zhuliuwu

生活 · 运动 · 户外

热爱生活。

热爱运动，热爱户外，喜欢探索新事物，不安于现状。我是 zhuliuwu，把还想记住的东西写在这里。

文章

全部归档

  1. 新轨迹里这一格的句子变了之后，把图上那一格改成新行里的那一句，再拿这份新轨迹问那三句

    新轨迹里这一格的句子变了，才把图上那一格改成这一行里的那一句，再拿这份新轨迹问那三句。 After this cell's sentence has changed on the new trace, change that cell on the picture to the sentence in the new line, then ask the three questions of this new trace.

    AI 系列

  2. 这一格指着的文件里，它读到的那一句才算改到；没读到的字，新轨迹里这一格仍是旧的

    同一文件里，这一格拼上页、并写进新轨迹的那一句变了，才算改到；注被丢掉，这一格仍是旧的。 In the file a cell points at, only the sentence it pasted onto the page and wrote into the new trace counts; a dropped note leaves that cell on the old sentence.

    AI 系列

  3. 对不上时，先改那一格指着的文件，再用新轨迹看它变没变

    对不上时，先改那一格右边指着的那一个文件，再走一圈；新轨迹里这一格的行变了，这一格才变了。 When the lines disagree, change the one file that cell points at and walk once more; the cell changed only if its line on the new trace changed.

    AI 系列

  4. 拿这一次的轨迹核对这张图还在不在

    轨迹是这一次按顺序留下的格子、文件和读到的那一句；拿它问三句，三句都过，图才还在这一次上。 A trace is the cells, files, and sentence this run kept in order; ask it three questions, and only when all three pass is the diagram still this run's.

    AI 系列

  5. 从这一系列里留下可以自己改的一张执行图

    执行图是仓库里一张你能改的图，每一格指着这一次真会走到的文件；只改图，这一次仍走文件。 An execution diagram is a picture in the repository you can edit, each cell pointing at a file this run will actually walk; editing the picture alone leaves the run on the files.

    AI 系列

  6. 把一个最小循环接到自己的仓库

    最小循环是你仓库里能指着的四样：拼页、放回、跑门、决定停；别人那一叠上的 87 接不上今晚的门。 The smallest loop is four things you can point at in your repository: assemble the page, put the model back, run the door, and decide whether to stop; an 87 on someone else's stack does not attach tonight's door.

    AI 系列

  7. 基准数字在说什么，又没说什么

    基准数字是一叠固定的页上，一套固定的对照判过了多少次；它不打开今晚的门，也不画出拒绝、确认和边界。 A benchmark number is how many pages in one fixed stack passed one fixed check; it does not open tonight's door, and it does not draw a refusal, a confirmation, or a boundary.

    AI 系列

  8. 对齐在执行里表现为什么：拒绝、确认、边界

    对齐不是一次调用里新长出来的良心；它落成三种能看见的动作：不往那边接、先等人回一句、以及这条线写在续写外面。 Alignment is not a conscience grown inside one call; it lands as three visible acts: do not continue that way, wait for a person, and keep the line outside this span.

    AI 系列

  9. 训练、权重和推理不是同一件事

    训练在调用外面写出新一版排序，权重是坐着不动的那一版，推理是对着这一版读这一页；三件事不能互相顶替。 Training writes a new edition of the ranking outside any call, the weights are the edition that sits still, and inference reads this page against that edition; one job cannot stand in for the others.

    AI 系列

  10. 改行为的三种办法：提示、检索、微调

    要换一种接法，改的是这一页、贴回来的材料，或排序本身；三处不是同一种「教」。 To change the next span, change this page, the passages pasted back, or the ranking itself; those are not three names for teaching.

    AI 系列

  11. 费用、延迟和步数是同一笔账

    多走一步，就要再读一遍前缀、再等一轮、再记一笔；钱、时间和圈数不是三本账。 One extra step reads the prefix again, waits another round, and adds one to the count; money, time, and loops are not three separate books.

    AI 系列

  12. 多个智能体分工时，谁拿目标，谁拿观察

    分工不是多长了几颗脑子，是几页不同的前缀：拿目标的那一页决定何时停，拿观察的那一页才看得见已经发生的事。 Splitting work is not growing extra minds; it is several prefixes: the page that holds the goal decides when to stop, and only the page that holds the observations can see what already happened.

    AI 系列

  13. 错误怎样留在历史里，并污染下一步

    历史是还在页上的前缀；错的观察一旦留下，下一步就对着它打分，窗口外的错账还会被取回来。 History is the prefix still on the page; a wrong observation left there is what the next step scores against, and a wrong line written outside can be fetched back.

    AI 系列

  14. 怎么判断它做对了：例子、对照和失败样本

    做对了不是模型说「我对了」，是页上先有例子和失败样本，运行时再用观察去对照。 “Did it right” is not the model saying so; examples and failure samples sit on the page, and the runtime checks the observation against them.

    AI 系列

  15. 哪些动作必须先问人

    会留下收不回痕迹的门，要写在循环外面的名单上，开之前等人回一句。 Doors that leave a mark you cannot take back sit on a list outside the loop; the runtime does not open them until a person answers.

    AI 系列

  16. 停止条件：完成、上限、重复失败、需要人点头

    循环什么时候必须断：做成了、步数到了、同一扇门反复失败、或下一扇门要等人点头。 The loop has to break for one of four reasons: the job is done, the cap is hit, the same door keeps failing, or the next door needs a person to nod.

    AI 系列

  17. 规划的三种安排：边想边做、先计划、做完再看

    计划不是另一种脑子，是页上字的三种排法：看完这一步再写下一步、先写下清单再走、做完再对照观察。 A plan is not another kind of mind; it is three ways to arrange the wording: decide the next step as you go, write a list first, or look back against the observations.

    AI 系列

  18. 记忆：留在窗口里的，和写在窗口外的

    循环一长，旧观察会堆满这一页；这一次能用的只是窗口里的字，要留下去就得写到窗口外。 As the loop grows, old observations fill the page; this call can only use wording still in the window, and anything that must last has to be written outside it.

    AI 系列

  19. 智能体：目标加上会停下来再出发的循环

    智能体不是另一种脑子，是一句还没做完的目标，加上会在一扇门之后停住、再把模型放回页末的循环。 An agent is not another kind of mind; it is an unfinished goal plus a loop that stops after a door and puts the model back at the end of the page.

    AI 系列

  20. 一次工具调用里，模型、运行时、观察各做哪一步

    一次调用拆成三步：模型只写出名字，运行时去跑，观察把结果写回下一页。 One tool call has three jobs: the model writes a name, the runtime runs it, and the observation is written back onto the next page.

    AI 系列

  21. 工具：模型接触外部世界的唯一方式

    工具是运行时在模型外面能真的跑起来的一件事；模型接触厨房、接口或文件，只有这一条路。 A tool is a named job the runtime can actually run outside the model; that run is the only way the model touches a kitchen, an interface, or a file.

    AI 系列

  22. 嵌入：用一组数表示像不像

    嵌入是把一段字收成一组数，用两串数的远近来挑哪几段像今天这句。 An embedding turns a stretch of text into a group of numbers; closeness between those groups is how retrieval picks stretches that look like this question.

    AI 系列

  23. 长资料进不了窗口：检索是在补这一刀

    窗口先按块数砍一刀；检索是运行时在拼页时，从窗口外面挑几段贴回来。 The window cuts by piece count first; retrieval is the runtime picking a few passages outside the window and pasting them while it assembles the page.

    AI 系列

  24. 让回答按约定的格式出现

    约定格式是气泡要留下的骨架：页上的字会推它，第 4 拍的门会挡住不合的块。 An agreed format is the skeleton the bubble must leave: words on the page can lift it, and a gate on beat 4 can block pieces that break it.

    AI 系列

  25. 它为什么会编造，以及编造和工具观察的差别

    编造是接得顺的下一段；观察是工具跑完之后写回页上的字。 Fabrication is a smooth next span; an observation is text a tool writes back after it actually ran.

    AI 系列

  26. 温度：为什么同一句会有不同答案

    温度改的是每一步怎么从排序里取下一块，不是让模型重新想一遍。 Temperature changes how each step takes the next piece from the ranking; it does not make the model think the question over again.

    AI 系列

  27. 对话里的三种角色：系统、用户、模型

    系统、用户、模型是同一页上的三段标签，不是三条通道。 System, user, and model are three labels on one page, not three channels.

    AI 系列

  28. 提示词：你交给它的全部材料

    提示词是运行时拼好的那一页：系统、旧轮次和你刚打的字，都在里面。 The prompt is the assembled page: the system line, older turns, and the line you just typed all sit on it.

    AI 系列

  29. 上下文窗口：这一次它能看见多少

    一次调用能看见的，是窗口里那串编号；系统、旧轮次、你刚打的字和正在接的回复，共用这一格容量。 One call can see only the numbered row that fits the window; system text, older turns, your line, and the growing reply share that one capacity.

    AI 系列

  30. 一次回复是怎么一个片段一个片段算出来的

    一次回复是个循环：给词表里每一块打分，取下一块接上，再对更长的前缀打分，直到停下。 A reply is a loop: score every vocabulary piece, append one, then score the longer prefix again until it stops.

    AI 系列

  31. Token：模型看见的不是字，是切好的片段

    便签先被切成词表里的片段并换成编号，模型每一步接上的是下一块。 The note is cut into vocabulary pieces and numbered first; each step appends the next piece.

    AI 系列

  32. 模型在做什么：接下去最可能的下一段

    一次模型调用给面前的文字接上最像的下一段，这一步并不打开冰箱，也不查资料库。 A model call appends a fitting next span to the text in front of it, without opening the fridge or a database.

    AI 系列

  33. AI 智能体的执行原理

    一篇教程：智能体怎样从目标出发，在推理、工具和观察之间循环，直到把一件事做完。

    AI

  34. Class Loader

    JVM ClassLoader 的层次，以及双亲委派是怎么往上找、再回到发起者的。

    Java

  35. Cloudera Manager Free Edition Installation Guide

    在 Ubuntu Server 上安装 Cloudera Manager 免费版，并搭起 CDH 集群的步骤记录。

    Hadoop

  36. hello world

    第一篇。确认这个博客能正常工作。

    随笔