X Chinese Hot Topics Brief¶
Translation
English translation of the Chinese original.
Today's top picks¶
1. Open-source projects bypass the Tensor throttle on consumer GPUs (matches 01)¶
Why it's worth doing: in two days the same story spread from the 3090 to the 4090, with two repos, multiple user re-tests and a batch of foreign media reports appearing at once. Most readers have a 30-series or 40-series card and can verify the gains themselves. This is not overclocking but "rewriting the operation issue rate before the driver binds", which leaves far more room for explanation than yet another "another card got more expensive" story.
Additional angles: explain clearly that "training and prefill get faster while token decoding and game frame rates stay roughly unchanged", then use the 225W / 360W / 450W power-limit tiers to explain why everyone reports different percentages. Add a risk and rollback checklist (a cold boot with power removed is mandatory the first time, sleep does not cover it, only the listed PCI IDs are recognized, and if the firmware has not been stopped you must cut power rather than unload the module), and finish with an unemotional judgment on "should you buy a 3090 now".
Format suggestion: if you can, re-test on your own card and write it as a standalone original post with data. Without a card, write a documentation-reading explainer, citing every number from the repo issues or third-party media and never presenting someone else's figures as your own measurements.
Status: awaiting testing (re-test numbers are not out yet; the explainer can be drafted now).
2. Upgrading old devices with AI has gone from one-off to a wave (matches 02)¶
Why it's worth doing: within a single day four threads appeared — Kindle jailbreaking, de-advertising Android, de-advertising TVs, and using old phones as small servers. What they share is replacing "knowing how to flash ROMs" with "knowing how to describe what you want", lowering the bar to something ordinary users will try. The 10-22 topic covered the Xiaomi TV launcher; this is not a repeat but the same approach spreading across four hardware categories: e-readers, phones, TVs and servers.
Additional angles: write a "what can old devices still buy you" ledger — the obtainable benefit, minimum cost and most likely point of failure for each hardware category. Stress that most of these projects disable rather than uninstall and can be restored in one click, removing the "what if I break it" worry up front. Then point out the genuinely new risk: "anyone who can ssh into the phone can also drive the web sessions already logged in on it", so use a dedicated old device.
Format suggestion: an original list post, giving one copyable way to describe the requirement plus one pitfall for each of the four threads, without reposting whole tutorials.
Status: ready to draft; the old-phone-tailnet repo could not be verified, so confirm the address before writing.
3. WorkBuddy starts putting a price on Skills (matches 03)¶
Why it's worth doing: this is the first time "can prompt assets be sold" has been put on a shelf by a platform, and the same day the opposing side pointed out that a Skill is essentially plain text, copies instantly when passed on, and gets bundled for one yuan on Xianyu. Both support and opposition come from real users, so the discussion does not need to rest on positions.
Additional angles: break down "what is actually being sold" — the file itself has almost no moat; what can charge money is maintenance, industry data and the connectors behind it. Treat the secondhand-source claim of a 1% platform service fee (no official text seen) as a starting point for discussion and ask whether the low cut is a subsidy window or a long-term strategy. Then distinguish "can be sold" from "can be sold sustainably".
Format suggestion: a standalone original opinion post quoting one point from each side, landing on "what kind of Skill is worth paying for".
Status: awaiting verification (the official fee schedule and open scope of the paid feature have not been seen in original form).
Candidate topics¶
01 Tensor throttling on the RTX 3090/4090 gets bypassed¶
Original posts and metrics: @lxfater (10-24 15:13, 317K views, 3.9K likes, 3.3K bookmarks, 238 replies, 350 reposts); @qwerty__cs (10-24 22:26, 11K views, 72 likes, 57 bookmarks, 9 replies, 5 reposts); @lxfater's re-test (10-24 23:37, 15K views, 113 likes, 99 bookmarks, 16 replies, 17 reposts).
Content and angles of interest: two open-source projects target the 30-series and 40-series respectively. The method is writing to the speed-select register before the NVIDIA driver binds, pushing the issue rate for FP16/BF16 inputs with FP32 accumulation to the maximum, without raising power, modifying the VBIOS or patching the driver. Discussion centers on three things: whether to stock up on 3090s now, whether prefill and image generation benefit, and whether secondhand prices will rise; quoted posts also note that game frame rates and token decoding are essentially unchanged.
Why it's worth doing / possible angles: gains readers can verify themselves are the most worth writing about, and the angle is to define the boundary of those gains precisely — matrix math, prefill and diffusion models get faster, while decode is limited by memory bandwidth, and INT8 and games see no difference. Then explain why the same card yields +25% for one person and +80% for another (different power limits and test shapes).
Timeliness and verification: within 24 hours. Verified: the repos are Zagitta/gpu-unlock (30-series) and JustAResearcher/gpu-unlock-ada (4090, 4070 Ti SUPER, 4060 Ti 16GB); the 4090's FP16 165.6→331.6 and BF16 165.5→330.6 TF/s are in the repo README, and the 3090 Ti's BF16 87.5→157.4 at 450W is in issue #1; it must be loaded on Linux as a kernel module plus initramfs, and the risk terms are in the README. Awaiting testing: numbers on your own card; "75→135 at 360W" currently appears only in user reports and foreign media retellings.
02 Upgrading old devices with AI: Kindle, Android, TVs, old phones¶
Original posts and metrics: @vista8 (10-24 17:01, 119K views, 532 likes, 601 bookmarks, 102 replies, 56 reposts); @vista8's follow-up (10-24 22:13, 21K views, 252 likes, 333 bookmarks, 49 replies, 24 reposts); @HiTw93 (10-24 20:46, 31K views, 376 likes, 539 bookmarks, 47 replies, 39 reposts); @geekbb (10-24 10:26, 30K views, 244 likes, 339 bookmarks, 37 replies, 37 reposts); @YuLin807 (10-24 22:44, 5.8K views, 54 likes, 81 bookmarks, 13 replies, 6 reposts); @yiqifacai (10-24 17:56, 11K views, 45 likes, 25 bookmarks, 13 replies, 2 reposts).
Content and angles of interest: a jailbroken Kindle running KOReader and a WeChat Reading plugin, an Android tablet stripped of ads and preinstalled apps, a TV trimmed by letting Claude Code drive ADB with a single prompt, and an old phone joined to a Tailscale private network as a server with a built-in battery. What they share is "state the requirement clearly and let the agent type the commands"; the disagreement is over risk — handing a device to an AI to operate, account security on secondhand phones, and having to re-enable wireless debugging manually after a reboot.
Why it's worth doing / possible angles: the 10-22 topic covered the Xiaomi TV launcher; this is a broader wave of old-device upgrades, and the four hardware categories differ completely in payoff and pitfalls, which makes a perfect comparison ledger. The angle is why designs like "disable instead of uninstall, one-click restore" make ordinary people willing to try, and what boundaries really matter when an old phone becomes a server (a dedicated machine, no banking apps, ACLs that only allow SSH).
Timeliness and verification: within 24 hours. Verified: the TV project is cleanyour.tv / Gathin23/cleanyourtv, which disables ad and telemetry apps over Wi-Fi ADB, supports backup and restore, and needs no root; the project site currently lists 280 TVs across 81 brands in 48 hours, larger than the 16 brands and 30 TVs in the post, indicating a fast-growing list. Not verified: the old-phone-tailnet repo, which returns no search results; the address in the post appears only in the replies, so confirm it before writing.
03 WorkBuddy opens sales of Skills and expert services¶
Original posts and metrics: @abskoop (10-24 20:30, 16K views, 40 likes, 29 bookmarks, 9 replies, 6 reposts); @Mileson07 (10-24 00:29, 67K views, 610 likes, 956 bookmarks, 16 replies, 107 reposts); @realchendahuang (10-25 00:40, 5.7K views, 23 likes, 28 bookmarks, 6 replies, 6 reposts); @op7418 (10-24 21:44, 15K views, 29 likes, 10 bookmarks, 19 replies, 0 reposts).
Content and angles of interest: the closed beta opens "monetization", letting developers price and sell the skills and expert services they write. At the same time a clear opposing view appeared: most Skills are a piece of Markdown with no anti-copy protection, they can be forwarded as-is, they get bundled for one yuan on Xianyu, and selling files is hard to make work. The core of the discussion is not the feature but "can a prompt count as an asset".
Why it's worth doing / possible angles: a platform starting to price these assets is a ready-made entry point for observing the business model of the agent ecosystem. The angle: separate "selling files" from "selling maintenance, data and connectors", discuss what the platform's low cut is subsidizing, and give readers a criterion — what kind of skill is worth paying for and what is faster to write yourself.
Timeliness and verification: within 24 hours (the tutorial post is 24—48 hours). Verified: Tencent's official WorkBuddy open platform explicitly opens five capability categories — Buddy apps, experts, Skills, connectors and hardware — and provides onboarding docs; secondhand sources say the beta has opened monetization, supports one-time purchase and subscription, with a 1% platform service fee (roundup, blogger's note). Not verified: Tencent's own text on fees, review and settlement rules; "upload a zip and it publishes" has not been confirmed point by point officially.
04 Gemini 4 Argon "dumbing down" and Pichai's compute remarks¶
Original posts and metrics: @DPIT_HQ (10-24 11:19, 53K views, 453 likes, 212 bookmarks, 29 replies, 31 reposts); @ohayou792 (10-24 07:00, 45K views, 169 likes, 39 bookmarks, 22 replies, 8 reposts).
Content and angles of interest: one side says its own tests found free Pro accounts routed to a weaker gemini-4-argon-trail, with Antigravity and the web app hit hardest; the other uses the 25 trillion tokens, 72 hours and compute overload from Pichai's interview as the explanation for the degradation, and mentions skipping 3.5 Pro. Public discussion has already split into "it really got worse" and "it is just peak-hour rate limiting".
Why it's worth doing / possible angles: it is rare for official numbers and user perception to collide, so a calm reconciliation is worth doing. What is confirmed is that the token surge and compute squeeze are real; what is unconfirmed is that "the system is secretly swapping models". The angle: give readers ways to check for themselves (compare the same prompt across accounts, look at the model identifier returned by the API) and note that the availability gap between Pro and Ultra inside Antigravity is itself a public controversy.
Timeliness and verification: the main post within 24 hours, the interview analysis within 24—48 hours. Verified: the 25 trillion tokens / 72 hours / roughly three times the same period in Gemini 3.0 come from Pichai's Big Technology interview (interview transcript, sources.news, The Paper's retelling); there is abundant user feedback that the Pro tier cannot get Argon in Antigravity (developer forum). Not verified: the model name gemini-4-argon-trail, which has no source in public channels; "degradation is caused by overload" is the blogger's inference, not an official statement.
05 The cost math of AI short dramas and the box office of the first Dragon Mark AI feature¶
Original posts and metrics: @jaker_Q (10-24 13:15, 50K views, 99 likes, 101 bookmarks, 24 replies, 8 reposts); @GitHub_Daily (10-24 12:00, 13K views, 227 likes, 282 bookmarks, 13 replies, 43 reposts); @abskoop (10-24 14:44, 22K views, 30 likes, 7 bookmarks, 18 replies, 1 repost); @oran_ge (10-24 16:23, 19K views, 27 likes, 9 bookmarks, 0 replies, 0 reposts).
Content and angles of interest: on one side, a short-drama company moves fully to AI, spending 5—6K yuan per half minute of sample footage and running the pipeline with Codex, leaving only management, technical and sales roles; on the other, the first AI-produced feature film to receive a Dragon Mark took 318K yuan on its first day against a cost in the tens of millions. In between sits a curated list of 83 short-drama Skills, showing that the tooling layer is already crowded.
Why it's worth doing / possible angles: the three data points map exactly onto both ends of "AI content industrialization" — the cost of vertical short dramas can already be calculated, while feature films have not yet found the formula. The angle: compare per-minute cost against revenue-share income to show that the divide between short drama and film is not generative technology but distribution and ticket-price structure; then add why the Skill list matters — it turns "can write a script" into a comparable capability.
Timeliness and verification: within 24 hours. Verified: the film opened on 10-23, took about 318K yuan on day one against a cost in the tens of millions, and the producers state that all visuals and part of the audio were AI-generated (21st Century Business Herald, Sina News, Tencent News); the short-drama list is awesome-short-drama-skills, with 83 Skills across 16 platforms. Awaiting verification: "5—6K per half minute" is the poster's own account only.
06 A close-reading site for podcast transcripts¶
Original posts and metrics: @oran_ge (10-24 22:10, 18K views, 293 likes, 417 bookmarks, 17 replies, 57 reposts); @vista8 (10-24 19:36, 7.9K views, 56 likes, 60 bookmarks, 32 replies, 6 reposts).
Content and angles of interest: eight Chinese tech podcasts are turned into an auto-updating close-reading site; the author stresses that these are fine-grained transcripts rather than summaries, so ten minutes yields the main information of an hour-long episode, and anyone interested can go back to the original audio; the plugin already has 3,000 users. Bookmarks clearly outnumber likes, suggesting it is saved as a tool rather than shared as an opinion.
Why it's worth doing / possible angles: it solves the concrete problem of "want to listen but genuinely have no time", and the difference in information loss between transcription and summarization deserves a clear explanation. The angle: contrast the situations where summaries lose key judgments, discuss the copyright and quotation boundaries of transcribing public podcasts, and estimate the cost of building such a site yourself; also explain why sites like this work better as a permanent tool than as one-off content.
Timeliness and verification: within 24 hours. Awaiting verification: whether the eight sites are available and update reliably, and the "3,000 users" claim, which is the author's own statement; neither has been independently confirmed.
07 How Computer Use works and agents operating phones directly (follow-up)¶
Original posts and metrics: @dotey (10-24 02:11, 27K views, 182 likes, 174 bookmarks, 89 replies, 17 reposts); @turingou (10-23 21:06, 55K views, 152 likes, 122 bookmarks, 13 replies, 12 reposts).
Content and angles of interest: one post explains "how AI clicks buttons without hands" — the model only sees and thinks, while another program does the actual work, looping over three paths: screenshots, accessibility features and software APIs; it also explains why it can work in the background on a Mac but takes over the desktop on Windows. The other lets DeepSeek inside Codex operate a phone directly, moving a year of trips from Umetrip into Flighty in 20 minutes. The reply counts are high, showing that the doubts center on "can you trust it not to click randomly".
Why it's worth doing / possible angles: pulling "AI can use a computer" from demo back to principle explains the experience gap that confuses ordinary users most. The angle: the trade-offs among the three paths (screenshots are universal but slow, APIs are fast but need integration, accessibility sits in between), why phones are actually easier to automate (fewer apps, well-specified controls), and how to draw permission boundaries around it. The turingou post continues a 10-22 topic of the same kind, so write the increment and do not repeat.
Timeliness and verification: 24—48 hours. Awaiting verification: the source of the speed comparison numbers in the post is unconfirmed; follow-up note: agent-operated phones were already covered on 10-22.
08 Performance of websites built by vibe coding¶
Original posts and metrics: @shao__meng (10-24 11:48, 16K views, 62 likes, 70 bookmarks, 37 replies, 3 reposts).
Content and angles of interest: a site scoring 47 on Performance, loaded with huge GIFs and large PNGs, is used as a cautionary example to complain that once "anyone can build a site", nobody manages performance; replies far outnumber reposts, making it a classic peer-resonance post.
Why it's worth doing / possible angles: this problem is bound to surface once "AI can build a website", and it lets you leave readers with an actionable checklist. The angle: the items to check after AI writes a site (image format and dimensions, render-blocking first-screen resources, fonts, third-party scripts, LCP and CLS, whether this framework is really needed), and why models do not optimize these proactively — they generate for "looks right", not for runtime cost.
Timeliness and verification: within 24 hours. Awaiting verification: the specific score of the critiqued site is the poster's own measurement and has not been rechecked.
09 Codex and Claude Code quotas are tight again (follow-up)¶
Original posts and metrics: @xiaohu (10-24 13:15, 42K views, 129 likes, 4 bookmarks, 113 replies, 0 reposts); @turingou (10-24 01:11, 37K views, 56 likes, 2 bookmarks, 11 replies, 1 repost); @tualatrix (10-23 20:16, 20K views, 61 likes, 13 bookmarks, 40 replies, 2 reposts).
Content and angles of interest: a few articles burn through the whole quota, half of a week's quota disappears in a day, and after dropping Max 20x three entry-level accounts quickly burn through the 5-hour window — all mixed with worries about being banned again and about dumbing down. Bookmarks are extremely low while replies are high, making it more of an emotional outlet than reusable information.
Why it's worth doing / possible angles: the 10-23 and 10-24 topics already covered quotas, and the increment this time is not "not enough again" but "even holding several tiers at once is still not enough", plus users starting to trade off bans against degradation. The angle: calculate the sustainable duration of each tier by usage intensity, and offer alternatives (local models, pay-per-use APIs, splitting long tasks, moving repetitive work into scripts); if there is no official quota policy change, write only a short update and not a standalone piece.
Timeliness and verification: one post within 24 hours, two at 24—48 hours. Not verified: any official quota adjustment announcement; all three posts are personal impressions.
10 DHH's "Seven years from China"¶
Original posts and metrics: @Michaelzsguo (10-25 00:27, 12K views, 54 likes, 24 bookmarks, 3 replies, 1 repost).
Content and angles of interest: DHH returns to China after seven years, visits Shenzhen, Hangzhou and Beijing, and writes down his impressions of electric cars, the Xiaomi factory lines, engineering culture and AI density, admitting that a one-week trip inevitably carries a filter; the Chinese post offers a fairly long excerpted translation, centered on the "as if forty years had passed" contrast.
Why it's worth doing / possible angles: an outside view of Chinese manufacturing and engineering culture naturally brings European and Japanese carmakers as reference points. The angle: do not re-translate the excerpt; instead pick three verifiable observations (the ride experience of domestic EVs, the degree of factory automation, large companies led by engineers rather than MBAs) and check each against public material, then discuss "what a one-week trip can and cannot see".
Timeliness and verification: within 24 hours. Verified: the original is Seven years from China, and its theme matches the translation. Awaiting verification: the specific factories and subjective comparisons are the author's impressions and cannot be externally verified.
11 eMudhra's private key was literally "4"¶
Original posts and metrics: @ruanyf (10-23 15:47, 19K views, 90 likes, 38 bookmarks, 16 replies, 5 reposts).
Content and angles of interest: the private keys of several intermediate certificates from the Indian CA eMudhra were extremely small integers, one of them 4, others 56 and 115; researchers signed with these keys to prove it, browser root programs promptly revoked trust, and the company shut down its certificate business. Discussion is almost entirely about how absurd it is, with little technical explanation.
Why it's worth doing / possible angles: this is the best case for explaining why public key infrastructure depends on randomness, and readers can take away transferable common sense. The angle: go from "private key = 4" to how revocation takes effect (the push paths of mechanisms like OneCRL and CRLSet), why the problem lies not in hacking techniques but in issuance processes and the lack of self-checks, and end with one self-check tip for site owners.
Timeliness and verification: 24—48 hours. Verified: multiple private keys were extremely small integers, Mozilla, Apple and Microsoft root programs initiated revocation, and the company stopped its certificate business (The Register, Help Net Security, Mozilla Bug 2083449, CyberScoop).
12 Anthropic's S-1: observe first¶
Original posts and metrics: @foxshuo (10-23 20:02, 21K views, 64 likes, 36 bookmarks, 5 replies, 4 reposts).
Content and angles of interest: it retells a The Information reporter's preview, expecting Anthropic to publish the full S-1 the next day, and excerpts translations on the valuation dispute, the gap between employees' and the public's AI expectations, and Wall Street confidence. The post itself is secondhand retelling plus a long excerpt, with no new document.
Why it's worth doing / possible angles: what is worth doing is not this preview but the document itself — the revenue, losses and compute commitments in a prospectus are the hardest set of numbers all year. The angle: observe first, and once the document is public write "what three numbers show", testing that 91% versus 39% expectation gap against concrete financial definitions; until then, write no preview.
Timeliness and verification: 24—48 hours. Not verified: as of 10-25 08:21 there is still no public S-1 on EDGAR, only the confidential submission on June 1 (official announcement, status check); "publishing tomorrow" in the post is an expectation, not a fact. Status: observe first, not a standalone topic.
Gaps: the old-phone-tailnet repo, WorkBuddy's official fee schedule and the gemini-4-argon-trail model name could not be verified; Anthropic's S-1 was not public as of publication.