X English Topic Brief¶
Translation
English translation of the Chinese original.
Top picks today¶
1. Goldman Sachs tightens Claude Code: traders and bankers are out, engineers only (card 1)¶
Why now: after Bloomberg's 11-02 report, @exec_sum's one-line post has 2,679,149 views, 7,490 likes and 900 bookmarks on X. In July, Goldman CIO Marco Argenti was still publicly encouraging non-technical staff to vibe code; by October the internal guidance restricted Claude Code and GitHub to engineers. It is the first time a large company has publicly pulled back from "everyone writes code."
Best angle to add: do not write it as "Goldman bans AI." Goldman did not ban AI; the internal GS AI Assistant stays open. What was pulled back are the tools that write and run code directly. Focus on how big companies trade off productivity, risk and cost, and where the boundary sits for non-engineers building tools with AI.
Format and status: industry commentary. Ready to draft; Bloomberg and Quartz reports verified. Headcount and Devin-scale figures in @BoringBiz_'s post come from that account's own compilation and were not individually checked.
2. levelsio decompiles the Battlefield 1942 demo with Claude Code, ports it to the browser with multiplayer (card 2)¶
Why now: the two main posts total roughly 270,000 views and over 1,700 likes. He says an attempt six months ago failed; this time Opus 5.5 did it: 57,000 x86 functions extracted, translated into 100,000 lines of C, compiled to WebAssembly, then about 300 Windows and Direct3D calls rewritten. In a separate post he says it used about 1 million output tokens, roughly 100 dollars at API prices, while his monthly Claude subscription is 100 dollars.
Best angle to add: this is not another "AI writes a small game" story. The real point is that AI can now do reverse engineering and binary porting, work that used to take specialists months. Worth also covering the true cost for heavy users under a subscription model.
Format and status: case breakdown. Ready to draft; the site bf1942.pieter.com loads and states it is the 2002 free demo, a non-commercial unofficial project. Technical details come from levelsio's own account, not independently reproduced.
3. Hume ships the Audio Decision API: ask questions of audio, get probabilities back (card 3)¶
Why now: @hume_ai's official post has 1,358,362 views, 1,225 likes and 1,326 reposts. One call answers questions like "how many speakers," "is anyone laughing," "is the caller upset" — questions that used to require stitching together a dozen models.
Best angle to add: make clear how it differs from speech-to-text: no transcription step, direct judgment of tone, emotion and speakers. Fits customer-service QA, meeting analysis and podcast editing.
Format and status: new product explainer. Ready to draft; features, pricing and benchmark numbers checked against Hume's official blog and docs, and the benchmark is vendor-run. Several large accounts posted similar promos in the same window, possibly a coordinated push (speculation), so discount the traction.
Candidate topics¶
1. Goldman Sachs pulls Claude Code and GitHub back to engineers¶
Posts and reach: @exec_sum, 2026-11-02 23:55, 2,679,149 views, 7,490 likes, 531 reposts, 163 replies, 900 bookmarks; @shortsqueeznews, about 66,000 views, 892 likes; @BoringBiz_ long breakdown, 11-03 02:34, 51,059 views, 380 likes, 209 bookmarks.
Content and what stands out: according to Bloomberg and Quartz, Goldman updated internal guidance in October and, after weighing productivity and risk, limited Claude Code and GitHub to engineers, with strats roles spanning trading and quant excepted. Traders had used these tools to monitor client flows, model positions and produce written and chart materials. A spokesperson said Goldman first opened AI tools company-wide to find uses, then handed them to the people best suited to long-term use. In July CIO Argenti was still encouraging non-technical staff to vibe code; in August partner Chris Churchman warned about eroding analytical skills.
Why it's worth doing and angles: enterprise AI adoption is shifting from "everyone gets access" to "access by role." Chinese readers want to know what goes wrong when non-engineers build tools with AI, and why a company would pull back.
Timeliness and verification: within 24 hours; verified (Bloomberg, Quartz).
2. levelsio decompiles Battlefield 1942 and ports it to the web with multiplayer¶
Posts and reach: first post, 11-02 22:17, about 193,000 views, 1,031 likes; technical post, 11-03 05:55, 73,786 views, 713 likes, 388 bookmarks; cost post, 11-03 06:50, 19,099 views, 144 likes.
Content and what stands out: Claude Code extracted 57,000 x86 functions and translated them one by one into roughly 100,000 lines of C, compiled with Emscripten to WebAssembly, rewrote about 300 Windows and Direct3D calls as WebGL 2, and built a WebSocket networking layer. levelsio says it was far harder than his earlier Quake and DOOM ports, because those have open-source engines and Battlefield 1942 has nothing. The project page states it is the 2002 free demo, non-commercial, and unaffiliated with EA and DICE.
Why it's worth doing and angles: the limits of AI at reverse engineering and porting old software, plus the cost math behind "a 100-dollar-a-month subscription producing 100 dollars of API usage." Mention copyright in a sentence, but do not make it the throughline.
Timeliness and verification: within 24 hours; site checked, technical process is his own account.
3. Hume Audio Decision API: ask questions directly of audio¶
Posts and reach: @hume_ai official, 11-03 00:00, 1,358,362 views, 1,225 likes, 1,326 reposts, 957 bookmarks; sample promo posts: @svpino, about 29,000 views; @DataChaz, about 23,500 views.
Content and what stands out: per Hume's official blog and docs, you feed in raw audio plus a question and get back a probability. Questions fall into four types — yes/no, choice, scoring, span location — with MP3 and WAV supported, 5 seconds to 5 minutes per clip, and billing at 0.47 dollars per million input tokens; access currently requires a waitlist. Hume says agreement with an average human review panel is 70.7%, versus a 56.5% single-reviewer baseline.
Why it's worth doing and angles: voice AI moving from dictation to understanding. Good to break down for readers in customer service, sales recordings and content moderation.
Timeliness and verification: within 24 hours; features and pricing verified, benchmark is vendor-run; multiple accounts posted similar promos in the same window, possibly a coordinated push (speculation).
4. When agents pick their own tools, they barely use MCP¶
Posts and reach: @stevelacey, 11-03 04:53, 49,775 views, 49 likes, 30 replies; @levelsio quote post, 11-03 05:03, about 52,900 views, 241 likes.
Content and what stands out: stevelacey says he built MCP integration into his product, but his own agents almost never choose it. His setup is one Claude Code session per agent, one project directory, one .env with keys filled in; he states the goal and leaves the method alone, and the agent usually just calls APIs or the command line.
Why it's worth doing and angles: is MCP overrated? Compare where MCP, direct API calls and CLI tools each fit, and give practical advice.
Timeliness and verification: within 24 hours; original post checked, personal experience, no controlled test.
5. Mollick: AI labs treat products as "a stopgap before AGI"¶
Posts and reach: @emollick, 11-02 23:35, 29,020 views, 612 likes, 208 replies.
Content and what stands out: Wharton professor Ethan Mollick says labs often ship abrupt software changes with poor communication, partly because their core belief is that products are temporary vehicles for raising money over the next few years, and that once real AGI arrives every product will be obsolete.
Why it's worth doing and angles: this explains the recent run of limit changes, features disappearing overnight and docs lagging behind. A Chinese piece could discuss how users and developers should respond: do not stake your business on one lab's product.
Timeliness and verification: within 24 hours; opinion post, original checked.
6. Switching from Codex to Claude Code: same 20 dollars, very different usage¶
Posts and reach: @akshdeeps_001, 11-02 23:15, 33,667 views, 240 likes, 62 replies; counterexample: @adxtyahq, 11-02 14:26, 4,470 views, 109 likes, complaining that Claude Code limits have tightened over the past two or three days.
Content and what stands out: the first says that on the same Figma implementation task, Sol 6.1 used up the 5-hour quota at about 70% done, while Opus 5.5 at medium thinking finished the whole thing using roughly half the quota. The second says he hit 40% without using it much.
Why it's worth doing and angles: real subscription-quota experience varies hugely by person and task. A practical piece on how to measure your own real usage and choose a subscription, tied to the past two days of quota debates.
Timeliness and verification: within 24 hours; both are personal accounts, not tested.
7. Exa puts full transcripts of 3.2 million podcast episodes into its search API¶
Posts and reach: @rileybrown, 11-03 03:35, about 44,000 views, 588 likes, 782 bookmarks.
Content and what stands out: per Exa's official blog, the index covers 119,000 shows and 3.2 million episodes, with speaker-labeled full transcripts, about 23.7 billion words; available through the existing search API at no extra charge. Exa's own test says the latest episode appears in the top five results 98.4% of the time, and 85.0% of 667 specific questions were answered correctly.
Why it's worth doing and angles: podcasts are where founders and investors say things first, and now agents can search the exact words. Fits a hands-on piece on "using AI for podcast research."
Timeliness and verification: circulating within 24 hours; Exa's original launch was 10-29, past the 72-hour window, so this is logged as a new wave of sharing. Features verified, benchmark is vendor-run.
8. Mollick built a US city routing map site with Opus 5.5¶
Posts and reach: @emollick, 11-03 04:01, 39,798 views, 567 likes, 243 bookmarks.
Content and what stands out: the site shows the fastest route to every US city, finds the fairest nationwide meeting point, and shows "road-network watersheds" for big cities, inspired by a defunct site called Arterial Mapping.
Why it's worth doing and angles: a good example of a non-programmer using AI to build a small data-visualization site, useful material for "one person shipping a complete product with AI."
Timeliness and verification: within 24 hours; post checked, site not opened.
9. Beacon: an open-source memory layer that lets different coding agents share experience¶
Posts and reach: @jqdsouza, 11-02 23:59, 153,719 views, 212 likes, 250 bookmarks; @svpino recommendation, about 9,600 views, 102 likes.
Content and what stands out: per the GitHub repo, Beacon records sessions from more than 20 tools including Claude Code, Cursor and Codex, distills debugging lessons and practices from them, and after human confirmation stores them as project memory or an Agent Skill, then serves them to later agents over MCP; local-first, MIT licensed. Product Hunt shows it at number one for the day on 11-02.
Why it's worth doing and angles: more people run several coding agents at once, and how to keep and share the experience is a real problem. Could be combined with the MCP topic above.
Timeliness and verification: within 24 hours; repo and Product Hunt checked. High views, few likes, so promotion is part of the reach.
10. Follow-up: Linux kernel CVE-2026-98432 still unfixed on Ubuntu, levelsio considers Debian¶
Posts and reach: @levelsio, 11-02 23:07, 113,845 views, 1,533 likes, 510 bookmarks; @IntCyberDigest, 11-02 03:57, 234,948 views, 2,344 likes.
Content and what stands out: this vulnerability was covered earlier. The new development: the Ubuntu security page was last updated 10-30, and kernels for major releases including 26.04, 24.04 and 22.04 are still marked "Vulnerable, fix in progress." levelsio says he may move all his servers to Debian, on the grounds that AI speeds up exploitation, so patching has to be just as fast.
Why it's worth doing and angles: AI shortens the time from disclosure to exploitation, making distro patch speed a selection criterion.
Timeliness and verification: levelsio's post within 24 hours, IntCyberDigest's within 24-48 hours; Ubuntu status checked.
11. Value watch: coding agents quietly pile up tens of GB of logs on disk¶
Posts and reach: @illyism, 11-02 11:56, 10,377 views, 62 likes, 90 replies. Below the main threshold, admitted under the practice-category relaxation: over 10,000 views and directly actionable.
Content and what stands out: after cleaning his Mac he freed 13.8 GB: 14,348 old session files from Grok CLI, 4,944 Cursor conversation records older than 30 days, 409 Codex sessions plus an 896 MB SQLite history database, and 44 stale keychain tokens. He includes a prompt that has the agent list sizes first, then delete anything older than 30 days, then add daily automatic cleanup.
Why it's worth doing and angles: beyond disk space, these records may contain code and keys. Fits a short tutorial on local data security and cleanup for coding agents.
Timeliness and verification: within 24 hours; figures are his own, not tested.
The account sweep groups (including sama, DrJimFan, demishassabis, ylecun, rasbt, deepseek_ai, drfeifei and others) returned nothing this round, and the levelsio, poteto and karpathy group was swamped by levelsio's posts, so those accounts may have gaps.