Skip to content

← English Topics

X English Hot Topics Briefing

Date13 October 2026 · TuesdayWindow2026-10-11T19:02:00+08:00 to 2026-10-13T19:02:00+08:00CoveragePartial

Translation

English translation of the Chinese original. This briefing picks English-language X discussions worth explaining to Chinese readers; "Chinese briefing" refers to the same day's Chinese topic briefing.

Look at these first today

  1. The harness decides the benchmark score (see item 1): almost no discussion in the Chinese circle this round, the largest information gap. Best angle: "fix your toolchain before choosing a model," using 8090 Labs' numbers as the hook. Format: standalone original post. Status: pending verification (need to find 8090 Labs' full report).

  2. Muse Wonder day-one testing (see item 2): release verified; English-circle experiences clearly diverge. Best angle: "how to judge in a new model's first week whether it is worth switching," noting the Full version is subscription-only. Format: quote Riley Brown or standalone original. Status: ready to draft; better after trying it yourself.

  3. Cursor joins X Subscription (see item 3): official information verified; complements item 1 of the Chinese briefing; the increment is that third-party models switch to API-price billing. Format: quote Cursor's official post. Status: ready to draft.

Candidate topics

1. Benchmarks untrustworthy? Same model, different "harness," scores differ twofold

  • Original posts and engagement:
  • @chamath Chamath Palihapitiya, original post, 10-13 01:52 Shanghai (within 24 hours), 937 likes, 85 reposts, 178 replies, 163K views
  • @ArtificialAnlys Artificial Analysis, original post, 10-13 01:51 Shanghai (within 24 hours), 915 likes, 43 reposts, 58 replies, 98K views
  • @emollick Ethan Mollick, original post, 10-13 07:10 Shanghai (within 24 hours), 396 likes, 15 reposts, 55 replies, 21K views
  • Content and focus: Chamath had 8090 Labs run 113 real coding tasks across 7 models × 5 harnesses, and says the same model's score can swing between 24% and 59%; e.g. Claude Fable 5.1 scored 41% in Claude Code and 61% in Codex. At the same time Artificial Analysis launched Cognition's FrontierCode 1.1 leaderboard, and Ethan Mollick also cautioned against taking composite indices too seriously. Quote discussion mostly agrees, e.g. @patrick_saner argues for evaluating "model + harness."
  • Why it's worth doing / angle: Chinese readers often choose models by leaderboard. Write "when choosing an AI coding tool, fix your toolchain first, then compare models," explaining what a harness is and why the official tool is not necessarily the best. Clear information gap; no high-engagement Chinese post on this topic seen this round.
  • Timeliness and verification: within 24 hours. Figures are claims from Chamath's post and 8090 Labs; no link to their full report found; pending verification. FrontierCode details come from the Artificial Analysis post.

2. Muse Wonder day-one testing: officially frontier, users say "went back to Codex"

  • Original posts and engagement:
  • @AIatMeta AI at Meta, original post, 10-13 01:02 Shanghai (within 24 hours), 1992 likes, 203 reposts, 180 replies, 195K views
  • @rileybrown Riley Brown, original post, 10-13 10:25 Shanghai (within 24 hours), 1158 likes, 13 reposts, 126 replies, 73K views
  • @emollick Ethan Mollick, original post, 10-13 02:34 Shanghai (within 24 hours), 276 likes, 13 reposts, 30 replies, 46K views
  • @RoundtableSpace 0xMarioNawfal, original post, 10-13 07:45 Shanghai (within 24 hours), 40 likes, 2 reposts, 22 replies, 50K views
  • Content and focus: Meta released Muse Wonder and Muse Wonder Full; after 5 hours of use Riley Brown said he went back to Codex and Claude Code, then a few hours later changed his tune to "maybe better than Claude Sonnet 3.7" (clearly a joke); Ethan Mollick, after limited testing, thinks it is a good model. Another post says Meta published docs for calling Muse models from Claude Code, Codex and OpenCode.
  • Why it's worth doing / angle: Continues item 15 of the Chinese briefing (Baoyu's rundown of availability). The increment is the English circle's divergent day-one real experiences and the trade-off that "the Full version is subscription-only"; write "how to judge in a new model's first week whether it is worth switching."
  • Timeliness and verification: within 24 hours, continuation. Verified: the Meta developer blog confirms Muse Wonder is offered via the Meta Model API ($4 / $20 per million tokens), and Muse Wonder Full is limited to Muse Code Power/Maximum subscriptions; the Meta research blog confirms a Muse experience for general users is "coming soon." Availability in mainland China not verified.

3. Cursor joins X Subscription: third-party models switch to API-price billing

  • Original posts and engagement:
  • @cursor_ai Cursor, original post, 10-13 07:10 Shanghai (within 24 hours), 5100 likes, 366 reposts, 450 replies, 401K views
  • @SawyerMerritt Sawyer Merritt, original post, 10-13 07:19 Shanghai (within 24 hours), 975 likes, 53 reposts, 130 replies, 88K views
  • @ckor Andrew Kordampalos, original post, 10-13 08:54 Shanghai (within 24 hours), 107 likes, 4 reposts, 38 replies, 21K views
  • Content and focus: Cursor officially announced it is joining X Subscription: existing users can keep their current plans; paid plans share quota across products such as Cursor and Grok Build; Grok model coding quota is twice that of the equivalent Cursor plan, but in the new plans third-party models are billed at API prices. Users such as @ckor focused on "Claude and Gemini no longer use included quota."
  • Why it's worth doing / angle: Continues item 1 of the Chinese briefing. The increment is Cursor's own words, "twice the coding quota for Grok models" and "third-party models moved to API billing," good for helping Chinese developers who use Claude heavily do the math.
  • Timeliness and verification: within 24 hours, continuation. Verified: the official announcement and pricing docs confirm price tiers, optional switching, and that third-party models do not count toward included quota.

4. Karpathy's original "AI awareness funnel" post: 5.83M views

  • Original posts and engagement:
  • @karpathy Andrej Karpathy, original post, 10-12 03:12 Shanghai (24—48 hours), 41331 likes, 5457 reposts, 1892 replies, 5.836M views
  • @rileybrown Riley Brown, original post, 10-12 07:18 Shanghai (24—48 hours), 2003 likes, 153 reposts, 63 replies, 154K views
  • Content and focus: Karpathy revisited his April post, saying the gap in awareness of large-model capability is widening into a funnel: about 6 billion people have barely touched it, 1—2 billion occasionally use free versions, and a minority of professionals use agents. Riley Brown used Opus 5.5 with a single prompt to generate a visual explainer video of the post.
  • Why it's worth doing / angle: Continues item 6 of the Chinese briefing (Baoyu's translation). The increment is the original post's scale and the trick of "using AI to turn a long post into a video explainer"; write "which layer of the funnel you're on, and one exercise to move down."
  • Timeliness and verification: 24—48 hours, continuation. The funnel figures are Karpathy's self-described rough estimates (napkin math), not statistics.

5. "Stop calling it vibe coding": has model code quality surpassed the average engineer?

  • Original posts and engagement:
  • @signulll signüll, original post, 10-12 00:17 Shanghai (24—48 hours), 5875 likes, 391 reposts, 462 replies, 398K views
  • Content and focus: @signulll argues the architecture and code coherence models produce are already better than the average engineer's and it should be renamed "precision engineering"; nearly 6K likes. Quotes include agreement and sarcasm, e.g. @liftplaycode joking that reading a vibe-coded codebase gave them a better understanding of the second law of thermodynamics.
  • Why it's worth doing / angle: Lay out both sides for Chinese readers: where AI code is already reliable and where it still needs review; contrast with the "code review" discussion in the Chinese briefing.
  • Timeliness and verification: 24—48 hours. Opinion post with no supporting data; discussion sample is small; cannot be treated as consensus.

6. "Software is dead"? Riley Brown: I'm downloading apps like crazy, and built an internal super-app for my company

  • Original posts and engagement:
  • @rileybrown Riley Brown, original post, 10-12 13:16 Shanghai (24—48 hours), 2286 likes, 85 reposts, 359 replies, 141K views
  • @rileybrown Riley Brown, original post, 10-12 22:51 Shanghai (within 24 hours), 347 likes, 12 reposts, 48 replies, 27K views
  • Content and focus: Riley Brown rebuts the claim that "AI will kill apps" and shares his experience building internal apps for his company: you can add any little toggle that only suits your own team.
  • Why it's worth doing / angle: Echoes the view that "development is not a moat" in the REA discussion in item 3 of the Chinese briefing; write about the new opportunities and limits of "individuals/small teams building their own software," suitable for non-programmers.
  • Timeliness and verification: 24—48 hours. Personal experience; internal app details not public.

7. Cost-saving routing: cheap models for simple judgments, Opus only for hard problems

  • Original posts and engagement:
  • @thegreatest_sv kiosa, original post, 10-12 02:06 Shanghai (24—48 hours), 956 likes, 132 reposts, 71 replies, 67K views
  • @Arindam_1729 Arindam Majumder 𝕏, original post, 10-12 22:23 Shanghai (within 24 hours), 107 likes, 18 reposts, 15 replies, 11K views
  • Content and focus: A widely spread post describes using Jev as a "fast brain" for classifying, scoring and checking, with Claude stepping in only for hard problems; the author claims about 30% lower cost on the same task. Another post shares a division of labor in Claude Code: Opus 5.5 plans, Sonnet 5.5 refactors, Haiku 5.5 filters.
  • Why it's worth doing / angle: Chinese users commonly complain about insufficient quota; write practical ideas for "tiered model use," stressing trying it on a small project first.
  • Timeliness and verification: 24—48 hours. The first post mentions a "Stanford team," for which no basis was found; a source claim, pending verification. "About 30% lower cost" is the plugin author's self-test.

8. AI is rewriting mathematical research: OpenAI's 722 manuscripts and "centaur" mathematicians

  • Original posts and engagement:
  • @emollick Ethan Mollick, original post, 10-12 00:56 Shanghai (24—48 hours), 1123 likes, 97 reposts, 133 replies, 74K views
  • @Rainmaker1973 Massimo, original post, 10-12 00:47 Shanghai (24—48 hours), 121 likes, 27 reposts, 8 replies, 26K views
  • @emollick Ethan Mollick, original post, 10-12 12:14 Shanghai (24—48 hours), 736 likes, 20 reposts, 127 replies, 47K views
  • Content and focus: Ethan Mollick says "centaurs" are coming to dominate mathematical research; another post describes OpenAI releasing 722 AI-generated mathematical manuscripts; Mollick also notes GPT-6 Pro finds errors in papers in everyday work, but academia lacks mechanisms to handle large volumes of corrections.
  • Why it's worth doing / angle: Good for students and research readers: what AI can do, where the verification cost lies, and what it means for academic publishing.
  • Timeliness and verification: 24—48 hours (about 42 hours). The 722 manuscripts are a claim from news and posts; OpenAI's original release was not opened this round; pending verification.

9. OpenAI retires GPTs: why big companies get whiplash from frequent AI product changes

  • Original posts and engagement:
  • @emollick Ethan Mollick, original post, 10-12 10:32 Shanghai (24—48 hours), 705 likes, 20 reposts, 149 replies, 64K views
  • Content and focus: Ethan Mollick says some leaders of large companies he has talked to, who relied heavily on GPTs as recently as last month, are unhappy about their retirement; he thinks GPTs are outdated, but labs' sudden changes make organizational planning hard.
  • Why it's worth doing / angle: Chinese companies and individuals also build processes around particular features. Write advice on "don't lock your workflow to a single product feature."
  • Timeliness and verification: 24—48 hours. "GPTs have been retired" is the post's claim; timing and scope of the retirement need checking against OpenAI's official statement.

10. AI makes targeted scams cheap: turn on two-factor authentication and agree on a code word with family now

  • Original posts and engagement:
  • @emollick Ethan Mollick, original post, 10-13 00:31 Shanghai (within 24 hours), 849 likes, 68 reposts, 88 replies, 60K views
  • Content and focus: Mollick warns that AI has lowered the cost of sophisticated attacks against individuals, and recommends enabling two-factor authentication on all accounts and agreeing on a code word with family to guard against voice scams asking for money.
  • Why it's worth doing / angle: Very practical for Chinese readers; combine with common domestic AI face-swap and voice-cloning scams into a family anti-fraud checklist.
  • Timeliness and verification: within 24 hours. The advice is general security common sense; do not cite unverified specific case data.

11. 19 phone-runnable models now beat GPT-4

  • Original posts and engagement:
  • @emollick Ethan Mollick, original post, 10-13 11:38 Shanghai (within 24 hours), 1241 likes, 93 reposts, 120 replies, 33K views
  • Content and focus: Mollick recalls that GPT-4 once led for a whole year, and now 19 models that can run on a modern phone are stronger than GPT-4 (with a chart).
  • Why it's worth doing / angle: Value observation: write "what local/on-device models mean for ordinary people," tied to privacy and offline use.
  • Timeliness and verification: within 24 hours. "19" comes from his attached chart; source and evaluation basis not verified.

12. Agents calling agents: Claude Code calls Grok Bot, Codex computer use

  • Original posts and engagement:
  • @rileybrown Riley Brown, original post, 10-12 02:37 Shanghai (24—48 hours), 1341 likes, 64 reposts, 105 replies, 130K views
  • @rileybrown Riley Brown, original post, 10-12 12:54 Shanghai (24—48 hours), 656 likes, 11 reposts, 95 replies, 45K views
  • @angeldot_ angel, original post, 10-13 03:59 Shanghai (within 24 hours), 471 likes, 13 reposts, 14 replies, 50K views
  • Content and focus: Riley Brown shares using a skill to have Claude Code hand X research to Grok Bot and wait for the result, and says he can't live without Codex's computer use; another post shows Bolter letting Grok, Claude and Codex collaborate in one group chat to fix a production issue (product promotion).
  • Why it's worth doing / angle: Echoes item 7 of the Chinese briefing, "multiple agents don't talk to each other"; write practical ways to "let different agents do what they're good at" and the permission risks.
  • Timeliness and verification: 24—48 hours / within 24 hours. The third post is a product promo video, not independently verified; skill configuration details need hands-on testing.

13. One-click AI influencer vlog: a ChatGPT + Seedance 2.5 multi-scene workflow

  • Original posts and engagement:
  • @geezmedia01 Magdy’s Production, original post, 10-12 02:28 Shanghai (24—48 hours), 361 likes, 77 reposts, 11 replies, 37K views
  • Content and focus: The post shows an AI vlog workflow using ChatGPT to write storyboards and Seedance 2.5 to generate multiple scenes, outfit changes and dialogue.
  • Why it's worth doing / angle: Value observation: can be written as a creator tutorial, but must add AI-generated content labeling and platform-rule reminders.
  • Timeliness and verification: 24—48 hours (about 41 hours). Results are the author's demo; AI influencer content risks impersonating real people; compliance reminders needed.

14. Simon Willison: practical notes on Python 3.15 + uv

  • Original posts and engagement:
  • @simonw Simon Willison, original post, 10-13 02:22 Shanghai (within 24 hours), 411 likes, 26 reposts, 68 replies, 39K views
  • Content and focus: Simon Willison shares notes on using Python 3.15 with uv, and on 3.11 features that can be adopted now that 3.10 is no longer the minimum supported version.
  • Why it's worth doing / angle: Value observation: for Chinese readers who write scripts, translate the key points and add usage conditions such as domestic mirrors.
  • Timeliness and verification: within 24 hours. For specific features, refer to the original blog post.

This round, searches for @sama, @ylecun and @rasbt returned zero results in the window and were not individually rechecked; some English posts' engagement was collected in a single snapshot this round. 14 qualifying candidates this round, short of 20.