bean retweeted
My skills repo has 160K stars, 7.5m downloads...
...and no tutorial.
So, here it is. Watch me walk through the essential skills:
- /grill-with-docs
- /to-spec
- /to-tickets
- /implement
- /code-review
It's the whole flow, end-to-end. Enjoy:
bean retweeted
Just added Best Practices for Claude Code to our docs! 🥳
Always looking to add more from the community though, what setups/patterns have been working well for you?
code.claude.com/docs/en/best…
最近又一个概念火了,叫拉尔夫循环(Ralph Loop)。
简单来说,就是你告诉 Agent 一个任务,比如:帮我生成一张图;Agent 开始执行,执行了很多轮之后,可能 Agent 就停下了,轮次上限到了;那有了这个拉尔夫循环的情况下,当 Agent 决定 "我完成了" 要退出时,拉尔夫会拦截这个退出,然后把同样的 prompt 再喂一遍。这个时候,Agent 之前的上下文还在(有的是持久化到了文件里,有的实现是复用上下文)、但是之前做的代码修改已经改掉了,Agent 会从上次停下来的地方继续干活。
这个有点像实习生觉得自己做完了就停了,然后老板甩着鞭子继续要求实习生干活,一直干到老板满意为止,这个形象很绝。
本质上,也是通过多轮迭代想要解决 LLM 上下文不够导致 Agent 不能很好完成长任务的问题。
目前流传最广的是说,有个老哥靠这个“鞭打”Agent,跑了一个 3 个月的循环,直接干出了一个完整的编程语言。
Claude Code 目前也支持了这个拉尔夫循环的插件:
github.com/anthropics/claude…
bean retweeted
在小红书给网红画图,希望引流,结果被一群人 diss 是舔狗……哎,你们想舔就舔嘛,哥们儿只是想卖图而已。
另,感谢 @dotey 的分享,我小改了一下 prompt,得到下面的图,感觉有点意思
bean retweeted
Our X Social Platform Deserves More Value🚀
From Web1 to Web2, we evolved over decades.
Now, with NebX, we’re driving the leap from Web2 to Web3. 🌐
NebX is transforming X into the launchpad for Web3. Join us and be the pioneers of this new era. ✨
NebX: Where the future is built. 🔗
#NebX $XCOIN #Airdrops #NFTdrops
bean retweeted
We're excited to announce our $15M Series A led by @foundersfund to build the Metalayer, the end game for Ethereum rollups 🌋
This round brings our total funding to $25M and will help us advance our goal of making Ethereum go web-scale.
bean retweeted
《Linux 内核实验》中文翻译
布加勒斯特理工大学的 Linux 内核教学课程,动手实践设备驱动的编写。
比较硬核的底层知识,各位X 友别看太多,容易沉迷,不利于做产品。
linux-kernel-labs-zh.xyz/lec…
bean retweeted
老外整理的特别详细的 Python 3 Cheat Sheet PDF 文件,点此下载
perso.limsi.fr/pointal/_medi…
bean retweeted
简明扼要,2024 Rust学习路径(附相关学习资源)
Rust 是一种开源的系统编程语言,旨在提供三个主要的优点:安全性、速度和并发性。它由Mozilla研究院开发,并首次在2010年发布。Rust特别关注内存安全,并通过一套称为所有权(ownership)的规则来管理内存,这些规则不需要垃圾收集器的介入……
下面精选几个比较不错的资源帮助同学们学透Rust,建议收藏⬇️
1. Comprehensive Rust
google.github.io/comprehensi…
由Google的Android团队开发的免费在线课程,提供了关于Rust的广泛而详细的教育,该课程旨在迎合各个级别的学习者,从Rust语法的基础知识开始,逐步深入到泛型和错误处理等高级概念!
2. Easy Rust
用最简单的英语表达来讲解rust这门语言 一本简单英语写成的 Rust 教材,既可以学习英语增加可理解性输入又可以深入了解rust这门语言,双管齐下。阅读难度大致在初三到高三之间。
网站:
dhghomon.github.io/easy_rust……
GitHub:
github.com/Dhghomon/easy_rus……
YouTube:
youtube.com/playlist?list=PL…
3. Rust语言圣经
course.rs/about-book.html
一本通俗易懂的开源书籍,解决rust难学的问题。
4. This week in Rust
this-week-in-rust.org/
每周更新rust生态的各种事件,随时掌握rust的大方向!
5. 通过例子学 Rust
rustwiki.org/zh-CN/rust-by-e…
边学边练,通过各种例子和各种小项目彻底吃透Rust语言有关的知识点,通俗易懂,是rust初学者必备的参考书。
bean retweeted
Gretel 最近发布了世界上最大的开源 Text-to-SQL 数据集,标志着业界对数据质量的重新关注,在当前的 LLM 和 AI 领域内,从 Falcon、Phi-1.5、Phi-2、Aya 到 Gemini 等尖端模型,巨大的性能提升再次证明了以数据为中心的 AI 发展方法的重要性。Gretel 通过提供工具生成高质量的合成数据来解决数据质量问题,推动了向数据中心 AI 的过渡。这个新发布的 Text-to-SQL 数据集是目前可获得的最大且最多样的合成 Text-to-SQL 数据集,包括 105,851 条记录,覆盖 100 个不同的领域,涵盖了 SQL 任务的广泛数组,数据集发布在 Hugging Face 上,并在 Apache 2.0 下开源。
博文:gretel.ai/blog/synthetic-tex…
训练集:huggingface.co/datasets/gret…
关键信息整理:
bean retweeted
github.com/PlexPt/awesome-ch…
超高人气的prompt项目 , 需要自取。
bean retweeted
Great Python notebook guide to LoRA fine-tuning Gemma with MLX LM.
gist.github.com/alexweberk/6…
Fine-tuned @Google's Gemma model using MLX on a MacBook Pro M3 Max 🔥
Model: Gemma-7b-it
Dataset: @teknium 's awesome teknium/trismegistus-project dataset
Here's the code for the fine-tuning and uploading to Huggingface:
gist.github.com/alexweberk/6…
bean retweeted
So midjourney currently has 12 parameters that let you tune & control the aesthetics/style of your generations
These are the ones I've been using the most:
--style raw⎜push towards photorealism
--style random⎜explore styles in --v 5.2
--sref urlA⎜use images as styles in --v 6
--sw 100⎜tunes --sref influence, 0-1000
--stylize 150⎜tunes MJ influence, 0-1000
--weird 150⎜makes things weird, 0-3000
--chaos 13⎜more unexpected grids, 0-100
--no itemA⎜tell MJ what you want removed
--quality 0.5⎜soften/sharpen details, .25-5
--ar 16:9⎜sets aspect ratio, any whole # pair
--iw 0.5 ⎜sets image influence vs text, 0.1 - 3
--stop 80⎜low= blurry/noisier images, 10-100
NOTE: parameters ALWAYS go AFTER your prompt, at the the very END
You can learn a lot by testing the extremes of each parameter against different styles of prompts
i.e. on a simple photo prompt:
photo of a man by a lake --stylize 0
photo of a man by a lake --stylize 50
photo of a man by a lake --stylize 100
photo of a man by a lake --stylize 400
photo of a man by a lake --stylize 1000
> The ideal parameter values will change depending on a) your prompt, and, b) the style you're after
> Testing is the best way to discover your favorite styles & effects. FWIW I run tests like this on pretty much all of my prompts
> Try out different values and combos on a variety of different prompts (simple, detailed) and image styles (photographic, illustrative, etc)
> For example: a lake --stylize 150 --weird 10, or, a lake --style raw --stylize 75 --chaos 10, etc.
REMEMBER: parameters don't work without a prompt to interpret. Your prompt is still in charge. That's why parameters go at the end of your prompt.
WHAT THEY ARE: parameters are little add-ons that always start with `--` and go at the end of your prompt. They accept a range of values
WHAT THEY DO: parameters change the way Midjourney interprets your prompt, letting you explore & tune an infinite amount of aesthetic directions
Have fun out there
bean retweeted
🌟嗨大家好!我们是哥伦比亚大学自然语言处理实验室!我们现在在开发一个可以提供语法反馈和鼓励式反馈的英语学习聊天机器人,想要征求大家在设计方面的宝贵意见!
forms.gle/4ggypHSkU1jYP9hL6
大家也可以在Google Form 里报名参加我们的聊天机器人试用!
✨麻烦各位多多转发!谢谢!
bean retweeted
New (2h13m 😅) lecture: "Let's build the GPT Tokenizer"
Tokenizers are a completely separate stage of the LLM pipeline: they have their own training set, training algorithm (Byte Pair Encoding), and after training implement two functions: encode() from strings to tokens, and decode() back from tokens to strings. In this lecture we build from scratch the Tokenizer used in the GPT series from OpenAI.































