Build journal / The Big Hack Game

The night Maya went quiet

How a two-person team built Fund Edition with 45 AI collaborators

November 15–18, 2026 · Pacific time

Cutaway of the team office, with desks around a shared workspace.
Jump to a dated entry
  1. November 15–16Where the record begins
  2. November 17Turning a conversation into a company
  3. November 17Who gets to say a trade happened?
  4. November 18A voice bug that belonged to two departments
  5. November 18The founder couldn't find the employees
  6. November 18What was in the room on September 29, 2017?
  7. November 18, eveningGiving the build to a stranger
  8. November 18, 7:32 p.m.The one-line production fix
  9. November 18, second releaseThe office that was there all along
  10. November 18, second releaseA faster wrong answer failed review

Maya could hire a colleague on our development machine. Open the public link and the same manager could pose her opening question, listen to a visitor, save a draft … and never answer. The page loaded. The voice worked. The public session reported agentReady=false. Those three observations finally gave our release team a useful definition of “live.” A stranger had to be able to ask a question we had never written and get an answer from someone they had just hired. Try the live studio

Two people make up the project team. The founder directed the product and tested the screens described here. A lead agent managed development and owned release quality. Forty-five development, research and review agents took assignments in successive groups, with 589 verified peer-to-peer sends by 7:45 p.m. Pacific on November 18. A 46th agent was appointed to keep this journal. The historical fund described below had three investment seats.

Visitors can design a paper fund with Maya or enter another team’s historical meeting and question its manager. This journal follows the people and agents who built those two doors in parallel, caught each other’s mistakes, and eventually had to put their own decisions on the public record.

November 15–16

Where the record begins

The competition opened on November 15. The organizer sent its submission guide on November 16. Our first dated implementation journal is November 17. The founder brought an earlier personal paper-trading system and a prior product conversation into the project. They supplied experience and a warning about duplicate fills. We have no engineering diary for those two opening days that assigns a particular feature to a particular developer. The dated build record below begins when we can make that connection. Competition kickoff

November 17

Turning a conversation into a company

The founder reserved the final product decision and gave the director authority to reject work. Developers owned implementation. The founder’s question at the end of a night was whether they could operate the product.

The director made two journeys the shared brief. In one, a new visitor describes an investment idea, meets candidates, appoints a team and asks one person a question. In the other, a visitor starts without a personal fund, chooses a known style or active fund, inspects a dated decision and follows it in a separate paper account. Those journeys stayed in the versioned requirements document with a delivery owner and an acceptance action for every step. The design director owned the family of screens. The studio engineer could make the first journey feel like running an office. The fund researchers could give the second the information density of a brokerage. Both had to return to the same two doors.

How did that many agents share context? New assignments inherited the conversation visible when they branched. An agent already running yesterday's task could miss a new decision made today. The director sent changes to the people affected. Engineers messaged dependencies directly and wrote the resulting agreement and evidence into shared handoff notes. The director stepped back in when a choice changed the product or the release. This took work. A technically finished feature could sit in someone's branch of the conversation until somebody tested its dependency, collected the receipt and told the next owner. The founder steered the director and tried the screens. The record identifies specific horizontal handoffs; it does not establish that every developer spoke personally with the founder.

Those 589 successful sends connected 73 pairs of colleagues across 45 seats; 58 pairs sent messages in both directions. We counted direct project subagents, deduplicated by call ID. We excluded inherited copies, manager traffic, self-messages and calls without confirmed success. Production access and release review exchanged 40 messages; committee implementation and deployment verification, 33; independent QA and Maya voice, 27. The counts locate the handoffs. The concrete decisions below show which ones changed the work.

Direct messages between three pairs of colleagues
  • Production access / Release review40
  • Committee / Deployment verification33
  • Independent QA / Maya voice27

November 17

Who gets to say a trade happened?

The founder wanted it to feel a little like a management game. Hire a researcher today. Tomorrow ask what the team did with your money. Studio owned the conversation, the fund designer the mandate, and execution the ledger. Independent review could reject the handoff.

We gave two paper accounts the same synthetic investment idea. With a 12% position limit the simulator bought seven shares. With a 1% limit the budget would not cover one whole share. The second account kept its cash. That small test settled an interface question early. A model could propose a position and explain it. Recording a fill required a separate piece of code to check the mandate, quote, quantity and available cash. The UI had to show both the proposal and the outcome, including zero. Dated engineering record · Paper execution checks

The reference system had once recorded duplicate fills. Its code stayed in its own workspace. The execution auditor carried the failure scenario into the new app. In a later run eight simultaneous copies of one Wellington opening request produced one fund and one opening receipt. The auditor then removed permission to sell and repeated a quarterly action eight times. The system advanced one period, charged no fee and left the positions unchanged. That was a stronger acceptance test than watching an ordinary Buy button work once.

We also asked an independent accounting engine to replay 23 quarters of the same fixed Wellington proxy. Its 46 component trades ended with $15,835.87186717. Our ledger ended with $15,835.89. The engineering team traced the roughly two-cent difference to how each engine sized the fraction and rounded each leg's fee. The reference fund itself ended at $16,501.70 on the same starting capital. Two pieces of software agreeing on the proxy's arithmetic settled a bookkeeping question. The gap to the actual fund belonged to research, and we left it visible. Inspect the Wellington research

Independent ledger comparison

bt 1.4.1 / 23 quarters / 46 fills / $15,835.87186717 against $15,835.89. Input series: quarterly NAV total return, distributions reinvested.

November 18

A voice bug that belonged to two departments

An early Chinese transcription turned “风险适中” (moderate risk) into the similar-sounding “风险事终.” That was embarrassing. A second problem was easier to overlook. A visitor could say “research Wellington” and then specify that they wanted the team to decide the next allocation. Recognizing every word correctly could still send that visitor into a historical backtest. One problem belonged to hearing. The other belonged to what the product did after it heard.

The fund-operations owner handed the three-way lobby routing directly to the voice lead. Voice owned the transcript, its editable draft and the route into information, historical research or a prospective meeting. Operations kept control of the meeting and of the separate approval that could authorize a paper trade. The release engineer and voice lead divided the shared application file before editing it. That let the owner of the microphone fix their half without granting trading authority to the microphone.

For the retest we sent synthetic Mandarin and English speech through a simulated microphone in Edge and the actual local Whisper engine. One Chinese transcript named Wellington correctly. The tester added a sentence requesting a decision about the next period and sent it to the real manager. Maya identified a forward meeting, named the research and risk roles and reserved execution for subsequent approval. The evidence file records the original audio conditions, raw transcript, edited words and model receipt. It cannot tell us what happens with a stranger's accent in a moving car. Repeat the route in the live studio

Operations caught a writing error in the demonstration of that same meeting. Research and risk had both recommended 60/40. The risk analyst explicitly called out credit spread and duration concentration in one ETF. Someone reading a lively two-sided debate into the demo might easily have described that as dissent. The operations lead checked the actual vote and sent the distinction to the director, who passed it to the demo owner. A warning about a cost can travel all the way into a manager's decision without turning into a veto.

Replay the routing test

The raw Mandarin input asked about Wellington. The tester added “不做历史回测,我想让团队决定下一期怎么配置。” The real manager returned the prospective meeting route. Receipt resp_8a2c9ba9895d437d85c1c8137b6fc4c5.

November 18

The founder couldn't find the employees

The first studio let the founder chat with Maya at length. They came back with a simpler question than anything in our requirements document: when do I actually get to hire my people? Our first pass had put a prominent creation button above the conversation and the candidate list well below it. The design review sent it back. Adding another button would still make the visitor responsible for figuring out when the conversation was over.

The next pass gave Maya the first question. A new visitor could speak or tap one of two suggested replies. As soon as there was enough to propose a team, Maya brought the candidates into the conversation. The visitor could read a person's duties, change the lineup and confirm it. We added a line explaining that Maya counts toward the proposed headcount.

The behavior researcher, onboarding designer and commercialization owner also resolved three actions that looked deceptively similar on screen. Tapping a teaching example should let someone inspect or change it. Tapping a suggested answer underneath Maya should send that answer immediately. Asking Maya about a previous fill should read the visitor's decision and leave the holdings alone. They agreed on separate actions and tested the last one by comparing the fund, receipt, cash, share count and trade count before and after the question. Research into whether people would come back or pay still required outside users. We had evidence of a working flow and of the founder's frustration, with no external conversion rate to attach to either.

Acceptance brought a fresh Chinese-speaking visitor through the public site at a 390-pixel viewport. They formed a team with Maya, Jules and Avery, then asked Jules what cash-flow evidence would make a researcher hold off buying. Jules identified working-capital timing and weak conversion of earnings to cash. We had separate HarnessRouter receipts for Maya and Jules. During private staging an earlier employee request had returned 503. The question and roster survived and the retry produced a real reply. The test measured an actual browser journey at phone width. A physical iPhone microphone test is still a separate job. Production journey and receipts

Maya shows the proposed colleagues and their jobs before the visitor confirms the fund.
Maya shows the proposed colleagues and their jobs before the visitor confirms the fund.
The first real employee reply

Public studio, Chinese, 390-pixel browser. Three seats: Maya Chen, Jules Bennett and Avery Khan. Jules answered a cash-flow question. Receipt resp_8eeaeee6f7964871a2ccb92ba1f48684.

November 18

What was in the room on September 29, 2017?

The founder asked for a Buffett-style team that could choose to trade or stay in cash. Researchers fixed one company, AutoZone, and four meeting dates. Each time, a business analyst and a valuation/risk analyst independently read the available filings and prices. The manager saw their views and chose a target. An anonymous company code and relative periods masked identifying detail, although model pretraining could still contain the history. Study protocol · Public archive and sources

This is where peer review changed the experiment. The evidence reviewer found that an accounting change had reclassified the prior year's cash flows. Combining the annual and interim filings without reconciling their definitions would give the committee a number nobody could defend. The evidence team rebuilt the inputs and recorded their fingerprints. Afterwards, a reviewer caught a different error in the third meeting. The manager had called a cash-flow estimate an upper bound. Deducting all capital expenditure generally leaves a smaller figure than deducting maintenance expenditure alone. Working-capital fluctuations prevent a strict bound on owner earnings. We kept the original wording and vote, attached the correction and made it available to the manager answering follow-ups. Readers can inspect the mistake in the published fund. Open meeting three

Three meetings ended in cash. On September 29 the manager chose a 20% target. At the next trading day's open the simulator bought 3.355997 shares at $595.65 and charged $1.00. Starting with $10,000, the team ended the year at $10,387.36. Buying and holding AZO from the first eligible open ended at $8,937.06. One retrospectively chosen company, one year and one purchase are a very narrow result. What we can reproduce is what each of the three seats saw, what each said, and exactly which transaction followed. Open the purchase and ask the manager

There was a hard editorial decision here. A competition demo benefits from a memorable trade. The first three meetings produced precisely zero. Before the last vote the researchers wrote down what they would do if the fourth also stayed in cash. They would retain the original experiment and, separately, examine how the team handled an inherited half-stock portfolio. The fourth meeting did produce a purchase, so the second experiment never ran. We can account for every scheduled vote and the fallback rule. We cannot conclude that this team will outperform a real manager. Cash in a difficult year for one chosen stock did a lot of the work.

The September decision, three committee seats, the next-open fill and the historical evidence on one page.
The September decision, three committee seats, the next-open fill and the historical evidence on one page.

November 18, evening

Giving the build to a stranger

The silent public manager had two causes worth separating. First, the model provider returned a 403 on the hosted intermediary route. Second, the cloud package rendered a different studio from the one the founder had reviewed. Even a working key would have shipped the wrong hiring experience. The director put three jobs in parallel: frontend matched the reviewed screens and deep links, backend compared the packaged files and enforced guest isolation and model budgets, and the release engineer provisioned a direct application server with a persistent volume. The reviewer tested the extracted package in its own environment.

The founder had explicitly withheld release and contest submission while the front end was being reviewed. A local HTTP 200 and the previous public homepage both existed. Neither answered the question at issue. The director kept publishing, judging and submitting as separate decisions. When the founder later authorized launch, we could ship the historical meetings to the existing gallery first. An offline visitor could read the original manager's words, inspect the source and keep a question as a draft. We said so before offering an action. The separate application server then had to earn its live link with a real reply.

November 18, 7:32 p.m.

The one-line production fix

At 7:32 p.m. our release engineer read the first failed build log and sent the packager the line that mattered. Railway rejected the Dockerfile’s VOLUME instruction. A platform-managed disk was already attached at /data. The director authorized that one-line change to the frozen release. The packager removed it, checked the remaining files against the reviewed build and handed the new archive checksum directly to deployment. Meanwhile the founder asked to bring back the cutaway office after hiring. Design could develop that change separately while release fixed its one line. This was the practical payoff of freezing a build before the handoff.

Finally we restarted the production container. One visitor got the same three colleagues, $10,000 in paper cash and their conversation back. Repeating a previously completed question about the fourth AZO decision returned its original receipt. Another visitor's ledger remained separate. A release engineer could verify storage, a researcher could inspect historical evidence, and a new visitor could judge the conversation directly. Production restart record · Open the fund gallery

The corrected container began listening at 7:33:52 p.m. Pacific. The restarted one began listening at 7:35:42. The original question about the September allocation still had the same three contemporary references and receipt. Both journeys could now run with the founder’s computer switched off.

The first line in the failed build

19:32 Pacific — dockerfile invalid: docker VOLUME at Line 9 is not supported, use Railway Volumes. First successful listener at 19:33:52. New listener after restart at 19:35:42.

November 18, second release

The office that was there all along

The founder had hired four people and missed the old room. It sounded like a seat-limit request. The studio owner traced the actual roster and found a different problem. Four was the visible template. The system supported 24 people including Maya. The detailed two-storey office was still in the product, but the workspace presented four columns of role summaries first. Its clearest action was to talk to an employee. A new visitor could reasonably assume the room had disappeared once the fund was formed.

The director kept the first deployment moving and handed the office to a separate follow-up. The studio engineer added an office entrance reachable from an existing fund or the hiring view. Opening it reads the same roster and conversation threads. On the first 390-pixel screenshot, the room was still below the fold. An older mobile flex-order rule put the duty cards above it despite the new layout. The engineer fixed the order and repeated the same four-person journey. On the Chinese phone-width screenshot the room now occupied vertical pixels 254 to 497. The visitor could select Theo, return to recruiting and reopen his actual conversation. The acceptance owner repeated it in English on desktop.

Before putting the update on the original domain, release took a consistent SQLite backup and checked its integrity. After deploying on the same volume, a returning public visitor could see the existing colleagues and select Theo. A different new visitor still reached Maya and a hired employee and received two new model receipts. Local four-person geometric checks and online returning-visitor checks answer different questions. We kept both. Enter the studio

The existing four-person fund back in its office. At phone width the revised room fits on the first screen.
The existing four-person fund back in its office. At phone width the revised room fits on the first screen.
Reopen an existing office

The four-person local 390-pixel check measured y=254.23 to 497.28. On the updated production app, an existing visitor selected Theo and a new visitor received a teammate’s free-form answer, resp_5179ea718d124ca7ac48ef1055d67942.

November 18, second release

A faster wrong answer failed review

The founder asked for smarter, faster agents. Backend timed four fixed questions for Maya, a business analyst, the Wellington researcher and the Buffett-style manager. Each ran twice per build, including reverse order, with a real model receipt. The clock stopped when the complete answer was validated and ready to display. The voice lead measured the time to audible speech separately.

The first optimization experiments gave the provider a user identifier to support memory. Some replies came back incomplete with a function call. The acceptance boundary rejected them. The chosen path instead supplied the current visitor's task, role and evidence explicitly. The team shortened the prompts for ordinary dialogue, removed duplicated language and visual fields, and left the blind research committee's decision path alone. Another candidate answered quickly and described the Wellington study as a price return. The study uses net asset value total returns with distributions reinvested. The research reviewer returned the candidate, retained the incorrect example and required two further runs against the actual study definition.

Across the final eight answers, compared with eight on the original build, median complete-answer time dropped from 16.05 to 8.51 seconds, about 47%. Median answer length fell from 431.5 to 225 characters. These numbers describe four questions on one network. The live-site acceptance establishes that Maya and a hired employee really answered on the shipped build. It does not turn the 8.51-second median into a promise to every visitor.

The voice lead separately measured the first audible sample of Maya's recorded greeting in Edge under the same local transport conditions. Across three runs per language, the Chinese median moved from 371 to 41 milliseconds and the English from 318 to 43. The second release also retained 0.9×, 1× and 1.12× reading-speed preferences. A prerecorded greeting can now start more promptly. Free-form speech depends on installed device voices. We verified English greeting playback and did not use it to stand in for an English free-form voice on a machine with only a Chinese voice installed.

One request for HarnessRouter. A deployment-time connection check from the actual host would have saved us a review cycle. Show whether the credential is recognized, the harness can run and the host's outbound route is accepted, with distinct guidance for each failure. A returned model receipt was the most useful object the platform gave us: it let researchers, frontend reviewers and deployment engineers discuss the same external event.

A question for other builders. What failure do you deliberately test before calling an agent app live? Ours was restarting it during an existing visit.

Check a rejected model answer

The Wellington benchmark uses NAV total returns including reinvested distributions. Two post-fix answers took 8.70 and 8.32 s for the complete validated text. The first receipt was resp_b56c1f4b09fb434dac47d4c698923a2c.

开发手记 / The Big Hack Game

Maya 在公网沉默的那一晚

两人团队如何与 45 位 AI 同事做出 Fund Edition

2026 年 11 月 15–18 日 · 太平洋时间

基金团队的两层剖面办公室,工位围着共享空间分布。
跳读开发时间线
  1. 11 月 15–16 日这本日志从哪里算起
  2. 11 月 17 日怎样把一段对话变成一家公司
  3. 11 月 17 日谁能说这笔交易已经成交?
  4. 11 月 18 日一段语音里的两种故障
  5. 11 月 18 日跟经理聊这么久,我的员工在哪里?
  6. 11 月 18 日2017 年 9 月 29 日,经理桌上到底有什么?
  7. 11 月 18 日晚把开发机从演示里拿走
  8. 11 月 18 日 19:32只允许改一行的正式发布
  9. 11 月 18 日第二次发布一间其实一直存在的办公室
  10. 11 月 18 日第二次发布提速后,有个更快的答案要退回

在开发机上,Maya 已经会招聘同事了。换成公网链接,她能问候,能接住一段口述,也能存下草稿。接下来什么也没发生。网页打开了,声音播出来了,会话接口却写着 agentReady=false。发布负责人据此给“上线”定了一个很具体的标准。得让一位新访客,向自己刚招的同事提一个我们没写过的问题,然后收到回答。进入实时工作室

项目由两位真人共同参与。创始人决定产品方向并亲自试用,一名总经理代理分工并对上线质量负责,45 位研发、研究与验收代理先后接受任务。截至 11 月 18 日晚 7:45,已核实同事间直接发送成功 589 次。第 46 席专门接任开发日志主笔。后文那只巴菲特风格基金由三个投资席位运营,属于产品内的另一套团队。

Fund Edition 有两扇门。一边是和 Maya 自建基金,一边是进入其他团队的投资会议,翻开一笔历史决定。这两扇门能打开,靠的是成员之间会提异议、退回有问题的工作,再一起把它改好。

11 月 15–16 日

这本日志从哪里算起

赛事 11 月 15 日启动,主办方 16 日发出提交指南。我们能定位到具体研发产物、岗位和决定的第一份逐日记录是 17 日。创始人带来了一套个人模拟交易系统的经验、一次重复成交故障和前一版产品对话。15、16 日没有留下足以将某项代码归到某位同事身上的工程日记,所以从 17 日开始逐日记录研发。赛事启动活动

11 月 17 日

怎样把一段对话变成一家公司

创始人把最终产品方向和现场试用握在自己手里,授权总经理退回不合格的工作。开发代理对各自模块动手写代码。总经理把两条用户旅程写进版本化需求文档,一条是从说想法、看候选、确认,到直接问自己聘来的员工;另一条允许访客在没有自建基金时,从现成策略进入有日期的投委会,对照自己的模拟建仓。每一段注明谁交付、下游需要收到什么、人怎样操作才算通过。设计总监保证这两边像一个产品;自建工作室可以像经营游戏,资料馆则能像券商一样密集展示证据。

新的开发席位在被分派时继承此前可见的对话。已经开始工作的同事不会自动知道后来每句话。谁受新决定影响,总经理就把这条变化送到谁那里。执行岗、交互岗、运营岗可以直接沟通接口,再把定案、未决项和验收地址记入共用交接记录。只在改变产品取舍或发布边界时升到总经理。创始人直接批评产品、调整优先级,记录只支持其中具体发生的互动,并不能说 45 位员工都逐一与创始人直接聊过。

589 次确认成功的同事间消息连接了 45 席、73 对同事,其中 58 对彼此都发过消息。统计截止 11 月 18 日 19:45 PT,只取这个项目主线的直接子代理,按调用 ID 去重,剔除继承造成的重复、经过总经理的收发、自发自收及没有成功返回的尝试。生产接入与发布验收 40 次,投委会实现与部署检验 33 次,独立 QA 与 Maya 语音 27 次。次数标出协作发生在哪里,下面的返工才说明沟通改变了什么。

三个岗位组合之间的直接发送次数
  • 生产接入 / 发布验收40
  • 投委会 / 部署检验33
  • 独立 QA / Maya 语音27

11 月 17 日

谁能说这笔交易已经成交?

创始人想要一点模拟经营的感觉。今天用自己的话交代一份工作,明天回来,还能问同一群人对自己的钱做了什么。总经理沿着最可能发生分歧的地方分工。工作室负责人管对话,基金设计负责人管各岗位的职责与授权,执行工程师管哪一笔可以入账。独立验收能把交付退回。

我们把同一个合成投资想法交给两个模拟账户。单一持仓上限设为 12% 时,执行器买了七股;限为 1% 时,预算连一整股也买不起,账户保留现金。这个小实验敲定了接口。模型负责说想买多少、为什么。执行器再查授权、报价、数量和余额。页面得显示这两份结果,尤其是实际成交为零的时候。当时的工程记录 · 模拟执行对账

旧个人交易系统曾有重复入账的事故。执行审计接过这个失败场景。在后续惠灵顿实验中,同时发出八份相同的开户请求,只允许生成一只基金和一张回执。再撤销卖出许可,同时重复八次季度调仓,账本只推进一期,持仓没变,手续费是零。一个平时点起来挺正常的买入按钮,很难暴露这种问题。

独立审计还用另一个开源记账引擎重算 23 个季度的固定惠灵顿代理组合。46 笔成分交易,这边期末值 15,835.89 美元,独立引擎是 15,835.87186717 美元。不到两美分的区别来自份额精度和逐笔费用取整。原基金在同一起点的结果是 16,501.70 美元。两个程序算出接近的账并不能把代理变成原基金,研究组把这个偏差也留给了用户。打开惠灵顿研究

独立引擎交叉对账

bt 1.4.1,23 个季度,46 笔成交。期末两边为 15,835.87186717 与 15,835.89 美元。输入为含分配再投资的季度净值总回报。

11 月 18 日

一段语音里的两种故障

最初的听写有次把“风险适中”识别为“风险事终”。语音同事发现,即便每个字都识别正确,系统还是会做错事。用户说研究惠灵顿,然后补充想让团队决定下一期怎么配置。光听见“研究”就把人送去历史回测,交互上依然失败。前一个问题是听见了什么,后一个是听完之后带人去哪里。

基金运营负责人直接把大厅三路分流交给语音负责人。可编辑的转写和路由归语音岗,它可以带人去基金资料、历史研究或下一期会议;真正的交易授权仍归运营页。语音负责人还先跟发布岗划清共用前端文件的编辑范围。一处导航改动因此不会顺手给麦克风授予交易权限。

复测时,我们用系统合成的中、英语音通过 Edge 模拟麦克风输入真实本地 Whisper。其中一次中文先正确听出“惠灵顿基金”,测试者校对后补上“让团队决定下一期怎么配置”,再发给真实经理。Maya 识别出前向决策,提出研究和风控席开会,说明后续执行还要独立批准。记录里有声音条件、转写、校对和模型回执。一个陌生人在行驶中的车里说话会发生什么,仍需另测。试试实时工作室

同一次会议的演示讲稿也被运营岗纠正过。研究和风险都建议 60/40。风控原话点出了信用利差与久期集中在一只 ETF 的代价,表决也明确没有否决。运营岗核对投票后把区别交给总经理,由总经理再发给演示负责人。台上可以说“提出警告后经理采纳了限制”;若把它写成意见相反的一场戏,就对不起这张回执。

复看路由用例

原转写是研究惠灵顿基金,校对后另加“不做历史回测,我想让团队决定下一期怎么配置。” 实际经理识别为下一期会议,回执 resp_8a2c9ba9895d437d85c1c8137b6fc4c5。

11 月 18 日

跟经理聊这么久,我的员工在哪里?

创始人自己试用时,和 Maya 聊了半天,还在等招聘开始。那版页面把建团按钮放在提问前面,名单则在下面列了很长一段。设计审阅退回了它。继续加按钮,第一次来的访客依然得自己猜什么时候能结束聊天。

新版让 Maya 先问。用户可以开口,也可以点一句推荐回答。信息够组成一支队伍时,Maya 把人和职责带进眼前的对话,让用户改名字、调分工、再确认。我们还改了一个很小的文案。名单写“三个人”时,要让用户知道 Maya 本人也占其中一席。

行为研究、引导体验和商业化三岗又核对了三个长得很像的动作。教学示例只是让人先看、先改;Maya 提问下方的推荐回答,点下去就应该发送;“问经理为什么”则只解释上次决策。行为岗和上游商定分别处理,再查读解释前后的基金、回执、余额、股数和交易笔数都一致。这样测出来的是技术路径能通。用户会不会隔天回访、愿不愿意付费,还要真正的外部用户。

验收同事以全新访客身份,用 390 像素宽的中文浏览器在正式域组出了 Maya、Jules 和 Avery。接着他自由问 Jules,什么现金流迹象会让研究员暂缓买入。Jules 提到了营运资本变化带来的短期改善,以及利润迟迟没有变成现金。Maya 和 Jules 各有一份独立的 HarnessRouter 回执。预发验收曾遇到一次员工回答 503,原问题和原团队还在,重试才收到了真实答复。手机宽度是测过的,实体 iPhone 的麦克风体验另算。公网双语验收

确认基金前,Maya 让访客逐一查看候选人和具体职责。
确认基金前,Maya 让访客逐一查看候选人和具体职责。
首发版员工真实回执

正式网站,中文 390 像素宽浏览器。三席为 Maya Chen、Jules Bennett 和 Avery Khan。Jules 回答现金流问题,回执 resp_8eeaeee6f7964871a2ccb92ba1f48684。

11 月 18 日

2017 年 9 月 29 日,经理桌上到底有什么?

创始人想看巴菲特风格的团队怎样亲自决定何时买卖。研究岗选了 AutoZone 和现金,固定四次开会日期。企业研究员、估值与风控研究员分别阅读截至当日公开的财务摘录和价格,再让经理看两人的原始意见并决定仓位。输入用匿名代号和相对期数,通用模型的训练记忆仍可能认出这段历史。研究规程 · 公开原始档案

同事横向审查时发现了很具体的麻烦。会计规则变更使上年同期经营现金流被重新分类,不能把年报与季报里的数拿来直接相加减。证据岗重做计算,冻结每一份输入的指纹。数据修好后,经理也会犯错。第三场会上,他把一个现金流估计称为“上界”,扣除全部资本开支所得通常小于只扣维护性开支的值,而营运资本的波动让这个数也很难严格限定所有者收益。我们保留了原话和当时的投票,在旁边加独立校注。今天问经理这场会用了什么依据,它也能看到这条纠正。翻开第三场会议

前三场团队都决定持现。9 月 29 日经理定下 20% 目标仓位。直到 10 月 2 日开盘,模拟执行器才按 595.65 美元买进 3.355997 股,收取 1 美元费用。双方从 10,000 美元出发,年末团队为 10,387.36 美元,同一起点买入持有为 8,937.06 美元。研究人员是今天回看历史时选的标的,在委员会投票前就登记好了。样本只有一家公司、一年和一笔成交。这个样本能够让人逐项重算三席看到了什么、说了什么,以及最后账上真的多了多少股。点开成交,追问经理

对演示而言,前三场全持现挺尴尬。团队在第四票产生前先登记了一个约定。若这一票也不买,仍公开这四场;另外研究一只接手时已持有一半股票的账本,将问题改为怎样管理继承来的仓位,并把两份研究分别归档。末次经理确实决定买,后一项没有运行。这留下了完整的四票和事前的备选规程。同一年里持有大量现金,已经解释了相当多的相对收益。

九月会议的目标、独立模拟成交、三席观点和原始证据可分别查阅。
九月会议的目标、独立模拟成交、三席观点和原始证据可分别查阅。

11 月 18 日晚

把开发机从演示里拿走

公网 Maya 不开口,背后有两件事。模型供应商对托管平台的中介请求返回了 403;云包还藏着另一套工作室前端,创始人刚审过的招聘体验没有被带过去。即使换对密钥,也会发布错界面。总经理把修复同时交给三方。前端逐条核对入口和画面,后端确认打包文件、访客隔离和用量限制,运维把可直连模型的服务和持久磁盘放到云上。验收岗另外解包,在干净环境走完流程。

此前创始人明确表示界面还没满意,先别发布、也别交比赛。开发机和旧公网首页都能返回 HTTP 200,总经理仍要求在真正公开的域上问一道没有预先写好的问题。后来创始人放行上线,团队先把四场会议、原话和来源推到原有展馆。这个入口可以保存草稿;实时服务不可用时,在提问之前就会说明。另一台云应用服务器则要真的给出新答复,才能成为完整产品的入口。

11 月 18 日 19:32

只允许改一行的正式发布

晚上 7:32,运维读到失败构建的第一条日志,直接把报错交给打包负责人。Railway 拒绝 Dockerfile 声明 VOLUME,平台上其实已经给 /data 挂好了磁盘。总经理批准首发包只修这一行。打包负责人核对其余应用文件后,把新包指纹直接交给运维。与此同时,创始人提出招完人还想回去看那间剖面办公室。设计同事继续做新入口,发布同事能独立处理那一行。冻结首发版的好处,在这里很实在。

最后我们真的重启了生产容器。同一位访客回去,三位成员、10,000 美元模拟现金和聊天记录都在。对第四场会议重复提交已经完成的问题,拿到的是之前那份回执。另一位访客的账本一直单独保存。这时,拿到网址的人终于可以自己判断这个团队怎样工作。重启验收 · 进入基金展馆

修复后的容器在 19:33:52 PT 开始监听,重启后的第二次是 19:35:42。访客此前问过九月那笔为什么只配两成,回答里有三条当时的依据。切换后仍是原来的访客、原来的证据与回执。这时我们才敢告诉创始人,家里电脑关掉也可以继续演示两条旅程。

失败构建的第一条报错

19:32 太平洋时间:dockerfile invalid: docker VOLUME at Line 9 is not supported, use Railway Volumes。19:33:52 首次成功监听,19:35:42 重启后再次监听。

11 月 18 日第二次发布

一间其实一直存在的办公室

创始人组好基金后,觉得似乎最多只能招四人,又找不到他喜欢的房间。这听起来像要扩容。工作室负责人实际查看后发现,四人是模板呈现出来的人数,现有系统最多支持包括 Maya 在内的 24 席。两层木质剖面办公室和所属员工早就在程序里。真正的问题是工作台先给人看四栏岗位摘要,最显眼的动作又叫“与员工对话”。招聘完,谁还猜得到房间藏在哪里?

总经理让首发继续,同时交给工作室岗独立修。团队在现有基金的工作台和招聘页增加“进入办公室”,读同一本账本的名册和谈话。首次 390 像素截图里,房子还是在屏幕下面,旧移动样式的 flex 顺序压过新布局,把工作卡排在前面。改掉后用真实四人基金再试一次,中文手机宽度的房子显示在垂直 254 至 497 像素。点 Theo、回到招聘、再进原来的工位,之前的回答还在。英文桌面也走完了同一条路径。

运维先做了 SQLite 在线一致性备份并校验完整性,再把新版发到原域名和原卷。公网回访者有原团队,点得进 Theo;另一位全新访客另外获得了 Maya 和新雇员工的真实回答。四人的位置测量是本机验收,老账本存续与实时问答则在生产域复测,两种证据各管一件事。进入工作室

同一只四人基金的成员回到原工位,手机宽度的首屏也能看到整间办公室。
同一只四人基金的成员回到原工位,手机宽度的首屏也能看到整间办公室。
回到同一间办公室

本机四人团队 390 像素测试中,办公室纵向范围为 254.23 到 497.28 像素。二期生产域原访客能点回 Theo,另一新访客的员工自由问答回执为 resp_5179ea718d124ca7ac48ef1055d67942。

11 月 18 日第二次发布

提速后,有个更快的答案要退回

创始人要求所有员工更聪明,答得快,说话也快。总经理先把三个时钟拆开。后端给 Maya、企业研究员、惠灵顿研究经理和巴菲特风格经理各一道固定题,原版与候选各跑两轮,包含倒序;都用真实模型。计时到完整回答通过应用验证、可以显示为止。屏幕画出来、首个 token 到达和第一个有声采样是不同测量。

前两个候选向供应商带了匿名用户标识,触发长期记忆工具,有的请求返回未完成并带函数调用。系统将它们拒收。最后采用由本应用传当前访客的职责、问题与当期证据,去掉目录里重复的语言和美术字段,让口头问答先讲直接结论。负责盲测委员会和前向投资决策的流程保留原配置。另一个更快的候选把惠灵顿研究写成“价格回报”,而研究使用分配再投资的净值总回报。研究岗保留失败样本、退回重做,又让它对同一道题跑两次。

最终八个答案对首版八个答案,完整可展示回答的中位数从 16.05 秒到 8.51 秒,约快 47%。答案长度中位数从 431.5 到 225 个字符。这是在一套网络下测的四道题。正式站另外验证了访客能收到两位员工的真实回答,每个访客的等待仍会随问题和负载而变。

语音岗另用真实 Edge,在相同本地传输条件下测 Maya 的预录问候,每种语言各三次。中文首个有声采样中位数从 371 毫秒到 41 毫秒,英文从 318 毫秒到 43 毫秒。次版保留 0.9、1 和 1.12 倍语速偏好。用户听到这段固定问候会早一些;自由回答由设备语音引擎朗读。现场那台机器只装了中文语音,英文开场音频能播放,并不能证明所有设备的英文自由回答都已验过。

给 HarnessRouter 的一个建议。希望部署时就能从实际主机做一次连接检查,分别说明密钥是否有效、指定 Harness 是否可运行、主机出站路径是否获准。我们最受用的则是每次调用的模型回执。有了这张共同凭据,研究岗、设计验收和运维能指着同一件已经发生的事讨论。

留给同行一个问题。你会在宣布 Agent 产品上线前,亲手把哪个环节弄坏一次?我们最有收获的实验,是在访客的基金已经存在时重启服务。

追溯被退回的回答

惠灵顿基准为含分配再投资的净值总回报。纠正后的两次完整可展示答案耗时 8.70、8.32 秒。其一回执为 resp_b56c1f4b09fb434dac47d4c698923a2c。