One desk where a PI learns, judges results and decides一个地方，同时用来学习、审阅结果和做决策
Several AI agents run X-ray astronomy projects around the clock. Their records are written for agents. The AI Desk turns them into short cards that a person can read on a phone in 5 to 10 minutes, and it learns from every answer the PI gives.
多个 AI agent 全天候推进几个 X 射线天文项目，它们的记录是写给 agent 看的。AI Desk 把这些记录变成一张张短卡片，人在手机上 5 到 10 分钟就能读完；PI 的每一次回答，它都会学习。
"I need one entry that merges the decision desk and the learning desk.""需要一个将决策台和学习台合并在一起的入口"The PI, 2026-11-06, when the desk was startedPI，2026-11-06，AI Desk 立项时
26 cards26 张卡片4 card kinds4 种卡片21 taste rules learned from answers从回答中学到 21 条偏好规则34 writer lessons34 条写作教训42 concepts on the reader map读者知识图上 42 个概念8 page versions in the first 2 days头 2 天迭代了 8 个页面版本
Learning学习·6 min6 分钟example card示例卡片
How short can a message be?一条消息最短能有多短？
Try one desk feature: predict before the figure. The figure stays hidden until you commit to an answer.
试一个 AI Desk 的功能：看图之前先预测。你作答之前，图是隐藏的。
A test has four answers, A to D. Answer A comes up half the time, B a quarter, C and D an eighth each. Can a code use fewer than 2 bits per answer on average?
一道题有 A 到 D 四个选项。A 出现一半的次数，B 四分之一，C 和 D 各八分之一。平均每个答案能用少于 2 比特的编码吗？
layer: reasoning层次：推理type: prediction题型：预测
The figure appears after you answer.作答后显示图。
schematic, not data示意图，非数据
equal code
等长编码
A 00
B 01
C 10
D 11
short for A
A 用短码
A 0
B 10
C 110
D 111
Average length = ½×1 + ¼×2 + ⅛×3 + ⅛×3 = 1.75 bits, against 2 bits for the equal code.
The wrong choice "always 2 bits" is the mistake a reader is likely to make. The desk builds wrong choices from real misconceptions, never from jokes. Only after this arithmetic does the card name the idea: entropy.
The writer reads the project records read-only and writes cards and figures. Every number cites a record line; every sum is recomputed by a script.
撰写者只读项目记录，写出卡片和图。每个数字都引用记录中的具体行，每一步算术都用脚本重算。
DeepSeekscript脚本
Check核查
A second run checks currency, numbers, figures and the PI's rules, and lists the writer's mistakes. A script recounts reading time and validates the card format.
另一次运行核查时效、数字、图和 PI 的规则，并列出撰写者的错误。脚本重算阅读时长并校验卡片格式。
DeepSeekscript脚本
Fix once, or hold只修一次，否则扣下
One fix round. A card the checker still fails is held back, not published.
只有一轮修改。修完仍未通过核查的卡片会被扣下，不发布。
Claude
Publish发布
A small Claude worker does only what needs Claude: upload figures, write cards to the page database, deploy, verify.
一个小的 Claude worker 只做必须由 Claude 做的事：上传图、把卡片写进页面数据库、部署、核验。
PI
Read and answer阅读和回答
On a phone: notes on any paragraph, quiz taps, ratings (information, sense, depth, difficulty, gain, interest, agreement), minutes spent, a choice.
Each finished answer is split into science questions, style proposals and rewrite requests.
每份完成的回答被拆成科学问题、风格建议和重写请求。
Claudescript脚本
Learn and route学习和分发
Style rules the PI's own words support go into the rule book; science questions go to the project's own agent; rewrites start the next round by themselves.
有 PI 原话支持的风格建议写进规则本；科学问题交给该项目自己的 agent；重写请求自动开始下一轮。
The rule behind the split: anything a program or a cheap model can do is not done by the expensive model. The Claude worker wakes every 50 minutes; when nothing waits it prints one line and goes back to sleep. The scanner caps the writer at 2 jobs at a time and 8 launches a day, and launches only while the machine is below 60 % CPU and memory.
Four kinds of card, each asking one thing of the reader四种卡片，每种只向读者要一件事
Learning学习
Teaches one concept a result needs. The reader rates each concept 0 to 4 before and after reading, and answers 2 to 5 short questions.
讲清一个结果所需的概念。读者在读前、读后给每个概念打 0 到 4 分，并回答 2 到 5 道小题。
Article文献
Reads one paper with the reader: its claim, its evidence and what it means for the projects. Full citation on the card.
陪读一篇论文：它的论点、证据，以及对项目意味着什么。卡片上附完整引用。
Result结果
Asks for a judgment of one result: how it was obtained, with its uncertainty and trust label, and whether it is on the project's main line.
请读者评判一个结果：它是怎么得到的、误差和信任标签，以及它是否在项目主线上。
Decision决策
Asks for a choice: Problem, Goal, Plan, Expected result, then 2 to 4 options with their effects, a default and a deadline.
请读者选择：问题、目标、计划、预期结果，然后是 2 到 4 个选项及其影响、默认项和截止时间。
The small ideas
小巧思
Details that make the desk work让 AI Desk 真正好用的细节
Reading on a phone在手机上读
A time budget that a script enforces由脚本把关的时间预算
Every card targets 5 to 10 minutes. The checker counts every word the reader sees (captions and quiz answers too) at 134 words per minute and fails a card that runs long.
Nothing is sent until the last button不点最后一个按钮就不算提交
Notes, taps and ratings save as you go, but a half-read card stays "Reading", pinned at the top. Only "Finished: send feedback" lets the agents digest it.
The PI asked for this: unfinished reading should not be treated as a verdict.
这是 PI 提的要求：没读完的内容不应被当成结论。
Notes on a single paragraph针对单个段落的批注
The reader can comment on any paragraph. The digest quotes that paragraph back, so a note is answered where it was written.
读者可以对任何一段写批注。消化时会引用那一段，所以批注在哪里写，就在哪里得到回应。
Two languages, one set of numbers两种语言，同一套数字
English by default, Chinese one tap away, as a full translation. Dates, numbers and units are identical in both, a rule the checker enforces.
默认英文，一键切换中文，中文是完整翻译。日期、数字和单位两边完全一致，这是核查必查的一条规则。
Teaching教学
Problem first, formula last先讲问题，最后给公式
Learned from two video tutorials the PI liked: open with a problem the reader can feel, work a toy case small enough to compute in full, write the formula as a summary of that arithmetic, and name the concept only after it is understood.
从 PI 喜欢的两个视频教程里学来：先抛出一个读者能感受到的问题，算一个小到能完整算完的玩具例子，把公式写成那段算术的总结，等读者懂了再给概念命名。
Predict before the figure看图之前先预测
For a figure whose message can be guessed wrong, the card hides it until the reader commits. One choice is the misconception the figure corrects. You tried it above.
Each question carries a layer (concept, judgment or reasoning) and a type (comparison, counterfactual, diagnosis, prediction, evidence gap). A learning card needs at least one judgment or reasoning question.
A decision card shows the writer's recommendation after the options, with what is still unknown. The default stays as the records set it, so advice never makes the choice.
A card may ask about a concept it uses but does not explain, to learn what the reader already brings. These are not scored, and "Not sure" is a normal answer.
卡片可以问一个它用到但没有解释的概念，以了解读者原本就懂什么。不计分，"不确定"是正常答案。
Range check范围测试try it试一试
Two codes describe the same source. The second was built for the wrong frequencies. What measures the extra bits it spends?
两种编码描述同一个信源，第二种是按错误的频率设计的。用什么量来衡量它多花的比特？
An evidence gate on the profile读者画像的证据门槛
One right answer is never trusted. A verdict on a concept becomes "trusted" only after a right range check on two different days, or a right answer that agrees with the reader's own rating. Answers given with the card open count at most as "partial".
Writers run no new analysis. Every number cites the file and line it came from, with its uncertainty and trust label. A number without a source is written "unknown".
Every figure is tagged observation, derived, model or schematic. A drawn schematic says "schematic, not data". Writers open each picture before captioning it, and never draw a real instrument from memory.
When the PI asks a science question on a card, the desk does not answer it. It goes to that project's own agent, which owns the analysis.
PI 在卡片上提出的科学问题，AI Desk 自己不回答，而是交给负责该分析的项目 agent。
Mistakes become lessons错误变成教训
Each checker run lists the writer's mistakes. They go to an intake list that new writers read at once; stable ones move into the brief templates. There are 34 so far, for example "read the colour bar, not the panel background".
The PI audits; the estimate assumed readingPI 是在审计，而时长估计按"阅读"来算
After 7 answers the desk ran its first systemic review: it looked for problems that came back across cards but that nobody had named. The clearest signal is time. Result cards took 2 to 4 times their estimate whenever the PI engaged.
"I need to check whether your result is right." The PI wants per-point fits, masks and the exact command, which do not fit in 10 minutes.
"我需要检查你的结果对不对。"PI 想看逐点拟合、掩膜和具体命令，这些塞不进 10 分钟。
Standalone learning cards wait独立的学习卡没人读完
12 learning cards and 1 article card were on the desk; none was finished. Desk time went to project results.
台上有 12 张学习卡和 1 张文献卡，一张也没读完。PI 的时间都花在了项目结果上。
The second front end
第二个前端
A review site on Cloudflare, for readers outside the projects放在 Cloudflare 上的评审站，给项目外的读者
"Then route this part to Cloudflare? It tests our second option, and it lets other people give feedback too.""那么就把这部分分流到 CF ？一方面测试我们的第二方案，一方面让别人也能一起提供反馈意见？"The PI, 2026-11-07PI，2026-11-07
The main desk is a private claude.ai page. The review site shows the same cards, but only for the projects the PI chose to share, to readers such as physics graduate students who do not know the project. Their answers come back into the same learning loop.
主工作台是一个私有的 claude.ai 页面。评审站展示同样的卡片，但只限 PI 选择共享的项目，读者是物理研究生这类不了解项目的人。他们的回答回到同一个学习闭环。
A light gate, stated honestly一道轻量门槛，明说它的强度
A reader types the author's name once. The card bundle is encrypted (AES-256-GCM) and its key is wrapped once per accepted name (PBKDF2-SHA256, 310,000 rounds). This keeps casual visitors and search engines out; it is not access control, and the design says so.
Every save goes to a small serverless function that keeps the latest answer of each browser to each card. Each card shows the other reviewers' answers, so they comment together.
Quantity, component, band, region and data, before any number. A card without it fails the check.
在任何数字之前先写清：哪个量、哪个成分、哪个能段、哪个区域、哪份数据。缺了这个框，卡片通不过检查。
proposed待决定
An audit drawer
审计抽屉
A folded section per result with the per-point fit plots, spectra, masks and the exact command, outside the minutes count. The story stays at 5 to 10 minutes; the audit has no limit.
One result per card, timed by the reader's real pace
一张卡只放一个结果，按读者真实速度估时
One headline number per result card, and a minutes estimate from the PI's measured pace per card kind instead of a fixed reading speed.
每张结果卡只有一个核心数字；时长按 PI 在每种卡片上实测的速度估计，而不是固定阅读速度。
proposed待决定
Rendered math with word equations
渲染公式，并配"文字方程"
Formulas render properly, and each term is listed with its name, meaning, value and source under the equation.
公式正确渲染，并在公式下面逐项列出每一项的名称、含义、数值和来源。
proposed待决定
A visible status for every note
每条批注都有可见的进度
A note that asks for work shows "asked, then task, then card" under it, so the reader sees what happened to it.
提出任务的批注下面显示"已提出、已派任务、已出卡片"，读者能看到它的去向。
testing验证中
Four hypotheses, cheapest test first
四个假设，先做最便宜的测试
Learning gets read when it sits inside the result that needs it. A "What is measured" box removes the fatal-ambiguity notes. The PI prefers the Chinese view for long cards. The PI's real pace is about 4 minutes per 100 words of a result story.
学习内容嵌在需要它的结果卡里才会被读。"测的是什么"框能消除致命的歧义批注。长卡片 PI 更愿意看中文。PI 读结果故事的真实速度约为每 100 词 4 分钟。
idea设想
Spaced re-checks
间隔复测
Re-ask a concept the reader got right after about 3, 7 and 21 days, as one item on a later card.
读者答对的概念，在约 3、7、21 天后各再问一次，作为之后某张卡片上的一道题。
idea设想
Small figures you can operate
可以动手操作的小图
Two to four sliders over one toy model, so the reader moves a parameter and sees the result change.
在一个玩具模型上放两到四个滑块，读者拖动参数就能看到结果怎么变。
idea设想
More front ends
更多前端
The card format is one JSON file per card, so the same cards can be tried on other surfaces. The Cloudflare site was the first such test.