Skip to content

← Chinese Topics

X Chinese Hot Topics Brief

Date23 October 2026 · FridayWindow2026-10-21 08:23 to 2026-10-23 08:23CoverageComplete

Translation

English translation of the Chinese original.

Top Picks Today

  1. "Claude Code cracks a hotel door lock in 90 seconds" is a translated viral post (see item 1). The Chinese version posted by @make_claude at 10-22 01:41 drew 424K views, 3.5K likes, and 4,355 bookmarks, telling in first person how "I was lying in an $85 hotel room, had Claude Code scan Bluetooth, and it opened the door lock in 90 seconds; the red light on the smoke detector was actually the door lock's handshake indicator." But this text is almost sentence-for-sentence identical to the English post by @starmexxx (10-21 03:50, 621K views), and several accounts on the English-speaking side copied it verbatim (e.g. @0xForem). So far there is only skepticism, with no independent reproduction or security disclosure: @xVT66 called it engagement bait outright, and a reply from @grok also said no independent reporting could be found. The best additional angle is a clarification post: explain why this looks fabricated (a first-person experience copied word for word by multiple accounts, key technical details that cannot be checked), then keep the part that is genuinely worth worrying about—hotel electronic locks have had publicly documented vulnerabilities, and guests should still throw the mechanical deadbolt. Format: quote the Chinese original post for the clarification. Status: ready to draft; do not repeat its technical details as fact.

  2. 1Password for Agents: let agents log in but not see passwords (see item 2). This is hard news: the 1Password official blog published it on 10-22, and The Verge and TechCrunch reported it the same day. The mechanism: when an agent hits a login form, it asks 1Password for credentials, the user approves on their own device, and 1Password fills the password directly into the login form, so the agent never gets the plaintext. The official English post by @1Password has 680K views and 6.7K likes. There is no decent Chinese-language analysis yet; @dotey only says general sites might be worth trying but he wouldn't hand over his main accounts, while @xmglab asks "would you connect 1Password to Grok Bot?" The best additional angle is to explain clearly to Chinese readers "what it solves and what it doesn't": the password never enters the model context, but what an agent can do after logging in is not restricted; then give tiered advice on which accounts to hand over and which not to. Format: standalone original post. Status: ready to draft; currently only personal and family accounts on 1Password.com are supported, Grok Bot and Meta Muse are "rolling out over the coming days," while Claude Duet, ChatGPT Work, and dots are "coming soon"—distinguish these when writing.

  3. After a Claude ban, how to get your money back (follow-up, see item 3). The 10-20 topic covered the wave of bans; the increment this round is the refund path. @sino_go_global 88K views, says a user on Max 5x ($125) who was banned got the money back through a chargeback with a China Merchants Bank Visa; @Lniosytest writes about the Apple in-app purchase refund route, with 175 bookmarks higher than 148 likes; @tualatrix disclosed his own ban, and the 121 replies are mostly about whether refunds are possible. The best additional angle is a table broken down by payment channel: App Store in-app purchases go through Apple refunds (Apple's official process: sign in to reportaproblem.apple.com, choose Request a refund, result within 24 to 48 hours), Google Play goes through Google, and direct credit card payment on the official site can only go to the card issuer for a dispute. Format: standalone original post. Status: ready to draft; the successful China Merchants Bank chargeback is one case and cannot be written up as a reliable method; other posts promote using multiple Apple IDs to get repeated refunds, which is refund abuse and should not go into a guide.

Candidate Topics

1. "Claude Code cracks a hotel door lock in 90 seconds": a viral post that was translated and copied

  • Posts and engagement:
  • @make_claude｜Original post｜10-22 01:41 Shanghai｜24–48 hours｜424K views, 3.5K likes, 397 reposts, 132 replies, 4,355 bookmarks
  • English source: @starmexxx｜Original post｜10-21 03:50 Shanghai｜48–72 hours｜621K views, 2.8K likes, 3,990 bookmarks
  • Chinese retelling: @sleepy0x13｜Original post｜10-22 11:18 Shanghai｜within 24 hours｜8.0K views, 27 likes
  • What it says: The post claims Claude Code scanned a BLE controller in the room whose firmware had not been updated since 2014 and which controlled the door lock, smoke detector, thermostat, and safe at once; the door lock used a 4-byte rolling key, and by the third attempt it generated the master key, concluding that anyone in the hallway could open the door with "one prompt plus 90 seconds." The Chinese version is almost identical sentence by sentence to the English one, and English-side accounts such as @0xForem and @Mr_Chonky also posted text with exactly the same opening. In the quote tweets, every language is passing around "hotels are no longer safe to stay in," while skeptics such as @xVT66 point out that Claude Code itself has no built-in ability to scan arbitrary Bluetooth devices.
  • Why it matters / angles: This is the highest-engagement AI post in the Chinese-language space in this window, and it spreads through fear. The value of a clarification post is to demonstrate how to judge this kind of "AI miracle": the same first-person experience posted verbatim by multiple accounts, no device model or video, and no security disclosure are all signals. At the same time, don't go to the other extreme—hotel electronic locks do have a history of publicly documented vulnerabilities, so the piece can land on practical advice such as "use the deadbolt when staying in a hotel and don't leave valuables in the in-room safe."
  • Timing and verification: 24–48 hours; the English source is now close to 72 hours old. No independent reproduction, vendor response, or security research disclosure was found; all technical details in the text are unverified and should be treated as a suspicious rumor.

2. 1Password for Agents: agents can log in but can't see your password

  • Posts and engagement:
  • @1Password｜Original post｜10-22 22:30 Shanghai｜within 24 hours｜680K views, 6.7K likes, 571 reposts, 4,095 bookmarks
  • @dotey｜Original post｜10-22 23:27 Shanghai｜within 24 hours｜13K views, 23 likes, 19 replies
  • @xmglab｜Original post｜10-21 17:18 Shanghai｜24–48 hours｜7.0K views, 14 likes, 18 replies
  • What it says: When an agent reaches a login page, it proves its identity to 1Password with a dedicated digital key; the user sees on their own device which credential it wants and why, and after approval 1Password fills the password or one-time code directly into the form, without handing the credential to the agent. First integrations are Meta Muse, Anthropic Claude Duet, OpenAI ChatGPT Work and dots, Grok Bot, and Town Townie, with Browserbase, Browser Use, Perplexity, Ramp, and others to follow. Chinese discussion centers on "dare you use it": dotey is only willing to hand over general sites, not his main accounts.
  • Why it matters / angles: Four proactive assistants just launched in a cluster, and login authorization is where everyone gets stuck; this connects directly to the 10-22 topic's question of "how to set authorization boundaries." It can be written as "three tiers of handing passwords to agents": don't hand them over at all, hand them to 1Password with per-request approval, or give the plaintext password directly—and note that per-request approval blocks password leaks but does not block mistakes the agent makes after logging in.
  • Timing and verification: Within 24 hours. Verified: the 1Password official blog, the Business Wire press release, and The Verge. Personal and family accounts are not charged extra; 1Password.eu and 1Password.ca accounts are not yet supported; integration progress varies by vendor. xmglab said the day before the official launch that Grok Bot could connect to 1Password; the exact timeline is unverified.

3. Refund paths after a Claude ban (follow-up)

  • Posts and engagement:
  • @sino_go_global｜Original post｜10-22 20:41 Shanghai｜within 24 hours｜88K views, 176 likes, 35 replies, 73 bookmarks
  • @Lniosytest｜Original post｜10-22 18:19 Shanghai｜within 24 hours｜11K views, 148 likes, 175 bookmarks
  • @tualatrix｜Original post｜10-22 08:44 Shanghai｜within 24 hours｜44K views, 131 likes, 121 replies
  • @Stargridcompass｜Original post｜10-21 09:20 Shanghai｜24–48 hours｜227K views, 611 likes, 135 replies
  • What it says: sino_go_global describes a user on Max 5x who was banned and got a refund through a China Merchants Bank Visa, and in a follow-up post summarizes by channel: US App Store goes to Apple, Google Play goes to Google, and direct credit card payment on the official site has basically no hope with Anthropic. Lniosytest gives the Apple refund entry point and an English template for the appeal reason. tualatrix, a well-known developer, disclosed his own ban, and the comments are about whether refunds are possible. Stargridcompass posted images claiming US users' DeepSeek accounts were also banned and the emails looked "familiar," mostly as a joke.
  • Why it matters / angles: The ban topic has already been covered, but "choosing your exit when you pay" is a new, actionable layer. The most suitable format is a channel comparison table, plus a reminder: before subscribing, think about who you would ask for a refund if you get banned.
  • Timing and verification: Mostly within 24 hours. The Apple refund process is verified against the Apple support document: choose Request a refund at reportaproblem.apple.com, updates within 24 to 48 hours, and approval is up to Apple. The successful China Merchants Bank chargeback and Google refund success rates are personal shares and unverified; Stargridcompass's claim that DeepSeek banned US accounts has only one screenshot and is unverified. There is also @zureA255062 (95K views) promoting "unlimited refunds after bans" with multiple Apple IDs, which is refund abuse and can only serve as a counterexample. This is a follow-up: item 4 of the 10-20 topic covered the ban wave, and this item's increment is refund channels.

4. Terence Tao and the mathematics world's "war footing"

  • Posts and engagement:
  • @foxshuo｜Original post｜10-22 14:08 Shanghai｜within 24 hours｜31K views, 262 likes, 36 reposts, 169 bookmarks
  • What it says: foxshuo's long post relays a New York Times interview with Terence Tao: he had been the most AI-friendly of mathematicians, but OpenAI's October 6 announcement of more than 350 mathematical results disrupted a large amount of ongoing research, and he felt OpenAI was just "using math to evaluate models." The post also says Tao knew about the release plan a month in advance and opposed it, and afterward published a statement on his blog and co-signed it with several Fields Medalists.
  • Why it matters / angles: This isn't "AI solved another problem," but a person in an industry describing the real situation after being hit by AI, which resonates with any professional. The 10-22 English topic covered the mathematicians' joint pledge; this item switches to a personal perspective; Tao's October 22 Nature commentary (arguing for building "Mathematics 2.0" and valuing human understanding rather than just piling up proofs) is the new increment, and the two pieces can be read together.
  • Timing and verification: Chinese spread is within 24 hours; the original NYT piece was published 10-16. Verified: the NYT report is available via the UCLA Newsroom summary and the full repost; "I'm on a war footing," the more than 350 results on October 6, and "at this point, I don't see a way to constructively engage" all come from the original; Tao's Nature commentary was published 10-22. Details foxshuo mentions, such as OpenAI sponsoring research institutes with tens of thousands of dollars and 24 Fields Medalists co-signing, were not confirmed in the original this round and remain unverified.

5. MemT-1.0 "memory-native" model: Chinese coverage got the parameters wrong

  • Posts and engagement:
  • @servasyy_ai｜Original post｜10-21 10:30 Shanghai｜24–48 hours｜40K views, 58 likes, 59 replies
  • @vista8｜Original post｜10-22 09:04 Shanghai｜within 24 hours｜10K views, 86 likes, 69 replies
  • What it says: The MemOS team (MemTensor) released MemT-1.0 on 10-20, with the idea of building memory into the model computation itself rather than bolting on RAG or a memory store outside. servasyy_ai's long post says "324B total parameters, 20B active," which does not match the official information.
  • Why it matters / angles: "AI starts from zero every time" is a pain point for all users, and this item can explain the difference between external memory and native memory. At the same time, it can correct the error along the way: the official release is two models, and the parameters are not the numbers being circulated.
  • Timing and verification: Mostly 24–48 hours. Verified against the arXiv paper, the Hugging Face model page, and GitHub: MemT-1.0-Nano has 36B total parameters and 3B active, based on Qwen3.6-35B-A3B; MemT-1.0-Flash has 288B total parameters and 17B active, based on DeepSeek-V4-Flash; at a 128K history length, persistent storage uses about 66 times less than a full KV Cache, and Nano's end-to-end inference is about 1.86 times faster. "The world's first" is the team's and the reteller's claim.

6. How to choose between Pi and DeepSeek Harness (follow-up)

  • Posts and engagement:
  • @xiaomovps｜Original post｜10-22 16:10 Shanghai｜within 24 hours｜19K views, 58 likes, 38 bookmarks
  • @leo114119｜Original post｜10-21 15:53 Shanghai｜24–48 hours｜37K views, 161 likes
  • @yetone｜Original post｜10-22 01:15 Shanghai｜24–48 hours｜137K views, 463 likes, 72 replies
  • What it says: xiaomovps compares the two: Pi keeps its core as simple as possible and relies on extensions and Skills for whatever is needed; DSH allows model adapters, tool registration, session storage, and the agent loop to all be replaced with plugins. He cites a set of "independent tests": on a Terminal-Bench subset using GPT-6 Astra for both, Pi scored 60.32% and DSH 52.38%. leo114119 posts his bill: calling DeepSeek V4.1 Flash, 700 million tokens cost $13 without DSH, while 900 million tokens cost a bit over $7 with DSH. yetone's line "dsh's user count has fully surpassed pi" has now reached 137K views.
  • Why it matters / angles: The 10-22 topic was about the DSH plugin ecosystem; this round readers are starting to ask "which one should I actually choose." The angle is "a thicker shell isn't necessarily better": the same model in a different shell changes both results and cost, so test your own tasks before choosing.
  • Timing and verification: Half within 24 hours and half 24–48 hours. The test scores give no source, the reason for leo114119's price difference (whether cache hits) is unexplained, and yetone's user numbers have no data; all three are unverified. Follow-up: item 1 of the 10-22 topic; this item's increment is the comparison and cost data.

7. The X Subscription: the $30 tier—is it worth it, and will it be "auto-merged"?

  • Posts and engagement:
  • @Zesee｜Original post｜10-21 21:31 Shanghai｜24–48 hours｜10K views, 97 likes, 71 reposts, 139 replies
  • @Gorden_Sun｜Original post｜10-22 10:55 Shanghai｜within 24 hours｜9.5K views, 15 likes, 16 replies
  • @abskoop｜Original post｜10-22 14:24 Shanghai｜within 24 hours｜19K views, 3 likes, 8 replies
  • What it says: X, Grok, and Cursor merge into one subscription. Zesee does the math: for people paying for all three, merging may be worthwhile; for those mainly using Codex or Claude and only wanting to try Grok Bot, it's extra spending; the second half of the post promotes a third-party group-buy platform, which should not be cited. Gorden_Sun says he did nothing and his Cursor and X subscriptions merged automatically, even merging chat history; abskoop asks whether there are bugs after the merge.
  • Why it matters / angles: Chinese users care most about two things with a new plan: what each tier includes, and whether old subscriptions will be forcibly changed. You can write a "what you pay now, should you switch" comparison.
  • Timing and verification: Mostly within 24 hours. Verified against the xAI pricing docs and the Cursor help center: Plus is $8/month, Premium $30/month, Super $100, Ultra $200; Grok Bot is only in Premium and above; Grok, Grok Bot, and Cursor share one weekly quota pool; third-party models in Cursor are billed separately; the official line is that switching is optional and existing plans remain valid. This differs from Gorden_Sun's "automatic merge" account; his case is pending verification.

8. Codex quotas shrinking, and the "reset" rhythm (follow-up)

  • Posts and engagement:
  • @xiaohu｜Original post｜10-22 11:34 Shanghai｜within 24 hours｜98K views, 210 likes, 82 replies
  • @dotey｜Original post｜10-23 00:01 Shanghai｜within 24 hours｜28K views, 119 likes, 44 replies
  • @imwsl90｜Original post｜10-22 19:41 Shanghai｜within 24 hours｜33K views, 19 likes, 22 replies
  • What it says: xiaohu teaches people when to use a "reset card" and how to tell which day the official reset happens; imwsl90 asks whether the official side keeps shrinking usage and then resetting like crazy; dotey quotes this and guesses that the 28-days-without-update reset event is meant to maintain the original one-week usage with a reset about every three days; he explicitly says it's a guess.
  • Why it matters / angles: Quota is a daily topic for Codex users, and the new emotion this round is "quotas are shrinking." You can write a calm piece: right now there are only feelings and guesses, no official numbers, so teach readers how to record their own usage to judge.
  • Timing and verification: Within 24 hours. There is no official explanation or comparable data for the quota shrinkage; it is all user perception and speculation. Follow-up: the 10-20 topic covered the 28-day plan, the 10-22 topic covered Codex caching, and this item's increment is the complaint about shrinking usage.

9. After Agent Skills goes to the foundation: where should skill files live (already covered in the daily)

  • Posts and engagement:
  • @dotey｜Original post｜10-23 02:02 Shanghai｜within 24 hours｜8.6K views, 29 likes, 21 replies
  • Quoted: @ClaudeDevs｜Original post｜10-23 01:55 Shanghai｜within 24 hours｜437K views, 5.8K likes
  • What it says: The news that Anthropic handed the Agent Skills spec to AAIF was already written in today's daily. dotey's response is one line: he hopes Claude Code can read the common .agents/skills directory instead of building its own .claude/skills.
  • Why it matters / angles: After the spec goes to a neutral body, the most practical question for users is whether the same set of skills can be shared across Claude Code, Codex, and Cursor. It can be written as a hands-on post: which directory each tool reads skills from now, and how to use symlinks to make one file take effect in multiple places.
  • Timing and verification: Within 24 hours. The news itself is in today's X AI daily; this item only takes the hands-on angle; there is no official word on whether Claude Code will support .agents/skills, and each tool's directory convention needs to be checked one by one.

10. Let agents read and reverse-engineer local software

  • Posts and engagement:
  • @geekbb｜Original post｜10-21 09:44 Shanghai｜24–48 hours｜226K views, 786 likes, 743 bookmarks
  • @tantanyuli｜Original post｜10-21 21:04 Shanghai｜24–48 hours｜8.8K views, 110 likes, 107 bookmarks
  • What it says: geekbb one line plus a video: software that stores data locally should all let agents read it, and if it can't be read, crack it and read it. tantanyuli introduces reverse-skill: after putting it into Codex, Cursor, or Claude Code, when the agent encounters an APK, front-end code, or firmware, it knows where to start and which tool to call.
  • Why it matters / angles: Bookmarks are close to likes, which shows people want to use it. The angle can connect to item 1: agents really do make reverse engineering easier, so it's all the more important to distinguish "reading your own data" from "cracking someone else's system"—the former is productivity, the latter has legal risk.
  • Timing and verification: 24–48 hours. geekbb's video content was not reviewed this round; the reverse-skill repository was not opened this round, and the capability description comes from a retelling and is unverified.

11. Claude Code cloud projects and remote control: authorization levels differ a lot

  • Posts and engagement:
  • @turingou｜Original post｜10-22 20:36 Shanghai｜within 24 hours｜16K views, 23 likes
  • @turingou｜Original post｜10-22 13:56 Shanghai｜within 24 hours｜15K views, 36 likes, 24 bookmarks
  • @dotey｜Original post｜10-22 08:40 Shanghai｜within 24 hours｜13K views, 28 likes, 41 replies
  • @dotey｜Original post｜10-22 23:49 Shanghai｜within 24 hours｜17K views, 71 likes, 61 bookmarks
  • What it says: turingou likes the Claude Code desktop app's "virtual project": a cloud computer hosts the project, automatically analyzes and watches PRs, and when needed he uses Remote Control to operate his Mac Studio in reverse; he also notices that Claude Code almost requires verbal authorization for every step during remote operation, while Codex with bypass mode enabled almost never asks. dotey adds that task cards can be thrown to the cloud for execution, provided the project is on GitHub and the connector is set up; another post says that when out, let the agent consult /advisor first when it has a question.
  • Why it matters / angles: Running in the cloud and remotely controlling your own machine have become standard; the difference is how much authorization you give. You can do a comparison of "how many confirmations the same task requires at each vendor," with the conclusion landing on who each is suited for.
  • Timing and verification: Within 24 hours. All are personal usage experiences, with no comparative testing; turingou says it is temporarily unavailable to download in Japan, and the available regions are unverified.

12. Making audiobooks with AI: what to do about polyphonic characters

  • Posts and engagement:
  • @dotey｜Original post｜10-22 08:47 Shanghai｜within 24 hours｜13K views, 29 likes, 33 replies
  • @dotey｜Original post｜10-22 11:34 Shanghai｜within 24 hours｜16K views, 64 likes, 88 bookmarks
  • What it says: dotey thinks AI audiobooks are still not as good as professional human narration, and polyphonic character errors will mislead children; if you really want to do it yourself, first have an LLM mark pinyin, pauses, and emotion, then have an agent operate the TTS. The second post introduces a detection method: feed the TTS audio to Kaldi for recognition, and instead of looking at the recognized Chinese characters, read the phoneme sequence along the decoding path and compare it with the target pronunciation to determine what was actually read.
  • Why it matters / angles: Parents making children's audiobooks is a very concrete need, and "annotate first, then synthesize, then use recognition to check" is a reusable workflow. Suitable for a tutorial.
  • Timing and verification: Within 24 hours. The original source of the detection method was not opened this round, and the effect is unverified.

13. "Learning ahead is wasted" and "1000+ / 10K+ / 100K+": how to prove yourself in the AI era

  • Posts and engagement:
  • @AlchainHust｜Original post｜10-22 16:26 Shanghai｜within 24 hours｜28K views, 107 likes, 40 replies
  • @lxfater｜Original post｜10-22 12:33 Shanghai｜within 24 hours｜16K views, 62 likes, 51 replies
  • @AlchainHust｜Original post｜10-22 11:24 Shanghai｜within 24 hours｜15K views, 94 likes, 72 bookmarks
  • What it says: AlchainHust argues that agents can already help one person make anything, so if you still don't have any one of a 1,000-star open-source project, a product with 10,000 users, or content seen by 100,000 people, you should reflect; in another post he says he checked Douyin AI bloggers' sponsored-post rates and feels there are too many "high-cognition poor people" on X, but one shouldn't be arrogant about it. lxfater says he learned to make videos in half a month, and after taking a break and coming back, his UI animation skills had been fully internalized by Opus 5.5.
  • Why it matters / angles: Both are asking "how does an individual's ability show in the AI era," with high reply counts and lots of disagreement. You can write your own judgment: results matter more than a skill list, but the three numeric thresholds themselves are debatable.
  • Timing and verification: Within 24 hours. Opinion discussion, no fact-checking needed; the sponsored-post rate screenshot is unverified.

14. Making learning websites with AI (follow-up)

  • Posts and engagement:
  • @vista8｜Original post｜10-23 00:17 Shanghai｜within 24 hours｜8.8K views, 142 likes, 152 bookmarks
  • @xiaohu｜Original post｜10-21 20:02 Shanghai｜24–48 hours｜9.9K views, 94 likes, 57 bookmarks
  • What it says: vista8 turned the previous day's learning prompt into a website, with content ranging from cooking and fishing to computer shortcuts and fitness, and every knowledge point comes with a quiz or interactive animation. xiaohu is building a 3D interactive textbook website with voice narration, and he says himself the workload is huge and it may not succeed.
  • Why it matters / angles: From one prompt to a site is a good sample of "making small products with AI." It can be written as a breakdown: how the prompt becomes a reusable page template.
  • Timing and verification: Mostly within 24 hours. The website was not opened this round. Follow-up: item 15 of the 10-22 topic covered vista8's learning prompt, and this item's increment is that it became a website.

15. Parts list for the 32 projects on OpenAI's hardware site

  • Posts and engagement:
  • @wquguru｜Original post｜10-21 10:51 Shanghai｜24–48 hours｜9.5K views, 141 likes, 166 bookmarks
  • What it says: wquguru says OpenAI launched a hardware site last week collecting 32 hardware projects mostly made with Codex; he dug through the parts lists, finding at least 14 use ESP32 (S3 the most), 8 use Raspberry Pi, and he picked 7 with open-source repositories, such as jailbreaking an old Kindle into a Codex/Claude usage display.
  • Why it matters / angles: Bookmarks higher than likes is the typical "save it for the weekend" pattern. You can pick one or two projects and actually build them, writing it as "little gadgets you can make with AI even if you don't know hardware."
  • Timing and verification: 24–48 hours. The OpenAI hardware site page and project count were not opened and checked this round; the retelling is the basis and remains unverified.

16. Graphify: turning a codebase into a knowledge graph, "saves 80% of tokens" to be verified

  • Posts and engagement:
  • @NFT_Chen｜Original post｜10-21 18:06 Shanghai｜24–48 hours｜9.3K views, 105 likes, 132 bookmarks
  • What it says: The post claims that large projects have to rescan files every time a new session starts, and Graphify turns the codebase into a locally queryable knowledge graph so Claude Code and Codex can query relationships directly, cutting token consumption by 80%.
  • Why it matters / angles: Saving quota is a hard need for Codex and Claude users, but numbers like "saves 80%" most need real testing. You can run a before-and-after comparison on your own project.
  • Timing and verification: 24–48 hours. The tool really exists: the Graphify GitHub explains that it uses Tree-sitter to parse code locally, generate a queryable graph, and has a Claude Code integration. "Saves 80%" is the post's claim and is unverified.

17. Fusion company says it turned mercury into gold (science crossover)

  • Posts and engagement:
  • @dotey｜Original post｜10-23 04:13 Shanghai｜within 24 hours｜18K views, 59 likes, 59 bookmarks
  • What it says: Marathon Fusion, a 28-person fusion company in San Francisco, posted on 10-21 that it had used fusion neutrons to turn mercury into stable gold, at 10 times the CERN 2025 record. dotey's long post does the math: according to the company's own paper, a gigawatt-scale plant would produce about 2 tons of gold a year, and catching up with global gold mine output would require building 1,800 of them, while currently there isn't even one; he thinks what the company really wants to sell is medical isotope feedstock.
  • Why it matters / angles: The headline is eye-catching and the math is clear, making it a good demonstration of "do the math first when you see a shocking science story." It has little to do with AI, so it is a science-crossover backup.
  • Timing and verification: Within 24 hours. Verified against the Marathon Fusion official explainer and the foreign media report BusinessToday: the evidence so far is only one gamma spectrum chart; output, purity, and cost have not been disclosed, and there is no independent verification.

This issue has 17 items, ranked by creative value; items 12 to 17 are worth watching rather than acting on, with lower engagement or incremental value. Coverage gaps: no video frames were reviewed this round, some sites and repositories were not opened, and the relevant claims are marked as unverified in the cards.