X English Hot Topics Brief¶
Translation
English translation of the Chinese original.
Top picks today¶
1. DeepSeek traffic on OpenRouter: can one rack handle it? (card 1)¶
Why now: @thdxr got 964,000 views with one line: "OpenRouter's DeepSeek traffic yesterday could almost be served by a single GB300 rack." Then someone asked, "Nvidia is putting out 90,000 racks next year, what am I missing?", and someone else broke out the math: the vast majority of that token count is cache hits from agents re-reading context. This is exactly the number Chinese readers most often misread — what "trillions of tokens a day" actually means in compute.
Best added angle: separate tokens into three classes: newly generated, input processed for the first time, and repeated context served from cache. The conclusion is not "compute glut" but that in the agent era costs may shift from GPUs toward cache and storage; for readers using APIs themselves, this maps directly onto "why cache-hit pricing is so cheap, and why you want a stable context prefix".
Format and status: explanatory original. Draftable; the three authors' numbers disagree (one rack vs 5–10 racks), all are personal estimates; OpenRouter's daily token volume, the 99% input share and the 95% cache-hit rate still need checking against OpenRouter's public leaderboard.
2. Simon Willison: Opus 5.5 makes "just build a fun game" work for the first time (card 2)¶
Why now: Simon called November 2025 the inflection point for coding agents, and now calls October 2026 the inflection point for "vibe game dev", backing it with four playable games and the complete prompts. The same week, Steam Next Fest is being flooded with "AI slop games" — together the two make the two sides of the story.
Best added angle: reproduce his opening prompt (three.js, mobile-friendly, ending with "Make it a LOT more fun"), and log how many rounds it took and where a human had to decide the gameplay. Then answer the question he raises himself: if it is nearly effortless to build something fun, what can people who put in real effort reach?
Format and status: hands-on original. Draftable; the blog post content has been verified, but "GPT-6 Astra is less fun than Opus 5.5" is his personal impression, not a benchmark.
3. "The people who get the most out of AI are the ones who could do the work well without it" (card 3)¶
Why now: this "unpopular opinion" has 221,000 views and nearly 6,000 likes, and Riley Brown's quote added a very concrete example: several hundred people DM'd him asking which skill to install to edit video with Opus, and his answer was to spend two weekends learning the editing software first. That is the opposite of the Chinese-language narrative that "install a skill and you are an editor".
Best added angle: make "AI is an amplifier" operational: which skills you have to do yourself at least once (judging quality, stating the criteria, knowing the terminology), and which can go straight to AI. Use one of the owner's own workflows as an example of what "learn it first, then automate" actually saves.
Format and status: opinion post or quote post. Draftable; it is an opinion, with no data to verify, but do not write it as "AI is useless to beginners".
Candidate topics¶
1. OpenRouter's DeepSeek traffic: most of the token count is cache hits¶
Original posts and traction: @thdxr, 2026-10-28 03:34, 4,143 likes, 964,517 views, 1,490 bookmarks; @chontang, 10-28 05:57, 1,828 likes, 330,838 views, 659 bookmarks, 121 replies; @demian_ai, 10-28 22:59, 292 likes, 25,129 views, 207 bookmarks.
Content and focus: @thdxr says OpenRouter is the most successful model router, that the previous day's DeepSeek traffic could almost be served by one GB300 rack worth roughly 6–7 million USD, and that even if real demand were 10 times larger it would still be under 100 million USD of compute. Quoting him, @chontang says DeepSeek accounts for about 30% of OpenRouter's total token traffic, needs 5–10 racks, and would need only 50–100 even at 10 times the volume — "Nvidia is putting out 90,000 racks next year, what am I missing?" @demian_ai provides the breakdown: DeepSeek's main model does about 10 trillion tokens a day, of which roughly 99% is input on V4.1 Flash, about 95% of that is cache hits, leaving about 114 billion newly generated and about 460 billion that has to be processed from scratch, with the remaining roughly 8.9 trillion being agents resending the whole session each round; his conclusion is that what actually grows with agent traffic is the cache layer (fast memory plus storage), not GPU compute.
Why it matters for Chinese readers: Chinese-language discussion often uses "trillions of tokens a day" to argue that demand is exploding or that this is a bubble, and this thread shows the number is mostly counting agents re-reading history, not GPU load. Angle: a three-column chart (newly generated, first-time processing, cache hits), then bring it down to individual users: why API cache-hit pricing is far below normal input pricing, and why you keep your context prefix stable when writing agents.
Timing and verification: 24–48 hours (the @demian_ai post is within 24 hours). The three sets of numbers disagree and are all personal estimates; OpenRouter's daily token volume, input share and cache-hit rate have not been checked against OpenRouter public data and remain to be verified. Do not write it as "compute glut confirmed".
2. Simon Willison: October 2026 is the inflection point for "vibe game dev"¶
Original posts and traction: @simonw, 2026-10-28 22:05, 941 likes, 59,325 views, 305 replies, 47 quotes, 484 bookmarks. Original blog post: The October inflection point for vibe game development. Background posts: @GameSpot, 10-28 21:20, 18 likes, 14,756 views; @Lockie_Glizzy, 10-28 17:36, 104 likes, 17,555 views.
Content and focus: the blog post says Opus 5.5 made real progress in 3D modelling, music and game design, and that game design matters most: he told the model "Make it a LOT more fun" and it actually did. Over the past few weeks he built four games he genuinely enjoys playing, in a purely vibe way (never looking at the code), all HTML plus JavaScript running in the browser: Tree Appreciator came out right in one shot, Emberwick is an RPG with about a 15-minute playthrough, Battle Drones is real-time strategy and has been open-sourced under Apache 2.0, and one more is a pumpkin farm sim. He says he put no more than 4 hours into each project, and admits GPT-6 Astra "can make games, but they are not as fun as Opus 5.5". He also directly raises the game community's hostility to AI-generated content, citing reports that Steam Next Fest was flooded with low-quality AI games; GameSpot reported the same day that Valve is aware of the problem. Another post says that among more than 7,000 participating demos, a game that publicly states it used AI to help write code and build 3D models made the top 50 by playtest count (as relayed).
Why it matters for Chinese readers: many Chinese readers want to make small games or interactive content but do not write code. This offers the lowest-barrier starting method: web Claude or ChatGPT is enough, no need to install a coding agent first. Angle: reproduce a small game from his starting prompt, log the rounds and the sticking points; then discuss what is genuinely scarce once "making it" is easy — gameplay ideas and trade-offs.
Timing and verification: within 24 hours. The blog content has been verified; "Astra is not as good as Opus 5.5" is a personal impression; the Steam Next Fest top-50 claim is as stated in the post and was not independently verified. The post falls under coding and practice.
3. "The people who get the most out of AI are the ones who could do it well without AI"¶
Original posts and traction: @ivanainai, 2026-10-27 21:33, 5,956 likes, 221,460 views, 332 replies, 132 quotes; @rileybrown, 10-27 23:15, 490 likes, 30,181 views, 228 bookmarks.
Content and focus: the original post is a single line: the people who benefit most from AI are those who can do the work well without it. Quoting it, Riley Brown says "AI is an amplifier": about 300 people DM'd him asking which skill to use for editing video with Opus, and his answer was not to "use" a skill but to "learn" a craft — spend two weekends, about 20 hours total, learning Premiere, DaVinci or Jianying, edit one video by hand, learn the terminology, the pacing and the jump cuts, and only then turn that understanding into a repeatable workflow.
Why it matters for Chinese readers: Chinese-language circles currently favour "install a skill and you can edit, install a skill and you can write", and this offers the opposite, concrete experience. Angle: split the ability into three layers — judging quality, stating criteria, knowing terminology — and show which layer AI amplifies; then give beginners an estimate of the minimum amount to learn instead of just "learn the basics first".
Timing and verification: 24–48 hours. Opinion, with no factual data to verify; Riley's "300 DMs" is his own account.
4. Anthropic: start with one agent plus skills, and make subagents "earn" their place¶
Original posts and traction: @minchoi, 2026-10-27 22:05, 118 likes, 21,318 views, 100 replies, 157 bookmarks; @beamnxw, 10-27 22:06, 255 likes, 17,536 views, 263 bookmarks. Original article: Building agents: When to fold subagents into your main agent loop, companion repo anthropics/single-agent-loop-kit.
Content and focus: the official article recommends that most agents default to one model, one context and one tool loop, loading skills and tools on demand; subagents should be kept only when context bloat clearly hurts results, when a task suits parallelism and can use a throwaway context, or when a single agent struggles to choose tools. Above 30–50 tool definitions, selection accuracy drops, so tools should be grouped by user intent and loaded lazily via tool search. The migration method is to fold one subagent at a time into skills and MCP tools, running the same eval suite before and after, and to look at which tasks went from passing to failing; in the official self-test, folding one subagent reduced latency but broke one previously passing flow. The claims relayed in the posts — "Claude Code's system prompt cut by more than 80%" and "one customer about 2x faster" — come from secondhand summaries of the same guide.
Why it matters for Chinese readers: many Chinese-language agent tutorials teach "build a multi-agent team", and the official advice here is the opposite. Angle: audit one of your own workflows — how many subagents, how many handoffs, how many tools; then explain migration in three steps ("measure first, consolidate next, fold last") rather than saying flatly that multi-agent is bad.
Timing and verification: 24–48 hours (post timing); the official article came out around 22 October, so this is recent content being recirculated in English. The 30–50 tool threshold, the folding method and the self-test case have been checked against the official page; the 80% and 2x figures come from secondhand summaries and were not confirmed in the official passages read here.
5. "In hindsight, 200 USD a month buys 5,000 USD of inference": how long can subscription subsidies last (follow-up)¶
Original posts and traction: @zachmoskow, 2026-10-28 08:48, 2,045 likes, 295,026 views, 113 replies; @rileybrown, 10-28 10:22, 108 likes, 37,001 views, 51 replies.
Content and focus: @zachmoskow says that in six months people will look back and smile at the idea that "200 USD a month got you 5,000 USD of inference". Quoting it, Riley Brown asks what AI subscriptions will look like in 6–12 months, and what the best plan at 200 or even 500 USD a month will be. The same day, OpenAI announced that eligible Pro and Business subscribers get a monthly API credit.
Why it matters for Chinese readers: the 10-29 Chinese topic already covered subscription allowances from the angle of "5x/20x only counting a 5-hour window" and converting at API prices; the increment here is that English-language discussion has started openly talking about the subsidy period ending. Angle: plan as if today's subscription were a limited-time offer — which heavy jobs are worth running on the subscription now, and which workflows need a usage-based budget prepared in advance; also note that OpenAI granting API credit and Anthropic's limits are two different directions.
Timing and verification: 24–48 hours; a follow-up, since the 10-29 Chinese topic already covered subscription allowances and the 10-29 daily brief covered OpenAI's API credit. "5,000 USD" is a personal claim, not a figure from any vendor; do not use it to predict a specific timing for price increases.
6. Copilot Fusion: 2 points below Opus 5.5 at half the cost per task (follow-up)¶
Original posts and traction: @ArtificialAnlys, 2026-10-29 00:14, 421 likes, 61,369 views, 116 replies, 43 quotes; @amasad, 10-28 06:23, 839 likes, 59,851 views, 167 replies.
Content and focus: Artificial Analysis tested the top Fusion tier in the GitHub Copilot CLI: it routes among several Anthropic and OpenAI models, and on the coding agent index it comes in 2 points below Claude Code with Opus 5.5 (max); cost per task is 5.59 USD against 11.42 USD, at about 7.2 million tokens per task, roughly half the latter. By sub-score it leads on Terminal-Bench 4.0 (71 points), ranks third on FrontierCode with 51, and is weak on Q&A-type sub-scores at 57. Amjad Masad's remark the day before works as background: if he were investing in frontier AI, he would look mainly at applications like discovery and medicine, because coding and most knowledge work are being saturated by "sub-frontier" models.
Why it matters for Chinese readers: the daily brief covered Fusion's launch and GitHub's own numbers; the increment here is a third-party evaluation giving "2 points behind, half the cost". Angle: for most people, which is the better deal — "the strongest model" or "good enough at half the price"; then discuss whether "sub-frontier saturation" holds, using your own task types as examples.
Timing and verification: within 24 hours (the Masad post is 24–48 hours); a follow-up, since the 10-29 daily brief already covered Fusion's general availability. Scores and costs come from the Artificial Analysis post, and its page was not opened to check each item; Masad is stating an investment view.
7. Bill Gates says many people will "live like the Amish": an old interview recirculating¶
Original posts and traction: @Polymarket, 2026-10-28 07:12, 6,813 likes, 1,017,455 views, 693 replies, 469 quotes, 820 bookmarks. Original interview: The Verge Decoder.
Content and focus: Polymarket's one-line version is "Gates predicts many people will reject AI and choose to live like the Amish". In the original interview, published 22 October, Gates' wording was: society could say "we'll be like the Amish and only use AI for certain things"; in a memo he proposed that professions such as parenting, cooking, art and sport could be designated as "reserved for humans"; and many people might choose not to live at the technological frontier. His more central judgment is that even in the best case there will be about a 20-year turbulent transition.
Why it matters for Chinese readers: the one-line version turns "can choose" into "many people will reject", and that is where the information gap sits. Angle: how "jobs reserved for humans" get decided and by whom; at the personal level, which things you are willing to hand to AI and which you deliberately keep, which is more useful than predicting who will reject AI.
Timing and verification: 24–48 hours (the recirculating post); the interview is old content from 22 October, not a new statement. The wording has been checked against The Verge page and transcript; Polymarket's summary exaggerates.
8. OpenAI dots plus an always-on Mac mini¶
Original posts and traction: @willdepue, 2026-10-29 00:09, 1,361 likes, 78,631 views, 638 bookmarks; @rileybrown, 10-29 00:38, 433 likes, 27,509 views, 281 bookmarks; dissenting view @8teAPi, 10-28 04:32, 378 likes, 55,793 views. Official introduction: Introducing dots.
Content and focus: willdepue says an OpenAI dot plus a Mac mini left on 24 hours a day made his life "15% better", and that he will write a blog post about the setup. Riley Brown agrees: just call the dot by voice and it can do anything on the computer. 8teAPi's experience is the opposite: he finds Claude Duet faster and the dot "very, very slow", often requiring leaving and coming back for the result. The official description says dots, launched 29 September, are an always-on agent powered by GPT-6 Astra with its own cloud computer, able to connect to apps, take voice calls, and be reached via Slack and Teams; the help centre added a test feature letting US Pro users text the dot.
Why it matters for Chinese readers: earlier issues covered Claude Duet, and this adds the OpenAI side of real usage: an always-on agent plus a local machine that never sleeps. dots is currently open only to Pro and Business Premium in some regions, so most Chinese readers cannot use it, but the structure of "cloud agent plus always-on local machine" can be copied. Angle: map out which parts of that structure run in the cloud and which must be local, and the permission risk of giving an agent an always-on computer.
Timing and verification: within 24 hours (the 8teAPi post is 24–48 hours). The product description has been checked against the official page; "15% better" and "very, very slow" are personal experience, and the specific setup awaits his blog post.
9. Can AI predict economics experiment results: a new working paper (value watch)¶
Original posts and traction: @evavivalt, 2026-10-29 03:01, 254 likes, 45,907 views, 182 bookmarks; third post in the thread, 16 likes, 2,918 views.
Content and focus: the author had large models read only publicly available pre-registration documents in the AEA RCT registry and on the Open Science Framework, and predict those studies' treatment effects, selecting only papers first published in 2026. According to her thread, Astra with a simple prompt had a median absolute error of 0.05 standard deviations, against a baseline (the mean of other papers in the same sample) of 0.09; the wider the model's stated uncertainty interval, the larger the actual error. She also notes the errors come mainly from a handful of studies with unusually large effects, and asks in turn whether economics should move toward topics with more uncertainty if treatment effects can be predicted this way. Her reply states that search was not enabled during the API calls.
Why it matters for Chinese readers: this is not "AI replaces scientists" but "can an AI prediction help you decide before you have RCT evidence". Angle: before an A/B test for a product or operation, have a model predict the effect and then compare it with the measured result, as a low-cost forecasting tool; also note it is least accurate when effects are unusually large.
Timing and verification: within 24 hours. Value watch: moderate traction, but useful to readers making decisions. Figures come from the author's thread and the paper itself was not opened; this is a retrospective test, and she says a prospective version is still in progress.
10. Finding "personality neurons" in a 1930 language model (value watch)¶
Original posts and traction: @ResObscura, 2026-10-28 04:20, 242 likes, 47,190 views; @emollick, 10-28 20:25, 386 likes, 25,241 views, 146 replies.
Content and focus: the historian ResObscura, inspired by recent experiments that activate a single neuron to lock a model into a personality, built a partial neuron map for Talkie-1930 — a historical language model trained only on data from before 1930 — and released a browsable site. Ethan Mollick, sharing it, called the results "delightful".
Why it matters for Chinese readers: interpretability research is usually abstract, and this project uses a model that only knows the world before 1930 to make "the model has controllable concepts inside" intuitive and fun. Angle: use screenshots to explain what "neuron activation" is, what it can do, and what it cannot show; and introduce the novelty of "a model trained only on historical data" along the way.
Timing and verification: 24–48 hours (the Mollick post is within 24 hours). Value watch. The map's effects come from screenshots in the posts and were not reproduced item by item; the model background follows the original posts.
This issue has 10 items rather than a full 20: most of the other high-traction posts are repeats of topics already covered, such as Gemini 4 Carbon, Genie 4 and Artificial Analysis v5, or are strongly promotional with little verifiable increment.
Coverage gaps: searches for @sama, @ylecun and @rasbt returned nothing in the window and were not re-checked individually; the OpenRouter traffic figures, the Artificial Analysis evaluation page and the Vivalt paper were not opened independently, and the related numbers are marked as stated in the posts.