即刻App年轻人的同好社区
下载
#AGI日思录
#1.AIGC时代个人生存/摸鱼探索指南.Dev gamma.app
- 尝试用AI 改变自己日常的100件事
- 通义听悟听博客确实不错，免费的时长挺多，能做翻译和总结
#2.Long Context模型的 Needle in A Haystack测试
twitter.com
- 准备一个fact
- 分别准备 1k ~ 128k 的 context文本
- 把fack 藏到不同长度的context里面，并且藏到不同的深度
- 丢给GPT-4，让GPT-4找fact
- 评估准确率
GPT-4-128K，context 超过73k之后，会出现找不到的现象，并且fact越靠近context的top，越容易找不到
#3.LeCun最新专访：为什么物理世界终将成为LLM的「死穴」？
mp.weixin.qq.com
点击链接查看和 Kimi 智能助手的对话 kimi.moonshot.cn
- 为什么自回归LLM有局限性
- 缺乏持久记忆
- 缺乏对物理世界的理解
- 语言的带宽比较低，缺乏对视觉的理解
- JEPA（Joint Embedding Prediction Architecture）
- 从输入数据中，提取有用的信息，形成抽象的表示
- 预测任务 预测的是抽象的表示，而不是原始数据
- 降低计算复杂性
#1.AIGC时代个人生存/摸鱼探索指南.Dev gamma.app
- 尝试用AI 改变自己日常的100件事
- 通义听悟听博客确实不错，免费的时长挺多，能做翻译和总结
#2.Long Context模型的 Needle in A Haystack测试
twitter.com
- 准备一个fact
- 分别准备 1k ~ 128k 的 context文本
- 把fack 藏到不同长度的context里面，并且藏到不同的深度
- 丢给GPT-4，让GPT-4找fact
- 评估准确率
GPT-4-128K，context 超过73k之后，会出现找不到的现象，并且fact越靠近context的top，越容易找不到
#3.LeCun最新专访：为什么物理世界终将成为LLM的「死穴」？
mp.weixin.qq.com
点击链接查看和 Kimi 智能助手的对话 kimi.moonshot.cn
- 为什么自回归LLM有局限性
- 缺乏持久记忆
- 缺乏对物理世界的理解
- 语言的带宽比较低，缺乏对视觉的理解
- JEPA（Joint Embedding Prediction Architecture）
- 从输入数据中，提取有用的信息，形成抽象的表示
- 预测任务 预测的是抽象的表示，而不是原始数据
- 降低计算复杂性
2 01
之前一直用werss 订阅微信公众号，后面这个项目不开通续费了，我的公众rss源都断了。推荐一个开源的微信公众号rss工具，原理应该是基于微信读书账号来订阅公众号，也给了免费基于huggingface + Railway 搭建的方案，非常方便。 我的rss源都复活了，mac/ios配合 Reeder食用更佳
项目：github.com
免费搭建教程：xiaoyublog.top
项目：github.com
免费搭建教程：xiaoyublog.top
0 00
春节过后，又要马不停蹄上班了，今年不知道我的脖子和老腰还抗不抗得住，最近打算系统性再了解一下腰椎、颈椎方面的治疗方法，先看了这本书，简单可尝试，先动起来！！
麦肯基疗法-7步告别颈椎腰椎烦恼
0 00
《Same as Ever》读书笔记，在当前洋溢的乐观向上的氛围中，保持一份冷静。
1. One is highlighting this book’s premise—to base predictions on how people behave rather than on specific events. Predicting what the world will look like fifty years from now is impossible. But predicting that people will still respond to greed, fear, opportunity, exploitation, risk, uncertainty, tribal affiliations, and social persuasion in the same way is a bet I’d take
基于人类的本性（对贪婪、恐惧、机会、利用、风险、不确定性、社群的反应）去预测未来更为靠谱
2. One is the constant reminder that wealth and happiness is a two-part equation: what you have and what you expect/need. When you realize that each part is equally important, you see that the overwhelming attention we pay to getting more and the negligible attention we put on managing expectations makes little sense, especially because the expectations side can be so much more in your control.
幸福 = 能力 - 预期，降低预期能更容易/可控的获得幸福
3.People don’t want accuracy. They want certainty.
The inability to forecast the past has no impact on our desire to forecast the future. Certainty is so valuable that we’ll never give up the quest for it, and most people couldn’t get out of bed in the morning if they were honest about how uncertain the future is
不能预测过去对我们预测未来的欲望没有任何影响。确定性是如此宝贵，以至于我们永远不会放弃对它的追求。大多数人如果诚实地面对未来的不确定性，那么他们早上就无法起床
4.The decline of local news has all kinds of implications. One that doesn’t get much attention is that the wider the news becomes the more likely it is to be pessimistic
新闻传播的越广，每天我们能看到各类事故的的可能性就越大，实际就是一个概率问题
5. Let’s say you’re a seventy-five-year-old economist. You started your career at age twenty-five. So you have half a century of experience predicting what the economy will do next. You’re as seasoned as they come.But how many recessions have there been in the last fifty years? Seven. There have only been seven times in your career that you’ve been able to measure your skills.
假设你是一位七十五岁的经济学家。你25岁就开始了你的职业生涯。所以你有半个世纪的经验来预测经济下一步会怎么走。你和他们一样老练。但是在过去的五十年里，有多少次经济衰退呢？七个。在你的职业生涯中，只有七次你能够衡量自己的技能
6.There is too much information in the world for everyone to calmly sift through the data, looking for the most rational, most correct answer. People are busy and emotional, and a good story is always more powerful and persuasive than ice-cold statistics.
世界上有太多的信息需要每个人冷静地在数据中筛选，寻找最理性，最正确的答案。人们都很忙，情绪化，一个好的故事总是比冰冷的统计数据更有力量和说服力
7.When a topic is complex, stories are like leverage.Leverage squeezes the full potential out of something with less effort. Stories leverage ideas in the same way that debt leverages assets.
当一个话题很复杂时，故事就像杠杆。杠杆作用可以用更少的努力把某样东西的全部潜力挤出来。故事以同样的方式利用思想，债务利用资产
8.The most persuasive stories are about what you want to believe is true, or are an extension of what you’ve experienced first hand
9.Pessimism is more intellectually seductive than optimism and captures more of our attention. It’s vital for survival, helping us prepare for risks before they arrive.
悲观主义在智力上比乐观主义更有诱惑力，更能吸引我们的注意力。它对生存至关重要，帮助我们在风险到来之前做好准备
Save like a pessimist and invest like an optimist.Plan like a pessimist and dream like an optimist.
像悲观主义者一样储蓄，像乐观主义者一样投资。像悲观主义者一样计划，像乐观主义者一样梦想
10.Same in investing. I wrote in my book The Psychology of Money: “More than I want big returns, I want to be financially unbreakable. And if I’m unbreakable I actually think I’ll get the biggest returns, because I’ll be able to stick around long enough for compounding to work wonders.”
投资也一样。我在《金钱心理学》一书中写道：“我不想获得高额回报，我更想在财务上坚不可摧。如果我是牢不可破的，我认为我会得到最大的回报，因为我可以坚持足够长的时间，创造奇迹
11.Nassim Taleb says, “My only measure of success is how much time you have to kill.” More than a measure of success, I think it’s a key ingredient. The most efficient calendar in the world—one where every minute is packed with productivity—comes at the expense of curious wandering and uninterrupted thinking, which eventually become the biggest contributors to success.
纳西姆·塔勒布说：“我衡量成功的唯一标准是你有多少时间可以消磨。”不仅仅是衡量成功的标准，我认为这是一个关键因素。世界上最有效率的日历——每一分钟都排满了生产力——是以好奇心的漫游和不间断的思考为代价的，而好奇心的漫游和思考最终成为成功的最大贡献者
12.This is one of the most useful life skills—enduring the pain when necessary rather than assuming there’s a hack, or a shortcut, around it.
这是最有用的生活技能之一，在必要的时候忍受痛苦，而不是假设有一个黑客或捷径可以绕过它
13.There are two types of information: permanent and expiring.Permanent information is: “How do people behave when they encounter a risk they hadn’t fathomed?” Expiring information is: “How much profit did Microsoft earn in the second quarter of 2005?”
有两种类型的信息：永久的和过期的。永久性的信息是：“当人们遇到他们没有弄清楚的风险时，他们会怎么做？”过期的信息是：“微软在2005年第二季度赚了多少利润？
14.Expiring knowledge catches more attention than it should, for two reasons.。One, there’s a lot of it, eager to keep our short attention spans occupied.Two, we chase it down, anxious to squeeze insight out of it before it loses relevance.
过期的知识引起了比它应该引起的更多的关注，原因有二。第一，它有很多，渴望让我们短暂的注意力被占据。第二，我们追逐它，急于在它失去相关性之前从它身上榨取见解
1. One is highlighting this book’s premise—to base predictions on how people behave rather than on specific events. Predicting what the world will look like fifty years from now is impossible. But predicting that people will still respond to greed, fear, opportunity, exploitation, risk, uncertainty, tribal affiliations, and social persuasion in the same way is a bet I’d take
基于人类的本性（对贪婪、恐惧、机会、利用、风险、不确定性、社群的反应）去预测未来更为靠谱
2. One is the constant reminder that wealth and happiness is a two-part equation: what you have and what you expect/need. When you realize that each part is equally important, you see that the overwhelming attention we pay to getting more and the negligible attention we put on managing expectations makes little sense, especially because the expectations side can be so much more in your control.
幸福 = 能力 - 预期，降低预期能更容易/可控的获得幸福
3.People don’t want accuracy. They want certainty.
The inability to forecast the past has no impact on our desire to forecast the future. Certainty is so valuable that we’ll never give up the quest for it, and most people couldn’t get out of bed in the morning if they were honest about how uncertain the future is
不能预测过去对我们预测未来的欲望没有任何影响。确定性是如此宝贵，以至于我们永远不会放弃对它的追求。大多数人如果诚实地面对未来的不确定性，那么他们早上就无法起床
4.The decline of local news has all kinds of implications. One that doesn’t get much attention is that the wider the news becomes the more likely it is to be pessimistic
新闻传播的越广，每天我们能看到各类事故的的可能性就越大，实际就是一个概率问题
5. Let’s say you’re a seventy-five-year-old economist. You started your career at age twenty-five. So you have half a century of experience predicting what the economy will do next. You’re as seasoned as they come.But how many recessions have there been in the last fifty years? Seven. There have only been seven times in your career that you’ve been able to measure your skills.
假设你是一位七十五岁的经济学家。你25岁就开始了你的职业生涯。所以你有半个世纪的经验来预测经济下一步会怎么走。你和他们一样老练。但是在过去的五十年里，有多少次经济衰退呢？七个。在你的职业生涯中，只有七次你能够衡量自己的技能
6.There is too much information in the world for everyone to calmly sift through the data, looking for the most rational, most correct answer. People are busy and emotional, and a good story is always more powerful and persuasive than ice-cold statistics.
世界上有太多的信息需要每个人冷静地在数据中筛选，寻找最理性，最正确的答案。人们都很忙，情绪化，一个好的故事总是比冰冷的统计数据更有力量和说服力
7.When a topic is complex, stories are like leverage.Leverage squeezes the full potential out of something with less effort. Stories leverage ideas in the same way that debt leverages assets.
当一个话题很复杂时，故事就像杠杆。杠杆作用可以用更少的努力把某样东西的全部潜力挤出来。故事以同样的方式利用思想，债务利用资产
8.The most persuasive stories are about what you want to believe is true, or are an extension of what you’ve experienced first hand
9.Pessimism is more intellectually seductive than optimism and captures more of our attention. It’s vital for survival, helping us prepare for risks before they arrive.
悲观主义在智力上比乐观主义更有诱惑力，更能吸引我们的注意力。它对生存至关重要，帮助我们在风险到来之前做好准备
Save like a pessimist and invest like an optimist.Plan like a pessimist and dream like an optimist.
像悲观主义者一样储蓄，像乐观主义者一样投资。像悲观主义者一样计划，像乐观主义者一样梦想
10.Same in investing. I wrote in my book The Psychology of Money: “More than I want big returns, I want to be financially unbreakable. And if I’m unbreakable I actually think I’ll get the biggest returns, because I’ll be able to stick around long enough for compounding to work wonders.”
投资也一样。我在《金钱心理学》一书中写道：“我不想获得高额回报，我更想在财务上坚不可摧。如果我是牢不可破的，我认为我会得到最大的回报，因为我可以坚持足够长的时间，创造奇迹
11.Nassim Taleb says, “My only measure of success is how much time you have to kill.” More than a measure of success, I think it’s a key ingredient. The most efficient calendar in the world—one where every minute is packed with productivity—comes at the expense of curious wandering and uninterrupted thinking, which eventually become the biggest contributors to success.
纳西姆·塔勒布说：“我衡量成功的唯一标准是你有多少时间可以消磨。”不仅仅是衡量成功的标准，我认为这是一个关键因素。世界上最有效率的日历——每一分钟都排满了生产力——是以好奇心的漫游和不间断的思考为代价的，而好奇心的漫游和思考最终成为成功的最大贡献者
12.This is one of the most useful life skills—enduring the pain when necessary rather than assuming there’s a hack, or a shortcut, around it.
这是最有用的生活技能之一，在必要的时候忍受痛苦，而不是假设有一个黑客或捷径可以绕过它
13.There are two types of information: permanent and expiring.Permanent information is: “How do people behave when they encounter a risk they hadn’t fathomed?” Expiring information is: “How much profit did Microsoft earn in the second quarter of 2005?”
有两种类型的信息：永久的和过期的。永久性的信息是：“当人们遇到他们没有弄清楚的风险时，他们会怎么做？”过期的信息是：“微软在2005年第二季度赚了多少利润？
14.Expiring knowledge catches more attention than it should, for two reasons.。One, there’s a lot of it, eager to keep our short attention spans occupied.Two, we chase it down, anxious to squeeze insight out of it before it loses relevance.
过期的知识引起了比它应该引起的更多的关注，原因有二。第一，它有很多，渴望让我们短暂的注意力被占据。第二，我们追逐它，急于在它失去相关性之前从它身上榨取见解
Reading《Same as Ever》with AI
0 00
今年上班开始骑自行车，路上听播客时间的少了， @硬地骇客 今年下半年开始关注，给我不少启发，希望明年个人项目上能有一些突破，能跳出大厂工作的束缚，继续向自由人生迈进
3 00
这哥们实在且厉害。
1.学会亏小钱赚大钱，比如知识付费，想让别人卸下防备跟你聊不花钱是不行的，不能指望为爱发电。
2.我希望别人快速看透我，因为快速别人看透你，他大概能量化出你的能力，当别人需要组织一些生产要素的时候，能够作为别人的资源被人盘活就好
1.学会亏小钱赚大钱，比如知识付费，想让别人卸下防备跟你聊不花钱是不行的，不能指望为爱发电。
2.我希望别人快速看透我，因为快速别人看透你，他大概能量化出你的能力，当别人需要组织一些生产要素的时候，能够作为别人的资源被人盘活就好
【浮生一日】三大坑牢牢捆住想翻身的普通人！商超大佬手把手教你，全网最实在的赚钱方法！！_哔哩哔哩_bilibili
0 00
#知识管理#
读了indigox的《构建外脑 / 智变时代的个人知识管理》，比较受启发。LLM时代，我们有了新的钥匙，它能打通自然语言与0/1数字语言，后面再逐步把多模态的理解能力解决后，我们可以更加放心的把音频、视频、PDF等等材料放到外部存储上，通过LLM作为桥梁，通过自然语言与这些知识交互。知识的存储也不用再按分门别类的排序，通过向量索引就能召回，做到足够扁平化，而且效率还高。www.indigox.me
读了indigox的《构建外脑 / 智变时代的个人知识管理》，比较受启发。LLM时代，我们有了新的钥匙，它能打通自然语言与0/1数字语言，后面再逐步把多模态的理解能力解决后，我们可以更加放心的把音频、视频、PDF等等材料放到外部存储上，通过LLM作为桥梁，通过自然语言与这些知识交互。知识的存储也不用再按分门别类的排序，通过向量索引就能召回，做到足够扁平化，而且效率还高。www.indigox.me
3 20
最近读了learningprompt.wiki ，觉得相比于照抄已有的案例，建立一个基础的使用框架更为重要，以下是自己一些粗浅的总结：
【基础部分】
1.Prompt 里最好包含完整的信息。
【badcase】 Write a poem about OpenAI.
【goodcase】Write a short inspiring poem about OpenAI, focusing on the recent DALL-E product launch (DALL-E is a text to image ML model) in the style of a {famous poet}
2.Prompt 最好简洁易懂，并减少歧义。
【case】Use a 3 to 5 sentence paragraph to describe this product.
3.尽量告诉LLM应该做什么，而非not to do
4.增加示例 【case】Prompt:Q1 -> A1；Q2 -> A2；Q3 ->
5.角色设定【case】You are a primary school teacher who can explain complex content to a level that a 7 or 8 year old child can understand.
【用法部分】
1.问答
2.推理
- 如果你想用 ChatGPT API 做点什么小应用，我建议可以从这个场景入手，相对来说没有其他场景那么红海
3.生成
- 写代码、写文章
4.改写
翻译：除了能翻译人类的语言外，还能翻译编程语言，比如将 Python 代码翻译成 Haskell 代码。
修改：修改内容的语法，甄别内容里的错别字。
润色：润色文章，将内容改成另一种风格
【case】You are a primary school teacher who can explain complex content to a level that a 7 or 8 year old child can understand. Please rewrite the following
http://5.信息解释
解释代码、解释论文
http://6.信息提取
让模型输出特定的结构化答案(比如JSON/Markdown等)
Models
1.CRISPE
CR： Capacity and Role。你希望 ChatGPT 扮演怎样的角色。
I： Insight，背景信息和上下文（坦率说来我觉得用 Context 更好）。
S： Statement，你希望 ChatGPT 做什么。
P： Personality，你希望 ChatGPT 以什么风格或方式回答你。
E： Experiment，要求 ChatGPT 为你提供多个答案
2.Chain of Thought
只需要在问题的结尾里放一句 Let‘s think step by step ，模型输出的答案会更加准确
也可以用 Let's work this out in a step by step way to be sure we have the right answer
仅在使用大于等于 100B 参数的模型时，才会生效
3.PAL Models (Program-Aided Language Models)
在 prompt 里引入代码，并引导模型使用代码来运算和思考
【基础部分】
1.Prompt 里最好包含完整的信息。
【badcase】 Write a poem about OpenAI.
【goodcase】Write a short inspiring poem about OpenAI, focusing on the recent DALL-E product launch (DALL-E is a text to image ML model) in the style of a {famous poet}
2.Prompt 最好简洁易懂，并减少歧义。
【case】Use a 3 to 5 sentence paragraph to describe this product.
3.尽量告诉LLM应该做什么，而非not to do
4.增加示例 【case】Prompt:Q1 -> A1；Q2 -> A2；Q3 ->
5.角色设定【case】You are a primary school teacher who can explain complex content to a level that a 7 or 8 year old child can understand.
【用法部分】
1.问答
2.推理
- 如果你想用 ChatGPT API 做点什么小应用，我建议可以从这个场景入手，相对来说没有其他场景那么红海
3.生成
- 写代码、写文章
4.改写
翻译：除了能翻译人类的语言外，还能翻译编程语言，比如将 Python 代码翻译成 Haskell 代码。
修改：修改内容的语法，甄别内容里的错别字。
润色：润色文章，将内容改成另一种风格
【case】You are a primary school teacher who can explain complex content to a level that a 7 or 8 year old child can understand. Please rewrite the following
http://5.信息解释
解释代码、解释论文
http://6.信息提取
让模型输出特定的结构化答案(比如JSON/Markdown等)
Models
1.CRISPE
CR： Capacity and Role。你希望 ChatGPT 扮演怎样的角色。
I： Insight，背景信息和上下文（坦率说来我觉得用 Context 更好）。
S： Statement，你希望 ChatGPT 做什么。
P： Personality，你希望 ChatGPT 以什么风格或方式回答你。
E： Experiment，要求 ChatGPT 为你提供多个答案
2.Chain of Thought
只需要在问题的结尾里放一句 Let‘s think step by step ，模型输出的答案会更加准确
也可以用 Let's work this out in a step by step way to be sure we have the right answer
仅在使用大于等于 100B 参数的模型时，才会生效
3.PAL Models (Program-Aided Language Models)
在 prompt 里引入代码，并引导模型使用代码来运算和思考
Prompt Engineering 基础
6 04