<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>harness on Docs</title><link>https://hiraethecho.github.io/docs/ai/harness/</link><description>Recent content in harness on Docs</description><generator>Hugo --0.165.0</generator><language>en</language><managingEditor>wyz2016zxc@outlook.com(Hiraeth)</managingEditor><webMaster>wyz2016zxc@outlook.com(Hiraeth)</webMaster><atom:link href="https://hiraethecho.github.io/docs/ai/harness/index.xml" rel="self" type="application/rss+xml"/><item><title>work with ai on a board</title><link>https://hiraethecho.github.io/docs/ai/harness/board/</link><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><author>wyz2016zxc@outlook.com(Hiraeth)</author><guid>https://hiraethecho.github.io/docs/ai/harness/board/</guid><description>Another way work with ai</description><content:encoded><![CDATA[<h1 id="board">
  <a class="anchor inpage" href="#board">#</a>Board</h1>
<p>Another way work with ai.</p>
<p>the center is files (pinned on a board), human and llm modify the files by diff. the files have higher priority in context.
the essential commutates from human to llm is the diff file.<br>
(say there is main branch and a dev branch, we work on dev branch. then there are two branches <code>human</code> and <code>ai</code> rebased on dev, human writes on human branch. there should be a orchestrator that give llm the diff file on human branch, and merge, rebase these two branches on dev branch. maybe <code>jj</code> is better on this workflow?)</p>
<p>there can be another channel where human and llm chat with each other, human can talk to llm with &ldquo;pointing some context in the files&rdquo;, but not the llm which write files.</p>
<p>The idea is, there should be a llm that &ldquo;do the actual work&rdquo;, and it get the information in a more efficient way, not using the chat history.</p>
<h2 id="motivation">
  <a class="anchor inpage" href="#motivation">##</a>motivation</h2>
<p>the human attention is poor, and we have to through away a lot stuffs that we consider not important, or dead end.</p>
<h2 id="context">
  <a class="anchor inpage" href="#context">##</a>context</h2>
<p>上下文应该是有不同层级的，它们在 llm 中地位应该不同。
Work Space 中的代码/文本应当是一等公民，由某个特化的 llm 来处理，并且应该是<strong>无历史</strong>的。
（人，或其他 llm/agent）对文件的修改也应该作为一个整体，同样被一个特化的 llm 处理，通过 diff 来理解 <em>修改</em>。显然人类的修改应该地位更高，对文本（workspace里那些东西）的修改，应该和在对话框里的 prompt 是同样地位的。
此外，还有一个传统对话的 llm，用来保留全部记忆。但它不直接参与修改，而只是负责统筹。</p>
<blockquote class="alert"><p>vibe coding 实践中的习惯：如果在agent操作时人类干预直接修改文件，这很难被 llm 感知。有时候简单改一些 typo 或小细节，为了嵌入到模型记忆中，需要通过 prompt 实现（还容易漂移，非常麻烦）</p>
</blockquote><p>学习一下 Aider 的思路。用 jujustu 来管理 diff.</p>
<h2 id="哲学">
  <a class="anchor inpage" href="#%e5%93%b2%e5%ad%a6">##</a>哲学</h2>
<p>从哲学理论中吸收一些（通常不科学的）概念。
在后现代语境下，什么是 <strong>文本/context</strong> ？生活材料本身也是文本，概念的解释只能在 context 中。
从语言哲学里学习一下，怎样划分原子语句。</p>
]]></content:encoded></item></channel></rss>