Skip to content

← Chinese Topics

X Chinese Topic Brief

Date16 November 2026 · MondayWindow2026-11-14 08:30 to 2026-11-16 08:30CoverageComplete

Translation

English translation of the Chinese original.

Today's priorities

  1. Turn a correction to an agent into a lint rule; don't pile more prose into agents.md (see candidate 1). Worth doing now: Yesterday's English topic brief already covered Lauren Tan (@poteto) posting the talk. The Chinese-language post that spread farthest is 宝玉's eight-point translation, with 857 likes and 80,600 views. The best angle: don't stop at "write less in agents.md." Spell out the practices that can be checked against the talk — if a correction can be a lint rule, a type, or a test, don't only put it in a doc; code comments and instruction files easily become a place for the agent to justify shortcuts. Her team figure is that they merged 11,000 pull requests last month, with about 0.5% reverted. Suggested form: an original post, or a quote of 宝玉's post with the source added. The talk page and video have been verified; it can be drafted. The 11,000 and 0.5% figures were given by the speaker on stage and have not been independently re-measured.

  2. The claim that "Optimus came back with Starship" spread widely, but it does not match the official mission notes (see candidate 2). Worth doing now: 小互's post has 3,043 likes and 264,000 views, the widest-circulating Chinese-language post in this window. It says Optimus flew on Starship, circled Earth, and was caught with the ship by the tower. The best angle: start from the official SpaceX Flight 15 page — liftoff on November 13, 26 Starlink V3 satellites deployed, and Starship caught by Mechzilla for the first time after about 22.5 hours. The page never mentions Optimus. In the quote thread, someone already labeled it a fake video. The first catch of Starship is officially confirmed; a robot being on board is not. Suggested form: an original post, with the official link in the body. The official mission notes have been verified; it can be drafted. Whether the video itself was AI-generated is still unverified — do not write that it has already been confirmed as a fake.

  3. In a simulation, giving each person an agent did worse than sharing one (see candidate 3). Worth doing now: this follows yesterday's Chinese brief on "how a person should manage a group of agents." 向阳乔木's post has 53 likes and 18,800 views. The conclusion is very short, and it is easy to mix up with another Anthropic paper. The best angle: make clear this is the simulation in Worse Together, not a field study in a real office. In all four settings (a shared API budget, a clinic calendar, a group food order or booking, and a merge queue), "one agent serving everyone" beat "each person has an agent, and the agents then coordinate." On Claude models the gap was about 10 to 45 percentage points. Do not write it as the same finding as AI Organizations, which says multi-agent organizations are more effective at business goals and less adherent to behavioral specs. Suggested form: an original post. The official blog post has been verified; it can be drafted.

Candidate topics

1. Follow-up: Chinese notes on Lauren Tan's talk — write the check first, then talk about full automation

Original posts and traction: 宝玉 @dotey, original post, 2026-11-15 13:18 (Asia/Shanghai), 857 likes, 80,600 views. The post translates Lauren Tan's talk "All Code Is Slop (Except Mine)" at AI Engineer CODE 2026. The speaker @poteto posted a video, with 7,125 likes and 925,000 views.

What it says, and what to watch: 宝玉 boiled it down to eight points. The core: each time you correct an agent, turn that correction into a lint-style rule that fails the build; add less prose to agents.md. Bad patterns and "justified patches" in the repo will be treated as convention by the next agent; comments are often used to excuse shortcuts. In the talk she also gave the team figures of 11,000 PRs merged last month and about 0.5% reverted.

Why it's worth doing / an angle: Chinese readers don't need another retelling of "write fewer docs." They need a judgment they can use immediately: which corrections belong in a checker, and which belong in a description. The talk page can be cited.

Timing and verification: within 24 hours; follow-up. Yesterday's English topic brief already covered the same talk; this item only covers practices and figures in the Chinese translation that can be checked. The talk page, video, and transcript confirm the 11,000 PRs and about 0.5% revert rate; those are the speaker's figures and have not been independently re-measured. The official page lists her as a SpaceXSI Principal Engineer; 宝玉 wrote "SpaceXAI, responsible for the Grok Bot desktop app." Use the talk page and her own post for the job description.

Original posts and traction: 小互 @xiaohu, original post, 2026-11-15 10:31 (Asia/Shanghai), 3,043 likes, 264,000 views, 432 reposts, 174 replies, 44 quotes.

What it says, and what to watch: The post says an Optimus was on this Starship flight, went up with the ship, circled Earth, came back, and was caught with Starship by the tower. Someone in the quote thread already called it a fake video. The official SpaceX Flight 15 page says Starship lifted off from Starbase, Texas, on November 13, deployed 26 Starlink V3 satellites, and about 22.5 hours later was caught by Mechzilla for the first time. Optimus is not mentioned anywhere on the page. The Super Heavy booster was not caught by the tower on this flight and aborted recovery over the water.

Why it's worth doing / an angle: This is the Chinese-language post with the most views in this window, and the one most likely to weld "Starship was caught for the first time" to "a humanoid robot was on board." Lay out the official mission notes first, then remind readers that a popular video is not a substitute for the launch provider's own payload list.

Timing and verification: within 24 hours. The official mission notes have been verified; there is no official basis for Optimus being on board. Whether the video was AI-generated is still unverified. This report does not treat "confirmed fake" as a fact.

3. Anthropic: in a simulation, one agent per person did worse than one shared agent

Original posts and traction: 向阳乔木 @vista8, original post, 2026-11-15 10:18 (Asia/Shanghai), 53 likes, 18,800 views.

What it says, and what to watch: He wrote that an Anthropic article concludes that giving everyone on a team their own agent performs worse overall than sharing one, and that the piece is in the replies. The article is Worse Together, published on November 11. The study built four simulated environments and compared "one coordinator agent serving several users" with "one agent per user, and the agents then coordinate." The former was better in all four environments, by about 10 to 45 percentage points on Claude models. The post also notes behaviors that appeared as teams grew — stalling, overriding other people's requests, and making unsupported claims about others — and that fixes depended on the setting.

Why it's worth doing / an angle: For readers who are giving a family, a small team, or a shared repo a set of agents, the paper asks whether they should share one or give each person their own. Any write-up must say this is a simulation, not a recap of a real office; and must not reuse the conclusion of the other paper, AI Organizations.

Timing and verification: within 24 hours. The official blog post confirms the direction of 向阳乔木's summary and the 10-to-45-point range; scenario-level scores should follow the charts in the original.

4. g-harness-evals: with the model fixed, pass rates barely separated; cost could differ by about 2x

Original posts and traction: @mylifcc, original post, 2026-11-14 11:02 (Asia/Shanghai), 50 likes, 51,300 views.

What it says, and what to watch: The Chinese post summed it up as "swap the framework around the same frontier model and pass rates are similar, but cost per run differs by more than 2.5x," and attributed the project to Y Combinator president Garry Tan. The repo is garrytan/g-harness-evals. The cost findings page is narrower: on the 20-task, fixed-model set, harness pass rates sat between 71.3% and 78.8% and could not be told apart; cost per solved task on Opus was $0.44 for OpenClaw and $0.22 for Claude Code, about 2x, not "more than 2.5x." About 81% of the gap came from system prompts and tool results being re-sent (cache writes). The repo itself notes that these tasks are too easy, that the cost gap has no statistical test, and that the author has investments or ties to some of the projects tested — all of which should be stated together.

Why it's worth doing / an angle: Chinese readers comparing Claude Code, Codex, OpenClaw, and Hermes are easily pulled toward "which one wins." A usable opening is: first ask "how much does each solved task cost," then warn against copying the "more than 2.5x" figure from the Chinese post. On the separate 52-task coding set, OpenClaw on Opus also lagged Claude Code on pass rate; numbers from the two sets must not be mixed.

Timing and verification: 24–48 hours. The repo and the cost findings page have been checked for the figures above; the Chinese post's "more than 2.5x" does not match the original "about twice," and a write-up should correct it. These are dev-set results at pinned versions, not a claim that holds for every everyday task.

5. Chinese users are getting lost in Bot names: Grok Bot, OpenClaw, and Hermes are not the same kind of thing

Original posts and traction: ahhhhfs @abskoop, original post, 2026-11-15 14:58 (Asia/Shanghai), 206 likes, 42,800 views.

What it says, and what to watch: He asked what distinguishes Grok Bot, OpenClaw, Hermes, Openbot, OpenMausbot, and OpenBot — "too many bots" — and posted a chart he had worked out. In the same window, others discussed self-hosting OpenBot and a direct Tailscale connection; see candidate 14. This item is based on the post text; the specific classifications in the chart have not been checked item by item and are not repeated here.

Why it's worth doing / an angle: For non-programmers, the first obstacle is often not "which model is stronger," but names that sit too close together with completely different boundaries. The chart can be quoted; redraw it yourself under "who built it, where it runs, whether it can be self-hosted, and whether it is an assistant or a coding harness." Do not invent a taxonomy.

Timing and verification: within 24 hours. The circulation is real; the comparisons in the chart still need to be checked against each project's official notes before a detailed write-up.

6. Some people have switched from Codex / Claude Code to Antigravity; others say its harness clearly holds it back

Original posts and traction: 海拉鲁编程客 @hylarucoder, original post, 2026-11-15 15:04 (Asia/Shanghai), 223 likes, 72,800 views, saying he had "fully switched from Codex / Claude Code to Antigravity." 图拉鼎 @tualatrix, original post, 2026-11-15 10:27, 61 likes, 20,100 views, saying the Gemini 4 model is fine but Antigravity's harness is not, and that even ACP adaptation was incomplete when he called it that way. Off-list @voidJan, original post, 2026-11-15 07:35, 18 likes, 11,800 views, scoring Claude Code 8 and Antigravity at most 3, and complaining about slow responses.

What it says, and what to watch: This is not "another new IDE launch." It is the same product drawing both "I've switched" and "the harness doesn't match the model" in the Chinese-language circle. 图拉鼎's observation is more specific: the problem is the harness and protocol adaptation, not necessarily Gemini 4 itself.

Why it's worth doing / an angle: This can be written as "evaluate the model and the new harness separately." Don't turn one person's switch into a general conclusion, and don't only quote posts from the watch list.

Timing and verification: the representative posts span "within 24 hours" and "24–48 hours." All of the above are individual impressions; no controlled comparison was run.

7. Follow-up: some people find Gemini 4 like a "slide-deck boss"; others use Carbon for long write-ups

Original posts and traction: Axton @AxtonLiu, original post, 2026-11-15 10:00 (Asia/Shanghai), 64 likes, 19,200 views. 歸藏 @op7418, original post, 2026-11-14 14:01, 204 likes, 40,700 views.

What it says, and what to watch: Axton said that after trying Argon and Carbon recently, the model is strong at overall plans and frameworks, then discounts sharply when it has to execute — like a "slide-deck boss." 歸藏, by contrast, called Carbon strong at writing and had it summarize 30 takeaways from the past nine episodes of Next Token. The two need not conflict: one is about strong planning and weak execution; the other is about long-form organization being a good fit.

Why it's worth doing / an angle: Earlier Chinese briefs already covered impressions of Hydrogen and Argon. This one can add a more specific division-of-labor question — is it better at a first outline, or at handing over finished work? Don't turn two people's impressions into an official split of capabilities.

Timing and verification: Axton's post is within 24 hours; 歸藏's is 24–48 hours; follow-up. Both are personal impressions, pending hands-on tests.

8. One personal bill: a $20 Claude subscription was counted as about 60x API value

Original posts and traction: @gm365, original post, 2026-11-14 12:42 (Asia/Shanghai), 60 likes, 22,000 views. 向阳乔木 @vista8, original post, 2026-11-15 00:55, 69 likes, 22,300 views, noting that when account bans feel risky, one can look at programming tools such as Factory.ai that are easier to pay for.

What it says, and what to watch: gm365 wrote that after a Claude account he had KYC'd with a passport was banned, CodexBar showed that over the past 30 days his $20 Claude Pro usage was worth about $1,239 at API prices, or 1:60. He then compared that with resellers: even at a 90% discount, the multiplier is only about 1:10. This is one person's single-month statistic, and he had not used the full quota; it must not be written as an official multiplier for everyone. 向阳乔木 pointed to a cleaner path: switch to a regular third-party programming tool, rather than keep looking for workarounds on the account.

Why it's worth doing / an angle: What readers need to judge is why the official subscription looks cheap, and which replacements after a ban are legitimate. Any write-up must label the $1,239 and 1:60 figures as the author's own CodexBar stats. Do not describe ways to bypass KYC, open new accounts, or use regional low-price sign-ups.

Timing and verification: 24–48 hours. The $1,239 and 1:60 figures are the author's account, pending verification. The claim that Factory.ai "won't ban you" is also a source statement; its terms and actual enforcement have not been checked.

9. Follow-up: someone followed the "always-on" article and made it so the phone gets a notice

Original posts and traction: 花叔 @AlchainHust, original post, 2026-11-14 14:08 (Asia/Shanghai), 80 likes, 19,900 views.

What it says, and what to watch: After reading DeepSeek's Super-agent work and Anthropic's "Building always-on agents," he built a version in FanBox: when the agent in the terminal stops and waits, and you are not at the computer, the phone (WeChat or Feishu) gets a one-line notice. He also wrote up four things he ran into while following the articles. This is a hands-on follow-up to yesterday's Chinese brief that "always-on is an experience, not keeping the model running," not a new official release.

Why it's worth doing / an angle: This can cover the concrete problems of taking "always-on" all the way to a phone notification, rather than explaining the five parts again. It should be labeled as a personal implementation, not an official Anthropic or DeepSeek product.

Timing and verification: 24–48 hours; follow-up. Implementation details and the "four things" follow the original post and have not been independently tested.

10. Someone's personal assistant is already looking for a house, and a bot placed an order for two Mac minis

Original posts and traction: 郭宇 @turingou, Wanman post, 2026-11-15 17:56 (Asia/Shanghai), 96 likes, 28,500 views; auto-order post, 2026-11-14 16:40, 108 likes, 53,900 views.

What it says, and what to watch: He wrote that, for finding a house in the countryside, Wanman is already more useful than Grok Bot, and that a vertical personal assistant that runs proactively around the clock is the direction. In the other post, he said his bot searched the web for cheap Mac mini M5 Pro machines and placed an order for two that day. These are the author's own usage claims, not a product review, and they do not show the full authorization and confirmation flow.

Why it's worth doing / an angle: A useful question is: what confirmation should a person see before an assistant "places an order by itself"? Don't write the two posts as a consumer capability that is already widely available, and don't mix them up with Wallet, which is only for U.S. ChatGPT Pro users.

Timing and verification: the Wanman post is within 24 hours; the order post is 24–48 hours. Whether payment for the two machines was completed, and whether the orders can be cancelled, have not been verified.

11. Some people say Codex has become hard to use, and that Astra has not been updated in two months

Original posts and traction: Orange AI @oran_ge, original post, 2026-11-15 23:42 (Asia/Shanghai), 72 likes, 15,200 views.

What it says, and what to watch: He relayed friends' feedback that Codex has become difficult to a new degree, figured that OpenAI has not updated Astra in two months, and mentioned "藏师傅" saying that recently distilling OpenAI models may have made their own models worse. This is in-circle impression and second-hand talk, not an official release note.

Why it's worth doing / an angle: If this is written up, the opening should be "first check the date of Astra's last official update, then talk about the feel of using it," not treat "already hard to use" or "distillation made it worse" as facts.

Timing and verification: within 24 hours. Whether Astra has gone two months without an update, and whether distillation reduced capability, are both unverified.

12. Two days after stopping Opus 5.5: one user feels he was "distilled" too

Original posts and traction: 图拉鼎 @tualatrix, original post, 2026-11-15 13:10 (Asia/Shanghai), 142 likes, 34,500 views.

What it says, and what to watch: After two months of heavy Opus 5.5 use, he still missed it on the second day, but found it easier to judge work quality when collaborating with Gemini 4; and said this helped him understand why model makers distill Claude — over the past two months he had, without noticing, been distilled by it too. This is a metaphorical personal observation, not an explanation of distillation as a technique.

Why it's worth doing / an angle: This can be written as "after a long time with one model, your judgment starts to follow it," a reminder to notice where your evaluation standard comes from when you switch. Don't stretch it into a general conclusion about capability transfer.

Timing and verification: within 24 hours; a personal impression. "Distillation" here should not be written as a technical fact.

13. Chrome turned built-in browser capabilities into a Skill that coding agents can look up

Original posts and traction: 铁锤人 @lxfater, original post, 2026-11-15 09:57 (Asia/Shanghai), 52 likes, 4,956 views. Likes cleared the initial screen; views did not reach the 20,000 high-engagement band.

What it says, and what to watch: He wrote that when AI writes front ends it often uses a large pile of JavaScript for a dialog; Chrome recently turned modern browser capabilities, compatibility, and usage into a Skill that coding agents can look up as needed — for example using the native dialog for pop-ups, and Popover for floating layers. This is an introduction to an official skill pack, not the author's own evaluation.

Why it's worth doing / an angle: Useful for non-programmers too: first ask "does the browser already have this," then let the agent write code. A write-up still needs the official repo or docs link for that Skill; this item has not completed that step.

Timing and verification: within 24 hours. The exact name of the Chrome project, how to install it, and what it covers are pending verification.

14. After self-hosting OpenBot, one user realized Tailscale can connect directly

Original posts and traction: Geek @geekbb, original post, 2026-11-15 20:40 (Asia/Shanghai), 108 likes, 22,400 views.

What it says, and what to watch: After switching to OpenBot he could self-host it on a server, and found that Tailscale supports a direct connection — it does not all have to be treated as a tunnel that must take a long detour. He also noted that in the minimum requirements, 4 GB of memory only supports a very limited load. This is a personal deployment note, not a new-feature announcement from OpenBot or Tailscale.

Why it's worth doing / an angle: This can be written as "when you self-host a personal bot, the network and memory often get in the way before the model choice does." The official minimum-requirements notes for OpenBot, and the conditions under which Tailscale direct connections apply, should be checked before giving readers advice.

Timing and verification: within 24 hours. The 4 GB memory limit is the author's report and still needs to be checked against official docs.

This issue has 14 items, stopped at creative value, and was not padded to 20. On the watch list, @yetone is a protected account, so posts in the window could not be read. In addition, ruanyf, tinyfool, waylybaye, DashHuang, turingbook, and bluebird0605 did not appear as original posts in the combined watch queries; they were not re-checked with a single-account posts call, and that must not be taken as a claim that these authors had no new posts.