Yiran's Blog
Recent content on Yiran's Blog
马上订阅 Yiran's Blog RSS 更新: https://zdyxry.github.io/atom.xml
Weekly Issue-tokui
文章 链接到标题
技术 链接到标题
Agents Need “State” Instead of “Memory”
使用 memory 一词的时候我们不妨想想,到底有什么是值得记忆的?比如,作为终端用户来说,我希望我的 chatbot 能记住我们之间每一次对话的每个细节,这是一种 memory。但是如果我的 agent 要处理一个特定的任务,它一轮一轮地调用 tools,它用 memory 记住每一个 tool 的每一个具体输出,这有价值吗?
因此相比于 memory,我们应该换一个词称呼它:state。Agent 在每一轮运行过程中,它首先根据自己的 state 做出决策、调用工具,然后根据 tool output 更新自己所处的 state,再进入下一轮运行。就像一个寻路机器人要知道自己在迷宫中的位置及先前探索过的路线,然后据此做出下一步移动。
我最近也在考虑,如何处理我与各个 LLM 之间的记录数据,直接全量保存没有意义,很多时候一些对话只是简单问题,手动保存不现实,可能也没有必要。现在会把一些长对话,在得到阶段性答案后手动记录下来。
我发现一个编译器贡献排前三的大牛打字使用二指禅,有天早餐时他坐我对面,我问起为什么只用两个手指操作键盘,会不会效率不高。他说两个手指的速度已经够了,敲键盘能跟上我的思考速度就行。
Typesense | Open Source Alternative to Algolia + Pinecone
To Raise VC, or Not. Choosing The Road Less Travelled. | Typesense
We are concerned that “launching” a new “product line” (selling stocks) and bringing on a whole new group of customers (investors), would cause us to lose our precious bandwidth that could have otherwise been spent on our core search product that our primary group of users and customers expect from us. After all, the “company stock” product line would not exist without the core search product.
When a large team congregates around work that cannot be parallelized, we’ve observed that new work gets created to keep everyone busy - new nice-to-have features get worked on. This ends up adding complexity to the product. New team routines are invented because there’s bandwidth available. This ends up adding communication overhead and layers between users and builders and slows down the pace of innovation. This sadly becomes a vicious cycle...
剩余内容已隐藏