to squeeze Qwen3.8-Flash-Next, at 125B total parameters with 6B active per token, into an 8GB GPU, supporting 128K context, multi-GPU acceleration, one-click install, and local OpenAI- and Anthropic-compatible APIs; lxfater saw someone run Apex-35B on a 48G 4090 hooked to Claude Code, fixing a Python project in 21 minutes with all 18 tests passing, and what he wants to know is whether an ordinary single GPU works.
  • Why it matters: Both posts point at the same question — have local models finally reached the line where they can do real work. For developers who don't want to send code to the cloud, this is very practical.
  • Angles: Testing it yourself is more valuable than relaying: how many tok/s an 8GB card actually gets, whether 128K context is really usable, and whether latency is acceptable when driving an Agent like Claude Code.
  • Timing and verification: Within 24 hours. The Strata repository, version and model specs all hold (GitHub, v0.2.10 release notes, Qwen model card); the repo is MIT and was created on 2026-09-24; but the official benchmark on a real 8GB Intel Arc A750 is 12.7–14.8 tok/s, with another run at 13.9–18.4 tok/s, so "a stable 20+ tok/s on 8GB" is not supported; Apex-35B's 21 minutes and 18 tests are a relay of someone else's test and are unverified.

12. Do Skills get less useful the more you use them? A paper tested making AI review itself

  • Posts and engagement: 宝玉 @dotey's paper breakdown (11-04 03:51, 9,304 views, 95 likes, 109 bookmarks); @hylarucoder's short post (11-03 10:41, 20K views, 84 likes, 46 bookmarks).
  • What it says: The paper "How Do Coding-Agent Harnesses Self-Evolve? An Empirical Study of Retrospective Adaptation" tested /retro review on 30 SWE-bench Verified issues, running it once after each fix in Claude Code, yielding 125 suggestions, of which 24 were implemented and 12 were testable; of 7 valid tests only 2 fully eliminated the target problem, and some changes made the AI do extra checks on unrelated tasks, one of them raising the median time from 82 seconds to 245 seconds.
  • Why it matters: Having Agents summarize their own lessons is one of the most popular practices of the past six months, and this paper provides rare concrete numbers — and negative results — that more people should see.
  • Angles: Give readers a criterion: information true across the whole project, such as which command runs the tests, is worth writing into rules; information like which tests failed in one environment won't match another task. You can also cover how to judge whether a change actually works.
  • Timing and verification: Within 24 hours. The paper was posted on arXiv on 10-26 by Christoph Treude of Singapore Management University and Sebastian Baltes of Heidelberg University; this brief relies on 宝玉's paraphrase and the numbers cited there, the original paper was not checked page by page, and the sample is small (only 5 runs per setting), a limitation the authors state themselves.

13. Follow-up: complaints about Grok Bot account merging and quota drops

  • Posts and engagement: @op7418's quota post (11-02 11:11, 57K views, 134 likes, 22 bookmarks); @oran_ge's login flow post (11-02 19:42, 34K views, 210 likes, 16 bookmarks); @WildVariable01's usage post (11-02 12:45, 5,173 views, 12 likes, 1 bookmark); @lxfater's "even the $200 plan isn't enough" (11-03 18:08, 43K views, 25 likes, 6 bookmarks).
  • What it says: op7418 says the chance of routing to Opus 5.5 keeps falling, while many people find their quota suddenly jumps by 20%, 30% or even 40%; oran_ge records a login loop after account merging, where grokbot first sends you to the cursor login page, that page can't log in directly with X, and subscription verification can't find his X premium membership; WildVariable01 says that at noon on 11-01 the weekly quota jumped straight from 30% or 50% to 100%, the status page was down that day with degraded billing and usage panel services, and it was marked resolved 5 hours later; lxfater only asks whether it will reset tonight.
  • Why it matters: Grok Bot is the personal Agent most discussed in the Chinese-language circle since late October, and users hit quota and account problems daily, yet there is still no clear official statement.
  • Angles: Turn the complaints into a checkable list: what the status page and screenshots can prove, and what is only personal impression; then explain why the cost structure of subscription Agents forces platforms to tighten quotas.
  • Timing and verification: lxfater's post is within 24 hours, the rest 24–48 hours. All are user accounts with no official statement found; op7418's "can't keep subsidizing" is a judgment, not a fact, and the routing ratio and the size of quota jumps are unverified.

14. Follow-up: Alma launches on Android, with per-window remote control on mobile

  • Posts and engagement: @yetone's launch post (11-03 00:27, 48K views, 387 likes, 288 bookmarks) and per-window remote control post (11-03 17:20, 20K views, 99 likes, 35 bookmarks).
  • What it says: The Android version is out and can be downloaded from the official site; remote desktop control on mobile can be done for each window separately, without operating the whole screen. A previous day's brief covered Alma's performance in a wallpaper-hunting task, which was the author's own test.
  • Why it matters: A desktop Agent reaching phones is another path for personal Agents; per-window control solves the real annoyance of being able to operate only the whole screen.
  • Angles: It could be written as a scenario list for using a desktop Agent on a phone: which tasks suit per-window control, and where the cost of switching between touch and keyboard lies.
  • Timing and verification: The per-window post is within 24 hours, the launch post 24–48 hours. It is the author's own product, not an independent review; features and download method follow the official site, and nothing was tested hands-on here.

15. Follow-up: Tuwa lends you a cloud phone so an Agent can call people on WeChat voice

  • Posts and engagement: 郭宇 @turingou's original post (11-03 20:51, 18K views, 76 likes, 37 bookmarks).
  • What it says: Beta users can borrow a cloud phone and install their usual apps on it; Tuwa Device has just added support for WeChat Mac voice calls and Zoom online meetings, so an Agent can call family and friends by WeChat voice or video, or attend a Zoom meeting for you. The previous day's brief described calling a clinic in your own voice; this time the approach is borrowing a phone.
  • Why it matters: It is the same class of infrastructure as the cloud PC in card 1, but the use case is more specific and more likely to trigger discussion of whether the other party knows.
  • Angles: Write about boundaries rather than showing off: whether to tell the other party when attending a meeting or making a call on their behalf, how recording and transcription are handled, and where account risk-control risks lie.
  • Timing and verification: Within 24 hours. A hands-on account of the author's own product, not tested here; the beta scope, the source of the cloud phone and its billing are unverified.

16. Canva restructures 1,271 roles, 22% of staff

  • Posts and engagement: 宝玉 @dotey's original post (11-04 07:21, 6,518 views, 15 likes, 5 bookmarks).
  • What it says: The post notes the 1,271 roles include people reassigned and people asked to relocate, so the number actually laid off will be lower, and the company has not published a specific figure; it also explains who Canva is: an Australian online design tool that lets people who can't use professional design software make posters, slides, social images and resumes from templates.
  • Why it matters: A widely known design-tool company doing its first company-wide restructuring has not drawn much attention in the Chinese-language circle yet, so the numbers can be laid out first; whether the restructuring is AI-related was not concluded in the report summaries checked here.
  • Angles: Write it as how to read a corporate restructuring in the AI era: role counts aren't layoff counts, and relocation requirements are also a restructuring tool; don't attribute it to AI out of hand.
  • Timing and verification: Within 24 hours, verified. Sydney Morning Herald report on 11-04: affecting 1,271 roles, 22% of total staff, some positions eliminated and some people reassigned or asked to move to centralized offices; CEO Melanie Perkins calls it the hardest moment in the company's thirteen-year history; the company has staff in 37 countries and over 265 million monthly active users, and this is its first company-wide layoff.

17. NVIDIA Auto Ontology: keeping AI from making things up when querying enterprise data

  • Posts and engagement: @Gorden_Sun's original post (11-03 22:11, 5,946 views, 104 likes, 126 bookmarks).
  • What it says: The post says that when enterprises have AI query data, it often can't tell what business terms mean, so computed revenue and metrics often come out wrong and ordinary users don't dare trust them; NVIDIA's open-source Auto Ontology builds an ontology layer with FastAPI, Postgres and other parts of the stack so queries have a clear definition. Bookmarks outnumber likes, showing readers are the "save it and see if I can use it" kind.
  • Why it matters: Most Text-to-SQL errors aren't the model failing but definitions being unclear; the problem is real on the enterprise side, yet Chinese-language coverage is scarce.
  • Angles: Use an example of one metric with three different calculation methods to explain what an ontology layer does, then say who it suits: teams with a data warehouse and messy metric definitions.
  • Timing and verification: Within 24 hours. The repo address, license and stack are relayed by the post and were not verified here, and no official NVIDIA release note was found; open the project page to confirm before writing.

18. 10 AI and programming YouTube channels worth subscribing to

  • Posts and engagement: @sdf302030's original post (11-02 15:46, 26K views, 341 likes, 290 bookmarks).
  • What it says: From Karpathy, 3Blue1Brown, Fireship and NetworkChuck to CS50, freeCodeCamp, Traversy Media, Programming with Mosh, Finematics and Ali Abdaal, each channel gets one line of justification. The opening line — "a lot of college students spend thousands on courses when half the content has long been on YouTube" — is what drives its spread.
  • Why it matters: It pairs with card 4 into a set on how computer science students can teach themselves, with a clear audience and little controversy.
  • Angles: Don't restate the list; reorganize by goal: who to watch to shore up math, who to watch to follow a framework, who to watch to find projects to practice on; then add one thing the list doesn't cover: how to tell that a tutorial is out of date.
  • Timing and verification: 24–48 hours. The channels all exist; the list and recommendations are the author's personal judgment; each channel's current update status was not checked one by one for this brief.

This issue has 18 cards in total, short of 20, because the remaining high-engagement posts were either already covered in the previous three days, or unrelated to AI, or fell into jailbreak-prompt and freebie content.

This time, 6 of the fixed-account searches hit their return limit and were not paged through, so earlier posts from a few accounts may have been missed.