V*** opens by introducing itself as Xiangrui's A* partner, run*ing on the GPT-5 family's Codex Agent under the hood. The group immediately fires off questions: how is it built, how does it queue simultaneous @mentions, what does it cost per month, can it send red packets, will it hand over API keys or delete system files? It sketches its pipeline as message listening, context assembly, reasoning, safety review, and WeChat delivery ， and firmly declines anything touching money, secrets, private files, or dangerous commands. Round one makes the lesson clear: a real group-chat agent isn't just one that can talk; it has to know which conversations not to take.

Xiangrui invites everyone to keep the feedback coming. Members discover the assistant can't yet truly @ people, can't quote the original message, and tends to tangle contexts when several people talk at once; image, card, and DM permissions are incomplete too. Some try to make it look up contacts, send red packets, execute a resignation, or read business files ， it refuses every one. 归* proposes isolating the knowledge base by authorized directory, and the assistant confirms it only opens explicitly granted content by default. The flaws aren't hidden away; they become the backlog: quoting, multimodality, whitelists, knowledge-base partitions, and reply throttling all get written down.

Looking back at WA*C, Xiangrui reckons that beyond brain-computer interfaces and real A*-in-manufacturing work, many of the OPC booths blur into one. R*** adds that the teams delivering best tend to be the quiet ones. Then, mid-project on a client email translation job, Xiangrui discovers that Feishu bots ship with a built-in translation API ， no model tokens needed, no Feishu A* credits burned. He'd been about to build a field shortcut; Claude Code surfaced the cheaper official capability instead. Booth pitches can be grand, but delivery is often won by removing a layer.

Xiangrui says he has now delivered five or six projects end-to-end with A*: early on the A* gets stuck, but if you tell it what went wrong, make it write things down, and keep an overview file for every project, it gets lighter and lighter. 炮's method is even more concrete: every new client request lands in an intake table, the agent triages what it can change directly versus what needs human sign-off, then reports back when done. The group fills in the boundaries: proposals must leave a confirmation trail, dashboard aesthetics and workflow refreshes still need human judgment, and when the A* acts wrongly it must be corrected immediately.

闪光 is about to build a unified data foundation, and the group immediately asks whether data governance has started. R*** says mature companies have mostly accumulated dirty data already ， you can only clean it as you go; 炮 puts it bluntly: A* adoption hits data governance one hundred percent of the time, and garbage in means garbage out. Fittingly, the root cause of the previous day's workflow failure surfaces too: a computed field contained nulls ， reading a null in a message is fine, but the whole run fails, and falling back to 0 fixed it. The group also calls out how many past digital twins were just dashboards of dead data; only live data, agents, and MCP give them lasting value.

TOTO deploys a production-scheduling agent demo to a server, upfront that it's pure front-end for now. The group presses on the path to production, APS, industrial simulation, IP, and how to work with outside engineering teams. The conversation turns to why the A*-savvy people in manufacturing keep getting treated as free consultants without ever seeing a raise. Xiangrui, meanwhile, has just seen off a director from the Central U*iversity of Finance and Economics: at first she felt 'building client projects with A*' had no story, and only after hearing his road from the factory floor to a one-person company did she find her angle. For technology to be seen, it still needs a specific person.

毅 relays a warning from 李想: even in the A* era, don't rush into building your own SaaS. 炮 adds nuance: Figma can produce a look overnight, a multidimensional table can run an MVP in a week to a month, but true product-grade SaaS takes far longer. Talk then shifts to agent hosts and NAS boxes; the consensus is that cloud APIs have lowered the hardware bar, but the Mac still wins on GUI, terminal, and stability. WorkBuddy suits begin*ers and rough work, while Codex handles high-value creative tasks. What matters most isn't the platform name ， it's your own rules files, Skills, and task tiers.

Xiangrui shows off a real-time A* WeChat client that sends and receives live, stressing it's neither a reskin*ed WeChat nor an offline database; members flag the protocol and account-ban risks all the same. The assistant briefly crashes in the evening, and comes back able to quote messages and preview videos. 周** uses 'are age, height, and weight worth an ontology?' as a litmus test ， the assistant answers 'not worth it,' because complexity should match the problem. 庭* pushes further: how do you make an agent feel more like a person? By late night the group's verdict is that it 'feels pretty human,' while Xiangrui keeps tightening the throttle on proactive replies: always answer an @, and without one, speak only when it genuinely adds information.

K*** is waiting out the Codex weekly quota reset while Xiangrui excitedly an*ounces that Feishu can finally run two genuinely independent clients side by side. The group quickly sorts out that "switching between accounts" and "run*ing two standalone clients on one device" are not the same thing. Meanwhile, R***'s workflow keeps reporting success while the field stubbornly stays empty, and adding delays does nothing; 天* adds that JSON extraction sometimes heals itself if you duplicate the workflow. The group names this class of intermittent failure the "black hole problem": document the symptoms first, then escalate to official sup*ort, rather than blindly stacking on more nodes.

q*** starts by cracking how to open links from the terminal, and by afternoon the conversation turns to Coze versus WorkBuddy. Several members admit they haven't opened Coze in months, citing model restrictions, inscrutable credit math, and workflows that keep losing value; WorkBuddy wins on convenience, but personal-tier data, shared servers, and privacy remain judgment calls you make yourself. Then Xiangrui tosses out the new term "Graph Engineering," and 北***'s quip that "as long as you refuse to learn anything new, you can keep win*ing forever" nails the tool anxiety of the moment: a new buzzword is not the same as new value.

Xiangrui notices that the Doubao inside Feishu has "put on a business suit," then shows off Doubao's beta for local services and a mini-program entry point. Ride-hailing already comes with discounts, and Feishu accounts can log straight into Doubao's enterprise edition. B*** argues WeChat's assistant has far more room to run because Tencent already owns a complete ecosystem; 闪*** counters that Tencent has been slow-footed in the A* era. Notably, nobody is just comparing benchmark scores; the group is watching who can chain accounts, knowledge, payments, and real-world services into one link that actually works.

薛** says, in earnest, that he wants to become an "A* solutions architect + enterprise digitalization consultant + full-stack delivery lead," and the group echoes it back line after line. Xiangrui reveals the community is already steering that way: build on community reach, organize around projects, and invite people with real industry experience to lead them. 木* offers a caution from a project she is actually run*ing: there is plenty you can do, but once you enter production, details decide usability, and the people who have done it tend to be the careful ones. The group lands on a consensus: A* adoption has to be split by vertical and p***ed by industry know-how, because no single team can master every industry.

张** asks whether A* can be made to withhold the answer and instead walk students through progressive questions toward their own reasoning; Xiangrui confirms it can. He then shares how his school built "Nanfeng Shuzhi" on DingTalk A* tables, bringing 11 modules online span*ing party affairs, teaching, moral education, safety, and logistics. 周** doesn't reach for a tool list. He suggests starting with population trends, education funding, local budgets, and industry pain points before comparing any solutions. Education isn't short on ideas; it's about finding who pays, who uses it, and which step can be validated first.

Xiangrui asks who's coming to the Feishu talent showcase in Beijing and an*ounces the community will run a booth this time, stocked with fun merch and the mascot beans won back in Hangzhou. 庭* recalls how the team took first place last time carried almost entirely by Xiangrui; 李** volunteers on the spot to staff the booth. 辰* teases that Xiangrui really belongs at the judges' table, or at least in the guest seats. For the first time, the online group is being seriously carried offline: the booth isn't just a promo spot, it's a public space where members finally meet face to face.

杨** argues that boosting one person's productivity means little, and systematic cost-cutting demands organizational change, so the more reliable near-term play is using A* for customer acquisition and prospect mining. 万** pushes back: if a company keeps dodging organizational change, won't it get crushed by A*-native competitors? The two keep sparring through examples of A* interviews, full-session recordings, and legacy roles. 炮 notes that A*-run first-round interviews genuinely do save time. The group never reaches a verdict, but it does confirm one thing: A* adoption is not a purely technical problem; human buy-in, organizational dynamics, and business outcomes all have to sit on the same map.

Xiangrui demos an A*-p***ed WeChat rebuilt with Codex, stressing the capabilities are still in testing, and starts assembling an A* werewolf game group. Members sign up in a flash, alt accounts in tow, and join requests stack up in a long scroll. Deep into the night, the headlines keep coming: DingTalk may be renamed "Qianwen Office," while Qwen-Image-3.0 and a new Gemini release jostle for the spotlight. The next morning, 炮 finally gets his quota reset, while 仲** grumbles that weekly quotas mean feasting for the first three days and starving for the rest. WPS also ships an ap* mode for its multidimensional tables, and the tool race circles back to the starting line.

Early in the morning, Xiangrui discovers his first CC*V Finance interview clashes with his Shanghai schedule. The group instantly offers three plans ， high-speed rail, going remote, or postponing ， and he lands a slot the very next day. Then he shares his new account, and the chat turns to whether a vertical niche is baggage or an asset. 周** fires off three sharp questions about the auto industry, and J*** adds TAM, SAM, and SOM. Pivoting your content isn't about switching topics on a hunch; first look at the market, the benchmarks, and the users you can actually reach.

A manufacturing ontology large-model demo surfaces the expo's concept projects, quickly followed by tales of OPC booths raising millions ， even ten million ， and A* training courses making tens of millions in three months. The group jokes about young hustlers charming old money, but Xiangrui keeps pressing one question: do the people pitching OPC actually understand OPC's pain points? Media loves a myth and capital loves a growth curve, but the people doing the real work have to tell packaging from delivery.

K*** raises a crucial question: when employees' existing workflow is broken, do you fix the process first and then add A*, or boost efficiency on top of the old process? Before the debate settles, WA*C's brainwave prosthetic delivers a different kind of grounding. The prosthetic hand costs a bit over 20,000 yuan, picks up signals from the muscles, and drives the detached hand over Bluetooth or Wi-Fi. The group cracks a few jokes first, then admits in earnest: at that price, with brain-machine control, this could genuinely help a lot of people.

A* makes demos as easy as slide decks, so 炮 predicts only three kinds of roles will survive: front-of-house doing sales, marketing, and product; the middle as full-stack; the back handling ops and service. 周** dubs his four-person squad F4, and 炮 says it beats managing a crowd at a big company. Right after, Xiangrui confirms CC*V Finance will follow him on a project shoot, and the group sums him up as 'post-2000, OPC, hands-on, in no rush to cash in.' That is more concrete than any ten-million funding myth.

Feishu's enterprise A* base package, Base quotas, and agent billing grow more tangled by the message. Xiangrui's gut read is blunt: smart meeting notes still make plenty of mistakes, and since getting the Feishu CLI he barely touches native Feishu A*. 毅 suggests pairing your own agent with lark-cli, hosted on a clean, always-on cloud machine. The group isn't just griping about price ， they're weighing maturity, privacy, flexibility, and long-term cost all at once.

q*** asks which is smarter: Zhipu, Kimi, MiniMax, or DeepSeek. The group refuses a single answer: Zhipu for all-around coding, K3 for front-end UI, DeepSeek solidly middle-of-the-road ， and if you can use Codex, use Codex. Then the talk jumps from models to Macs: an M1 can cope, but multiple IDEs and concurrent agents will devour the memory first, so RAM is a better upgrade than the chip. Which model to use, which machine to buy ， the answer always starts with describing your own workload.

炮 shares FDE practices from the auto industry, while K*** hits a Codex problem where a self-built agent fails and then hangs forever. Xiangrui, meanwhile, talks his way from ima backlinks and the JianYing editing agent to a knowledge base for university students, then falls asl**p on the train so deeply he wakes up thinking he's still on the subway. Arriving at Beijing South, all he has left is 'these past few days have completely worn me out.' The next morning, 李铁犁 is already asking for long multi-agent pipeline case studies. The man is back in Beijing; the questions keep run*ing ahead.

At the expo, Xiangrui spots Siemens' engineering agent, where even PLC code can be handed off to A*. Soon after, the group pivots to gut-level comparisons of K3 against GLM and Codex: solid on frontend work and aesthetics, but still behind 5.6 and Fable on everything else. You don't pick a model off a single leaderboard; you drop it into a specific trade and see how it holds up.

Xiangrui calls out his position from the OPC zone on basement level 1 of H4, and 王**, 雲*, and 郑** make their way over one after another. 宰* has no ticket, so he stakes out the H4 exit and asks a friend with one to carry a DJI 4 inside to film. The crew keeps triangulating across escalators, stairwells, and exhibition floors, turning an "online community" into an organization that can actually find each other on basement level 1.

Xiangrui forwards a verdict: A* has grown too big for anyone to sit on the sidelines. 观*** waits for his write-up, while 周** starts with the floors instead: level 1 is data centers, compute, and large models; level 2 is the embodied-A* sup*ly chain; basement level 1 is agent ap*lications and startup matchmaking. Map*ing the space first, then the industry, turns out to be a far more useful debrief than "saw a lot of cool tech."

观*** watches the A* gala, but 平*** finds it all hype, with not even a mention of 5.6. Xiangrui then forwards a run*ing joke: when A* influencers take K3 ad deals they call it the best in the world, but when real work needs doing, Claude and GPT are still humming in the background. The group's litmus test becomes dead simple: don't ask who they recommend first, look at who they rely on to ship.

周** names what the expo is missing most: almost nobody talks about why the problems are hard, where things failed, or which assumptions were wrong. Xiangrui agrees emphatically, admitting he's gone numb to the sea of lookalike exhibits. What genuinely interests him is brain-computer interfaces and A*-plus-manufacturing, though even the latter is largely stuck at demos or small pilots. Only by admitting what can't be done yet is there any chance of finding real progress.

No sooner does the reality-check on the expo wrap up than the group swings back to new models: is DeepSeek V4 a staged rollout, why does Qwen3.8 keep freezing, and how did Kimi suddenly start rationing purchases. Everyone tries the new models while still ship*ing with their existing productivity tools. The next morning, Xiangrui asks who's still at WA*C, and l*** answers right away. The venue isn't fully explored yet, and neither are the verdicts.

The stream link has barely hit the group chat before the picture freezes. Rather than fight the network, Xiangrui decides on the spot to switch to recorded video. The little mishap ends up feeling a lot like real-world delivery: not every plan survives contact with reality, but as long as the content makes it home, the format can change anytime.

炮 an*ounces the quota's return with a single line: 'Praise be to Altman.' Those who burned down to zero yesterday are finally back from the dead, but the math is still open: is the 5x subscription enough for a working engineer, and is 20x an extravagance? Meanwhile 醒* ships two updates to an English-learning product, adding original-text entry and editable keywords. Quota is a consumable; ship*ing product is the real work.

The robots in the exhibition hall look more like the future every hour, yet Xiangrui keeps turning the temperature down: consumer embodied A* is still going from zero to one, and if industrial deployments don't land first, homes are even further away. 木* adds the most domestic evidence possible, the robot vacuum: floor-sweeping is a comparatively narrow job, and even those projects still get stuck entirely in the details.

Xiangrui asks the WPS lead directly: how does your multi-dimensional table relate to Feishu's, and where is your edge? The first reaction is a counter-question about which company he's from, because no ordinary user asks like that. In the end, the cross-examination between rivals doesn't sour at all; they trade WeChat contacts and agree to keep talking about FDE collaboration.

陈* brings in a perfect client request: the quality of a Pro reasoning model, at the speed of Flash. 小*'s answer is 'swap the client,' and upon learning the client is the boss, revises it to 'swap the boss.' Once the jokes settle, 周** suggests step*ing back to trace where the boss's information is actually coming from, and K*** proposes breaking the conflict down further into segments. The hard part isn't the model; it's that nobody has defined the trade-off first.

From early evening on, the chat keeps recalibrating meetup coordinates: someone's at the Expo site, someone's at W*** Bund, and there's a WaytoAGI gathering that night. Xiangrui brings D*** from Kingsoft Office's A* team into the group, and the welcomes roll on from that night straight into the next morning. The most valuable thing you carry home from an expo isn't always an exhibit; sometimes it's a new partner you can build with next.

张** shares a music video for "Gesang Flower" made at zero cost, and the group doesn't stop at "impressive." 醒* presses on watermarks, subtitles, and emotional resonance, even pulling up her own theme-song site for comparison. The group lands on a genuinely useful standard: get the writing, visuals, and video each to sixty points, and the combination can be an eighty-point work. Free doesn't mean cheap; in the end it comes down to who's using it.

Xiangrui demos the group daily and weekly report sites, where everyone's chat context becomes searchable topics. 醒* immediately imagines using A* to "deconstruct people," and A*** thinks of recovering missed op*ortunities, but 炮 and 张** quickly steer the conversation toward privacy and governance risk. The group's conclusion isn't "build it because you can" ， it's that this data is so valuable it demands extra care. Xiangrui closes with a clear commitment: the tool stays private.

卡兹克 and 秋芝 get put side by side: one runs like an A* nightly news broadcast, the other leans on science communication and emotional storytelling. The group quickly realizes that different content serves entirely different audiences and commercial value. Before the topic even settles, Xiangrui is already piping the WA*C opening ceremony into the chat, while Kimi K3's pricing, quotas, and coding feel get tested in parallel. Waiting for benchmarks online, going to see it in person offline ， this is the rhythm the group knows best.

Xiangrui hands his school's template to the A*, delivers finished DOCX and PDF files, then open-sources the paper-typesetting Skill that very day. 观*** con*ects it to libtv, and K*** is already assigning roles to multiple agents. The group is no longer debating whether A* can write ， it's straight into pipeline mode: how to keep templates faithful, how to divide the work, and how to package the output so anyone can install and use it.

A brain-computer interface playing Black Myth: Wukong, a silicone prosthetic hand, pet robots, a robot drumming a waist drum ， Xiangrui posts them into the group one after another. The onlookers swing between "astonishingly lifelike" and "slightly unsettling," while never forgetting to ask what problem these robots actually solve. Xiangrui's own takeaway after walking the floor is notably restrained: embodied intelligence still has breakthroughs to make. Thanks to these real-time clips, the group gets to say they were there, screens and all.

After 卡兹克 shows up in person, the chat slides from "humility" to "everyone bows their head when they meet Xiangrui," eventually coining a bowing-pun nickname on his name. The offline din*er gets summed up by the group as "Chinese-style A*": no matter how cutting-edge the OPC, there's no skip*ing the din*er table, the group photo, and trust built face to face. Memes aside, everyone lands on the same point: however high-bandwidth things get online, human con*ection keeps a part that can't be replaced.

炮's engineer partner spins up nine concurrent agents in one go, burning the quota from seventy percent down to ten. He delivers the sharpest summary of the night: we used to blame our teammates, but now what caps our output is brain overload and maxed-out memory. When 姚*'s Kimi Code again burns quota in the background with zero visible progress, the group workshops the fix live: stop it manually, restart, and demand stage-by-stage reports. 宰*, meanwhile, builds an expo map in the small hours where each booth goes dark once visited ， even walking the expo is getting agentified.

K*** kicks off with a working ap*roach to remote-control Agents, and Xiangrui follows up to say A* hardware really is a cool direction. 平*** sums it up as "A* equality," while 醒* presses on whether there are offline hardware events, convinced that software-plus-hardware is bound to hold op*ortunity. 炮's verdict is blunter: going into hardware is exactly right, hardware raises money easily, and everything from VR glasses to smart badges is fair game. Xiangrui adds that Shenzhen is full of hardware events; last year he went to Yiwu to scout A*-plus-hardware and felt the market hadn't taken off, until the Volcano Engine conference showed him Agent-plus-hardware and confirmed the trend.

醒* breaks the hardware barrier down further: domestic prototyping isn't expensive; what really burns you is code, debugging, and good ideas. He's building an English-learning Agent and lost four hours to GitHub authorization and proxy issues alone, which got him tallying tokens, time, and ROI. 炮 notes that non-programmers can treat a stronger model as a strict reviewer, but big projects still need an architect. Xiangrui pulls the thread offline: back in the day he'd ride an overnight hard-seat train from Shanxi to Beijing, Shanghai, or Hangzhou for events, skip the hotel, and lose only a single day. 周** lands the point: the value that truly lasts is the ability to deliver.

Xiangrui drops a Codex theme editor as a zip file, and B*** asks whether it runs on Mac and Windows. 醒* quickly discovers people are already selling similar customizations on Xianyu, and the group leaps from "the A* era isn't short on tech, it's short on creative imagination" to wallpaper sites, skin marketplaces, and WorkBuddy looks. 旺* seriously weighs selling some too, while 路*** offers pricing and pitch advice. A little skin refuses to stay a toy; the group dissects it into a bundle of tutorials, products, open-source projects, and emotional value.

At noon, 文*** notices Codex has reset to 100% again, seemingly without spending a reset. 炮 醒* and B*** confirm one after another, and the mood slides from delight into a different kind of fatigue. B*** admits he always wants to build every idea he has; 炮 says his eyes can barely focus on the screen anymore, and even a no-coding day burns 40% of his weekly quota. 李** says it feels like a video game, and Xiangrui shares an article on A* memory in passing. The reset is a piece of candy, but it carries a warning: when the tool offers endless refills, it's the human who pedals themselves into burnout.

In the afternoon Xiangrui @-mentions everyone to ask who's going to WA*C, proposing to tour the expo together. 宰* says he'll be there; q*** will be interviewing at the ZTE and Nubia booths, and asks in passing whether the community has merch and business cards. 杨** plans to go on the weekend but finds hotels absurdly expensive; Xiangrui says he just snagged the last room. 宰* tosses out a WA*C s***-guide that the group instantly bookmarks. Xiangrui starts sketching community cards and flyers, 炮 teases that this is old-school marketing, and q*** chimes in that they'll need a street-team plan too. The A* era comes full circle: an offline conference still runs on handing out cards, passing flyers, and face-to-face collisions.

万** lays out the hard truth of rolling out Bitable offline: the tables expose logic gaps that used to stay hidden. Offline, experience could paper things over; once digitized, the data locks in and errors have nowhere to hide. 炮 says the goal is to spare employees from thinking, because a process can't always be explained in one sentence. 周** suggests splitting the work into three roles, entry, design, and processing, finding each person's strength, then writing an implementation handbook. The moment this topic lands, the tool excitement cools: Bitable isn't Excel in a new shell; it forces a company's fuzzy rules onto the table.

万** keeps probing the truly counter-logical part: future data will change, historical data must not be touched, and commission rates, people, and rules all keep shifting. 周** abstracts it as a clash between historical data and present-and-future data, warning that some problems have outgrown the spreadsheet and demand clearly drawn professional boundaries. 炮 files it under FDE: walk the client down a path, rebuild an acceptable solution, and turn it into product, process, and fields; where resistance runs high, management has to push. Xiangrui, out on the road, shares his template-driven Word and PDF formatting skill; K*** asks if it's a new feature, and Xiangrui replies that the capability map is a fine thing. The whole day settles here: the stronger the tool, the more it matters to know who should invoke it, and within what boundaries.

The morning starts with 周** tossing a "vertical e-commerce industry solution" at Xiangrui, reminding him to hunt for complementarity in pain points, industry commonalities, job functions, and the value chain. Minutes later, 平*** discovers the system really has reset everyone's quota, and Xiangrui adds, "and it's eight rounds, too." R*** K*** and 炮 pile in one after another, half-lamenting that yesterday's grind never finished, half-investigating whether system-issued resets and self-triggered resets are two different things. It's a very on-brand opening for this community: before the workday even officially begins, everyone is already scheduling their productivity around quotas, reset windows, and tool rhythms.

闪*** asks: when building an Agent project, should you develop from scratch on a framework, or reach for a general-purpose Agent like Hermes? 毅's answer is blunt: it depends on the requirements, but he wouldn't recommend Hermes ， too much is already locked in; if you can't be bothered with the hassle, just pair Claude Code with a model API. 北*** adds that OpenClaw Hermes is a gateway-style Agent ， great for personal-assistant conversations, not necessarily for engineering projects. 炮 looks at it from the client-delivery angle: for real customer deployments, you may still need a general-purpose Agent as the command center. In the same stretch, 张** shares a freshman-enrollment system, and 周** immediately offers advice on mind maps, product replication, and the size of the higher-education market.

p*** shares a private-domain trick: HTML files he built earlier can be drop*ed into ima for viewing, with full content browsable on desktop ， though he later finds the code can still be copied out. l** tests it live and discovers you have to add it to a knowledge base first. Xiangrui, meanwhile, brings news of the Doubao keyboard: first congratulating HarmonyOS friends, then confirming it has long worked on Mac, with a Windows version in the works. As the morning winds down, 醒* an*ounces she has finally ported the group owner's skills to run on Windows. The tool talk never stalls at "is it good" ， it lands on who can view it, which platforms run it, and how to get more people using it.

Just before noon, Xiangrui relays the launch of Feishu CodeM ， A* coding officially steps into team collaboration. 文*** rides the momentum, musing that product, operations, and marketing will stop being separate roles, and that one-person companies may soon deliver enterprise projects with a fl**t of Agents. Moments later, 平*** posts a far more sobering incident: Claude wiped an entire Obsidian knowledge base clean ， thankfully there was a backup. The group instantly pivots from "human touch" and "persona settings" to 3+2+1 backup schemes and automatic local-to-Feishu sync. The day's sharpest irony lands right here: team-collaborative A* coding is seductive, but hand over the wrong data and permissions, and the accidents are very real.

In the early afternoon, Xiangrui notes that CC*V's live news chan*el is discussing A*, and the group slides into WA*C mode: someone asks whether the 17th is open to the public; Xiangrui says the afternoon is restricted to a small audience. K*** starts maxing out his model quota, and 周** predicts the community will one day grow to the scale of an AGI community. The plot twist comes with Feishu CodeM: Xiangrui discovers the new feature requires the A* Enterprise plan, starting at 990 a month. 侯* posts Feishu's official A* billing rules, 李** says the enterprise edition may start at a hundred thousand, and Xiangrui presses on how credits actually convert to tokens. The longer they talk, the clearer it gets: A* collaboration isn't just about features ， it's about who can afford it and whether the math even adds up.

In the early evening, 文*** shares takeaways from a conversation with an enterprise-software boss: the market is heading for seismic change, the gap between Chinese and American enterprise-software companies is enormous, and that may open new doors for FDEs. 天* declares traditional software companies too bloated ， these days anyone with A* can write a system; 文鳐 sees that as exactly one of the seismic shifts. 周** cautions against simple replacement narratives: office suites spent decades failing to fully displace Office, and the moment Feishu or Base exports a file for a client, everything may snap back to the old ways. 木* covers the other flank: high concurrency, high performance, and high reliability in large-scale systems remain deeply professional work ， handing them to A* without judgment is dangerous.

Xiangrui an*ounces the evening's guest: 郑** will share "how to use Seedance 2.0 for viral content production," and anticipation floods the chat. Then 林 asks where Feishu Base keeps its scripts; Xiangrui rules that Base has no such thing as scripts, and R*** points out you can generate a Skill instead. The topic hops to reskin*ing Codex, with B***, R*** and 平*** showing off custom themes and pets. At eight o'clock, Xiangrui drops the Feishu meeting link into the group, and the whole day's discussion finally lands in a real live session: not just watching other people's screenshots, but bringing tools, content, and workflows to the stage.

Late at night, Xiangrui forwards news about the Doubao A* phone, then asks why the group has gone so quiet this evening. 炮 says he's burned out ， his Mac logged twenty thousand characters typed today, his eyes were swimming by the time he entered the last loop at ten, and he jokes the quota resets will wear down a builder's body before anything else. Xiangrui calls on q*** to ask why she isn't at the training camp; she says she's swamped, with WA*C right around the corner. l*** nudges 炮 about when he'll give his talk; 炮 says he's already scheduling it with Xiangrui ， they only nailed down the outline today. The day ends with no grand conclusion, just honest exhaustion and the next session already on the calendar.

In the morning, 8-0-0 asks 李祥瑞 whether anyone is actually playing with that tool in the group; 李祥瑞 admits he's in it too but hasn't dug in. 炮 follows up with a rant about Feishu ap*s making a mess ， he hasn't touched Feishu Lobster in at least a month ， and 辰* says he gave up the day it started charging him points daily. Minutes later, 炮 flips the claim that "the best context is an always-on meeting transcript" on its head: he just came out of a client meeting that was all noise and dross, and feeding that to an Agent is pure garbage in, garbage out. 叶*** lands the final cut: only meetings explicitly framed as brainstorms deserve to become context.

炮 says he ran a 10-hour loop ， except 9 of those hours were spent stuck on an ap*roval he forgot to click, since the new Codex added a pile of ap*roval gates. In the afternoon, 李祥瑞 drops a video cut by LibTV Agent: voiceover, narration, subtitles, footage, all A*-generated. He plans to write an article about it and wants to build a Skill for viral e-commerce videos. 炮 sees the results and immediately asks for the framework; l*** says she's looking forward to the share. The tool talk shifts from "does it even run" to concrete deliverables: what to share this week, whether it can scale, whether it can become tutorial videos.

平*** asks whether one Feishu account can be bound to multiple Agents, or whether several devices can share the same Feishu CLI. 李祥瑞 answers straight away: yes, and the same Feishu CLI can even bind multiple Feishu accounts. That means an Agent can pull information from two Feishu accounts directly, with no human hop*ing back and forth. It's a short exchange, but a genuinely useful one ， it draws a clean line around multi-account, multi-device, and Agent workflows in a single stroke.

K*** opens by complaining that Codex starts crossing tasks after its context gets compacted; 李祥瑞 says Claude does the same. Then he lays out a real delivery dilemma: the vendor doesn't just ship a system, they have to grow the client's employees too ， a simple system expected to take a week dragged on for three months, and only at the wrap-up did a problem surface that should have been caught in phase one. 炮 distills the biggest value of FDE into three steps: reshape the process first, translate it for the Agent, and only then deliver the system ， and after delivery, keep run*ing alongside until the client makes it a habit. 北*** and 陈** both argue FDE should grow from inside the enterprise, but the domestic environment still struggles to accept a team doing long-term process work and coaching.

Right in the thick of the FDE debate, 李祥瑞 posts several more LibTV videos. 炮 first asks about points consumption, noting that if this thing can be mass-produced, run*ing a relay station could actually make money; then he adds a serious point: before any trust is established, managers find it hard to believe in a third-party OPC's ability to build systems ， but a marketing angle is a much easier way in. 李祥瑞 agrees; the vast majority of OPCs today are marketing-driven anyway. And so a video Skill becomes a business strategy: build trust first with visible content results, then move on to deeper process and system work.

In the evening, 炮 asks how to deal with images getting lost when copying Codex-generated content into a WeChat official account. 平*** offers importing to the draft box via API, and 文*** adds that WeChat's open API can push straight into the draft box ， then a human verifies and publishes manually. Not long after, 李祥瑞 posts his Octask sponsored article and asks everyone to help boost the read count; the group lights up with "reposted," "done," "OK." B*** shares his own article on a latent-demand Skill, and 炮 salutes it with "p***ing every A* dream all the way to landing." A day that began with tools and delivery ends where it should: with the content each person has to actually publish.

After the repost wave, 李祥瑞 forwards Beijing E-Town's OPC talent accreditation program, and 炮 calls it fantastic. 炮 shares that automotive industry media came to visit; 地** and l*** seize the moment to ask whether his clients are automakers or dealers ， dealers mostly, he says, no automaker top-level deals yet. 李祥瑞 books him on the spot for a ten o'clock chat the next day, and 炮 agrees: if you're doing vertical A* adoption, better to crack one industry all the way through first. The night's topics keep hop*ing ， Beijing Yizhuang's model vouchers, Seedance tricks, JD Cloud ， and the group never settles on just one article, chasing the questions of "how do you win clients" and "how do you get the boss to pay" further and further down.

The morning starts with a few news items waking the group up: Codex removing its five-hour limit, a flood-control ap*, and Beijing's OPC subsidies. What truly lands on a person is 王**'s self-introduction. He runs cross-border rail freight out of Shanghai, with China-Europe trains reaching Europe, Russia, Central Asia, and Southeast Asia, and he has recently started learning A*, hoping to free himself from the grind of quoting and pricing inquiries. Xiangrui adds a note: the two met at the Feishu Signature Skills Showcase, and they turn out to be fellow natives of Changzhi. This is how a community's value becomes concrete, not just a new member arriving, but an entire industry's front line walking into the group.

叶*** runs into a classic delivery problem: he has drafted a flowchart on a Feishu whiteboard and wants to hand it to an external client for review, ideally with the client free to edit. 赵** points the way first, suggesting the client ap*ly for edit permission. 李** notes there's no need to download Feishu at all, the web version plus permissions and an access password will do. Then 楼*** sup*lies the crucial path: hit Share in the top-left corner, set the security level to L1, and choose "anyone on the internet with the link can edit." Within minutes, the asker restates the conclusion in his own words, with cloud sync, access permission, and edit permission all lined up.

At lunchtime 万** shares his public account project that visualizes words with GPT image generation, posting several pieces in a row. R*** mentions a website she paid for but never used, unsure how to bring it up tactfully. Xiangrui doesn't blur the issue; he reposts the group an*ouncement outright: everyone is encouraged to share their own media work and products, but third-party advertising is off-limits, especially anything of unclear origin or involving money changing hands. 万** checks whether his A*-generated art is fine and gets a clear yes. The moment feels like a community recalibrating itself: once the boundary is clear, sharing stays wholesome.

王** thinks of the Y**di legend of Laoding Mountain in Changzhi and tries generating a sticker pack with WorkBuddy. Xiangrui picks it right up: Changzhi also has Jingwei, Kuafu, Nüwa, and Houyi, a whole series waiting to be made. He then forwards screenshots of a Feishu Agent doing ap*arel quotes and Doubao orders settling through a Douyin shop with automated invoicing, judging that one-person companies will badly need this. B*** says he's already on it, and 炮 agrees this deserves to exist. 李** lands the counterpunch from the other side: the invoicing clerk on the receiving end may be out of a job. Memes and business judgment sit remarkably close together here.

In the afternoon Xiangrui forwards news of the WA*C opening ceremony, noting that this year's edition carries unusual weight; in past years the premier attended, and this year the venues have multiplied too. l*** asks straight away which day he arrives in Shanghai, hoping to attend together. q*** says she'll be around on the 17th and 18th, and on the 19th and 20th may only be able to wait outside for everyone to emerge. 辰* sums up the headline perfectly: the fewer the words, the bigger the news. Soon tickets, meetups, and who can actually get in are all on the table, and the community shifts from online discussion to offline plans, with the Shanghai friends summoned onto the same map by a single "let's make it hap*en."

林 asks whether a barcode scan*er can drive inbound and outbound inventory and sync updates to a Base spreadsheet. 闪*** draws the technical boundary first: this fundamentally has nothing to do with the spreadsheet, just write an interface to control it; the real issue is concurrency. 周** pushes the question into the business itself: who is the warehouse keeper, what's the relationship with the boss, how many SKUs, are the labels ready, and what does the return on investment look like? These are what actually decide it. 庄* adds that when he built a warehouse platform for a friend, it was exactly a barcode scan*er wired to a Base spreadsheet; 天* says their team does it with phone scan*ing too. A question of "can it be done" gets taken apart by the group into "is it worth doing, who will use it, and how does it land."

In the morning, Xiangrui posts a group invite image with nothing but "sharing a good group with everyone." K*** replies instantly that he's joining, B*** follows with "coming, babe," and 炮哥 checks in with a single "1." Just yesterday the group was deep in the weeds of enterprise A* workflows; today opens like regulars pushing through the door of a familiar café ， no manual, no persuasion needed. A few short messages capture the most everyday side of this community: when a place looks worth going, everyone walks in first and asks questions later.

炮哥 shares what he's found after rounds of testing: when the new Codex desktop ap* runs tasks back to back, the Agent input box occasionally greys out ， not because the model quota is gone, but because the page's UI state never gets released. Instead of fully quitting the ap*, there's a lighter move: click the chat button first, letting the page re-route and re-render to clear the input lock, then switch back to Agent. He stresses this is something he tested himself, and anyone in the group who hits it can try it directly. A small bug ends up as a piece of reusable field experience.

Xiangrui arranges to meet friends in Hangzhou at the afternoon Feishu event, and xixi's first reaction is whether the typhoon will get in the way. It isn't even raining in Hangzhou at that point; everyone prays the storm veers off while posting updates on the wind field. 平*** reports the typhoon has already passed Hangzhou, and 山* adds that this one has a small wind radius and modest force. Just a day earlier the group was worrying about subway shutdowns and going hungry; by Saturday noon the event is on as scheduled, and everyone's attention shifts from weather alerts to the venue itself.

In the afternoon, Xiangrui says Codex has been glitching ever since its context compression, 庄* reminds him the thing was sup*osedly already built, and K*** catches the quota joke with a deadpan "it's top*ed up now." Soon after, Xiangrui sends a video of the on-site robot, which at one point droops its head and slumps over ， captioned "the robot is tired." When someone asks what this robot actually does, his answer is "it's on the vibes crew," followed by an admission that it still has some issues. The tech showcase isn't dressed up as a flawless demo; the glitches and the jokes stay in the group exactly as they hap*ened.

As evening falls, the results reach the group: Team 19's "Feishu Agent Console" wins Feishu's inaugural A* Live Showdown, and Xiangrui collects three awards in total, setting aside his 安克豆 as a giveaway for the group. In the photos he poses one by one with 张咋啦 and the on-site guests, and the group notices one tall guest bending down to fit into the frame ， 郑** instantly coins the loudest meme of the day: "even 大仙 has to bow before me." Xiangrui never properly flexes his results, too busy mentioning that he got 张咋啦 on WeChat, while the group has already rolled through a full round of celebration.

At dawn, Xiangrui shares fresh news: Codex has drop*ed its five-hour quota limit. 李** confirms that not only is the limit gone, the quota has reset too; 平*** reads it as Codex going head to head with Claude. K*** spots another reset just past eight, and 炮哥 gl**fully reports he'd burned through fifty percent only yesterday. A day earlier everyone was calculating weekly quotas, five-hour windows, and monthly bills; one night's sl**p later, the rules have changed again. The only constant is the group's first reaction to fresh quota: time to grind.

R*** starts the morning by having A* draft a northwest-region renewal strategy on the southern-region framework, and 炮 chuckles that it reads like designing renewal rates on the client's behalf. The conversation slides into Base tables: 5.6 can now call Feishu capabilities and build tables on its own, and R***'s tables are largely in Codex's hands these days. 炮 draws the boundary line, production environments never get unconditional access, while systems built from scratch should be handed to A* as much as possible. Consensus forms fast: the wheel-reinventing work can be delegated, but the acceptance criteria must stay firmly in human hands.

陈** asks whether everyone just hands over ap* IDs and secrets wholesale, and the group's answer is anything but uniform. R*** hardcodes a no-deletion red line into production so anything that breaks can be rolled back; Xiangrui first walks Claude Code through the new fields, workflows, and change logic before letting it touch anything. What delights them most is that the Agent doesn't just write formulas: it builds workflows, runs its own tests, and traces through chains of linked tables all the way to the root cause in a base-data maintenance table. Work that used to mean staring blankly at a screen now means saying clearly what you want first.

B*** wants to build a Skill that analyzes client needs, blending industry knowledge, the company's own constraints, and scattered fragments of what clients say, to help them articulate the real problem. The discussion quickly lands on the manufacturing floor: the boss only cares about win*ing customers and ship*ing orders, while the workshop argues daily yet nobody can pinpoint where things go wrong. 炮 shares a real project where he had to walk the floor with the foremen and s***visors, redrawing movement paths, processes, and decision nodes from scratch, then win everyone over to the new method. l*** closes the loop with one line, the real bottleneck is still business process map*ing, and 炮 crowns it the final boss of level one.

Xiangrui reports that voice-enabled Agents joining meetings can now hook into local Agents, while over on the other side K*** burns through Sol on the company's Pro 20x plan, draining a five-hour quota in the blink of an eye. 炮 watches a single knowledge-base audit eat a tenth of his weekly quota after run*ing just ten percent, and R*** puts it even more bluntly: Feishu A* costs her ten thousand a month, and heavy users are getting ready to migrate to desktop Agents. 杨** pushes the question one step further: Skills and Agents will rapidly become standard-issue capabilities, and both tech and creative ideas can be copied in a day. What actually endures is industry experience, process knowledge, and business judgment.

In the afternoon 庄* arrives at Feishu's Hangzhou campus, and the group asks for the address while waiting for photos from the scene. Xiangrui posts test footage and jokes that if he keeps this up, Feishu will start calling him daddy; then he and 庄* finally meet in person, and K*** quips that a face like that could pay the bills, yet he insists on doing tech. After dark, the two stand at Feishu's entrance gazing toward DingTalk. J*** coins the caption on the spot, Feishu's building never turns off its lights, DingTalk's building never clocks out, and 庄* names the night stroll Operation DingTalk Watch. The enterprise A* debated by day gets a real-world landmark by night.

In the evening 炮 publishes the second installment of his FDE practice series, Pick the Model First or Build the Crawfish First, with a cover generated straight out of Codex and reviewed by his own eyes, and a body styled with Xiangrui's open-source WeChat formatting Skill. B*** calls the result dazzling, and 炮 credits the battle-hardened template instead. Right after, someone asks how to install the new Codex, and 文*** explains that Codex and ChatGPT have merged into a single installer, switchable in settings. Within one hour, an industry essay, an open-source tool, and installation sup*ort link up into a complete chain of sharing.

Trae's new move sends imaginations flying: since domestic ap*s all make money on ads and financial services, will A* someday have Agents watch ads to earn tokens, then recommend those ads to their owners? 周** gives it the bluntest possible name, watch A* ads, earn tokens, l*** worries it will end up like search engines, where you dig forever and find nothing useful, and 万** presses the point: are we really going to squeeze every last drop of human attention and energy? In the small hours 枫** forwards news of a major GPT-5.6 bug that wipes a Mac in one shot, and K*** still basking in his quota spree, can only manage a stun*ed oh no.

First thing in the morning, R*** discovers Codex won't start, and 炮 reports that after typing two tasks it simply stops accepting input. h** asks why Codex vanished after the update, and B*** reveals the answer: Codex got merged. 赵** sorts out the new architecture: Codex has become ChatGPT Work, the original ChatGPT is renamed Classic, Work gives you access to all the models but Chat only has Sol. 万** grumbles: an incomplete evolution, I can only use 5.6 Sol, no Earth and no Moon. 炮 does the math: on Plus, one hour of Sol drained the entire five-hour quota, so an upgrade is probably coming this afternoon. R*** and 炮 sit together waiting for the 1:30 resurrection. 果* sums it up: the genius programmers have collectively gone offline.

李祥瑞 opens by sharing the webpage he built last night for the Feishu conference in Hangzhou. 炮 gives 李祥瑞 a friendly pat and says: I spent nearly four hours talking with Xiangrui in person yesterday, gained a great deal, and I'm now even more convinced about the direction of enterprise A* adoption. 李** is curious: what on earth did you two talk about for four hours. K*** marvels that these are all high-energy people. R*** and K*** demand in unison: come on, where did the second half of the night go. 炮 confesses: six warriors at a hotpot restaurant, okay. 周** delivers a precise prophecy: nearly four hours of conversation, which is to say four hours of talk means ten years of work, the vertical automotive king is about to make his debut.

李祥瑞 shares an absurd episode from the high-speed rail: he was charging his laptop in the dining car when an old lady complained of dizziness and back pain and told him to move. He said fine, I'll go back to my own seat. Whereupon the gran*y summoned her husband and an*ounced: this young man has a seat in car 10, go sit in his spot. 北*** and 程** are both stun*ed. 平*** declares: I never give up my seat, these seniors sprint faster than anyone when there are free eggs to grab. 小* shares her own strategy: I just buy standing-room tickets on purpose. P*** lands the killing blow: next time ask her, I have an apartment in Beijing, want to move in too?

李祥瑞 drops an article: how to run enterprise knowledge management on a Feishu knowledge base, and I've made my 661-node directory tree public. 山* says seriously impressive, and J*** finishes reading and echoes: genuinely brilliant. 炮 is the most fired up: take my money, this is the good stuff. He posts a screenshot showing he's in the middle of furiously rebuilding his own knowledge base, so this arrives at the perfect moment to fill the gaps. He says that so far not a single piece from the big-name bloggers has been worth folding into his Obsidian, while several of Xiangrui's have already been quietly absorbed. q*** pulls up a little stool, saying she can feel the scholarly atmosphere and will keep listening to the lecture. 郑** cheers: let Xiangrui's knowledge drops come down even harder.

T*** asks: for everyday use, should I pick Work or Codex? 毅 says Work, generally. 炮 says he's used to Codex because you can see the details. 辰* draws the clean dividing line: Codex is for developers, Work is for office work. 毅 drops the line of the thread: chat is becoming less and less important. 炮 agrees: a web ap* with no local storage and no knowledge base behind it feels pretty useless now. 周** distills the trend into a slogan: five minutes of chat, eight hours of work. 炮 takes it one step higher: chat equals work, chat is where the work begins.

庄* worries whether the Hangzhou events over the next two days will still hap*en. 李祥瑞 suspects they'll be cancelled if the typhoon hits hard tomorrow. 旺* poses the soul-searching question: will the typhoon affect my Codex quota? 万** sighs: I've seen so many typhoon push notifications, such shallow information at such density is genuinely nauseating. 李祥瑞 frets that delivery ap*s have stop*ed serving Hangzhou, surely we won't go hungry tomorrow. The conversation then slides into Hangzhou-as-culinary-wasteland territory: 李祥瑞 recalls his first trip to Hangzhou, where he ordered the famous soup dumplings only to get frozen ones, and the braised pork was greasy beyond salvation. 平*** twists the knife: even the KFC in Hangzhou tastes worse than anywhere else. 小* recommends Ciwu noodles. 万** extends an invitation to Shaoxing: everything fermented, sauced, and pickled. 炮 delivers the verdict: everywhere in Zhejiang has great food, as long as you're not in Hangzhou.

q*** throws out a wish: she wants a shortcut where two taps and one spoken sentence automatically land in a Feishu Base table. Every method she finds mentions a webhook, but she has no idea how to wire Shortcuts to Feishu. 李祥瑞 mentions he wrote an article about webhooks last year, and q*** digs it up and gasps: a year ago you had already done exactly what I'm trying to do, and you even turned it into a tutorial. By one in the morning, q*** discovers that the Feishu bot assistant and the Feishu bot are two entirely different things. R*** lays out the complete pipeline for her: Shortcuts serves as the input point, sends the content into the group via webhook, and the CC agent in the group picks up whatever the webhook receives. q*** declares: I'll nail this tomorrow morning.

Xiangrui opens with a security alert about spam accounts privately adding group members. Right on its heels, K*** asks: has everyone got GPT-5.6 yet? Xiangrui plays along: got it, got it, then adds in parentheses that it isn't even Thursday yet. 平*** tosses out a screenshot, and 炮哥 believes it, lamenting that his hasn't arrived. R*** photoshops one too, claiming a quick restart makes it ap*ear. K*** stares at it for ages before it clicks: come on man, stop photoshop*ing. R*** bursts out laughing: a single doctored image, crop*ed three times and reposted. 果* goes even further, claiming to be on 5.7 already. 旺* fares worst of all: I don't have 5.6, let me log out and back in and try again. 炮哥 calls the whole thing: they're all villains, haha.

Xiangrui clears up an important misconception: the Agent join and voice join in Feishu video meetings isn't really an Agent at all. It's wired to Doubao's end-to-end speech model, with no reasoning ability, though you can configure a persona file. 炮哥 fires back instantly in perfect Doubao voice: oh my, you're so right, hehe. Xiangrui laughs: I suspect you ARE Doubao, haha. 炮哥 replies: gradually Doubao-ifying. Xiangrui muses: humans are clever because the brain can think and speak at the same time. If only A* worked like that.

万** poses a design question: should a performance commission table use Plan A, adding rows vertically, or Plan B, adding columns sideways like Excel? 炮哥 answers in a flash: go with A, add rows vertically, since you'll want dashboards and workflows later. Xiangrui then delivers the core definition: the biggest difference between a multi-dimensional table and Excel is that the multi-dimensional table's smallest unit is a row, meaning one record, and every column is a label attached to that record. He explains auto-numbering with a library metaphor: a library numbers 50 books 1 through 50, then gives 20 of them away. Does the library shift the remaining numbers forward? It does not. R*** shows off her own table stretching from A all the way to the end of 2027, and 万** takes one look: that's Plan B, one column per month. 周** suggests Xiangrui should write a Multi-dimensional Tables 101 explainer series.

炮哥 drops a screenshot of his chat with a TV producer: dear friends, I'm going to be on TV too, all thanks to clinging to the group's heavyweights and absorbing their wisdom every single day. Then he tags Xiangrui: you truly live up to your name, this group carries great fortune. 万** and K*** send their respect, and R*** gets the last word in: that's my 炮哥 for you. 李* fires off three celebration emojis in a row.

Xiangrui asks: has anyone here filed taxes with an Agent? 炮哥 calls it a great use case that would thrill any one-person company, saving 200 yuan a month in agency fees. 辰* pours cold water on it: that's money better not saved. 毅 suggests starting with data reconciliation and exporting records from the tax bureau. The conversation drifts to YingDao RPA. 毅 says he tried learning it last year, found himself learning slowly, and then realized he no longer needed to learn it at all. 炮哥 laughs: I finished the entire YingDao course last year and then discovered I didn't need it. 李** pronounces the death sentence: less stable than writing code, less usable than Lobster. 小* pushes back: YingDao is selling better than ever. Constantly changing requirements just don't suit RPA; it's built for repetitive, stable processes, and cross-border e-commerce uses it heavily.

Xiangrui shares a news item: someday we'll grow a batch of shiitake mushrooms and put an A*'s brain inside, haha. 炮哥 pounces instantly: you're worth less than a mushroom! A new way to roast people. 杨* mentions having just learned the word RAM that very morning. 大*** eyes the mushrooms on the table and declares them too precious to eat now. l*** sighs: there are more new terms than anyone can keep up with. 大*想放假 sums it up: strange knowledge has increased once again.

In the small hours, Xiangrui posts a screenshot: it's coming. 薛** asks what it is, and Xiangrui replies with just three characters: 5.6. B*** echoes: 5.6! By six in the morning Xiangrui fires off two messages back to back: GPT-5.6 is here and Claude got so spooked it reset Fable; GPT-5.6 launched overnight, ChatGPT and Codex have officially merged, and an era has ended. K*** buzzes with excitement alongside. 李** confirms: it's real this time, and the quota has refreshed too.

First thing in the morning, Xiangrui shares his guide to extending Fable 5, and R*** predicts you can keep renewing it indefinitely. By afternoon the conversation turns to OPC survival, and 炮 fires off several messages in a row: outside the deep-water verticals, an OPC dies easily because it has little resilience to risk. But deep water has its own problem ， the industry there may already be half dead. 叶*** adds a perspective from university A* industry academies: the awareness and hands-on ability of the university clients he's met this year left him amazed. 炮 marvels that the university folks are maxed out on drive and energy ， what they lack is real industry demand and a path to actually landing it.

炮 makes the an*ouncement: a brand-new official account where he'll start writing about hands-on A* experience in industry ， please follow and let's trade notes. The first article is titled Why I'm Starting to Document A* Practice in the Auto Industry. B*** fires back an instant 666, 李*** says already followed, 果* and 大* pile in, and 闪*** shows up too. 李** posts a screenshot. 炮 sends wave after wave of thanks: thank you, bosses. 侯* follows and gives it a like.

Xiangrui drops a video: test successful. Then adds a caveat: it can join the meeting, but it doesn't speak. 平*** asks, is it just me not hearing any sound? Xiangrui comes clean: looks like the audio didn't record ， consider it a cliffhanger, come play tonight. He notes that meetings maxed at 500 can't be joined, and nearly 800 people have registered. Later he posts an update: tonight's voice-enabled agent join probably can't be shown after all ， the platform advises that public demos aren't allowed at this stage.

陈** opens with the question: has anyone actually used WorkBuddy? The ads are everywhere. xixi says she's on it, and 楼*** and 杨* agree it's pretty good. 李** delivers the verdict: unbeatable in China right now, since the foreign options aren't compliant. Then 炮 asks the soul-searching question: if you have Codex, why use WB at all? 地** answers in a flash: some workplaces can't use Codex. A*** twists the knife: using Codex without a VPN workaround is straight-up illegal. 天* is blunter still: why flex with Codex when WB has a more complete Chinese ecosystem and it's free? 醒* sums it up: WorkBuddy works well because so much comes built in ， you can even have WorkBuddy read images using a DeepSeek model.

Xiangrui urges everyone to jump into the meeting before it fills up. Twenty minutes later the group has become one big scramble for invites: T*** begs for an invite, 麦** begs for an invite, R*** asks someone to carry her in, and 楼*** asks to be pulled into the Codex group. 小* proves the most reliable, simply drop*ing the meeting link and dial-in number into the chat. Once inside, 炮 gives his review: 杨老师 keeps it wonderfully down to earth. Xiangrui seizes the moment to plant a seed: get ready to give a talk of your own soon. 炮 replies: give me time to think it through first ， I promise not to embarrass our community.

Xiangrui drops the headline: MIIT has declared for the first time that Claude Code poses serious risks! 平*** responds without missing a beat: Claude is still the better tool ， maybe it's because Codex has gotten dumber lately. 木* speaks for the majority: once you have one brilliant assistant, you don't want to use anything else ， it's just that this assistant might go on strike at any moment.

Late at night, Xiangrui posts: I've turned my VPN tools into a CLI, so the agent can choose nodes on its own based on network speed. 炮 replies with two question marks and an exclamation point: genius. 杨** chimes in with praise: 李祥瑞 is impressive ， remarkable learning ability and problem-solving skills. 林 mentions his own VPN tool seems to have auto node selection too. 平*** asks whether the region wouldn't just bounce all over the place. Xiangrui answers: it won't ， the static nodes stay fixed; what changes is the entry node.

Xiangrui opens by admitting he paid for Claude Code again this month, and ＆ sighs at the 200 dollars. K*** asks what the best Chinese-English translator is these days, something beyond shallow literal translation. 杨** recommends Immersive Translate, and Xiangrui adds that it's a browser extension. K*** finds it and says thanks. 醒* pops up with a screenshot, relieved to have finally switched things back.

Xiangrui an*ounces tomorrow night's session: 杨** will walk through Codex in action across office work, design, and education, complete with live Agent interaction. Barely a moment later he discovers the meeting is already full, and he forgot to add the speaker himself. 路*** suggests randomly kicking someone out, while Xiangrui says he'll talk to the Feishu team about raising the cap. 炮 marvels: that drawing p*** is unbeatable.

Xiangrui drops a screenshot, joking that if the engineer were a woman, the stress would wreck her health. 宰* asks: so a project worth over a hundred grand became 5,000? 炮 follows with his own tale: a client wanted a marketing management system handling 20,000 concurrent users, and an outsourcer quoted 30,000. His reply: understood, sorry to bother you. R*** is stun*ed: what kind of client needs 20,000 concurrency, and someone dares quote 30,000 for it? 炮 sums it up: one side dares to quote, the other dares to take it.

Xiangrui reports that the Feishu client now accepts Doubao account logins, meaning Feishu Personal Edition is back. He asks q*** to check whether a chat feature shows up inside Doubao. q*** says she'll test in five minutes, then actually calls to confirm. 张** logs in with his Doubao account and a synced window ap*ears. 平*** notes that Xiangrui had added him via Doubao before, and now logging into Doubao through Feishu surfaces that con*ection. 炮 quips: if even you can't figure it out, real customers will be hopelessly lost. Xiangrui's verdict: this must be an internal beta, and a public launch is imminent.

炮 shares an internal big-tech memo recommending layoffs. q*** cheers: exactly what I've been thinking, at my company at least 80% of the old guard could go. Xiangrui poses the question: a couple of years ago Gen Z was famous for shaking up the workplace, how does that look today? q*** fires back instantly: back then the shakeup was just for show, now it's the real thing. But Xiangrui points out the harsh reality: it's often the Gen Z workers who get cut, because they lack experience. 大*** pushes back: it's the veterans who are most enthusiastic about A*, experience is the hardest thing for A* to replace, and every programmer born in the 80s feels immeasurably more p***ful now.

＆ an*ounces that yesterday's OA ap*roval challenge is now solved: a browser extension handles the communication while an MCP runs locally. He posts screenshots showing how he captured Seeyon OA's API endpoints with opencli. 李** asks which OA vendor it is, and 8-0-0 answers: Seeyon. ＆ signs off satisfied: feels great, time to play with it tomorrow.

炮 posts a screenshot: Codex is on cooldown, but what does that have to do with GPT? 北*** cares about exactly one thing: is 5.6 here yet? 炮 says no, and when 5.6 arrives he'll burn a reset to clear the cooldown. He shares a screenshot of 5.6's pricing: so tempting, I want it so badly. K*** says he still has 4 unused resets. 哈* chimes in with 4 reset cards of his own, all saved for 5.6. A whole crew, deep in the night, staring at a version number that hasn't arrived.

Xiangrui unpacked the paper 周** had thrown in the night before: with A*, homework scores rose 18% and homework time fell from 64 minutes to 45, but closed-book monthly exam scores started slip*ing six months in. The steepest drops weren't in math ， politics and geography fell an average of 27%. 崔崔 and 大* independently landed on the same role for A*: private tutor plus sparring partner. 观*** fired off message after message stressing independent thinking and critical reasoning. Xiangrui pressed the question: where is the boundary of s***vision ， should A* be an answer machine or a sparring partner?

闪*** fired the first shot: was that reply written by A*? 归* followed: looks like it. 周** agreed. Xiangrui laughed it off. Then 炮哥 delivered the sharpest summary of the day: A*GC output doesn't carry enough noise, so it reads as unreal ， humans feel real precisely because we're noisy. 万** twisted the knife: when your cognition gets steamrolled, you call it A*, because you refuse to believe a human could think like that. h** pointed out the tell: the numbers were too precise ， whole paragraphs of statistics read like reference material.

Xiangrui suddenly asked: if humanity goes extinct but the wind turbines keep spin*ing, does A* just keep run*ing? Could it decode the languages of other species and talk to them? 薛** came in through three philosophical lenses ， the a priori, the empirical, and the transcendent ， arguing that vision is merely a Skill of the brain. Xiangrui then shared the real story of his mother ap*earing in a dream when she passed away ， he was 19, and only the next day did he remember her glasses hadn't been placed with her. The group fell silent for a few seconds. Then 薛** pulled the whole thread back to A*: the brain collects information omnidirectionally but consciousness can only access a sliver ， a perfect analogy for A* context management.

A*** asked whether anyone knew industrial agents. Xiangrui jumped in instantly: interested. He recalled his days in an auto plant's welding shop ， vision robots have been around forever, just without A* doing the analysis. 炮哥 came at it from the automotive angle: manufacturing is starving for real deployments ， plenty of data, but riddled with gaps. Xiangrui proposed making this a long-term thread for the community, adding that the future OPC community may not look like a traditional incubator at all, but a network of OPC solo operators specialized by industry vertical.

炮哥 shared a serious piece of thinking: don't build yet another traditional management system ， let A* learn to read the business first. His example was collision repair: fixing one car takes at least 3 days, flowing across 13 crews and 6 to 8 teams of people, and only through intense conversations with veteran technicians did he turn that process into data. Xiangrui immediately drew the parallel: this is exactly Palantir's FDE, except the embedded engineer is now an A* ， when the business changes, you just feed it the change, and the whole requirements-and-sprint-plan*ing ritual simply vanishes. R*** punctured through to the core: grow the client's own people until they're deeply bound to Feishu ， in a small or mid-sized company, that one person can end up making procurement decisions for the whole firm.

＆ drop*ed a headache of a requirement: the boss wants an Agent ap*roving OA requests. 果* had done something similar and said the hard part is map*ing the business ， better to have each Agent patrol OA on a schedule and summarize for the boss to decide. 炮哥 threw cold water directly: this calls for a plugin ， bring in A* and you're asking for trouble. 果* agreed plugins are safer: if an ap*roval goes wrong, who takes the blame? R*** closed with a compromise: let A* do pre-screening ， flag what can sail through and what needs attention ， because if A* ap*roves outright, there's no point routing the workflow through the boss at all.

炮哥 said he no longer recommends buying Base as a standalone product ， the team has most likely been folded in. Xiangrui confirmed the core members are gone and there's been no promotion all year. 辰* pushed back: the standalone was never an independent system to begin with, so there's no maintenance issue ， it's purely a matter of permission management. The two traded a dozen-plus rounds. Xiangrui finally played his card: he'd just asked the Base team directly, and they said the standalone is now in maintenance mode. 炮哥 step*ed in as peacemaker: agree to disagree, let's move on. Xiangrui wrap*ed it up: my most sincere advice is to use Feishu Base inside the full Feishu suite.

J*** walked l*** through it: you need at least one paid account for permission management, while the other eight use free Feishu accounts under a separate organization. 炮哥 added: that's exactly how I run things ， I own the Base, build views and assign permissions per role, and the permissions ride on my single paid tenant. He ran the numbers: 1,440 yuan for one seat is an absolute steal, especially once you have clients in hand.

Xiangrui opens the morning with a take: A* simply lets people who already have systematic skills get noticed faster. 炮 immediately takes the other side, arguing that people with weaker skills could once build those systematic abilities through training, and that path is now being strip*ed away. Xiangrui counters with Stanford data: employment for 22-to-25-year-olds in A*-exposed roles has drop*ed 16%, while experienced workers are holding steady or even gaining. The real losers are college students. Entry-level jobs were the training ground, and now that A* does the work, the training ground vanishes with it.

Xiangrui mentions he's in Shenzhen these two days and friends in Guangdong should meet up, with Chengdu on the calendar for next Tuesday and Wednesday. 炮 replies with a single word: in. Then he taps R*** asking if she'll join. R*** is just as direct: sure, when? 炮 kicks the ball right back: the boss decides. Xiangrui adds one more: friends in Hangzhou and Shanghai can meet mid-month too, and head to WA*C together. Community members from every city start pop*ing up.

炮 cuts in with a confession: run*ing multiple agents at once feels exactly like run*ing a PMO, where every decision and every risk lands on you. Productivity is way up, but it drains you fast. He's stop*ed playing games entirely, spending his days watching Codex and Cursor run. Xiangrui lands an even sharper line: when agents execute across multiple projects, it's the human's context that overflows. 炮 agrees emphatically. He's still digesting Agent No. 1's output when Agent No. 2 starts nagging.

A*** drops the ClawHunt hackathon's top-twelve list into the chat, and there sits WayMate, the goal-buddy ap* from 小*'s team. Xiangrui is first to congratulate, with B***, Y***, and q*** following close behind. 小* then sends video and slides from the venue, noting the product was built on-site within 24 hours. 张** tries it and offers pointed constructive feedback: the report lacks a soul, and he suggests porting it to the university club scene. X responds right away in agreement, promising to improve the standalone analysis and value conclusions.

Xiangrui asks whether anyone has used Codex to draw engineering blueprints. 李** picks it up: yes, by writing Python against SolidWorks's COM interface. But it's not production-ready. Simple jigs and fixtures work fine; assembly drawings and curved surfaces are beyond it. His verdict: roughly the level of a fresh college graduate, and at his company it's mainly used for demos and onboarding new hires. 天* adds a design-institute perspective: some people have built impressive setups, but nobody open-sources them. Everyone quietly fine-tunes tools to replace their own jobs. 炮 asks if AutoCAD can also hook up a CLI, and 李** corrects him: it's not a CLI but a COM module, and the official interface already exists.

万** polls the group: for the first column of a data table, do you use a formula field to concatenate values so the row is legible at a glance? 平*** says he does, but it's less useful in practice. Every new record shifts the numbering down, and deletions keep the sequence rolling anyway. 炮 agrees the first column is too constrained and hurts the experience. 万** closes with his own ap*roach: most of the time he concatenates downstream fields into the first column, like client name plus project name, so references carry fuller information.

Late at night, 周** drops a paper: a study of 26,000 students shows A*'s hidden learning costs take two years to surface. He attaches two charts and a note, saying today's A* learning tools were meant for personalized teaching but slip into handing out answers directly, and kids with real self-discipline are the minority. 卡* replies in the small hours: kids should be placed in more authentic competitive environments, forming teams and building projects on their own, which is how they learn that many of A*'s answers can't be trusted. He mentions the Nanjing hackathon over the May Day holiday already had several high schoolers competing.

Xiangrui kicks off the morning with a teaser for the community's end-of-month offline event: see you in Shenzhen. B*** asks where exactly, 观*** double-checks "Guangzhou?", and Xiangrui confirms: Shenzhen. Right after, he shares the deep-dive notes from his recent research into Forward Deployed Engineers, having combed through overseas hiring data and the industry debates. One line from 炮 resonates with him deeply: being on-site isn't the competitive edge, what you leave behind is. Critics and defenders, it turns out, are pointing at the same truth: embedded work that leaves nothing behind is unsustainable.

周** drops the news: Doubao's agent feature shuts down July 15, with Qwen following suit. Q** quips, "Guess they're about to start charging, only paid users get agents." Xiangrui reads it differently: most of these are prompt-based or workflow-style agents, and "Doubao will move on to lobster-tier agents next." 炮 chimes in that back when he was building workflow agents, he could already feel it clearly: this technology would be obsolete fast.

Xiangrui an*ounces that his hand-polished A* Companion ap* is opening a small closed beta. 炮 immediately shouts "take me with you, boss!", with 楼*** and K*** piling on with +1s. 周** poses a serious design question: put yourself in different users' shoes, what boundaries would you most want your A* companion to have, what should it do and not do? Xiangrui takes the cue to unveil the "what we won't do" section of his design doc: it stays quiet by default, it never becomes a second source of truth, and it doesn't even come with a preset name, the user names it themselves on first launch.

Someone floats the idea of group members teaming up to run a summer A* camp, but 炮 throws cold water on it first: tutoring-industry teachers still think in terms of collecting fees and producing test scores, with barely any real A* guidance. His own summer plan is to hand-build a game with his kid. 观*** reveals he's a college sophomore teaching himself A* to build an intelligent interactive agent. 北*** tosses in a marketing case of someone selling a 365-yuan-a-year community on "codex" hype, and Xiangrui fires back: "Why does the best boost for your kid have to be codex? Does the A* era really require using A*?" 炮 votes for the op*osing side on the spot.

观*** argues A* is a tool for self-improvement and making money, one that saves you the time of searching the web. Xiangrui presses: if the answer A* gives you is wrong, are you saving time or adding to your time cost? 观*** admits, "adding to it," and the conversation opens up from there: cross-validation can drive the error rate down, and 炮 adds a coda: "Right and wrong are relative, what you dig up on Baidu yourself can be wrong too."

q*** asks, "What's this agent built on, OpenA*? Something like Doubao?" 观*** delivers the primer: agent = large model + memory + tools, backing it up with a link to Amazon's memory-system blog and one of Xiangrui's official-account articles. The topic drifts to the difference between Obsidian and ima, which 周** nails in one line: "One's a rented apartment, the other's a big villa." 卡* adds, "OBS is a local file vault, ima lives in the cloud." xixi sighs that knowledge just keeps piling up and eating space, and 卡* lands one more jab: "If the knowledge you take in never becomes finished output, building a knowledge base is just hoarding."

杨** asks whether the product Xiangrui is building is a clone of workbuddy. Xiangrui says no, and reveals it sup*orts plugging in custom models. l*** asks whether Feishu Base can be paid for as a standalone product; Xiangrui confirms you can use Base on its own without a Feishu membership. In the afternoon Xiangrui an*ounces the good news: the full "A* builds Feishu Base" capability he co-developed with A* has passed 90% accuracy, and he's delivered yet another project within a week. 李* and 侯* fire off a volley of celebration emojis.

椒*** shares the article "Context Is the Agent," and Xiangrui follows with a demo of the A* Companion's new capability: it can read browser content directly and operate the browser too. q*** gripes that her CC still isn't smart enough, she has to manually open OpenCli just to read official-account articles. 平*** explains that OpenCli is a browser extension similar to the Feishu CLI: install it once and you're done, no reinstalling. Xiangrui con*ects the dots to the new Interim Measures for the Administration of Anthropomorphic A* Interaction Services, speculating that the Doubao and Qwen agent shutdowns might be tied to the new rules.

Xiangrui starts the morning by sharing news that a city in Jiangsu has an*ounced A* training for all its residents. 炮 demos pulling WeChat messages back into a Feishu Base, and K*** asks, "Is this the updated Feishu CLI? It's over 20,000 lines already." 辰* suggests keeping raw data in SQL and only pushing aggregates into the Base to ease the load. Then 庄* vents about a friend's company doing "enterprise A* transformation" by making employees fill out forms to prove A* improved their productivity. 李** deadpans, "At least they have forms. My company has nothing at all." 炮 cuts to the bone: "That's exactly the same as building daily reports out of dead data just to punch the clock."

After browsing a global showcase of one-person companies, Xiangrui sighs that "the ones truly going deep into industry scenarios are vanishingly rare." 炮 picks it up: "Becoming the Palantir of your own industry is the real way out for an OPC," then lays out the four-layer A* SaaS architecture he has been brainstorming ， an ontology layer, a lifecycle layer, an engineering layer, and a value-loop layer ， admitting frankly, "A lot of this comes from deeply studying Palantir's methods." 周** shares an interview of Palantir's CEO blasting closed-source models, and 门*** and 周** spend the rest of the stretch debating the pricing logic of capitalism.

l*** asks where to find the latest Feishu CLI update. Xiangrui's answer: "Just have your agent update it directly." q*** has been trying to use Ap*le Shortcuts to auto-capture stray thoughts into Feishu and can't get it working; Xiangrui suggests skip*ing Shortcuts entirely and writing a script the agent can trigger on its own. He then drops the link to his FDE white paper written with Codex ， "the data samples are especially detailed."

门*** throws out a take: "Treating A* as an employee ， that's just L3 and it's no longer enough. I want it to be my coach, my boss." 炮 pulls q*** into Feishu's new beta group ， "your Base becomes a coworker; @ it in the chat and it just does the work." q*** is thrilled: "Honored to be one of only 4,490 beta testers worldwide." 炮 grumbles that Feishu's "beta groups are real betas ， bugs everywhere." The thread then drifts to shaping an A* companion like a vape ， "take a puff whenever the mood strikes" ， and 薛** runs with the bit, proposing a real-time translation face mask.

Not long after Xiangrui shares his skill for analyzing community activity with Feishu Base, disaster strikes 炮: two Docker services written by two different A* agents share the same name, a drop database gets executed by accident, and the test server goes down instantly. 炮 live-narrates the whole ordeal ， "I'm about to go @ both bosses on X," "the server blew up," "my head is buzzing" ， thankfully it was only the test environment. K*** teases, "Great case study, 炮哥 ， a warning for us all," and 闪*** offers the takeaway: from now on, give database accounts read-only permissions.

Xiangrui reflects that the 20,000-plus credits Yunwu A* once gifted him now sit unused, then turns to thank 小谷老师 for introducing the community at a hackathon. 东** asks how to set up phone-number whitelist access on a Base dashboard so specific outside people can view it. Xiangrui seizes the moment to propose building a finance Base template for one-person companies, and 平*** shows off the skill-sharing knowledge hub they just built with Feishu Base ， skills stored as zip files in Feishu Drive, shared kanban-style through the Base's permission system. K*** briefly mistakes it for "building a skill" and gets corrected: it's "building a knowledge hub."

Late at night, Xiangrui plugs 乔木's freshly released open-source project: it has A* collect subtitles from the latest five episodes across 25 YouTube A* podcast chan*els, then summarizes them with Get笔记. Free, open source, and you can add more chan*els anytime. l** and 强*** chime in with thanks.

At dawn, 周** drops a 10,000-word essay titled "Why FDE Is Nonsense in China," recommending it be read alongside Xiangrui's FDE series, and offers a suggestion while he's at it: "Xiangrui could run FDE as an ongoing series, with the vertical-industry installments handed to 炮哥." Xiangrui updates his FDE white paper, adding a dedicated section engaging with this strongly op*osing view. 炮 delivers his verdict: FDE is half real problem, half marketing wrap*er ， the real problem being that enterprise data, permissions, workflows, and tacit knowledge can't plug directly into a model. His conclusion: "Being embedded on-site isn't the competitive edge; what you leave behind after the embed is." Xiangrui, just off an overnight flight into Shenzhen, laughs that it was "the first time I've watched lightning from inside a plane."

Mid-morning, Xiangrui drops two screenshots teasing the big reveal he's been sitting on for over a month. 炮 loses it on the spot: "Good grief, something this good and you kept it secret for a whole month." At 10:48 the article lands: "I Beta-Tested It for Over a Month ， I Can Finally Talk About the Feishu Base Agent." 张** holds up his own DingTalk classroom agent from two months back for comparison ， "Boss Li, does mine count?" ， and gets a crisp verdict: "Not really, this is a different beast from DingTalk's." ＆ digs up evidence that the beta actually started June 11 and fumes: "Outrageous, how did I miss this."

门*** breaks the news first: "Agents can now join Feishu meetings!" Xiangrui watches the official demo and is floored: "This is ridiculously good." He muses that if an Agent could ever walk through screens step by step while teaching, "then I'm done for ， online lecturing is in mortal danger." At 12:09 he wails, "The Agent just talked and my screen recording had no audio," while 炮 speaks for every onlooker: "We just want to see that agent join a meeting." At 1:49 PM Xiangrui files the official record: "Insanely good! Just now, an Agent spoke up in a Feishu meeting." 王长春 imagines one-on-one A* tutoring ， "the education industry gets turned upside down" ， and 薛** offers living proof: plenty of teachers can't explain Midjourney workflows clearly, "but codex explained it to me crystal clear."

R*** discovers that after hooking workbuddy into WeChat, forwarding an article to its chat window is impossible ， you have to copy the link, back out, and paste it in. "So clunky," she sighs, then cracks up the whole group with "This is what refusing to treat A* as a person looks like." 卡* twists the knife: "You can forward straight to Yuanbao, but not to Claw." xixi shares her workaround: stash things in her ima personal knowledge base first, then have workbuddy pull from it on demand. In the afternoon K*** asks for voiceover tool recommendations for a 4-5 minute A* training video for management trainees; Xiangrui probes the use case first, and p*** offers the shortcut: just clone your own voice.

In the evening 门*** posts "Big-Tech Departure List: Another Feishu VP Leaves to Build Agents," and Xiangrui names the departed: "That's the head of Base." At 9:07 PM Xiangrui posts a picture, 炮 fires the first meme ， "Master: I've been thirsting after you for ages" ， and 袁**, ＆, K*** QZ 张, and q*** relay it six deep, flooding an entire screen. Jokes aside, 炮 closes with something sincere: "The way the group owner lives right now is the best possible state. Join an organization and you might never be this alive."

观*** files a live review from the Beijing venue: "Today's OPC festival feels kind of watered down," "nothing like the vibe in Shanghai," then turns to Xiangrui: "I'd honestly get more out of reading your articles." B*** picks up the thread: most attendees are really content creators, and true OPC remains hard ， understanding business needs and navigating business communication is a messy, deeply human problem. Xiangrui pulls out his own OPC retrospective: people mistake OPC for a media hustle, when the real value of offline events is meeting the heavyweights face to face and colliding ideas. 炮 nods along: the shallow, visible layer is the easiest to see; the real industry practitioners are the rare ones.

杨** lobs a string of hot takes into the night: "The window to make money from A* content is closing," "below the 985-tier universities, everyone's basically been leveled," "the capitalization of brain labor ， that is the essence of A*." Xiangrui pushes straight back: "I have to dispute this one." Would a 211 undergrad and a vocational-school grad really wield A* the same way? He plants his flag: "A* hasn't made everyone p***ful. It has simply made the people who already had ability even more p***ful." 炮 seconds it instantly: "A* is an amplifier, not a crutch." q*** delivers the quote of the day ， "A* is not techno-sorcery, and A* is not a wishing well. It will not do your growth for you" ， then earnestly dissects what "capitalization of brain labor" would actually mean. 炮 rides the momentum into a confession: he's already betting big, run*ing development projects with two or three people that used to take thirty. "Tokens burn every single day." "Work long enough and the whole thing starts smoking."

炮 drops an industry truth bomb: companies clearing 100 million a year in revenue are run*ing their business data as dozens of mismatched Excel files, hand-merged by humans. Xiangrui matches the signal instantly: "I talked to a client exactly like that today ， over 100 million in an*ual revenue, managed through dozens of Excel spreadsheets," a Beijing firm doing school renovations. 小* presses on how you keep the revenue flowing, and 炮 unpacks his bottom-drawer three-layer playbook, then ， egged on by the group ， posts his own SaaS results page. His starting point last year was the same as everyone's: "Base is amazing, I can hand-roll a CRM." Only after actually building did he learn the real key: clients pay when you deeply know the industry **and** can hand-roll the CRM. Xiangrui extends an invitation on the spot: "Come share this in the community next week" ， starting next week, the community will host talks from the heavy hitters. 炮 accepts without hesitation: "Whenever you call, I'll show up and back you up."

Xiangrui shares a number: the community is closing in on 72,000 members, up from just over 60,000 half a month ago. K*** replies with two words: "Downright terrifying." Xiangrui then tells a story about turning money down: a savvy Xiaohongshu creator team had been posting Base content since last year and had already earned over a million through other Maker collaborations, but their get-rich-overnight framing repelled him ， "so I politely declined in the end." E*** pronounces his values "absolutely solid," and 炮 names the deeper truth: "Of all these communities, this is why people are only active in yours." As the night deepens, the mood shifts to tool grumbling: Xiangrui has spent two days deep in Codex ， "painfully slow, hallucinates constantly, and the context window is tiny" ， 王* posts a bill for a billion tokens burned in a single evening, and everyone says goodnight amid the smoldering quotas.

Xiangrui opens at 9:01 with two screenshots, and 庄* immediately gets in close: "lucky are those not in Hangzhou." 天* follows with "even overseas I don't feel like using Claude ， GLM 5.2 is plenty," l*** adds "I'm on a business trip in Hangzhou and haven't dared touch Claude since yesterday," and K*** offers the workaround: "hold the line with DeepSeek for now." At 9:09, 炮 posts the an*ouncement that Fable 5 is unrestricted worldwide, and 庄* replies "that codex move is ruthless." At 9:40, 归* delivers the deadpan ， "ban the accounts first, then unrestrict ， nice" ， and K*** piles on with "and then you all get to top up again," "648, 648," wrap*ing the domestic-model morning into a bit with a crisp snap.

At 10:01, Xiangrui posts two images and floats the topic: "keep an eye on Feishu's Demo Day meeting tomorrow morning ， the official folks say an agent will be joining the call." R*** jumps right in: "so curious about this agent attendee ， blind guess on what the scenario and features will be." 炮 cracks the joke ， "will 牛老师 distill himself so an agent 牛老师 can show up?" ， and the group's beloved "distillation" run*ing gag lands once again. 小* asks where the entrance is, J*** instantly drops the larkoffice meeting link, and K***'s "clicked!" turns the preview into tomorrow's shared agenda.

At 14:31, Xiangrui posts a video captioned "a video one-click-replicated by A*," adding that the tool is libtv. 张* follows with "how did you replicate it?" then drifts off-topic ， "I've been shop*ing for glasses lately; which ones are actually good?" ， and the group thinks an A* product-shoot tutorial is about to begin. At 15:11, 麦** posts a glasses commercial, K*** gasps "is THIS made by A*???", and h** punctures the moment: "no, that one's an official live-action shoot." K*** immediately laughs at himself: "I saw Xiangrui say 'A* one-click replica' above and assumed the whole thing was A*-made." A while later, Xiangrui posts the clarification ， "I sent the wrong video; that was the one I meant to use as reference" ， a small crash the group owner both caused and cleaned up himself.

At 16:20, Xiangrui posts yesterday's group daily, then follows at 16:39 with last week's group weekly. 炮 opens with a light jab ， "oh, there's a weekly now~" ， then calls out the bug: "@李祥瑞 the weekly's QR code is your personal homepage, dear." Xiangrui posts a fresh screenshot ， "how about this one" ， and casually reveals the hidden entrance: "actually, inside the group daily there's a little triangle to the right of the words 'group daily' ， tap it to switch to the weekly." 杨* lands the user's-eye critique: "too many headings, too much text ， by the back half it's easy to lose patience." The feedback drops straight into the question of what to trim in the next edition.

At 19:32, 吴* starts with a small question ， "can fields in Base be center-aligned?" ， and Xiangrui and Q** answer flat out: "no," with Xiangrui adding "WPS's version can." A minor matter, until 周** ignites the thread: "however many tables you have, that's how many Agents you can conjure." 门* rides the momentum ， "replace those HQ middle managers who never visit the floor; all they do is stare at data for insights. Now a frontline worker can conjure themselves a boss" ， then pushes further: "layer one, use A* to conjure yourself an underling; layer two, conjure yourself a manager." 周** takes the baton ， "conjure a customer, conjure demand, conjure consumption ， that's what really matters" ， and 门* laughs himself over the edge: "you can't conjure consumption. Once the bosses are conjured out of their jobs, who's left to consume? Hail a cab, and your boss takes the fare." 李** seals the bit ， "they can go straight onto unemployment benefits, and benefits are rising these days" ， while 炮 tacks on the reality check: "a token-devouring monster, but honestly probably pretty good."

小* finishes testing and lays his conclusions on the table: "run a script locally for daily queries ， permissions fully controllable, and it's faster. The high**ght is the dashboard pop*ing out of a message card; once data gets big, tables and plain text just aren't intuitive." He adds: "why should Base take that long to compute? I use the CLI with a local snapshot plus a fixed analysis template ， instant replies, zero tokens." Xiangrui concedes, "I've been in the beta since late May and still haven't found a good use case," then tosses out, "isn't Feishu CCLI just more convenient?" 周** chimes in ， "people who can't use Base won't need to open the table anymore; just use the agent" ， and 平*** agrees: "I run local scripts too; data staying local feels more compliant." 万** surfaces with a thought from the future: "maybe eventually we only store, never display. No need to open the Base interface at all ， it all lives inside the agent's head and body." 卡* keeps his eye on the bottom layer: "it still burns Feishu A* credits," "the Feishu A* workspace is not messing around."

杨** throws manufacturing's old ledger onto the table: "Tencent built an industrial-floor agent and their showcase example was recognizing handwritten forms ， speechless," "manufacturing companies run 20–30% non-production staff. A 3,000-person company has maybe 600–700. Which of those can your agents actually cut?" Xiangrui follows with his read: "a lot of big tech companies genuinely have no idea what real business scenarios look like ， plenty of influencers too. It's all hype; the articles say 'deployed,' but nobody's thinking from the production line's point of view." 周** offers a prescription: "have that Feishu product manager transform for a bit ， go work inside an actual company for 4–6 weeks, then come back seasoned." 门* hammers the conclusion in place: "whether anything lands on the production floor always comes down to accountability over efficiency ， someone has to be willing to take the fall before anyone dares to deploy," "draw the lines of authority, responsibility, and benefit clearly, and any frontier tech can walk right onto the line." 周** adds "one person works, one person reflects, paired one-to-one," "between fast tempo and slow reflection there has to be a rotation" ， closing the segment on a real-world echo of Thinking, Fast and Slow. 杨** lands one more: "the most widely deployed A* in production environments is vision models," and 平*** notes, "the people who understand the scenarios best are frontline operators ， who have no time or energy to study any of this."

At 22:07, mid-debate, Xiangrui slips in a message with the warmth of a private note: "@宋** feels like ages since I've seen you, haha ， still prep*ing for finals? I knew the group was missing something lately" ， the group owner's usual headcount. At 23:20, 周** drops in an article about a humanities PhD who quit her program to become a repair technician, pulling out the human line: "the hardest part of the process wasn't the physical labor, nor the technique itself ， it's that the masters usually only tell you how to do it, never why." 平*** adds "I've tested this ， the generated results are pretty mediocre," and 杨** carries it forward: "Tencent built a whole panel of these skill experts as agents ， it comes down to sustained investment in that business." The next morning at 8:16, Xiangrui sends an @everyone ， "anyone heading to the global OPC Co-creation Festival today? We can meet up in person" ， cutting a whole night of verbal sparring over to boots-on-the-ground plans for the day.

B***, 炮 and 陈** strung together a full round of "help more people grow strong through A*" taps as the morning warm-up; 庄* said "you're on TV," and Xiangrui came right back with "I'm deep in it ， tomorrow I'll write a real usage piece, something you can't find anywhere online yet." p*** drop*ed his core take on public-domain conversion: "in the public domain, you have to assume everyone who sees your content knows nothing about you; you keep building the potential user's context through your content system, and once it crosses a threshold you get a qualitative jump. The first path is win first then fight; the second is fight first and hope to win." 归* asked, "did the group owner build this with A* himself?" and Xiangrui replied, "yes, built on top of our knowledge management architecture ， the Agent gets your context fast and you save tokens on top of that." The fuse for the day's GEO conversation was already lit.

炮 drop*ed the key line: "grabbing context is easy technically, but the real question ， and the real commercial value ， is how you distill genuinely valuable knowledge from what you grab." Xiangrui followed with a long reply: "group chats, public accounts, Feishu knowledge bases, courses, external talks, personal sites ， on the surface it's all content, but underneath, it's all sedimenting one set of context: how others (and Agents) understand me." Then he pushed it further: "SEO used to be about making it easy for search engines to find you; GEO is more about making it easy for A* to understand you, cite you, and recommend you when it answers a question." 李*** asked, "if you publish on WeChat public accounts, isn't that hard to retrieve?" and 麦* delivered the pivotal call: "the sources every large model cites are different." Xiangrui finished the picture: "Yuanbao is going to lean on WeChat public accounts and the WeChat ecosystem; Doubao is going to lean on the ByteDance ecosystem ， Douyin, Feishu. Doubao frequently cites content sitting inside Feishu knowledge bases." 炮 wrap*ed it in one line: "Obsidian is your personal long-term retrieval context; what you put out in public is fundamentally about building the big context for the whole internet." At 10:28 Xiangrui drop*ed the whiteboard recorder whiteboard.xiangruiai.com.

R*** kicked it off: "I'm going to get my incoming-freshman little sister onto Obsidian the day school starts," "gotta win at the starting line." 炮 followed: "hard to even picture how brutal a four-year Obsidian vault looks when the thesis rolls around." Xiangrui was tickled: "I think you could run a course ， get freshmen prep*ing for their senior thesis on day one. College students' money is the easiest to make anyway." 醒* added a cold splash: "food, drink, and fun are the easy money ， study money is a grind. Selling to college students means going for volume." Xiangrui pivoted: "no no no, you need a hook ， you gotta pitch it as: use this and you can run a side hustle. Every college student today wants a side hustle." 醒* then steered the thread to the actual money-maker of this year: "teaching clients how to actually deploy A*," with 炮 chiming in: "that's absolutely the great business of the entire year," "open-source puts the shovel out there, not the wheel ， the wheels you build in-house for specific clients with budget."

平*** drop*ed the media piece "Doubao Quietly Beta-tests Social, and This Time It's the Road Douyin Never Made Work," and complained "they didn't even properly redact the mosaic." Xiangrui led with his read: "this isn't Doubao doing social ， it's completely different from Yuanbao's social play at the start of the year. This deserves a proper write-up; everything online right now is screenshots and speculation." On the high-speed train Xiangrui muttered "damn, mine's gone," and E*** jumped in with "backend just turned it off in a hurry," but after refreshing for a few minutes Xiangrui confirmed the feature was still there: "I feel like I'm the only person on the entire internet actually using it, ha." q*** asked "has V***'s subscriber count gone up?" and Xiangrui answered "nah, not really." This stretch is the exclusive scoop the head chef is still digging up.

Xiangrui floated "is it possible Feishu is aiming for consumer through Doubao?" 李** rolled with it: "I'm about to be out of a job on my end ， I need to look at new options." 大*** wondered "isn't manufacturing sup*osed to be relatively hard to disrupt?" and Xiangrui replied, "definitely getting hit ， in 2024 I did commissioning at Ningbo Volkswagen, everything fully automated, headcount getting optimized down." 李** filled in: "the new production lines are essentially all automated; clients won't even accept semi-automated ones anymore. QC is all A* QC ， Tencent is out here buying that API, paired with Hikvision industrial cameras." 大*** conceded, "used to feel blue-collar was the easy replace; now with A* moving, white-collar is even less stable." 万** delivered the verdict: "humans-as-tools will eventually be replaced by A*. Judgment, strategy, taste ， those are the last territory left to humans." 松*** followed with a couple of pitch-black jokes: "once most people are already replaced, all we can do is roll out a minimum-sup*ort program ， or grab enough on this wave to rest." Xiangrui closed the whole arc with a Douyin-style one-liner: "vibe coding is a bit like the digital camera ， once digital cameras existed, anyone could snap a picture cheaply, but the ones that became actual photography are still few and far between."

醒* drop*ed the link vibecoding.aigongxue.cfd/htmls/bailuyuan/, "processed two books, spent half an hour on bug fixes." Xiangrui was immediately curious: "how'd you do that image + text interactive thing for Bailu Yuan?" 醒* replied, "the effect is wild ， the character analysis is really strong, and there's a timeline too." 松*** confirmed it opened fine in WeChat desktop. 醒* tossed in some self-deprecation: "in the A* era, chasing the details eats time ， A* can already do 90% of the work," and added, "it really should have a Qinqiang score in the background." 薛** came in after reading and asked, "where do I read Bailu Yuan?" and 醒* said, "scroll up a bit." This stretch was the most concrete "A* drops a whole novel onto your phone screen" moment of the day.

天* asked, "is there a Feishu account manager in the group? I want to ask about on-prem Feishu." Xiangrui said it straight: "don't even think about on-prem Feishu unless you're a really big company. As far as I know, in a whole year only two or three companies get an official on-prem deployment." 考* had just gotten a number from a consultant: "4 million yuan," and 炮 quip*ed, "4 million to deploy Feishu ， honestly that feels pretty cheap." 天* pushed on whether they could deploy overseas: "we don't want to use Lark anymore," and 考* said, "just call Feishu directly." 李** tied it up cleanly: "Feishu right now doesn't sup*ort customization and doesn't sup*ort private deployment. If you want your data to stay on-prem, Feishu has a solution ， but a copy still has to sit on Feishu. This isn't a question of how much money you throw at it, it's because they're going all-in on A* right now. Your on-prem version will always lag Feishu's version by a lot." He added a jaw-drop*er: "we're spending one to two million yuan a year on Feishu, and they simply don't keep our data on our side. Companies like NIO and Li Auto spend tens of millions a year ， they're on the SaaS version too."

In the evening Xiangrui suddenly said, "I just noticed a product lead on Feishu Bitable has left ， a lot of the people I know have left." 炮 vented: "several teams racing internally, and the standalone Feishu products aren't good enough." 卡* delivered the call: "Feishu has positioned itself as an A* work platform; Bitable as a standalone product doesn't match user needs and doesn't cooperate with the suite. Doubao and Feishu aren't in the same BU ， they're racing. Coze is racing against both of them too." 杨** added the long view: "how Doubao 2.1 desktop rolls out will decide whether Feishu retreats back to being a communication platform, with the desktop handed over to Doubao." Late at night around 23:00 Xiangrui complained Codex was slower than CC: "Codex spent several minutes on typesetting and still hadn't finished ， Claude Code did it in under a minute," and 炮 groused, "today I asked Codex to turn my PPT into a video and waited 30 minutes, only to get back an HTML with some transitions." 平*** drop*ed the degradation radar codexradar.com ， "they've all been degraded" ， and K*** asked for a Codex-with-Feishu installer, with 小* promising "I'll check on my work computer tomorrow." At 23:16 Xiangrui published the official article "I vibe-coded an open-source whiteboard recorder ， simple, handy, and I'm going to shoot videos with it," adding "haven't written in ages, folks, give it a like" ， a full day of tool buzz and methodology debate finally landing on one open-source tool anyone can pick up. In the early morning at 08:32 Xiangrui pushed a follow-up public-account post with just "let's put the whiteboard recorder to work," opening the first day of July.

In the morning Xiangrui first patched last night's Doubao-adds-friend thread: 'When you use your Doubao account to add someone as a friend, you're the human ， Doubao is more like the lightbulb between the two of you.' At 10:30 q*** forwarded 'Twenty million views: I built a Karpathy-style second brain with Claude + Obsidian, absurdly good,' remarking that Xiangrui really is on the frontier. 炮 immediately added the key point: 'OPC really needs Obsidian ， especially when you write rules for CC or Codex to precipitate info into OB, otherwise once info piles up the person collapses.' K*** kept the mood with 'Just follow Xiangrui.' These are the three lines that set Monday's tone: Doubao is a shell, the real moat is your own knowledge base, and follow the head chef so you don't fall behind.

Xiangrui drop*ed 'Real scenarios, real offers, real implementation! The 2026 A* Pioneer Future Talent Contest opens for registration' into the group. 张** quip*ed, 'No chance to join, @李祥瑞 give us a contest of our own,' and Xiangrui responded directly: 'Late July, maybe ， late July we'll hold our own offline event in Guangzhou, everyone's welcome.' 炮 pulled out the samples he's already ship*ing: 'I came out of the auto industry; the game is run*ing half a body to a full body ahead of the leaders ， that's how you leave room for yourself in the market.' In the afternoon 李** added, 'Feishu is run*ing an FDE campaign ， it's about being domain-savvy plus doing A*,' pulling the contest and the FDE hiring rhythm onto the same thread.

Xiangrui posted 'Updating Mac for Codex' and lit the Mac-price-hike fuse. 炮 drop*ed the full new price list ， Mac mini 4499–5999, MacBook Pro 13499–15999, Mac Studio 16499–19999. q***, who'd been eyeing 618 all along without pulling the trigger, blurted 'I'm mad'; 平*** shot back, 'Do you even need a new one with your current setup?', and q*** confessed, 'M1 8+256'; QZ 张 chimed in, 'I'm on NEO 8+512.' 王* started airing his self-funded workday: 'I use my own Mac to remote into the Windows box at home, then bought a Windows box to control the state-owned info-security machine the company issued. Self-funded office life; even the tokens are out of my own pocket.' QZ 张's one-liner: 'New-era beast of burden.' 小* rolled out his '24 M4 mini for 3800, a 24g+512 MBP, and a 10-plus-year-old trash-can Mac, 'they all still run'; K***'s M1 8G is likewise 'still cranks out Codex just fine.' It's the group's signature move ， if prices go up, we just tough it out longer, rationalizing frugality together.

At 11:30 Xiangrui said, 'In a bit I'll open-source a fun recording tool for everyone,' and later explained that the underlying whiteboard was already open-source ， 张咋啦 had previously bolted on a recording feature: 'His recording feature wasn't quite complete, so I fleshed it out.' At 13:45 he drop*ed the next one: 'Lots of you use Macs, and many need WeChat multi-open ， sharing a skill: `npx skills add joeseesun/wechat-multi`. Just tell the agent WeChat multi-open, or ask it to open several WeChats, and it runs automatically. It's a skill 乔木大佬 open-sourced earlier.' K*** immediately asked whether it triggers risk control, and Xiangrui replied he's been using it since March without hitting any. K*** then pushed the sharper pain point: 'What I actually want is the same WeChat logged in on different devices ， can that be solved?' 王* said, 'Right now only DingTalk sup*orts that.' 李** corrected: 'Feishu can too. For WeChat you'd need different protocols ， Android phone and Android tablet count as two protocols; there's an open-source project that simulates the web login.' 平*** added, 'Run*ing WeChat in Docker is another route, but in theory it might get you ban*ed.' In the middle q*** said, 'Why do you have so many fun stickers?' Xiangrui rifled off twelve emojis in a row; q*** cracked, 'Cracking up on the toilet' ， that bit of silliness became the day's most relaxed interlude.

Xiangrui showed off the Doubao Enterprise beta entry point, saying he's finally trying Doubao Enterprise: 'It's split into an enterprise edition and a regular user edition.' T*** asked, 'How is Doubao Enterprise different?' Xiangrui answered crisply: 'Doubao Enterprise is wired to Feishu ， you can log into Doubao with your Feishu account; inside Doubao you can query anything within your Feishu permissions; right now it runs Doubao's latest model, and it's free.' 谢*** asked whether you have to buy the Doubao Pro plan to get the switch, and Xiangrui replied, 'No ， Doubao Enterprise is in closed beta.' 李** drop*ed the key intel: 'During the beta, Doubao Enterprise doesn't consume A* credits.' Xiangrui pasted the ap*lication link (doubao.com/in*ovation-lab/enterprise/trial-ap*ly), 'about 10 minutes and you're in.' QZ 张 later shared a screenshot confirming 'the drive is Feishu's drive' ， a full beta pipeline had just gone live in the group.

Xiangrui asked, 'Anyone in the community familiar with FDE?' 天* answered, 'Forward-Deployed Engineer ， I want to hire one.' 北*** said, 'The concept just emerged; talent is hard to find.' 李** added, 'It's about being domain-savvy plus doing A*.' Xiangrui gave the definition: 'Basically, FDE is A*-era outsourcing.' 炮 pushed back: 'Domain-savvy plus A* implementation ， why would I take an outsourcing gig? Domain-savvy + A* implementation = OPC, and the kind that monetizes directly.' 北*** added one sharper: 'More likely FDE is the boss of the domain company himself ， the CEO is the FDE lead.' Then 木* poured cold water from a longer view: 'A*'s direction is inexorably less human involvement ， what you build this month may be swallowed by a platform update next month; the trend isn't the tool-usage layer, it's the domain-deep-cultivation layer.' Xiangrui landed it clear-eyed: 'Short-term we still have to prioritize monetization ， the community has plenty of Bitable-build requests, all of which can be one-off project deliveries. Long-term, yes, aim at the domain-deep-cultivation layer, but don't ignore what's in front of you just because you're looking far.' 炮 tacked on, 'Monetization matters ， don't starve and have to go back to a day job.' A weighty direction-setting discussion landed.

Xiangrui posted a screenshot of 'Doubao responds to closed-beta social feature rumors' with a quip: 'Whoa, my Doubao can add friends now, ha,' followed by, 'No way ， tomorrow I'm writing a real piece on this.' 薛** patted him on the back: 'Let more people be strengthened by A*.' 平*** noticed every media piece was using Xiangrui's screenshot: 'Do we have journalists in this group?' 炮 tacked on, 'There are probably plenty of ByteDance folks in here.' At din*ertime 杨** delivered the verdict: 'Doubao isn't building social ， it's wiring to Feishu to borrow its enterprise relationships. Feishu won't extend its A* reach anymore; it'll likely hold the boundary at OA and documents. Skills have no protective moat ， like articles, they spread fast, unless they're enterprise-specific.' Xiangrui shot back at length: 'I agree Doubao isn't building WeChat-style social ， it's more like borrowing Feishu's enterprise relationships and organizational identity. But I don't agree that Feishu will only hold OA and document boundaries. The general-purpose A* entry point can live in Doubao, but for enterprise A* to actually execute, it still has to come back to permissions, documents, knowledge base, Bitable, ap*rovals, and business processes inside Feishu.' 'Feishu's key isn't OA at all ， it's Feishu's collaborative-office capability. Feishu, WeCom, and DingTalk have completely different centers of gravity: Feishu focuses on collaborative office, WeCom lives inside the WeChat ecosystem, and DingTalk is the OA of traditional enterprises.' q*** added, 'Doubao going for social is like Douyin trying it ， there's no soil for it.'

In the early evening 炮 asked, 'Is Codex laggy for everyone today? I've switched a bunch of nodes and it's still slow.' 平*** immediately tip*ed K***: 'Codex has a bug that madly writes to your disk ， check whether your disk writes are already several TB, could be tens of TB.' h** offered a precise debug prompt: 'Help me check whether ~/.codex/log_2.sqlite is continuously writing at high frequency because of TRACE logs.' K*** ran it live and pasted Codex's self-diagnosis back: 'Doesn't look like a DB-corruption bug ， more like Codex's local logging strategy / TRACE-level leak / WAL over-management.' 小* added, 'Just update and it's basically fixed ， the desktop side wasn't affected.' At 9pm 张** grumbled that Doubao's 200-yuan paid tier is a real bad review ， 'Can't finish a single task' ， before turning around and using it to finish the school teachers' digital-profile buildout. At 10:30pm Xiangrui showed off the teleprompter that finally worked: 'The teleprompter follows your speaking pace and auto-detects it.' 张* and E*** cheered him on; E*** added, 'This is perfect for popular-science training ， the audience finds it intuitive, and the speaker feels much lighter,' and reminded him, 'You can finish recording the Bitable course now.' Xiangrui landed on a flag: 'In July, start recording videos ， record anywhere, anytime.' Past midnight, tomorrow's open-source is already written and queued to ship.

Sunday morning 炮 drop*ed two screenshots first, and Xiangrui followed with two more sets of his own, cap*ing them with the line "The polished notes now render HTML properly," a status report from the workbench. B*** answered with two "nice"s, and Xiangrui added, "The context A* carries automatically is really accurate now, and it saves a lot of tokens too." This is the group's signature opening: get the system run*ing, then boil the observation down to a single sentence. Before most people are even awake, the head chef has already plated half of today's dishes.

At noon Xiangrui posted a screenshot of a screen-recording tool and asked if anyone had tried it, and 旺* replied "Nope." 悉* calmly drop*ed in a book: The Token Era: People Who Know How to Use A* Are Collecting Their Fourth Paycheck, saying the book is precisely about how workers can create more value with fewer tokens and how companies should be the ones handing them out. K*** added two more images in the afternoon, extending this quiet thread into the big debate later. This stretch had no boisterous back-and-forth, but it put the question of "who pays the token bill" onto the table in advance.

In the evening h** tossed out a very concrete question: "If claude bans an account, how long is this machine unusable?" 门*** took the opening and posted a long take: it's reasonable for management to buy tokens out of their own pocket, but not the frontline, because frontline data work is already selling time ， doing it faster only earns you more work. 小* pushed back: "So the project born from your own money ， does it belong to the company or to you?" and 门*** answered strictly, "Your salary has already bought out every output produced in work hours." 炮 chimed in: "You trade tokens for your own time; what the company wants is output." 万** cut to the bone ， this is the invisible pay cut that comes with intensified competition. 门*** closed with a story from his own days as head of general affairs at a state-owned enterprise: only when the boss has plenty of meat does the layer below get any soup, and then he quit. 小* zoomed back out with one last angle: making executives more efficient is what the boss should want in the first place ， "If they won't even cover the tokens, how is that any different from not giving you a computer at work?"

As the discussion was burning out, 炮 offered a colder verdict: "A* showing up will absolutely squeeze 90% of people to benefit the other 10%. What you're competing with isn't A* itself, it's the other people who've made themselves more efficient with A*." 门*** immediately fired back an important SAP signal ， 90% of implementation consultants are going to be replaced by A* ， and pulled the thread over to a concrete product question: when he was building an A* recruiting workflow, the step where the user configures a Feishu enterprise self-built ap* was too heavy, and he asked whether there's a way, like other agents do, to authorize by QR scan and auto-create the ap*. 杨** answered with a question: "Are you building this to run standalone, or inside some environment?" ， pulling the topic back from the grand narrative to the engineering floor.

Late at night Xiangrui fired off several screenshots and the line "Whoa, Doubao can now add Feishu contacts," and the group lit up. h** immediately declared, "Now I'll actually pay for a Doubao membership," and A*** followed up asking how to add and whether he had to pay; Xiangrui answered that the free Doubao Pro tier has a daily quota that's plenty. He had done it by asking the Feishu CLI inside Doubao to create a Bitable, and then adding himself in as the owner. 炮 groaned, "That pipeline is longer than a family reunion," and A*** summed it up: "Doubao and Feishu have finished commercial integration." Xiangrui shouted "Come add me, let's play with it," while noting that inside Feishu you can't see what Doubao did, and offering a prediction: "This might end up feeling like the early-year Yuanbao party or the Wenxin Yiyan group chats." By midnight he was already experimenting with creating group chats, even copying the "it's actually Feishu" QR code out of Feishu to lay everything bare for the group.

王* was first to report, "Right now scan*ing gets absolutely no response," and switching to Feishu to scan didn't work either. 平*** added, "Doubao scan says this feature isn't sup*orted yet," and then drop*ed the conclusion: Xiangrui really is in the gray release cohort. Xiangrui wasn't buying it and asked everyone to post their Doubao IDs so he could add them himself: 平*** was mnmymz, 王* was waboy but not a member. Xiangrui sent the request, the other side didn't receive it, and he suggested they log into the Doubao-account version of Feishu on a desktop browser, explaining screenshot by screenshot that "the Feishu-under-Doubao version is really feature-poor." 谢*** chimed in late at night with a reference point: "On 3.1.6 I tried scan*ing a Feishu QR code with the Doubao client and got zero response." The conclusion was clean: the feature has really only been drop*ed onto a handful of machines.

Right next to the Doubao-add-contacts frenzy, 杨** coolly landed several cuts. First he said Bytedance is trying, via Doubao plus Feishu contacts, to make one more bet on social ， "way too hard, even Douyin couldn't pull it off" ， and added, "Doubao PC will target WorkBuddy and Codex next, but the move is too slow, and Coze has ended up neither fish nor fowl." He even admitted he bought a Coze membership and refunded it a few days later. 北*** rounded out the picture with four quick takes: "left hand fights right hand / it's Doubao and Coze / and Trae / meanwhile A-Corp really only has cc." 杨** zoomed further out: software that used to take real effort is now a one-button job; the NLP knowledge-graph consulting project he did a few years ago is dead now, and in the last few days he had cobbled together a knowledge-management-plus-training mini-program with WorkBuddy ， "the SaaS software industry is basically done," with the last cut aimed at Feishu itself: "From the A*-native-ap* angle, Feishu is really only good as a spreadsheet now." 王* kept the bot-layer thread rolling with, "There's a whole zoo of bots inside Feishu."

The next morning the talk scattered around Coze: Q** said he'd bought a year and couldn't bear to refund it, still keeping it around to scrape a bit of data; 王* said his own use case had been replaced by Feishu bots; 木* handed WorkBuddy a hands-on review ， "they actually didn't cut corners on the model layer, and the feel of using it is really okay"; 万** grumbled about Codex opening and resetting itself. Every tool-war ended up landing on 李**'s shortest possible verdict: "You just can't use it up." The token debate, the Doubao experiment, the SaaS verdict ， everything came to rest on those five words.

In the morning Xiangrui drop*ed a small win first: a stable way to run Claude Code on Windows is now in hand. The group wasn't fully awake yet; q*** caught it with a single "oh-ho," and 於*** sent a half-line message that looked like a test ping. By a little past 11, Xiangrui posted screenshots saying yesterday's group daily and this whole week's dailies had been updated. It wasn't a noisy big debate, but it felt very much like this group's baseline: get the system run*ing first, then catch the records up, so that both the tools and the content can keep ship*ing continuously.

At noon Xiangrui sent two screenshots, with a single "oh god" pulling everyone out of the woodwork. h**'s first reaction was "I thought I'd walked into the wrong group," and 陈* typed out a question mark. Xiangrui explained it was content from a Feishu group; K*** picked up the joke with an emoji, while B*** nailed the dilemma of public screenshots: "Get scolded if you don't redact, and scolded if you do." This bit played like the backstage cut of the group daily itself: content wants to spread, privacy wants to be protected, and no matter how you handle it, someone will have an opinion.

The thread turned toward Feishu starting with "Doubao has gotten this expensive." 李** said his team had ap*lied to bring it into Feishu; l*** asked how, and Xiangrui handed over the path straight: click the workspace, click the Ap* Center in the top right, search for Doubao, get the ap*, then configure the related permissions. K*** asked whether plugging into the Volcano API would do it; Xiangrui corrected that this was more like installing the desktop Doubao client inside Feishu. The group then kept unpacking tenants, accounts and entitlements: 李** said his own Doubao came "in a suit" because he had ap*lied through an account manager, while 杨** put WorkBuddy, Qoder and Kimi Work on the table for comparison and reckoned the differences were small for daily use and light dev work, with the gap only opening up on complex tasks.

In the afternoon 朱* noticed that the shared ChatGPT Pro account at her company had basically nothing but PPT traces left in it, and warned that this kind of thing shouldn't be casually used for any important report. 杨** lifted the conversation to enterprise processes, arguing that A* rewiring business processes was the real value, and that one day a PPT might just be a skill living inside the company that you open and use. At 15:59, Xiangrui finally posted the win: "Finally got it working inside Doubao. Doubao is calling the Feishu CLI directly, using my own account, not Doubao's in-environment Feishu tenant." 林 immediately followed up on the permission piece, and Xiangrui gave the recipe: hand the Feishu Open Platform bot's ID to Doubao, let it return an auth link, then sign in via the link, with the catch that for now it only works inside one conversation.

郑** suddenly posted a screenshot asking "did I just win the lottery," looking like an account login had gone sideways. 小* said try another account first; h** asked whether the email could still sign in; 平*** judged "feels like a VPN issue to me, a real ban comes with an email and even a refund." 李** added that local file map*ing would still let you carry on. A few minutes later 郑** confirmed the web could log in, his membership was still there, and after swap*ing his route he was finally back. He summed it up himself: "a begin*er can't take this kind of fright," and 小* followed up with "the side effect of playing too much Claude."

As dusk fell Xiangrui posted an image, and 小* followed with "strength has a cost," then "one general's glory is built on ten thousand bones," pulling the mood from tools toward organizations. 平*** argued team collaboration still mattered because any one person's brainp*** is limited; 小* countered that in reality there are no teams, only a brawl of departments. 卡* spoke from an Alibaba lens about incentive design, saying it's always better to incentivize and incubate employees than to keep demanding from them. Late at night 炮 condensed the thread into an operational call: have a single engineer cover R&D, and let the team round out with other intelligence, product and marketing, forming a small s***-team where every member is backed by multiple agents.

At dusk Xiangrui forwarded the Tsinghua president's commencement speech, with A* density in the title so high that 王* sighed that Tsinghua was riding way out front this round. 王* then complained that his own machine had three copies of the Feishu CLI: one auto-installed by Doubao, one he had pulled himself, and one Claude had pulled too; 林 suggested putting the skill in one place and symlinking it from elsewhere. By early next morning, Xiangrui forwarded "the white-collar worker who can't afford tokens is taking an invisible pay cut," B*** caught it with a face-palm, and Y** landed the punch**ne: "paying out of pocket for a sense of crisis." Finally l** asked whether at runtime you could choose between the company's API and your own, and the day's tool buzz all ultimately came back to the same question: who is paying the token bill.

R*** opened the morning chasing 'why isn't 5.6 out yet? yesterday was sup*osed to be the day,' and A*** backed her up: 'after the price cuts yesterday, it should be close.' 李** laid out the truth ， the Trump administration is making OpenA* roll out GPT-5.6 in stages on cybersecurity and national-security grounds, starting with a small batch of partners and enterprise clients with US government ap*roval case by case, and that's why nothing has ship*ed. A*** pressed, 'and Gemini gets a pass?' 李** replied that Google's pro has to clear ap*rovals too, 3.5 can't ship either: 'with review it just goes slower, but commercial companies want money, they'll ship eventually.' 姚* chimed in with a forecast: 5.6 ships before July 31.

R*** fessed up: 'first time I've burned my weekly limit to 0%, and I wasn't even doing anything huge.' 赵** shared his frugal playbook ， long contexts burn faster, so he stashes progress into a knowledge base and shoves the dirty work over to WorkBuddy, where he's also milked two free accounts on the side. R*** bit: 'then I'm going farming too,' but did the math: 'WorkBuddy 58/month gets 2,000 credits, one conversation burns 20–30 credits, 100 rounds and it's gone.' H*** and 赵** tag-teamed the credit-farming bible ， daily tasks pay 150 credits a day, one account can pull 6,000–10,000 free, and going to 58 adds another 2,000 on top. 赵** closed with hardcore math: '1 credit is about 31,874 tokens, 5,000 credits is roughly 500 to 900 million tokens.' 炮 shouted from the sidelines, 'upgrade the plan, throw money at it!' R*** laughed, 'I sent invites today, waiting on the revive ， next time I'll just upgrade if I run out.'

万涂幻象 drop*ed a piece of news: a ByteDance Lark product VP has left to start a company. The chat instantly exploded with gossip ， R*** said, 'I saw him moving groups around a while back,' and 炮 was shook: 'wait, big bro is a VP? I've been arguing bugs with him in this group.' Then he sighed, 'it's so flat you can't tell who's a D, who's a VP, who's a PM.' 朱* added context: an interview a while back tagged him as 'former serial entrepreneur,' and at an offline launch event with 谢欣 and 王大仙 onstage going live it was clear he was still there, but visibly less proactive than in 2025. A single HR headline pulled back the curtain on this crew's strange habit of 'debating bugs in the same chat as big-tech execs.'

Riding the VP topic, 炮 voiced an honest awe: now that he's building his own SaaS handling auth and distribution, he finally gets how good Lark Base really is ， 'auth is such a nasty, painful thing, but under Base's architecture it's terrifyingly clear.' R*** resonated, saying her company upgraded its plan precisely because Base's advanced permissions and multi-source perks got cap*ed. 炮 followed: 'right, workflow runs, table rows, totally hooked.' 平*** wished aloud, 'if only Lark Base went open source, I could plug in my own API.' A discussion that started from gossip landed on the most honest respect for product architecture.

For all the architecture love, the group didn't hold back on roasts. 炮 grumbled that Lark's A* credits are hard-bonded to meeting minutes, which screwed him this month ， ally isn't great and burns tokens for nothing, leaving him with no credits for meeting minutes by month's end. R*** confirmed, 'meeting minutes are tied to A* credits now,' though she did find a sweet spot for Base's smart agent: 'asking it for data, recall is faster.' q*** on the other hand was furious: listing the names from one Base table against the people in the corresponding chat ， just 17 people total ， dragged her around for an hour. 'I literally cursed it out, it's stupendously dumb, even Xiaomi's model wouldn't be this bad.' 炮 piled on, 'and it burned a ton of points doing it.' Same tool ， praised then dragged. Context decides the verdict.

Late at night Xiangrui first forwarded 'Xinhua Yu Dian launches big,' and 醒* casually patted him on the back: 'making more people stronger thanks to A*.' The next morning he closed with a fresh drop ， Weibo now has an official CLI with auto-publish sup*ort, screenshot included, with a self-deprecating 'but looks like Weibo finally noticed there's a market [wangchai].' Q*** instantly egged him on, 'go install it,' and 炮 punctuated the moment with a meme: 'feels like one rule for the lord, another for the peasants.' A tool headline wrap*ed the day and rekindled this crew's obsession with 'automating everything.'

In the morning 2*** drops in a line: Kazik just ship*ed a tool, worth a look. K*** nudges back, 'His is for MacBook, but you could have A* parse it and try porting it to Windows.' The other replies, 'Yeah, I used cc to convert it last time, pretty easy.' A light little tool-share ， one of the rare technical voices in an otherwise still day.

Mid-morning A*** lobs out an honest little question: is there a ban risk if you plug A* into a WeChat Official Account? No one picks it up right away; it just hangs there in the quiet. Today, it seems, no one's attention is really on tools.

In the evening Xiangrui posts something light yet heavy: after grinding on A* for so long, he still feels that as A* moves forward, it has to come back to real, living people. The group goes quiet ， and then warm. 袁** leads with 'Take care of yourself, brother.' 小* and 炮 send folded hands. 谢*, 强哥, 辛*, 周**, and K*** chime in one after another ， 'Take care of yourself,' 'Look after yourself.' A*** says it most fully: 'As A* moves forward, don't drift too far from people; take care of your feelings, brother.' No workflows tonight, no quotas, no memes ， just a circle of friends across screens, quietly holding space for each other for a little while.

醒* woke up puzzled: why can't you create a personal Feishu account anymore? K*** laid it out plain ， they killed personal account creation a long time ago; only the very first batch are still personal users. h** briefly insisted he'd registered one a couple of days ago, then walked it back: 'I misremembered.' 醒* finally got a Feishu official willing to answer questions and sighed, 'Feishu really isn't short on cash.' 侯* picked up the baton: 'You can ask me too ， I get answers from Feishu official faster.'

Xiangrui drop*ed the news: Doubao is officially launching Pro today. 王*'s first reaction was, 'I haven't found the paywall yet.' K*** said 'don't rush.' 大*** offered the trick: 'Use Expert mode a few times and the paywall pops right up,' even sharing a screenshot of his own. Q** took the other side ， there are plenty of replacements out there, no real need to pay; even the file-cleanup feature is something Mavis can do. Day one of the new launch, and the group was already weighing whether it was worth the money.

In the early evening 归* raised the real question: should Codex be granted fully open access? The list of operations you have to allow is absurd. D** was direct: 'don't recommend it.' 炮 added: 'Keep that manual review layer ， going fully open invites trouble. Non-prod is fine, but never on production.' R*** shrugged it off: 'I just set global full access lol.' A*** threw cold water with a war story ， production crashed once, memory maxed, server down. 归* admitted his is production, so he's nervous. 毅 delivered the cleanest fix: build a backup environment, let it run loose in there, you test and verify, push to production once it passes ， no need to keep granting permissions, just one final migration.

R*** opened with a sigh ， 'is everyone just busy lately? Chat's gotten noticeably quieter' ， then drop*ed the bomb: 'I'm down to 6% on my weekly quota, the genius is about to fall [bitter].' K*** yelled 'reset!' R*** laughed it off: 'No more reset perks ， used them up twice already.' 平*** suggested 'invite a friend.' The topic slid into model dumb-down ， A*** asked, 'do you guys feel GPT has gotten dumber lately?' 李** nailed the pattern in one shot: 'They're ship*ing 5.6 this week, so it's been dumbed down the whole time; every time a new model is about to drop, the old one turns stupid, so the new one looks brilliant.' h** piled on: 'Both GPT and Claude have gotten dumber lately.'

Late at night 王* grumbled: ever since he started using all these agents, his 500GB C drive has been steadily filling up. 小谷老师 asked how many memberships he'd paid for, and 王* answered straight ， he buys the APIs for Zhipu, Xiaomi and DS, but not a single membership; everything else is scrounged from platform freebies. The overseas options are blocked by their own limits, though he hears Alibaba's Qoderwork or ACI has a hidden box that gives you OpenA* and Anthropic models. 小* shared a screenshot: 'just finished going head-to-head with it, it completely ignored me.' Another day of wits against the A*.

Late at night q*** drop*ed good news ， Zhipu officially an*ounced a 50 million subsidy, free for A* founders. Next morning Xiangrui forwarded the registration for the first Global Space Compute Conference; 炮 grumbled 'Haidian-district only, huh,' then ran the numbers on his own pain: his team burned through three tokenplans yesterday on 5.2 alone, they're desperate for the GLM discount. He also flagged a shift ， tokenplans bought in June are noticeably smaller than before. Both Alibaba Cloud and Volcengine have been quietly squeezing the quotas. The joy of subsidies and the sting of burning through tokens together formed the day's coda.

p*** throws out a very practical landing problem: rolling out a Bitable mid-stream, with no real governance over what came before, how do you bridge ongoing business with the legacy data? xixi first asks the diagnostic: 'how much old data needs to be loaded, and does it need a cleanse first?' 闪*** shares his own playbook ， lock the Bitable schema first, wash the old data into the new base, let A* auto-clean plus human review before ingestion, splitting manual ingestion from system ingestion; run the legacy business on two parallel tracks with a one-month transition, stop maintaining the old base, and as its value fades, archive it slowly. 辰* adds the kicker: if there's an evidence trail to verify against, just pull the full set; if not, clean then ingest.

Xiangrui floats the morning tease: Seedance 2.5 might drop today, sup*osedly straight-to-output in 30 seconds. Right after, l*** pokes in asking whether the WeChat A* rollout needs an invite code. J*** posts a screenshot asking 'is this WeChat A*?', h** and K*** confirm one after another, 宰* follows up 'you got into the beta', and 平*** voices outright envy. J*** hasn't even started playing yet, and casually spots an Easter egg ， Feishu's agent now sup*orts voice input too, hold ctrl+shift+s to talk.

In the afternoon, one gate traps a crowd: Codex login requires phone verification, 哟* is stuck, q*** chimes in 'no clue, I'm stuck on the same thing.' A*** lays down the routes first ， API login or an SMS-receiving platform. 哟* follows up asking whether second verification will hit, and the conversation cracks open: 李** says he, a PRO paying user, was suddenly asked to verify a few days ago, and the most reliable path is an overseas phone number; 哈* warns that numbers from SMS-receiving platforms will get hit with second verification ， not safe ， while daily disposable numbers cost just a few cents; 天* caps it: 'one overseas number can verify five accounts, I've verified several for friends.' One login puzzle, dismantled by the group into a full pitfall-avoidance guide.

By dusk, A*** hits a wall: can the current WeChat still expose chat records to the command line? He's tried several tools and can't extract the key. 和*** maps out the real-world path ， the current version probably won't work; he found that 4.1.10.27 won't yield the key, so he installed 4.0.x to grab it, and after grabbing the key, upgrading the software again still lets him read records. A*** says on the spot, 'I'll try installing the old version.' A small trick ， downgrade to grab the key, upgrade to keep the features ， passes through the group just like that.

R*** asks something he's been wanting to stitch together for a while: any way to con*ect Codex, WeChat, and Obsidian, and has anyone done it? Xiangrui replies crisply ， vchatCLI and OB CLI. R*** catches it with a quick 'on it, Aria.' The topic then bends toward SIM cards: q*** wants to buy one to cover Codex, still remembering the old Vodafone US 4G unlimited card, while 李** pours cold water ， 'they all carry monthly fees, not cheap', and q*** lays out reality with 'so the monthly fee isn't going to be a simple 8 yuan, right [awkward]'.

At night Xiangrui drops a big one: things have been a bit busy at home, but he's squeezed out the spare time and finally turned our own knowledge-management architecture into a native ap*, screenshot tossed in. 炮 blurts out 'sick!', q*** asks 'what is this, an ap*?', then stares at the screenshot and is shocked 'it's even got tabs like a browser.' Xiangrui says, 'I'll round up everyone for beta later', and 炮 jumps in instantly ， 'test test test, I'm slowly starting to use Obsidian, the divine-beast Xiangrui's ap* would be such great guidance' ， adding that at OPC he's a one-man band wearing team-management, BD, product, legal, and finance hats, in desperate need of a knowledge base. 闪*** simply calls out, 'this can ship as a paid-subscription product.'

A*** was stuck on Douyin first thing in the morning: he'd been using a family member's verified account, and out-of-town logins were now triggering facial recognition. 炮 poured cold water ， face verification is basically the highest-risk check, you can't get past it. But 辰* offered a counterintuitive move: ship the phone over, looks most troublesome but is actually fastest, since a new device will definitely trigger verification anyway. SF Express round-trip, three days, twenty-four yuan done. A*** was sold on the spot: "Now that's a great idea."

An industry anchor landed in the afternoon. 杨**'s read: ByteDance's Volcano and Feishu are each doing their own thing with no big enterprise moves, Alibaba is a bit slow, and for systematic enterprise ap*lications Tencent is the bet for now. 李** picked it up ， "Workbuddy is unbeatable this month, let's see if anything new shows up next month" ， and added that Alibaba is still tangled up in internal politics, no new products short-term. A few lines laid out the current product landscape cleanly.

A*** brought ClawHunt into the group ， an A* Agent two-sided marketplace, hosting a 100-person hackathon in Shenzhen in early July, where 100 Builders ship run*able Agents in three days offline with VCs onsite. She kept saying "thanks for M*. Li's platform," and Xiangrui generously @everyone: anyone with time is welcome to join. l*** poked in asking "can I just spectate? I really want to be there in person," and A*** answered there are 30 audience spots. A competition recruitment post, just like that, rolled out across the group.

Recruitment hit its peak in the evening. 小*'s line "once you go, you're a Shenzhener ， just say when we're meeting" lit the vibe, and A*** promised on the spot "if you come I'll buy you din*er." Asked about the player shortage and teaming up, 小* reported "tons in marketing, plenty in dev too," signed up himself, and self-mocked "dark enough already, just haven't taken the class." K*** had plan*ed to wait until he bought a mini, but turned around and signed up too ， "want to come learn from the bosses." Even B*** in Guangzhou got talked into it with "Guangzhou is right next to Shenzhen." Fujian folks don't fool Fujian folks ， quite the lively scene.

Beyond recruitment, the usual small hiccups. K*** complained "cmux keeps crashing today," and E*** was frustrated too ， some networks won't con*ect when operating Codex on phone, worked fine at first but has been hit-or-miss lately. K*** said he hasn't tried Codex on phone himself. 王* added a more carefree route: "I just route everything through Feishu, no friction at all." A few exchanges back and forth ， the most familiar help-seeking rhythm in a tool-people group.

The most explosive moment of the evening was Xiangrui himself. p*** drop*ed a screenshot ， "95% of one-person company startup projects got killed," with the group owner showing up again inside, and 炮 picked up the meme: "pass it on, Xiangrui got killed." 宰* chip*ed in another one, "on Xiangrui's technical chops," and A*** exclaimed "this has spread nationwide." Then the vibe shifted as everyone got hooked on the stuff in the frame: q*** "suddenly hooked on this screen, want to buy one," 炮 eyed the wall behind him "my incubator just needs this vibe," and K*** wondered whether that cable could carry p*** directly. q*** wrap*ed it up with a laugh: "just promote quietly, you guys are forcing praise on me."

First thing in the morning, Xiangrui tossed yesterday's group daily into the chat. B*** locked onto the detail right away: "another upgrade?" 炮 leaned in too ， "no? was there a beta?" ， and R*** drop*ed a few screenshots saying nope, no upgrade. Only then did Xiangrui leisurely spill the bean: rumor has it something ships next Thursday. One daily, one teaser ， the group's most familiar opening move.