X Chinese Hot Topics Briefing¶

Date17 October 2026 · SaturdayWindow2026-10-15T08:30:17+08:00 to 2026-10-17T08:30:17+08:00CoveragePartial

Translation

English translation of the Chinese original.

Look at these first today¶

sofatap ("couch slouch"): how far can you get from your computer while agents run tasks (see item 1). Guo Yu released his 19th product of the year, which turns an iPhone into a touchpad and keyboard so you can control the TV-connected computer from the couch. 620K views, 1,713 likes and 207 quotes; it is the highest-viewed Chinese AI post this run found. The quote posts are almost all about "the launch video is better than the product" rather than the product itself. Two good angles: first, place it in the shift where "in the agent era, the human just taps once and replies once" — interaction moves from "sitting at a desk writing" to "remote approval"; second, break down why this launch video spread (a real-life scene, the pain point stated in one sentence). Format: standalone original post, or a quote-post breakdown. Status: ready to draft; product features checked against the GitHub repo description; voice control of agents is the author's next plan and not yet released.
"Neural networks are already clean": the AI black box may not be so black (see item 2). Former OpenAI interpretability researcher Nick Cammarata published a long essay arguing that, after studying at scale with agents, most neurons turn out to have clear explanations. The English original has 3.465M views; the Chinese summary by @MaxForAI has 217K views, 2,598 likes and 2,253 bookmarks. The best angle is to explain to non-technical readers how the classic "cat face plus car front" example was reinterpreted, and why this matters for AI safety, while making clear that the author himself calls it a "working hypothesis" — do not write it as "the black box has been cracked". Format: standalone original post. Status: ready to draft; key numbers checked against the original.
Expensive models as the controller, cheap models do the work: dividing labor across subscriptions and agents (see item 3). Tw93's long post explains how he uses up 2 Claude Max plus 1 ChatGPT Pro (60K views, 516 bookmarks); in the same window, ideas such as "on a $20 Pro, let Opus 5.5 only do scheduling and hand the grunt work to Chinese models", "Claude plans, other agents execute, Claude accepts", and "I don't recommend local deployment" appeared at the same time. The best angle is a division-of-labor checklist for ordinary users: which steps are worth the most expensive model, which can go to cheap or free models, and the cost and risk of multiple accounts and devices. Format: standalone original post, possibly with your own division table. Status: ready to draft; each approach is personal experience with no independent comparison.

Candidate topics¶

1. sofatap (couch slouch): operate your computer lying on the couch, and the launch video is hotter than the product¶

Original posts and traction:
@turingou Guo Yu, 10-16 12:33 Shanghai, 1,713 likes, 189 replies, 207 quotes, 1,304 bookmarks, 620K views
Discussion samples: @ivanalog_com "launch marketing that truly understands human nature is this plain"; @xiaohu Xiaohu quoted it saying "great product demos often use the plainest expression", 36K views
Content and focus: The author says most of his vibe coding over the past two weeks happened in the living room on the TV-connected computer; while agents ran tasks, every time he needed to tap or reply he had to get up for the keyboard and mouse, so he built sofatap. It supports Mac, Windows, Linux and some Google TVs, is open source and free, and is in TestFlight public beta. The one page of quotes this run read (about 20) mostly praise or joke about the launch video; one person said "I watched it 5 times and still don't remember what the product does", and another said they had already vibe-coded an Android / HarmonyOS version.
Why it's worth doing / angle: Write about "where the human sits in the agent era": people move from operators to approvers, and what they most need may not be a stronger IDE but a lighter remote and confirmation method. Or write a separate breakdown of "why this launch post got 600K views".
Timeliness and verification: within 24 hours. Checked against the GitHub repo chekusu/sofatap: the description reads "turns your iPhone into a touchpad, air mouse and text remote for Windows, Mac and supported Linux desktops"; the license page was not retrieved. Voice control of agents is what the author says comes next and is not live.

2. "Neural networks are already clean": good news for interpretability¶

Original posts and traction:
@MaxForAI Max For AI, Chinese explainer, 10-16 10:41 Shanghai, 2,598 likes, 2,253 bookmarks, 217K views (off-list)
English original (fact source and background): @nickcammarata, 10-15 23:40 Shanghai, 18K likes, 3.465M views
Discussion sample: @rickawsb extends it into "the dawn of AI alignment", 14K views
Content and focus: Nick did early mechanistic interpretability work at OpenAI; the tools back then could read only about 2% of neurons, which made him pessimistic. Now he uses agents to observe neurons, form hypotheses and run experiments automatically, and concludes that most neurons actually have clear functions; the famous polysemantic "cat face plus car front" neuron is really detecting a shared V shape. rickawsb pushes further to "deceptive alignment has nowhere to hide" and "the weak can control the strong"; that part is speculation.
Why it's worth doing / angle: Ordinary readers strongly associate AI with "a black box"; this is a rare scientific good-news story that can be explained clearly. Use the cat-and-car example to explain "why we used to misread it", then why being able to read a model's internals helps safety. Separate the original conclusions from the Chinese community's extended interpretations.
Timeliness and verification: the Chinese post is within 24 hours; the English original is 24–48 hours. Checked against the original essay "Neural Networks Are Already Clean": the author labels it a "working hypothesis"; confirmed that in preliminary audits of Qwen3-32B roughly 90% of neuron poles have one main job and about 7% have two related jobs; about 99% of the ConvNeXt last layer is interpretable (the original 2014 InceptionV1 is about 50% with new tools, about 70% retrained); the cat-car neuron 4e:55 corresponds to a V shape; removing the weakest 75% of activations drops accuracy from 72% to 71%, while removing the strongest 7% drops it to 4%. MaxForAI's "99%–100% in the last layer" refers to modern vision models, not all models.

3. Expensive models as the controller, cheap models do the work: how to divide labor across subscriptions and agents¶

Original posts and traction:
@HiTw93 Tw93, 10-16 09:12 Shanghai, 561 likes, 167 replies, 516 bookmarks, 60K views
@wquguru WquGuru "use Opus 5.5 as the scheduling brain on a $20 Pro", 10-16 01:02 Shanghai, 185 likes, 227 bookmarks, 58K views (off-list); @dotey Baoyu adds "let Fable be the Advisor, GPT 6.1 Sol does the chores", 10-17 00:38 Shanghai, 18K views
Discussion samples: @howie_serious "Claude plans, other agents execute, Claude does final acceptance", 9,868 views, 52 replies; @lxfater Tiechuiren "I don't recommend local deployment; a dozen cobblers can't match Claude the Zhuge Liang", 21K views; @York_981 uses a US host plus tailscale to connect several devices
Content and focus: Tw93 uses Opus 5.5 medium for ordinary tasks and high for complex ones, GPT-6-Astra medium in Codex, and has Claude call Codex's computer use to operate a VM, a real Windows machine and an Android tablet for multi-environment testing, with 7–10 agent tabs running in parallel. Another camp argues for using the expensive model only for planning and acceptance and handing execution to cheap models.
Why it's worth doing / angle: Readers care most about "how much to spend and how to spend it well". Write a division checklist: use the strongest model for planning, task breakdown and acceptance; hand highly deterministic execution to cheap or free models; factor in the hardware cost of local deployment and the capability loss after quantization.
Timeliness and verification: Tw93, howie, lxfater and Baoyu are within 24 hours; WquGuru is 24–48 hours. All approaches are personal experience without comparable cost or quality measurements; do not write it as "the cheapest approach has been verified".

4. Anthropic internal report: agent hours are 6x employee hours, junior technical hires fell from 27% to 9%¶

Original posts and traction:
@dotey Baoyu, Chinese summary, 10-17 06:30 Shanghai, 69 likes, 44 replies, 8,648 views (about 2 hours old)
English release: @PeterMcCrory Peter McCrory, 10-17 04:48 Shanghai, 471 likes, 41K views
Content and focus: Anthropic published two reports: one on how AI has changed its own hiring and ways of working, the other comparing employees' and ordinary Americans' expectations of AI. The Chinese post is about 2 hours old and still in early spread.
Why it's worth doing / angle: Very direct for working readers. Write "what human work becomes at the company that uses AI most deeply": the report says about 80% of Claude session time is spent waiting for human replies, agent-management work is rising, and hiring increasingly values autonomy and judgment. Also note this is data from a fast-growing AI company and cannot be applied directly to ordinary businesses.
Timeliness and verification: within 24 hours. Checked against Anthropic's report "Employment at Anthropic": confirmed the junior share of technical department employees fell from 21% to 14% and the junior share of technical hires fell from 27% to 9%; in August interactive agent hours were about 6 to 1 against employee hours (employees counted at 40 hours per week); code commits per engineer are 5x six months ago; among 192 engineers tracked after adopting coding agents, agent-written code showed no difference in removal or revert rates from human-written code; sessions spend about 80% of time waiting for humans; agent-management work rose from 10% to 22%. Numbers in the second report on the expectations gap were not checked item by item.

5. Growing accounts with Opus 5.5 videos (continuation): from one-sentence clips to matrix digital-human lead generation¶

Original posts and traction:
@ai_xiaomu Huang Xiaomu "I was among the first to discover Opus 5.5 can grow accounts", 10-16 18:23 Shanghai, 136 likes, 28K views (off-list)
@qihang_zeng6688 Qihang "28 videos, 176K new followers after Opus 5.5 launched", 10-15 18:37 Shanghai, 7,468 views; @Gorden_Sun connected Opus 5.5 to TTS, ASR, image and video generation to make a skit and published the prompt, 10-16 23:05 Shanghai, 88 likes, 6,923 views
Risk sample: @ameliax111 says she was penalized on Douyin several times, partly for not labeling AI-generated content, and asks whether the account can be saved, 10-15 15:03 Shanghai, 19 replies, 5,966 views
Content and focus: Continues the video topic from the 10-15 and 10-16 issues. The increment is that discussion has shifted from "can it be made" to "how to monetize": some say several creators are using "matrix digital humans plus Opus 5.5" to get leads, others share one-sentence prompts. At the same time a real request for help about platform labeling and penalties appeared.
Why it's worth doing / angle: Write "read the platform rules before growing an account with AI video": labeling of AI-generated content, down-ranking risk for reposted and templated content, and the limits of mass matrix accounts. That adds more than yet another prompt.
Timeliness and verification: Huang Xiaomu and Gorden_Sun are within 24 hours; Qihang and Ameliayu are 24–48 hours. Follower and lead numbers are creators' claims and unverified; the penalty reasons are self-reported. Another post in the window used "full-strength, not dumbed-down" to promote a relay group and was excluded.

6. Tencent Hunyuan open-sources Hy Memory 2.0: a 3B-active small model dedicated to memory¶

Original posts and traction:
@TencentAI_News Tencent official (English), 10-16 17:22 Shanghai, 981 likes, 902 bookmarks, 64K views
Chinese spread: @NFT_Chen, 10-16 17:58 Shanghai, 150 likes, 14K views (off-list); @AYi_AInotes, 99 likes, 46 replies, 11K views (off-list); @shao__meng meng shao technical breakdown, 3,532 views
Content and focus: An open-source trio: the memory-writing model Hy-Memory-30B-A3B, the memory framework hy-memory SDK, and the local-first desktop tool MemHub, which claims to let Cursor, Codex, Claude Code and other tools share one memory. AYi writes that what he can stand least is "manually re-explaining the same project background five times in different editors" and asks readers, drawing 46 replies.
Why it's worth doing / angle: "Switch tools and the AI forgets everything" is a real pain point for many people. Write "why memory should be an independent layer", and note that the 3B model "beating Opus 5" applies only to the single task of memory writing.
Timeliness and verification: within 24 hours. Checked against the Hugging Face model card: confirmed about 3B active parameters and that it can directly replace the writer model inside Mem0 and MemOS; the four-framework macro average is 55.49 versus 54.62 for Claude Opus 5, and on EverMemOS, which was not used in training, 56.89 versus 54.04; scores are on a fixed 20% benchmark subset and are self-reported. The GitHub repo page exists, but its content was not retrieved. The 51.8% storage reduction and multi-hop recall rising from 50% to 90% are claims from the official post.

7. Magpie goes fully plugin-based: one gateway for multiple subscriptions and models¶

Original posts and traction:
@yetone yetone, 10-16 05:09 Shanghai, 397 likes, 41 reposts, 195 bookmarks, 45K views
Same author on the "slim Claude system prompt" toggle, 10-17 03:34 Shanghai, 79 likes, 12K views
Content and focus: Magpie opens subscription integration, gateway middleware, routing policy, usage callbacks, tools, agent support and pages to npm plugins. The second post explains it used to run requests through real Claude Code, adding about 1K tokens of system prompt per request; a new toggle replaces that, and the author measured about 35% less input over 3 turns for a lightweight client, while warning that "it looks less like Claude Code itself; weigh the account risk yourself".
Why it's worth doing / angle: Can be tied to item 3 to explain "model routing" for ordinary people: who decides how requests are spread across multiple accounts and models. Make the author's own account-risk warning clear; it is not suitable as a money-saving tutorial.
Timeliness and verification: the plugin post is 24–48 hours; the prompt-toggle post is within 24 hours. Features and token-saving numbers are the author's own claims; the toggle ships "in the next version" and is not live yet.

8. Orca: an open-source workbench where one agent directs a team of agents¶

Original posts and traction:
@lxfater Tiechuiren, 10-15 21:57 Shanghai, 236 likes, 383 bookmarks, 25K views
@poolmoon6 Hantang "4 steps, up and running in 5 minutes", 10-16 14:16 Shanghai, 93 likes, 117 bookmarks, 9,349 views (off-list)
Content and focus: Orca puts Claude Code, Codex, OpenCode and others in one workbench, with an isolated git worktree per task; a coordinating agent can split tasks, dispatch work and collect results, and there is a mobile app. Bookmarks clearly exceed likes, suggesting more people are saving it to try.
Why it's worth doing / angle: Write what "multi-agent parallelism" means for non-programmers: hand the same job to several agents at once and pick the best plan, and when parallelism makes things messier. Good to pair with a screenshot of your own setup.
Timeliness and verification: Tiechuiren's post is 24–48 hours; Hantang's is within 24 hours. Checked against the Orca website: confirmed it runs Claude Code, Codex and other coding agents in parallel, each in its own worktree, has an iOS / Android companion app, and states "Open source under MIT".

9. Claude account bans (continuation): banned on a business trip, store refunds and a "post-mortem"¶

Original posts and traction:
@blackanger AlexZ "an American friend from Hong Kong came to Shenzhen for a Rust conference and woke up with a banned Claude account", 10-16 02:01 Shanghai, 473 likes, 112K views (off-list); @dotey Baoyu quoted it: "Americans on business trips to China get banned from Claude too", 10-16 07:29 Shanghai, 108 replies, 59K views
@dasungta_zhang Guiji Zhongnian "post-mortem seven days after the ban", 10-16 18:14 Shanghai, 61 likes, 62 bookmarks, 15K views (off-list); @pengchujin Jiangzibiao "the App Store refunded my Claude subscription", 39 replies, 28K views (off-list)
Content and focus: Continues item 1 of the 10-16 issue. Two increments: first, it is not only Chinese users inside China who are banned; foreign users on business trips to China are hit too; second, refund paths diverge — the post-mortem says store subscriptions were refunded "instantly" while the website subscription refused a refund, and another person received an App Store refund. The post-mortem's conclusions are personal experience.
Why it's worth doing / angle: Keep avoiding anti-ban tutorials; instead write "how costly it is to stake your work on an account that may fail at any time", adding the refund differences between subscription channels after a ban (noting these are personal experiences).
Timeliness and verification: AlexZ and Baoyu are 24–48 hours; the post-mortem and refund posts are within 24 hours. Ban reasons and refund rules are participants' claims with no official explanation seen; "we only hire people who can use Claude reliably", already cited yesterday, is not repeated.

10. "Dumbed down" doubts: Opus 5.5's taste slipping, GPT quota cuts, and which model Dot uses¶

Original posts and traction:
@leo114119 Leo "netizens report GPT is again dumbed down at scale and quotas cut", 10-16 10:16 Shanghai, 82 likes, 30K views (off-list)
@dtzy_88 Darren "Claude's UI is actually worse than GPT's", 10-16 20:26 Shanghai, 22K views (off-list); @leo114119 says Pro users' Dot uses GPT-6.1 Sol Low, 9,843 views
Counter-sample: @llqoli "GPT 6.1 Sol feels really comfortable for UI now", 12K views (off-list)
Content and focus: On the same day, some said Opus 5.5 has been dumbed down and suspected a new model is coming, some said GPT was dumbed down with quota cuts, and some found GPT 6.1 Sol pleasant for UI. Most posts have only screenshots and personal impressions.
Why it's worth doing / angle: Write "how to tell whether the model got dumber or it's your luck or prompts": keep a fixed set of your own test questions, record tier and time, and separate the model itself from the product wrapper (e.g. the model behind Dot). More useful than joining the "it's dumbed down" chorus.
Timeliness and verification: all within 24 hours. All "dumbed down" claims are user impressions or hearsay with no official confirmation; Dot's model configuration is unverified.

11. Antigravity opens Gemini 4 Argon to paid Google AI Pro¶

Original posts and traction:
@dotey Baoyu, 10-16 07:36 Shanghai, 128 likes, 122 replies, 50K views
@leo114119 Leo's quota test "back-calculated weekly limit is about $10, pitifully small", 10-16 11:32 Shanghai, 11K views (off-list); @Gorden_Sun adds that student plans, Jio and Pixel channels don't qualify, 25 replies
Content and focus: The 10-15 issue covered Argon hands-on tests; the increment is that paid Pro users can now use it in Antigravity, with a separate weekly quota. Chinese discussion centers on "the quota is too small" and "which Pro channels don't count".
Why it's worth doing / angle: Good for a practical info post: who can use it, from which entry point, and how the quota works, so readers don't buy the wrong channel.
Timeliness and verification: Baoyu's post is 24–48 hours; Leo's and Gorden_Sun's are within 24 hours. Checked against the Antigravity models doc: confirmed Gemini 4 Argon on Google AI Pro is limited to non-trial subscriptions, unavailable on Free and Google AI Plus; not in the Antigravity IDE, but available in the Antigravity App, CLI and IDE extensions. The specific weekly quota is not officially published; "about $10" is a user's estimate.

12. OpenCode free Flash models (continuation): is the free GLM 5.3 Flash any good¶

Original posts and traction:
@mranti Michael Anti "did OpenCode just raise money? It opened DeepSeek 4.1 Flash and GLM 5.3 Flash for free back to back", 10-16 15:28 Shanghai, 15K views (off-list)
@geekbb Geek "the free GLM 5.3 Flash is useless; 20 seconds, by then the moment's gone", 10-16 22:53 Shanghai, 8,804 views; @xueyu1125 Xueyu compiled this week's list of free models, 4,733 views (off-list)
Content and focus: Continues item 2 of the 10-16 issue; today's daily already covered OpenCode's free GLM 5.3 Flash. The increment is the first wave of user feedback: some find responses too slow, others are compiling which free channels still work.
Why it's worth doing / angle: Write "what free models are good for": slow but cheap models suit background batch and non-urgent execution, not real-time chat. Can be merged with item 3 into one post.
Timeliness and verification: within 24 hours. Speed impressions are personal feedback; the free period and limits have no official statement seen.

13. Microsoft Copilot head Lamanna: "All in Claude" earns him a successor-candidate spot¶

Original posts and traction:
@foxshuo Lanxi, 10-15 14:32 Shanghai, 159 likes, 35K views
Content and focus: Lanxi relays a The Information report: Copilot head Lamanna proposed last summer replacing Copilot's built-in GPT model with Claude, which was controversial internally; he convinced the CEO by demoing Claude editing Excel and making PPTs. Copilot's enterprise reputation then turned around, he now also oversees Office and Teams, and is seen as one of the successor candidates.
Why it's worth doing / angle: Write "why big companies are willing to pay more for a better model": how much business difference choosing the right model can make for a product, a persuasive case for readers pushing AI tools at work.
Timeliness and verification: 24–48 hours. All details come from Lanxi's relay of The Information report, which was not read this run; numbers such as 200% user growth and 30M+ paid seats need verification; when citing, state that it is a media report.

14. Tesla's new Roadster: "I thought it was AI, but it's real" (tech crossover)¶

Original posts and traction:
@xiaohu Xiaohu, 10-16 10:00 Shanghai, 2,061 likes, 204 reposts, 278K views
@dotey Baoyu's long post, 10-16 10:22 Shanghai, 313 likes, 124K views; @abskoop ahhhhfs "I thought it was AI, turns out it's real", 295 likes, 95K views
Content and focus: The launch demo car used compressed-air jets to flip upright mid-air and land. Among reactions in the Chinese AI circle, a recurring line was "my first reaction was that it's an AI video". Baoyu's long post clarifies that the "flight" is just a few seconds of jet-assisted jumping, far from manned flight, and lists deposits, versions and the repeatedly delayed delivery dates.
Why it's worth doing / angle: Not a car review, but "after AI video floods everything, real demos have to prove themselves first": what it means for brand launches and content credibility when people's first suspicion is AI.
Timeliness and verification: within 24 hours. Vehicle specs and the 2028 production date follow Baoyu's summary and were not separately checked this run; remote-controlled driving and the 260 kg air tank are media reports not confirmed by Tesla.

15. JEV 1.14.0 adds multimodality: a fast decision model turned into a "browser hand"¶

Original posts and traction:
@op7418 Guizang, 10-15 10:56 Shanghai, 298 likes, 267 bookmarks, 47K views
@GitTrend0x GitTrend roundup of JEV apps, 10-15 13:28 Shanghai, 180 likes, 220 bookmarks, 12K views (off-list)
Content and focus: Guizang says JEV 1.14.0 supports multimodality, adds a box type that returns target positions, raises context to 262K, and improves multilingual ability. GitTrend rounds up community uses such as a "browser hand that clicks the next cell in 200ms", a local decision core, and skill routing. Today's daily noted that Liquid AI's d1-ultra surpassed the JEV series on a decision-model leaderboard.
Why it's worth doing / angle: Value watch. Write "not every AI needs to chat": small models that make fast choices suit clicking buttons, classification and routing, working alongside large models.
Timeliness and verification: 24–48 hours. Version features come from Guizang's relay; official release notes were not read this run; the projects GitTrend listed were not opened one by one.

16. Apple subscriptions support cross-developer bundles: a new play for indie developers¶

Original posts and traction:
@waylybaye Baye, 10-15 13:14 Shanghai, 113 likes, 58 bookmarks, 26K views
Content and focus: Baye says he had been thinking of making a "Panda One" that unlocks all his apps with one purchase, and didn't expect Apple to officially support this kind of feature, even across developers.
Why it's worth doing / angle: Value watch. AI lets indie developers ship a dozen small products a year (like Guo Yu in item 1), and bundled subscriptions address the problem of "many products, weak individual payment"; write it up as a monetization idea for indie developers.
Timeliness and verification: 24–48 hours. Apple feature details come from the post Baye quoted; Apple's official documentation was not checked this run; needs verification.

17. Connecting AI to the BaoTa panel MCP: grant only low-risk permissions¶

Original posts and traction:
@abskoop ahhhhfs, 10-15 18:24 Shanghai, 6 likes, 21K views
Content and focus: The author connected AI to the MCP of the server panel BaoTa and stresses not to tick write and command-execution permissions.
Why it's worth doing / angle: Value watch. Interaction is very low but views are decent, and it is the point ordinary people most easily miss when wiring an agent into a real system. Write "the minimum-permission principle for agents", with a screenshot of your own permission choices.
Timeliness and verification: 24–48 hours. Only one experience post from the author; BaoTa MCP permission items were not checked.

18. An agent recommended a podcast with zero likes on YouTube but 20K plays on Xiaoyuzhou¶

Original posts and traction:
@oran_ge Orange AI, 10-16 16:11 Shanghai, 76 likes, 77 bookmarks, 11K views
Content and focus: The author says the podcast an agent recommended had not a single like on YouTube yet 20K plays on Xiaoyuzhou, and he got a lot out of it.
Why it's worth doing / angle: Value watch. A personal agent's value is not only doing work but also finding, across platforms, content the algorithm won't push. Write "how having an agent find information differs from scrolling a recommendation feed".
Timeliness and verification: within 24 hours. Single-person experience; the author did not explain how the agent found the podcast.

Original posts and traction:
@geekbb Geek, 10-16 09:25 Shanghai, 33 likes, 66 replies, 12K views
Content and focus: The author says the new extension's page summaries work very well, like moving ChatGPT into the browser, but it only allows signing in with an official OpenAI account and can no longer call local Codex, and asks whether it, like Claude, no longer supports third-party APIs. Replies outnumber likes, indicating the limitation sparked discussion.
Why it's worth doing / angle: Value watch. Write "in-browser AI assistants are getting stronger but also more closed", helping readers choose between official extensions and third-party tools.
Timeliness and verification: about 23 hours, within 24 hours. Update contents and limitations come from the author's experience; official release notes were not read.

This run has 19 items, short of 20. In the account patrol, three groups — abskoop / ruanyf / tinyfool, DashHuang / geekbb / tualatrix, and shao__meng / AlchainHust — were crowded out by high-frequency accounts and not paged, so the earliest roughly 4–10 hours of the window may be incomplete; deduplicated against the 10-15 and 10-16 Chinese topic briefs, the 10-16 English topic brief and today's daily; topics without substantive new angles, such as Handwired, Kling 4.0 / OmniHuman 2.0, AI fake-human replies and DeepSeek's product department, are not listed separately.