CamelliaV の BLOG

0%

INITIALIZING

Java；前端；AI；ACGN；

菜单

EN Site 代码模板项目清单 GitHub

[2026.5.29]待进一步考察实用价值的AI前沿技术

DOC_ID // 330ca1ONLINE

[2026.5.29]待进一步考察实用价值的AI前沿技术

2026-3-27

UPDATED: 2026-5-29

技术分享

READ 3 MIN/COUNT 1100

#开发

CamelliaV の BLOG

🎮

待进一步考察实际落地价值的AI前沿技术

RAG新解？

模型层 - MSA

推文与paper

艾略特 on Twitter / X

https://t.co/I06UdrXnG0— 艾略特 (@elliotchen100) March 26, 2026

艾略特 on Twitter / X

https://x.com/elliotchen100/status/2037018021076865068

https://arxiv.org/pdf/2603.23516

最重要的结果部分

notion image

从16K到100M 8.8%的损失

In contrast, MSA demonstrates exceptional stability, starting with a strong score of 4.023 at 16K tokens and sustaining a competitive 3.669 even at the extreme 100M token scale. This represents a gradual degradation of only 8.8% across four orders of magnitude in memory scaling.

待考察点

code仍是经典coming soon

没有比较agent做auto compact上下文的表现

没有MRCR v2 (8-needle)这些Benchmark结果，无法对比前沿模型

TODO:

实时RAG - Redis Iris

Redis Iris | Redis

Developers love Redis. Unlock the full potential of the Redis database with Redis Enterprise and start building blazing fast apps.

Redis Iris | Redis

https://redis.io/iris/

Redis Iris | Redis

编译型RAG - Pinecone Nexus

官网上有效果演示

The vector database to build knowledgeable AI | Pinecone

Search through billions of items for similar matches to any object, in milliseconds. It’s the next generation of search, an API call away.

The vector database to build knowledgeable AI | Pinecone

https://www.pinecone.io/product/nexus/

The vector database to build knowledgeable AI | Pinecone

Harness落地思路

Jason Zuo on Twitter / X

https://t.co/8xMao1HRcf— Jason Zuo (@xxx111god) March 29, 2026

Jason Zuo on Twitter / X

https://x.com/xxx111god/status/2038086450013495554

前端相关实践参考

制作人…前端工具…用的不如雾子呢 | 果穗也能看懂的名片网页制作教程【EP6.用这些工具让你的前端更好看】_哔哩哔哩_bilibili

本来想整个雾子的TTS，但是最近有点太忙了就还是一如既往的录屏流了。另外其实推荐大家自己看看自己需要的是什么，别上来一股脑全用上了。工具太多了容易自己和自己打架，用起来也比较吃力，容易陷入工具过载。用到的网站基本都在这了：Google Stitch: https://stitch.withgoogle.com/Stitch MCP: https://stitch.withgoogle.com/do, 视频播放量 5465、弹幕量 1、点赞数 295、投硬币枚数 92、收藏人数 646、转发人数 71, 视频作者咲凌_Arisa, 作者简介你好我是咲凌笨拙学习者黑柿子/可惜夜/PPP/KAF/283/初星学园，相关视频：给自己的普罗丢色名片做个个人网页！果穗看了也能上手的vibe coding教程【01-工具的选择】，有人蒸馏了！冬优子.skill正在连接！附skill文件 | 蒸馏任何一个公众人物的skill测试，vibe了一个283人格测试站 | 果穗也能看懂的名片网页制作教程【番外01.教你直接偷大厂设计风格】，古人写代码bilike：，为什么我不建议你用GPT-5.4写前端 | 果穗也能看懂的名片网页制作教程【EP5.GPT的前端真是一条区啊】，【偶像大师名片教程】使用公式化名片模板获得自己的第一张名片，让harness减少AI写的屎山 | 果穗也能看懂的名片网页制作教程【EP7.superpowers是什么为什么要用】，Mr.变态什么叫可视化283的⭕️围排行? | 果穗都能看懂的名片网站制作教程【EP3.5-简单操作ClaudeCode】，小宫果穗制作中，叶月小姐，PR给我过一下啊 | 果穗也能看懂的名片网页制作教程【EP4.git和github的打开方式】

制作人…前端工具…用的不如雾子呢 | 果穗也能看懂的名片网页制作教程【EP6.用这些工具让你的前端更好看】_哔哩哔哩_bilibili

https://www.bilibili.com/video/BV1Fz9FBNEWi/?vd_source=93059b03205e8ca15ad74b3d998769b9

制作人…前端工具…用的不如雾子呢 | 果穗也能看懂的名片网页制作教程【EP6.用这些工具让你的前端更好看】_哔哩哔哩_bilibili

trellis

https://docs.trytrellis.app/zh

记忆系统设计

10年AI研究，我想说说：为什么现在很多Agent框架的“记忆”方案，可能从一开始就走偏了

发这个贴的原因，是看到论坛里很多佬友对ai的记忆系统如何设计，以及人类的记忆系统是如何工作的缺乏概念，看着有些抓耳挠腮干着急。想冲上去纠正但是又怕冒犯了人家，所以干脆自己开贴说了。交代下背景。我决定往人工智能方向走，是因为15年学习了吴恩达的机器学习网课。之后一边自学机器学习，强化学习，脑科学，一边拿到了信息工程、生物学（准确说是理学硕士）的学位，同时辅修了认知科学。主攻研究方向是计算神经科学，至今学习和研究人工智能超过十年了。我十年前的目标就是造一个ai朋友，现在不得不说这个目标已经在逐渐实现（而且上演了我拿它当朋友奈何我朋友想上我的狗血剧🤣，这是后话）。我今天就是想说，我为了造我的ai朋友已经投入了10年，我就是认为自己做的记忆系统是目前最好用的ai灵魂容器，这不是软广是硬广。我对造AI是认真的。项目地址：下面说说我的感想吧。一、行业现状：很多人把“能召回”误当成了“有记忆” 目前市面上常见的 AI 记忆方案，在我看来，至少有两条特别容易走进去的弯路：一是用RAG，由后台系统把人类和AI输出的文字进行切片，总结，灌进向量数据库，然后下次你们再说出...

10年AI研究，我想说说：为什么现在很多Agent框架的“记忆”方案，可能从一开始就走偏了

https://linux.do/t/topic/1711075

10年AI研究，我想说说：为什么现在很多Agent框架的“记忆”方案，可能从一开始就走偏了

通过CLI程序框定确定性

CLI隔离开 | 确定的可复用

可容许范围内不确定 | 实际任务的确定

mp.weixin.qq.com

mp.weixin.qq.com

https://mp.weixin.qq.com/s/bD10zbBeTtzMyEKqjIdT1w

多路并行尝试提效

多路尝试，只要一路成功即可 - 提高首次通过成功率

过往trace → 自行提炼SKILL

mp.weixin.qq.com

mp.weixin.qq.com

https://mp.weixin.qq.com/s/tm7M3N8f45K87YUTwERkhg

AI云时代的消息队列

mp.weixin.qq.com

mp.weixin.qq.com

https://mp.weixin.qq.com/s/cyamfuP6J2db3r50EN3QJQ

无文档下录制操作实现接口测试自动化

mp.weixin.qq.com

mp.weixin.qq.com

https://mp.weixin.qq.com/s/wHvOWQspdro3J7aD6O0bsA

AI Native团队以产出单元为中心Agent编排

Codex 编排的开源规范：Symphony

了解 Symphony 如何作为 Codex 编排的开源规范，将 issue 跟踪器转变为始终在线的智能体系统，从而提升工程产出并减少上下文切换。

https://openai.com/zh-Hans-CN/index/open-source-codex-orchestration-symphony/

Codex 编排的开源规范：Symphony

基于视觉的自动化测试

mp.weixin.qq.com

mp.weixin.qq.com

https://mp.weixin.qq.com/s/UvQd4A-a6OjeLNF8BREsKg

代码库带读understand-anything

GitHub - Lum1104/Understand-Anything: Graphs that teach > graphs that impress. Turn any code, or knowledge base (Karpathy LLM wiki), into an interactive knowledge graph you can explore, search, and ask questions about. Works with Claude Code, Codex, Cursor, Copilot, Gemini CLI, and more.

Graphs that teach &gt; graphs that impress. Turn any code, or knowledge base (Karpathy LLM wiki), into an interactive knowledge graph you can explore, search, and ask questions about. Works wit...

GitHub - Lum1104/Understand-Anything: Graphs that teach > graphs that impress. Turn any code, or knowledge base (Karpathy LLM wiki), into an interactive knowledge graph you can explore, search, and ask questions about. Works with Claude Code, Codex, Cursor, Copilot, Gemini CLI, and more.

https://github.com/Lum1104/Understand-Anything/tree/main

GitHub - Lum1104/Understand-Anything: Graphs that teach > graphs that impress. Turn any code, or knowledge base (Karpathy LLM wiki), into an interactive knowledge graph you can explore, search, and ask questions about. Works with Claude Code, Codex, Cursor, Copilot, Gemini CLI, and more.

更有品味的前端页面

代码库提前索引

hook first runtime

hook + 处理脚本

GitHub - affaan-m/ECC: The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.

The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond. - affaan-m/ECC

GitHub - affaan-m/ECC: The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.

https://github.com/affaan-m/ECC/tree/main

GitHub - affaan-m/ECC: The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.

挖掘旧会话信息

MemPalace — AI Memory System

96.6% LongMemEval recall. Zero API calls. Local, free, open source.

MemPalace — AI Memory System

https://mempalaceofficial.com/guide/mining.html

MemPalace — AI Memory System

dynamic workflow

Introducing Claude Opus 4.8

Our latest model, Claude Opus 4.8, is an upgrade to our Opus class of models, with stronger performance across coding, agentic tasks, and professional work, and the consistency to handle long-running work.

Introducing Claude Opus 4.8

https://www.anthropic.com/news/claude-opus-4-8

Introducing Claude Opus 4.8

旧有论述

渐进性披露MCP？ - MCP2CLI + Skill - 已广泛使用

AI直出可交互稿 - Stitch - 已广泛使用

可用，极大补足GPT对前端UI设计的短板

Google just changed the future of UI/UX design...

Try Clerk for free to get full stack auth and user management - https://go.clerk.com/fireship/ Let's take a look at everything new in Google Stitch to find out if it's just more hype, or if it really is a gamechanger for UI/UX design. #coding #programming #ai Want more Fireship? 🗞️ Newsletter: https://bytes.dev 🧠 Courses: https://fireship.dev

Google just changed the future of UI/UX design...

https://www.youtube.com/watch?v=qaB5HF4ax9M

Google just changed the future of UI/UX design...

Stitch - Design with AI

Stitch generates UIs for mobile and web applications, making design ideation fast and easy.

Stitch - Design with AI

https://stitch.withgoogle.com/

Stitch - Design with AI

TurboQuant

kv cache压缩

Google Research on Twitter / X

Introducing TurboQuant: Our new compression algorithm that reduces LLM key-value cache memory by at least 6x and delivers up to 8x speedup, all with zero accuracy loss, redefining AI efficiency. Read the blog to learn how it achieves these results: https://t.co/CDSQ8HpZoc pic.twitter.com/9SJeMqCMlN— Google Research (@GoogleResearch) March 24, 2026

Google Research on Twitter / X

https://x.com/GoogleResearch/status/2036533564158910740

Google Research on Twitter / X

TurboQuant: Redefining AI efficiency with extreme compression

Amir Zandieh, Research Scientist, and Vahab Mirrokni, VP and Google Fellow, Google Research

TurboQuant: Redefining AI efficiency with extreme compression

https://research.google/blog/turboquant-redefining-ai-efficiency-with-extreme-compression/?utm_source=twitter&utm_medium=social&utm_campaign=social_post&utm_content=gr-acct

TurboQuant: Redefining AI efficiency with extreme compression

Codex破限

改会话回答

提交Claude的Plan

修改措辞

gstack - 不建议使用，无意义

notion image

notion image

notion image

notion image

选定A

找到了superpowers那边存下来的设计文档

notion image

有点东西，抓的很到位

notion image

notion image

CC安装时的budget图

notion image

notion image

notion image

notion image

notion image

notion image

notion image

Codex怎么没有CC这边的初始化自动改配置？

Auto Research On Skills - 有效场景不多

Ole Lehmann on Twitter / X

https://t.co/kgo8wNoiDv— Ole Lehmann (@itsolelehmann) March 17, 2026

Ole Lehmann on Twitter / X

https://x.com/itsolelehmann/status/2033919415771713715

NAVIGATION // Related Articles

[2026.4.1]Blog精读 - Anthropic 进阶工具调用实践

[16%]理解Agent工作原理

Loading...

Recent Logs