X Chinese Hot Topic Brief¶
Translation
English translation of the Chinese original.
Today's Priority Reads¶
Priority 1 | Fable 5.2 "can paint, can make 3D video"? Unpacked, it is writing programs and calling tools¶
Worth doing now: on day two after Fable 5.2's release, discussion in Chinese-language circles moved from quotas to output. 歸藏 (Guizang) @op7418 only said "make a project that shows off your ability," and the model handed over an interactive piece where typing generates ink-wash fluid (494 likes, 52K views); 花叔 (Huashu) @AlchainHust asked it to "recreate with a drawing tool" a photo of Bora Bora, and said it "may mean a language model can absolutely be an image and video model"; Orange AI @oran_ge also showed off a photo Fable drew.
Best angle to add: break "it can draw" down into "how it draws." Gorden Sun @Gorden_Sun pointed out that in one viral video, the white-model animation in the first half was made by Fable, while the character animation in the second half was actually generated by calling a video model; when he reproduced it himself, Fable proactively called the Seedance API he had configured. Some in the quote posts say Huashu's photo was the model writing a NumPy plus OpenCV program to plan brushstrokes and then replaying them (not confirmed by the original author). Add cost on top: someone relayed that for the same 3D voxel task, Fable 5.2 cost $108.87 and GPT-6.1 Sol cost $13.02. The conclusion can be written as: the language model's new ability is "organizing tools and code," not replacing image and video models; ordinary users should first check whether they have these tools hooked up and whether they are willing to pay this money.
Format suggestion: a standalone original post with two or three "what you'd assume / what actually happens" contrasts.
Status: ready to draft. Case details are all the posts' own accounts or relays, the cost figures need verification, do not write them as an independent test conclusion.
Priority 2 | No more LeetCode; now it's "read an unfamiliar codebase with AI"¶
Worth doing now: a Reddit interview rant translated by 宝玉 (Baoyu) @dotey got 426 likes, 65K views, 130 replies: Bloomberg's second round, an "AI codebase interview," gives you a huge Django project, the first 30 minutes are a behavioral interview, leaving only 30 minutes for coding, the interviewer does not say how far AI may be used, and the candidate used planning mode and ask mode, passed a few tests, and was still rejected. The same day Baoyu also relayed Manus co-founder 张涛 (Zhang Tao): engineer interviews are changing to "spend four hours solving a problem with Manus," then come back with the full human-machine conversation record and debrief with the interviewer.
Best angle to add: the two interview formats actually test the same thing — whether others can follow the process of you working with AI. You can give job seekers a prep checklist: first ask clearly where the AI boundary is, practice "reading the codebase to find the problem" instead of having AI do it all in one sentence, and keep a reviewable conversation record. The disagreement in the quote posts is also worth writing: some feel this is just a different trap, others feel an AI interview is more subjective than grinding problems.
Format suggestion: quote Baoyu's post and add an interview prep checklist.
Status: ready to draft. The Reddit post is a personal experience, and Zhang Tao did not spell out all the Manus interview details in the interview, so do not fill them in.
Priority 3 | Lenny's new jobs report: PM roles at a four-year high, frontend and mobile roles nearly halved in a year¶
Worth doing now: 宝玉 (Baoyu) @dotey put together Lenny Rachitsky's tech jobs report published on November 10 that same day. Verified against Lenny's original post: open PM roles are 75% above the March 2023 low, the most in nearly four years; open engineering roles rose only 2.5% in a year, stopping growth for the first time; the US hired about 113,000 software engineers in the past 12 months, less than half of the 2022 peak of 239,000; open product design roles are about 1,900, a record low; frontend and mobile demand nearly halved in a year, while open FDE (forward-deployed engineer) roles rose to 4.6 times.
Best angle to add: the report uses two different measures — "how many people companies want to hire" and "how many people they actually hired" — and the two often do not line up, which is the easiest place for ordinary readers to misread it. You can write that "in the AI era jobs are not disappearing, they are merging: one person covers a wider scope," and note that this is US and global tech company data and cannot be applied directly to China.
Format suggestion: a standalone original post or a quote of Baoyu's post, with a "up / down" comparison.
Status: ready to draft, the main figures are verified. The report's seventh conclusion is in the paywalled part, do not guess at it.
Candidate Topics¶
14 selected this round, not padded to 20.
1. The Fable 5.2 output wave: behind the ink wash, the photo and the 3D short are code writing and tool calls¶
Posts and traction: - 歸藏 (Guizang) @op7418 post, 2026-11-10 17:24, 493 likes, 51910 views, 115 replies, 246 bookmarks (interactive ink-wash fluid piece). - 花叔 (Huashu) @AlchainHust post, 2026-11-10 16:06, 131 likes, 32025 views (recreating the Bora Bora photo). - Orange AI @oran_ge post, 2026-11-10 22:55, 255 likes, 85515 views ("elegantly killed the competition"). - Gorden Sun @Gorden_Sun post, 2026-11-10 23:22, 86 likes, 7955 views (clarifying the division of labor between the white model and the video model). - _FORAB @_FORAB post, 2026-11-10 14:52, 257 likes, 50273 views (relaying that a browser shooter game cost $550). - AI Will @FinanceYF5 post, 2026-11-10 17:42, 15 likes, 10201 views (cost comparison between Fable 5.2 and GPT-6.1 Sol). - Discussion sample: Shenjingape @Shenjingape quote post, 2026-11-10 18:20, 5 likes, 75 views (says the photo was drawn by a NumPy plus OpenCV program).
This item corresponds to "Today's Priority Reads 1."
Timeliness and verification: within 24 hours, a follow-up on the Fable 5.2 topic (yesterday's piece was about quotas, this one only covers output and cost). The _FORAB post says the author is an Anthropic engineer, identity not verified; both $550 and $108.87 are relays, pending verification.
2. AI coding interviews: Bloomberg tests a codebase, Manus tests four hours of human-machine record¶
Posts and traction: - 宝玉 (Baoyu) @dotey post, 2026-11-10 16:03, 426 likes, 65419 views, 130 replies, 368 bookmarks. - 宝玉 (Baoyu) @dotey post, 2026-11-10 23:27, 84 likes, 19180 views (Manus's Zhang Tao on hiring and interviews). - Discussion sample: kenshinji @kenshinji quote post, 2026-11-10 17:01, 7 likes, 1162 views ("I just did an AI coding interview last week, it felt quite subjective").
This item corresponds to "Today's Priority Reads 2." Source: Reddit original post.
Timeliness and verification: within 24 hours. All are personal experiences and interview relays.
3. Lenny's tech jobs report (second half of 2026)¶
Posts and traction: - 宝玉 (Baoyu) @dotey post, 2026-11-11 02:44, 44 likes, 9438 views, 50 bookmarks. - Related discussion: meng shao @shao__meng post, 2026-11-10 21:04, 81 likes, 52355 views (a 36-year-old embedded programmer who was laid off, gave up his rental and went back to his hometown).
This item corresponds to "Today's Priority Reads 3." Fact sources: Lenny's Newsletter original post, Live Data Technologies data.
Timeliness and verification: within 24 hours. The main figures have been checked against the original; the meng shao post is a personal anecdote and can only serve as a mood sample, not as employment data.
4. Follow-up: mainland-China iPhones getting overseas Apple Intelligence, hands-on feedback and costs start to appear¶
Posts and traction: - 歸藏 (Guizang) @op7418 post, 2026-11-10 14:05, 163 likes, 43608 views, 74 replies (listed three pitfalls: both the Apple account and the App Store must be on the US region, system and Siri language must be English or another supported language, and after enabling it you have to download about 10GB of local model). - ahhhhfs @abskoop post, 2026-11-10 15:03, 477 likes, 56201 views, 482 bookmarks (after enabling, ChatGPT says it is unavailable in the region). - Colinclean2 @Colinclean2 post, 2026-11-10 17:07, 254 likes, 91844 views (says after "changing two lines of code," eSIM and FaceTime work too). - kazegogogo @kazegogogo post, 2026-11-10 07:48, 25 likes, 7099 views (says some have reported overseas eSIM does not work, and breaking it may require a firmware reflash).
Content and focus: we covered GestaltX itself yesterday, so this item only covers the increment. Testers say it can be enabled, but only after switching to a US region account, switching to English, and downloading the 10GB model; on eSIM and FaceTime, the most widely spread post says "everything works again," another post says eSIM does not work, and the claims contradict each other. Another easily overlooked fact: the highly engaged Chinese-localization fix post from yesterday links upstream to jailbreakdotparty/Erosion, whose repository description states it relies on a sandbox escape vulnerability, so the advertised "no jailbreak needed" does not equal "no risk."
Why it is worth doing: lay out clearly in three columns for readers "the feature you want, the cost you pay, the claims nobody has confirmed yet," and remind them that the official mainland-China version is still on track to ship with iOS 27.3 in December.
Timeliness and verification: most posts within 24 hours, the kazegogogo post 24–48 hours; a follow-up. The Erosion description has been checked on GitHub; the eSIM and FaceTime claims are unverified; do not provide instructions or tutorials for bypassing regional restrictions.
5. Follow-up: Fable resets granted by subscription length, official rules verified¶
Posts and traction: - 歸藏 (Guizang) @op7418 post, 2026-11-10 10:01, 51 likes, 19204 views, 78 replies. - 小互 (Xiaohu) @xiaohu post, 2026-11-10 06:45, 23 likes, 11630 views. - Same topic: Baye @waylybaye post, 2026-11-10 11:11, 29 likes, 10620 views.
Content and focus: verified against the Claude Help Center article: Max, Ultra and team premium seats get Fable resets by paid tenure — 1 reset under 6 months, 2 for 6–12 months, 3 for more than 12 months, expiring March 8, 2027; Pro and team standard seats can use Fable this week for up to half the weekly quota, plus 1 extra reset. One easily misunderstood point: a Fable reset only refunds the weekly quota spent on Fable, usage by other models is not refunded, but the five-hour window refills entirely.
Why it is worth doing: yesterday we wrote about the Pro trial week, this item adds "what a reset actually refunds," and can be made into a short post answering "should I click it now."
Timeliness and verification: the Xiaohu post 24–48 hours, the rest within 24 hours; a follow-up. The rules are verified.
6. homelab-diagram: turn your home network's YAML inventory into a topology diagram¶
Posts and traction: Geek @geekbb post, 2026-11-10 17:42, 1091 likes, 80283 views, 1593 bookmarks.
Content and focus: the most engaged post on this round's list. The repository Goblinmine/homelab-diagram description confirms: renders r/homelab-style network diagrams from a YAML inventory in Python, outputs SVG or PNG, MIT license, author notes it was "written with Claude," repository created on November 7, 551 stars at capture time. Bookmarks are 1.5 times the likes.
Why it is worth doing: "write the inventory first, then let a program draw the diagram" is an AI use case very suitable for non-programmers, and can extend to having AI first help you organize your home devices into YAML.
Timeliness and verification: within 24 hours. Repository information checked, not tested hands-on.
7. 向阳乔木 (Xiangyang Qiaomu) open-sources xnews: a bot scrapes overseas long posts on X every day and rewrites them into a Chinese site¶
Posts and traction: 向阳乔木 (Xiangyang Qiaomu) @vista8 post, 2026-11-10 18:21, 357 likes, 22942 views, 87 replies, 477 bookmarks.
Content and focus: the author says it can be repointed at accounts you care about, such as links posted by researchers and model companies. The comments give a demo site and the open-source repository, whose description is "overseas founders' posts and AI videos, rewritten into Chinese."
Why it is worth doing: "use an Agent to automatically build an information site every day" is exactly a long-term automation scenario ordinary people can get started with; you can spell out the steps needed — scraping, filtering, rewriting, publishing — and the copyright boundaries of reposting overseas content.
Timeliness and verification: within 24 hours. Repository confirmed to exist, not tested hands-on.
8. 李继刚 (Li Jigang): looking at three kinds of Agent products — ChatGPT, Projects and Bots — through "product atoms"¶
Posts and traction: 李继刚 (Li Jigang) @lijigang post, 2026-11-10 12:40, 211 likes, 23264 views, 203 bookmarks.
Content and focus: the author divides Agent products into three kinds whose smallest unit is the conversation, the project, or the role, and argues the biggest difference is who bears "continuity": with conversations, the person is the memory pool; with projects, memory is written into documents; with Bots, it settles into individual roles; splitting Bot responsibilities can follow 西蒙 (Herbert Simon)'s "nearly decomposable systems."
Why it is worth doing: the concepts are clear and directly reusable, good for explaining "why the same memory feature feels so different in practice."
Timeliness and verification: within 24 hours. Opinion piece, no figures to check.
9. 宝玉 (Baoyu): the bottleneck for software is no longer building it, but making it usable by Agents¶
Posts and traction: 宝玉 (Baoyu) @dotey post, 2026-11-11 03:29, 83 likes, 8905 views, 72 bookmarks.
Content and focus: the author argues complex GUIs are nearly worthless to Agents, and what is most Agent-friendly is CLI, MCP and code; human work becomes previewing, marking and fine-tuning; software will be broken into atomic capabilities that Agents compose on demand. He calls out Manus's new video feature for still shipping a complex timeline editor, saying they "haven't figured it out either."
Why it is worth doing: it can be paired with card 8 to explain "what software looks like in the Agent era," and can also be turned into a question: where are the "preview and mark" tools for ordinary users.
Timeliness and verification: within 24 hours. Opinion piece.
10. "The Bitter Lesson" gets misread: Sutton himself says raw compute scaling is not enough¶
Posts and traction: 宝玉 (Baoyu) @dotey post, 2026-11-10 14:24, 223 likes, 27830 views, 86 replies.
Content and focus: Baoyu relays Sutton's new tweet "brute-force compute alone is not enough for true intelligence, we need clever new methods," and a reply's reading: the essay argues against hard-coding human knowledge into systems, not against algorithmic innovation.
Why it is worth doing: "compute-only-ism" spreads widely in Chinese-language circles, and this is an evidence-based correction, good for a short conceptual clarification post.
Timeliness and verification: within 24 hours. Sutton's words are a relay; open the original tweet to check before quoting.
11. Value watch: switching from Claude Code to Pi, and Pi 1.5.1's background cache cost¶
Posts and traction: - 小墨同学 (Xiaomo) @xiaomovps post, 2026-11-10 17:33, 55 likes, 8824 views. - 小墨同学 (Xiaomo) @xiaomovps post, 2026-11-10 15:09, 14 likes, 9453 views (below the usual threshold, passed the initial view screening).
Content and focus: the first introduces Goodbye Claude, Hello Pi, original read: the author says Claude Code shipped 244 versions this year, more than three quarters of them changed the system prompt or tool descriptions, and the default Bash tool description is nearly 3000 tokens. The second warns that Pi 1.5.1 changed cache keep-alive so it refreshes even when idle by default, which may incur costs while you are not talking to it, and can be switched back to streaming or turned off.
Why it is worth doing: traction is average, but "tools getting heavier, switch to a lighter one" and "a setting that quietly spends money in the background" both have real value for users paying out of pocket.
Timeliness and verification: within 24 hours. The Pi 1.5.1 default setting change is a post relay; check Pi's changelog before writing.
12. Mirage: a 22MB vision model that runs in 74 milliseconds on a phone¶
Posts and traction: Gorden Sun @Gorden_Sun post, 2026-11-09 23:14, 545 likes, 28648 views, 671 bookmarks.
Content and focus: the post says the model has 51.3 million parameters, about 22MB after 2-bit quantization, 74 milliseconds for the full pipeline per run on the NPU of a Samsung S26, and can do scene understanding, circling objects, reading text and outputting structured data. The Hugging Face model page confirms it is a "image understanding and structured extraction" vision model released by Cactus Compute, callable directly from a Python package.
Why it is worth doing: on-device small models matter for privacy and offline scenarios; you can write about what small tools "looking at images locally on a phone" enables.
Timeliness and verification: 24–48 hours. The model page does not list parameter count, size or latency; these three figures are post relays, pending verification.
13. Baye: cloud services used for years going offline one after another, indie developers' dependency risk¶
Posts and traction: Baye @waylybaye post, 2026-11-09 12:58, 96 likes, 47566 views.
Content and focus: the author says several of his services that had run continuously for years hit failures one after another in the past year, for reasons including cloud platform companies shutting down and Alibaba Cloud and Tencent Cloud discontinuing certain cloud services.
Why it is worth doing: AI lets everyone build products, but few people think ahead about "what if the service I depend on disappears one day." You can write a dependency checkup checklist for personal projects.
Timeliness and verification: 24–48 hours. Personal experience, specific service names not listed.
14. Value watch: domestic Personal AI "proactiveness for proactiveness' sake," and one indie developer's one-month ledger¶
Posts and traction: - Orange AI @oran_ge post, 2026-11-10 21:51, 56 likes, 11692 views, 26 replies. - Tw93 @HiTw93 post, 2026-11-10 15:05, 50 likes, 11551 views (one month after the Mole Windows release: 15 official builds, 1084 commits, sold in dozens of countries).
Content and focus: after trying several domestic personal AI products, Orange AI argues proactiveness is a means, not an end, and excessive proactiveness only annoys people. Tw93 disclosed the development figures for the Windows version's first month.
Why it is worth doing: both are real firsthand feelings from frontline product people, suitable for a short take on "how proactive should personal AI be" or "how much can an indie developer ship in a month with AI."
Timeliness and verification: within 24 hours. Average traction, included as value watch; Tw93's figures are the author's own account.
Engagement numbers are an API snapshot from the morning of 2026-11-11; a single snapshot cannot prove a trend.