Skip to content

← English Topics

X English Topic Brief

Date15 November 2026 · SundayWindow2026-11-13 08:56 to 2026-11-15 08:56CoverageComplete

Translation

English translation of the Chinese original.

Today's Top Picks

  1. ChatGPT Wallet goes live: agents can pay for you, but you confirm every purchase with Face ID or a fingerprint (candidate 1). Worth doing now: this is the product story that most needs its scope explained first. The official X post has 4,847 likes and 793,061 views; the official blog is verified: Wallet is built with American Express, Mastercard and Visa, rolls out first through dots to US ChatGPT Pro subscribers, requires users to be 18 or older, and requires a card issued by an eligible US card issuer. For Chinese readers the point is not "AI can shop now" but three boundaries: the agent receives a payment token, and the user's real card number is not shared with the merchant or the agent; every purchase requires the person to review the item, total price and card used, then approve; and for now you cannot set a limit and hand purchasing over to the agent entirely. Suggested format: a standalone original post that puts the four conditions — US, Pro, 18+, eligible US card — up front, then explains which sites may wrongly block agents. A live purchase was not completed.

  2. Grok Bot can now join meetings: it works in Google Meet, Zoom and Teams, but only on Team, Enterprise and Ultra plans (candidate 2). Worth doing now: the official post has 6,672 likes and 2.58 million views, one of the most widely spread feature updates this issue; the companion poteto post (1,084 likes, 59,735 views) explains the usage more concretely than the official one. The official documentation verifies the platforms, tiers and admin switch. You can contrast it with Claude Meetings, covered in an earlier daily: when it entered beta it supported only Google Meet. Use that only as a "previously reported" comparison; this round did not re-verify Claude's current platform list. Three things must be stated clearly: the plans are Team, Enterprise and Ultra; using it means pasting the meeting link, approving a confirmation card, and being admitted by the host; and Enterprise admins can turn it off under Bot Capabilities in the admin console. Joining a meeting with the Bot was not hands-on tested.

  3. Reason: X says "20x cheaper", the official blog says 22%, 15–23% and "up to 30%" — the three numbers cannot be mixed (candidate 3). Worth doing now: with 2,123 likes and 3.26 million views, this is the most widely spread coding-related launch post this issue, and the cost multiple is exactly what is easiest to misreport. The launch post says "up to 20x lower cost"; the official 1.1 blog says 22% lower cost per completed task and 15–23% lower cost per attempt on GPT-6 Astra; the launch blog post also says "up to 30% lower cost per task than Codex, Cursor or Devin". All three are vendor claims, and they differ by an order of magnitude. More important, that 1.1 blog evaluated Coding Agent Index v1.5, and states that Artificial Analysis moved to v2.0 on October 22, so the scores cannot be compared directly with the current leaderboard. Better to write "four questions to ask when you see a cost multiple" than to declare an efficiency revolution on the vendor's behalf.

Candidate topics

1. ChatGPT Wallet: the first consumer entry point for agent payments — explain the four conditions first

  • Post: ChatGPT @ChatGPT post, 2026-11-14 00:44 (Shanghai time), 4,847 likes, 793,061 views.
  • Discussion sample: TheRealAdamG @TheRealAdamG post, 2026-11-13 23:43, 935 likes, 113,805 views. This one relays the official blog; it is not an official account.

What it covers: the user transacts directly with the merchant; the agent pays with a payment token and does not share the real card number with the merchant or the agent; the user first checks the item, total price and card used, then approves each purchase with a security check such as Face ID or a fingerprint; eligible disputes keep the card network's and bank's existing rewards and protections. OpenAI also released tools and guidance to help businesses support agent payments. This item corresponds to "Today's Top Picks 1".

Why it's worth doing / possible angles: this item corresponds to "Today's Top Picks 1". Chinese users most easily read "the agent buys things for you" as globally available on every plan, when in fact there are four limits: region, subscription tier, age and card-issuing country. Concrete angles: how Wallet differs from "automatic ordering", why merchants may block it, and whom to turn to in a dispute.

Timing and verification: 24–48 hours. Verified: the official blog (2026-11-13) confirms the partners, the token mechanism, per-purchase security checks, the initial rollout scope and age, the card requirement, and that purchasing cannot yet be fully delegated to an agent, with pre-authorization by limit, category or time only a future possibility. OpenAI separately published merchant guidance. A live purchase was not completed; do not turn unrelated user-scale marketing figures into usage data for Wallet.

2. Grok Bot joins meetings: three platforms, three plan tiers, and the host still has to admit it

  • Post: Grok Bot @bot post, 2026-11-14 03:11, 6,672 likes, 2,582,685 views.
  • Companion: poteto @poteto post, 2026-11-14 03:25, 1,084 likes, 59,735 views, 131 bookmarks.
  • Discussion sample: XFreeze @XFreeze post, 2026-11-14 04:13, 190 likes, 9,584 views. This item falls under the relaxed threshold (190 likes, below the general cutoff; cited because it is a product practice walkthrough); note when citing that it is not the main spread post.

What it covers: the Bot can join Google Meet, Zoom and Microsoft Teams, listens and takes notes, and can answer in the meeting chat after being called by name; after the meeting it posts a summary and action items in the Bot chat, and if Pages is available it puts a longer report on a page and links it in the chat. Meetings is available on Team, Enterprise and Ultra plans; Enterprise admins can turn it off under Bot Capabilities in the team console. To use it, paste the meeting link into the chat with the Bot and ask it to join, approve the confirmation card it raises, and the host admits it once the meeting starts; Disable Interactivity can be set so it only records and does not speak in the meeting. The documentation states that it will not share user credentials in a meeting. This item corresponds to "Today's Top Picks 2".

Why it's worth doing / possible angles: this item corresponds to "Today's Top Picks 2". Chinese teams more often use Feishu, Tencent Meeting and DingTalk, so this is a chance to explain the boundary of "overseas meeting platforms plus enterprise plans" and keep readers from assuming it will join their own weekly standup once installed. Possible angles: how a Bot as a participant differs from a meeting-notes tool, who can approve its entry, and who is responsible for recording and compliance.

Timing and verification: 24–48 hours. Verified: the official documentation confirms the platforms, plan tiers, admin switch, approval and admission flow, Disable Interactivity, and that credentials are not shared; the Changelog already lists the capability in 1.0.1 on 2026-11-11, so November 13 is the public spread post, not the earliest appearance in the docs. Joining a meeting with the Bot was not hands-on tested; the docs give no price or region list, so do not invent them.

3. Reason: a coding harness that sells "cost per task" — sort out the three sets of numbers first

  • Post: Adi Singh @xadisingh post, 2026-11-13 23:34, 2,123 likes, 3,256,626 views.
  • Discussion sample: testingcatalog @testingcatalog post, 2026-11-14 04:25, 241 likes, 16,029 views. Relaxed threshold; may be cited for coding and agent topics with a note.

What it covers: the launch post claims "Frontier coding performance at up to 20x lower cost" and says it beats top harnesses on Terminal-Bench and DeepSWE while dropping cost per task to a fraction. That is the launch post's wording, not a verified independent evaluation. The official 1.1 blog (2026-10-10) uses a completely different scale: on GPT-6 Astra, Reason claims a Coding Agent Index of 65.1 against Codex at 62.5, 22% lower cost per completed task, 15–23% lower cost per attempt and 23% fewer input tokens; it states explicitly that this is its own reproduction following Artificial Analysis methodology, "not an official Artificial Analysis result, and Artificial Analysis has not reviewed or endorsed it". That post evaluated v1.5, and notes that AA moved to v2.0 on October 22, so the scores cannot be compared directly with the current leaderboard. The launch blog post separately says "up to 30% lower cost per task than Codex, Cursor or Devin". This item corresponds to "Today's Top Picks 3".

Why it's worth doing / possible angles: this item corresponds to "Today's Top Picks 3". Chinese circles are already passing around "20x cheaper", but three vendor scales that differ by an order of magnitude make a good demonstration of how to read vendor cost numbers: ask which benchmark version, which comparison system, per task or per attempt, and who reproduced it.

Timing and verification: 24–48 hours. Product facts follow the official site and docs: the launch blog (dated 2026-11-09 at the top, and saying it has been live for about four weeks, so November 13 is launch promotion rather than the first launch), the 1.1 benchmark blog, the usage docs, and the repository. Safe to write: there is a free open-source engine; it works alongside Claude Code, Codex, Cursor and Grok Bot; you can use your existing keys and subscriptions; there is a web app, a CLI and a cloud computer. No benchmark was independently reproduced; the discussion is thin and mediocre, so do not write that "the community has broadly validated it".

4. Obsidian Skills reaches seven: letting agents read and write your vault directly

  • Post: Steph Ango @kepano post, 2026-11-15 02:27, 1,315 likes, 43,607 views, 1,434 bookmarks.

What it covers: the maintainer announces that Obsidian Skills now has seven entries that teach agents to fully operate vaults and wikis: obsidian-markdown, obsidian-bases, json-canvas, obsidian-cli, obsidian-headless, defuddle and knap. obsidian-headless is for running Sync and Publish from the command line without the desktop app. This is a "now there are seven" update, not the first release of the whole project today.

Why it's worth doing / possible angles: Obsidian has a large following in Chinese personal knowledge management circles, and many of those users are not programmers. The value here is moving "the agent helps you tidy your notes" from a chat window to direct file edits. The risk must be stated alongside: these skills let agents read and write the local vault, with a risk of mangling files, so back up first and try them in a test folder. Possible angles: what each of the seven maps to, which are for daily use, and which only matter on a server.

Timing and verification: within 24 hours. Verified against the GitHub repository: kepano/obsidian-skills redirects to the official Obsidian organization, and the README lists the seven above; installation should follow the current README (the marketplace address is obsidianmd/obsidian-skills), so do not copy the old kepano/obsidian-skills from older tutorials as the only official path. The Obsidian Headless documentation confirms the command-line Sync/Publish use. A reliable GitHub star count was not obtained this round, so do not cite one; for licensing, follow the repository page, and if unsure write "see the repository license". Installation was not hands-on tested.

5. CodeRabbit Deep Scan: 1,215 high-severity findings across 1,125 open-source repos, with conditions on the free first scan

What it covers: the vendor says attackers already probe public code around the clock with AI agents, and maintainers should too. The published result: between 2026-10-23 and 2026-11-03 it ran Deep Scans on 1,125 public repositories, producing 1,215 high or critical findings (1,200 high, 15 critical), at least 75 of which are already fixed. The first Deep Scan is free for organizations with two or more people, capped at $1,000.

Why it's worth doing / possible angles: Chinese indie developers and open-source maintainers are interested in free security scanning, but the framing matters most here: both the scanning and the severity grading are CodeRabbit's own, not an independent audit; the 75 fixed is also the vendor's statement. Possible angles: how to use the free allowance, how to handle false positives, and why the exposure of an open-source project differs from that of a company repository.

Timing and verification: 24–48 hours. Verified: the official blog (2026-11-13, Santosh Yadav) gives the date range, repository count, finding count and severity split, and fixed count, and closes with "Your first Deep Scan is on us", pointing to app.coderabbit.ai/security or oss@coderabbit.ai. Eligibility and allowance are governed by in-product terms; They were not confirmed on anyone's behalf. Note that a separate earlier "$1 million scan and fix" campaign is marked as ended in the blog index; many reposts conflate the two, so do not reuse "$1 million" as this new offer. The discussion thread contains many promotional posts with highly similar wording; do not treat them as independent user tests.

6. poteto: asking agents for the "smallest, quickest" fix may teach the next agent to patch

  • Post: poteto @poteto post, 2026-11-15 00:49, 2,673 likes, 110,604 views, 889 bookmarks.

What it covers: his argument is that asking an agent for the smallest or quickest change looks like a harmless band-aid, but the next agent copies that pattern, and eventually it spreads across the codebase. This is experience sharing, not an evaluation, and not a feature launch.

Why it's worth doing / possible angles: Chinese users naturally tell an agent "just change this one line, don't touch anything else", because they fear it will run wild. This offers the reverse cost: short-term convenience, long-term code that is a pile of patches. Possible angles: when to demand the smallest change (production outages, someone else's code), when to demand "refactor first, then change", and how to state the boundary in a prompt.

Timing and verification: within 24 hours. Spread metrics recorded; the view is the author's practical experience with no experiment or data behind it, so do not present "it makes code worse" as a proven conclusion. It can be folded into coding-practice content rather than run as a standalone news item.

7. Budget3D: one model, several token budgets — 40% of the tokens still yields 3D

  • Post: AK @_akhaliq post, 2026-11-14 08:01, 436 likes, 30,965 views, 500 bookmarks.
  • Author's note: ozgurkara99 @ozgurkara99 post, 2026-11-14 06:41, 52 likes, 7,732 views. Below the general threshold; use only as the author's explanation, not as a representative post.

What it covers: the project extends compute-adaptive tokenization from image diffusion to image-to-3D: one checkpoint, with different token budgets at inference — a low budget computes less and returns faster, a high budget adds detail. The project page has comparison videos, and the weights are on Hugging Face.

Why it's worth doing / possible angles: Chinese 3D creators care most about VRAM and generation time. This can be written as a quality-versus-speed dial, but "any budget" must not be confused with "any hardware": the official examples use one H100, batch 8 and optimized inference, and the numbers change on consumer GPUs or at larger batch sizes. It is also worth separating the weight license from the runtime dependencies.

Timing and verification: 24–48 hours. Verified against the Hugging Face model card: this is an unofficial community implementation extending BudgetPix's method to image-to-3D (built on Sperid Labs' Idol / Idol-Turbo), not the official release of the BudgetPix paper, which itself covers only pixel-space image diffusion; the weights are labeled Apache-2.0, but running it also requires DINOv3 and the FLUX.2 VAE, which carry their own licenses. Speed figures are the project's own measurements: Budget3D-Idol on a single H100, batch 8, optimized inference, takes 5.67 s per object at 100% tokens (8192) with 9.9 GB peak memory, and 2.66 s at 40% tokens (3277) with 7.8 GB; Turbo on a single RTX PRO 6000 takes 0.48 s per object at 100% and 0.25 s at 40%. Decoding Gaussians and rendering take additional time, about 3 extra seconds per object regardless of budget according to the project page. Related links: project page, Turbo weights, code. Not tested.

8. SUS-Bench: 250,000 people playing a Mafia-style game with AI — the results look good but are not yet a paper

What it covers: the project turns social deduction into an online game: 10-player rounds, find the 2 Opps, and players do not know who is human and who is AI. The project page claims more than 250,000 people, 167 countries and more than 500,000 rounds, played voluntarily and without payment; the author says these are the first results and more detail will follow. The author's post uses labels such as "most likely to lie", "most loyal" and "biggest snitch" — that is spread rhetoric, not a general personality or safety conclusion about the models.

Why it's worth doing / possible angles: Mafia-style and scripted murder games are a ready-made frame on the Chinese internet, and readers are drawn to "can AI lie". The right approach is "the test design is interesting, but the conclusions must wait for a paper": explain the sampling bias of voluntary participation, the human-to-AI ratio in 10-player rounds, and what "snitching" actually maps to in the rules. Do not cite specific scores that the project page has not confirmed.

Timing and verification: 24–48 hours. Verified against the project page for the self-reported figures and gameplay description; no formal paper or independent reproduction has been seen, so scores and methodological details cannot be treated as peer-reviewed scientific conclusions. "A fruit-fly brain beat 17 of 28 models" comes from the author's follow-up remarks and relayed posts, and the project page text captured this round did not give the full leaderboard; if you write it, label it as the author's or a relayed claim, and mark the full leaderboard as to be verified.

9. Value observation: SatBench says 63% of 173 benchmarks are already saturated

  • Post: Ethan Mollick @emollick post, 2026-11-15 07:46, 199 likes, 13,834 views. Only reaches the relaxed threshold.

What it covers: the author says he "threw together" SatBench to measure how quickly the major benchmarks cited by the labs since 2023 become saturated: 63% of 173 benchmarks have been effectively beaten, and saturation is accelerating each year. The core argument is that "AI is getting better faster than our measurements are".

Why it's worth doing / possible angles: Chinese readers are used to judging models by whether "the leaderboard has been smashed again", and this offers a more fundamental angle: the leaderboards themselves are expiring. Good as a short methods piece reminding readers to ask which year's benchmark a new model topped and how much discriminating power it still has.

Timing and verification: within 24 hours, value observation. It is the author's self-described personal project; no independently verified paper or reproducible repository was seen this round, so mark it as to be verified; when citing 63% and 173, attribute them to the author.

10. Tripo Agentic Canvas: one sentence lets an agent plan the whole 3D workflow

  • Post: Tripo @tripoai post, 2026-11-13 23:00, 112 likes, 144,624 views, 50 bookmarks, 26 quotes.

What it covers: it positions itself as letting an agent plan, connect and run a 3D workflow: create by chat, build reusable pipelines, batch-process assets, without switching among tools or arranging steps by hand. The vendor says it produces "production-ready" assets.

Why it's worth doing / possible angles: the pain of 3D workflows is too many nodes and too many steps, which Chinese e-commerce modeling, game art and short-video props all share. Possible angles: which steps an agent can genuinely plan and which still need a human to set the style; and what "reusable pipelines" mean for batch production. The discussion thread has many promotional posts with similar wording (including tutorials with referral links); do not present them as extensive independent testing.

Timing and verification: 24–48 hours. Spread metrics and the official positioning are recorded; pricing, commercial terms and the free allowance were not verified this round, so mark them as to be verified. Not tested.

11. Follow-up: Interior AI imports the full IKEA catalogue and turns user furniture photos into 3D

What it covers: only the increment relative to the English briefs of November 12 and 13. What was covered before was floor-plan and video-to-3D; the additions here are two: importing the IKEA 3D catalogue into Interior AI, and letting users upload a photo of their own furniture and turn it into a placeable 3D object in under 30 seconds with an image-to-3D model. The author says more furniture brands will follow.

Why it's worth doing / possible angles: renovation, interior styling and rental makeovers are frequent scenarios for Chinese readers, and whether the furniture "looks right" matters more than render quality. Possible angles: when a ready-made brand catalogue helps versus user-created furniture, and the accuracy limits of photo-to-3D. Note that the IKEA catalogue differs by region; the author's "entire catalogue" refers to the data he could obtain and does not mean every market is complete.

Timing and verification: within 24 hours, marked as a follow-up. Both posts are the author's own product progress updates, i.e. a personal demo and vendor claims; Interior AI's subscription tiers and pricing were not verified this round. Not tested.

12. Follow-up: Riley Brown makes a 2-minute 4K film from one Fable 5.2 prompt

  • Post: Riley Brown @rileybrown post, 2026-11-15 00:33, 1,020 likes, 66,909 views, 947 bookmarks.

What it covers: the author says the 2-minute 4K film was made entirely in Blender from one prompt on Fable 5.2 and took about 4.5 hours; Claude built every asset in Blender by writing Python. The post itemizes characters, props and scenes, but the text was truncated in the API, so the full list was not visible.

Why it's worth doing / possible angles: earlier briefs already covered the general "Fable 5.2 makes games and videos" topic, so this can only be a follow-up, and should carry only the new information: 2 minutes, 4K, about 4.5 hours, assets generated by Python inside Blender. The angle is that "one prompt" still sits behind human decisions on storyboard, materials and editing; do not read 4.5 hours as "4.5 hours from nothing to finished film".

Timing and verification: within 24 hours, marked as a follow-up. It is a personal demo, not an independent validation of Fable's capabilities; judge the finished film from the video in the original post, which was not re-tested here. The post body was truncated in the API; use the visible text and the attached video for the asset list.

This issue has 12 items and does not reach 20. The full SUS-Bench leaderboard and paper have not been seen; the CodeRabbit and Tripo discussion threads contain many promotional posts with similar wording, so use the official posts for engagement figures. Some long posts were truncated in the API; key facts were cross-checked against official blogs and documentation.