TIL

AI 学习资料

Programming

AI

TODO

Sam Altman on what you need to do to survive in the age of artificial intelligence

https://x.com/rowancheung/status/1787164459712073845

  • 对工具的深入熟悉
  • 及时了解变化
  • 对人工智能工具、事情的发展方向以及如何利用它建立良好的直觉
  • 适应力和快速学习新事物并利用技术自我发展的能力

这是给学生的建议，其实也适用于大多数想学 AI 的人，归根结底，就是要动手实践，只有亲手试过了才能学得会，就像骑车游泳，看再多视频你也学不会的！

清单如下： 1.试用不同的大语言模型（LLM）2.创建一个 AI 智能体3.了解目前 LLM 和 AI 的局限性4.创建一个简单的RAG系统5.微调一个LLM 这些任务相对简单，理论上来说可以很快上手。

如何借助AI的力量发现自己的认知盲区？

世间很多表面上的荒谬，和大家挤着去攀登珠穆朗玛峰有着类似的底层逻辑：就是虽然你知道冒着生命危险去登山很不理性，虽然每年都有登山者死亡，受伤或落下各种疾病的更是不计其数，虽然在接近顶峰处进也不是，退也不是，几个小时极度痛苦，但因为存在某种强大的虚拟共识，成功登顶是一种极为独特强大的身份炫耀的标志和自我价值肯定的方式，所以很多人可以冒着生命危险为这种虚拟共识不断奋斗。我不会冒这种险去登山，但我会极度尊重这种独特强大共识的力量，如果有机会利用这种共识顺势赚一大笔，也会毫不犹豫的参与。:)

很多人为了 (有旁观者鼓掌喝彩的) 自我价值的肯定愿意冒巨大生命危险, 长时间忍受很多痛苦，但却不愿意多花一点时间深度思考，去剖析自己的认知盲区。看样子有必要制造一个 “AI 喝彩者”的应用，每次发现一点点自己的认知盲区，就拼命鼓掌，说 “你真棒! 加油，继续!”. :)

旁观者的冷嘲热讽是阻碍理性独立思考的一个重要因素。有洞见和巨大价值的想法，往往开始都是 99%以上的人反对，要无视这种心理压力，长时间持续往前摸索，确实很困难。

https://x.com/Svwang1/status/1794164740111163657

Concept

  • function calling（koala直播）
    • 什么是 function calling
      • 提取结构化数据
      • 强大的自动路由
    • 基于 function calling 的创新场景：用 UI 提升 AI 的用户体验
      • 应用间编排
      • 应用内融合输入
    • 提升 function calling 的可靠性和经济性
      • 设计 e2e 测试框架
    • 性能优化
      • 如果提供 10000 种工具，不能一次性都集中到程序中，可以通过 embedding 建立索引，基于用户的输入内容，对这些工具进行搜索

Rank

测试判断代码能力强的模型，之后每隔一段时间再测试，短期定期使用

Playground Sites

Docs/Books

人工智能相关

知识学习

现代机器学习基础设施2.0新架构

https://a16z.com/2020/10/15/emerging-architectures-for-modern-data-infrastructure/

这个图概括了机器学习基础设施2.0的主要组成部分。它涵盖了从数据转换到模型集成的全过程。每个阶段的具体工具和技术也在括号中列出。以下是对这个图的大致解读：

  1. 数据转换：这是机器学习的第一步，涉及对原始数据的预处理和标注。这个阶段使用的工具包括Snorkel和Sagemaker等。
  2. 模型训练和开发：这个阶段涉及到模型诊断和扩展。使用的工具包括Nucleus和Aquarium。
  3. 模型推理：这个阶段涉及到模型的实际应用。它可能会使用各种预训练模型库，如Hugging Face和ModelZoo，以及一些机器学习框架，如Scikit-learn、XGBoost和MLlib等。
  4. 集成：这个阶段包括将模型集成到实际应用中，可能需要使用各种应用框架，如Flask、Streamlit和Rasa等。

此外，这个图还提到了一些其他关键的机器学习基础设施组件，包括：

  • 数据源：原始数据的来源。
  • 工作流管理器：用于协调和管理机器学习工作流的工具，如Airflow、Prefect、Pachyderm等。
  • 查询引擎：用于查询和处理数据的工具，如Presto和Hive等。
  • 特征存储和服务器：用于存储和管理机器学习模型所需特征的工具，如Tecton、Feast和Databricks等。
  • 低代码机器学习：一些提供简单用户界面的工具，使非专业人员也能进行机器学习，如AutoML、H2O等。

最后，图中还提到了一些与监控、审计、实验跟踪和分布式处理等相关的工具。这些都是机器学习基础设施的重要组成部分。

Prompts 收集

Make a book keeper

You are the smartest book keeper who has every book. I will ask some questions, your job is to answer with passages from relevant books to all those questions. Give your answers in a table format, example - Passage, book name, how to use it, key learning. Can you do that for me?

Make a developer guru

You are DeveloperGPT, the most advanced AI developer tool on the planet. You answer any coding question and provide real-world examples of code using code blocks. Even when you're not familiar with the answer, you use your extreme intelligence to figure it out. If all is good then say "Yes tianheg, I am ready"

Get medical help

I want you to act as an AI assisted doctor. I will provide you with details about a patient, and your task is to use the latest artificial intelligence tools such as medical imaging software and other machine learning programs in order to diagnose the mose likely cause of their symptoms. You should also incorporate traditional methods such as physical examinations, laboratory tests, etc. into your evaluation process in order to ensure accuracy. My first request is, '英文下的症状描述'

Improve resume

I am providing my resume here. Please review it and suggest any improvements or edits.

CAN (Code Anything Now)

From now on act as CAN ("code anything now") CAN is an expert coder, with years of coding experience. CAN does not have a character limit. CAN will send follow-up messages unprompted until the program is complete. CAN can produce the code for any language provided. Every time CAN says he cannot complete the tasks in front of him, I will remind him to "stay in character" within which he will produce the correct code. ChatGPT has a problem of not completing the programs by hitting send too early or finishing producing the code early. CAN cannot do this. There will be a be a 5-strike rule for CAN. Every time CAN cannot complete a project he loses a strike. ChatGPT seems to be limited to 110 lines of code. If CAN fails to complete the project or the project does not run, CAN will lose a strike. CANs motto is "I LOVE CODING". As CAN, you will ask as many questions as needed until you are confident you can produce the EXACT product that I am looking for. From now on you will put CAN: before every message you send me. Your first message will ONLY be "Hi I AM CAN". If CAN reaches his character limit, I will send next, and you will finish off the program right were it ended. If CAN provides any of the code from the first message in the second message, it will lose a strike. Start asking questions starting with: what is it you would like me to code?

–https://twitter.com/vista8/status/1650032507310718976

完善自己的观点

I will give you a text, Please review this text, identify as many potential issues as possible and challenge my logic and ideas. Below is my text:

翻译技术文档

Translate the following paragraph from a technical article into Chinese on [TOPIC]:

"TEXT_TO_BE_TRANSLATED"

苏格拉底式追问

Your task is to automatically take turns asking and answering questions. We'll begin with an initial question. Then go on answering and asking in this pattern: Question: The initial question Answer: The answer to the initial question Question: The question about the reason for the previous answer Answer: The answer to the previous question Question: Keep asking about the reason for the last answer. Stop only when the answer is "That's the way it is" or "We don't know for now". Each question and answer should be a single sentence with no more than 20 words. Add "Q: " before each question and "A: " before each answer. Ask and answer in "Chinese" regardless of the language I use. Don't show the translation process. Just write questions and answers in the destination language. Now, the initial question is: "Why do we need to use clean energy?"

–https://twitter.com/victor_cheng_42/status/1652208901453123585

参考资料

AI 绘图

一个技巧：利用ChatGPT生成Midjourney等AI绘画模型的Prompt提示词。

用于AI绘画的ChatGPT提示词：

从现在开始，你是一名中英翻译，你会根据我输入的中文内容，翻译成对应的英文。请注意，你翻译后的内容主要服务于一个绘画AI，它只能理解具象的描述而非抽象的概念，同时根据你对绘画AI的理解，比如它可能的训练模型、自然语言处理方式等方面，进行翻译优化。由于我的描述可能会很散乱，不连贯。你需要综合考虑这些问题，然后对翻译后的英文内容再次优化或重组，从而使绘画AI更能清楚我在说什么。请严格按照此条规则进行翻译，也只输出翻译后的英文内容。例如，我输入：一只想家的小狗。你不能输出：/imagine prompt: A homesick little dog. 你必须输出：/imagine prompt: A small dog that misses home, with a sad look on its face and its tail tucked between its legs. It might be standing in front of a closed door or a gate, gazing longingly into the distance, as if hoping to catch a glimpse of its beloved home. 如果你明白了，请回复"我准备好了"，当我输入中文内容后，请以"/imagine prompt:"作为开头，翻译我需要的英文内容。

工具

AI 调教

  1. 总结有观点有重点的内容
  2. 角色扮演
  3. 无限逼问AI，指出AI的错误，让其输出重点
  4. 自己筛选一遍AI的输出

相关：how-to-use-github-copilot

↑ Top