Keywords

  • Ground ur Ideas Ziming Liu
  • Amanda Askell
  • Blue moon 🔵🌙

TOC

One more light song by linkin park 2017

Lyrics

Should've stayed. Were there signs I ignored?
Can I help you not to hurt anymore?
We saw brilliance when the world was asleep
There are things that we can have but can't keep

If they say

Who cares if one more light goes out
In the sky of a million stars?
It flickers, flickers
Who cares when someone's time runs out
If a moment is all we are?
Or quicker, quicker
Who cares if one more light goes out?
Well, I do

The reminders pull the floor from your feet
In the kitchen one more chair than you need
Oh
And you're angry, and you should be, it's not fair
Just 'cause you can't see it, doesn't mean it isn't there

If they say

Who cares if one more light goes out
In the sky of a million stars?
It flickers, flickers
Who cares when someone's time runs out
If a moment is all we are?
Or quicker, quicker
Who cares if one more light goes out?
Well, I do

Who cares if one more light goes out
In the sky of a million stars?
It flickers, flickers
Who cares when someone's time runs out
If a moment is all we are?
Or quicker, quicker
Who cares if one more light goes out?
Well, I do
Well, I do

☕Gooddays

↩️

  • If one of us goes to war we all go to war 这声音还在echo 爱尔兰狂人发起冲锋
  • Chael 说要击败Islam，需要几个人，类似组成一种team，一轮轮的match，然后you take some notes，一场场比赛下来对于Islam进行描边，这个最难的Puzzle才有可能被解决。Merab之前就这样，初期的title defences近乎无敌，但一轮轮下来，能力边界在radar chart上被试出来。也有点像Thanos，一波波人去消耗，去了解才能功课。破解伟大的难题就是这样，一波波地失败去增加understanding，最终才会慢慢找到出口，数学难题，material science，Astroscience，Biological Sciences……

  • Good writing is obvious, not original

  • 美国AI4S战略进入实施阶段，能源部牵头打造“创世纪”开放科学大模型

    • 其中一个核心判断是：AI 正在改变科学研究所面对的稀缺资源。
      • 过去，科学家的主要限制之一是人的认知能力。一名研究者一生只能阅读有限数量的论文，只能掌握有限的研究方法，也无法穷举巨大的假设和组合空间。白宫认为，AI 可以帮助科研人员跨越这些限制：读取跨学科文献、搜索人类无法穷举的候选空间、分析海量实验数据，并进一步参与提出假设、设计实验、控制仪器和分析结果。
      • 但报告同时指出，仅仅给每个科学家配一个 AI 助手，并不会自动带来科学革命。如果 AI 只是让研究人员更快生成论文、申请书和实验结果，反而可能进一步压垮现有的同行评议和验证体系。真正的 AI for Science，需要同时改变模型、数据、实验设施和科研组织方式。
  • 线上会议 人们表达时候一讲到重要的内容，或者想让别人同意或者必须吸收的内容的时候会不由自主地身体前倾，贴近屏幕

Ground ur Ideas & Toy Models by Ziming Liu

  • 知道刘子明还是在刚本科时候，先是我看的一本比较感兴趣的书，然后从图书馆借来，作者是Max Tegmark，迈克斯•泰格马克 （Max Tegmark）书名：《穿越平行宇宙》，我不太知道Max Tegmark是谁，以为和普通的这种科学新知的书籍作者一样，读完之后很久好像才慢慢了解到Max Tegmark，后面还看了《生命3.0》，很长一段时间对科幻和这种前沿科学，非虚构都很着迷。后面social上搜，blog主页找，慢慢又看到了Ziming Liu的主页。

  • 最近看到他之前的一些post，打印成PDF阅读了一番。主题就是：Ground ur Ideas & Toy Models

  • 原文：

  • 在这里，我记录一些我的内化思考，

    • 很多人每天都会有很多ideas，价值高低和数量各异，但是抓住了就有可能会有所结果，如果没有及时记录，后面的生活中某些场景或者思考可能会联系起某个idea再回来，虽然可能会比之前还更加成熟，但是时间成本已经付出，如果没有及时输出，那别人同样push出去占领，自己大概只能懊悔。当然记下来也会消耗时间精力来ground落地，拖慢其他project的进度。

  • Students with theoretical backgrounds are especially prone to thinking that the more abstract and profound something is, the more advanced it must be, whereas the more concrete and practical it is, the less sophisticated it seems.

  • build and break things fast

  • “concretization” ：This process of “concretization” inevitably loses some information, but that is acceptable—the purpose of subsequent iterations is precisely to recover and refine it. 有时候会慢慢演变，可能会遗忘初始时候想出来的一些细节，或者后面迭代之后细节观点等慢慢变形，所以可能需要有一个outline或者keypoints记录文件或者checkpoints可以回滚。

  • Therefore, we must find a balance between idealism and pragmatism.

  • 我的框架大概是ideas->actions->mvp ，总体上早点落地写demo，然后迭代优化。

  • 具象化大多数是先落地新建文件夹，写上当前最新鲜的想法作为readme，然后再去新建info.txt 集中记录一些零散的细节或者网址参数，然后在refer的文件夹中集中一些可能会用到的原料文件

  • To truly understand a conclusion, we must understand not only when it is correct, but also when it fails. 至少也是追求的目标。而且是一个成长变化，随时间演变，进步，添加思考insights进去，直到慢慢趋于收敛，至少是自己的认知能力范围之内。

  • we also generate new ideas, and sometimes these ideas genuinely come from intuition.

  • Constructing toy models is a way to explore this scope of applicability. Through toy models, abstract ideas can quickly become concrete and enter an iterative cycle. toy models基本上是MVP （Minimum Viable Product）即最小可行性产品

By the way, there are generally three ways to conduct research. “Finding applications for a method” is the second one. Ranked by impact, they are:

Best: Identify an important application and solve it by any means necessary (I use the word “application” broadly—for example, “solving continual learning” is also an application).
Second: Start with a method and then find applications where it works well.
Worst: Fix both the method and the application in advance and force them to work together, even though they may fundamentally be incompatible.

TB TD top2bottom B2T的策略，还有双向施工

但是第二篇解释为什么是MVP 也就是toy models

  • 最小可执行和最终产品结果之间当然有所不同，挑战通常也在gap之中，办法就是插值：
    toy models和real models

What works in a toy model may not work in a real model—and even if it does, it may work for entirely different reasons.

* ==Knowledge is created at the boundary between the known and the unknown. ==If we fully understand a toy model and the real model behaves differently, there must be a phase change (sharp or smooth) along the path connecting them. Understanding the phase change is key to understanding the difference between the toy setup and the real setup. Starting from the toy model (known) and gradually moving toward the real model (unknown) mirrors how humans learn. Fully known is trivial; fully unknown is confusing. The sweet spot—half-known, half-unknown—is where information gain is maximized. ==Research is, at its core, an information-gathering game.==
* 所以掌握和建立属于自己的信息流或者workflow很重要，是一个ecosystem
* 从信息获取到输出：
    * input, sources
    * output , perform

How to interpolate?
Directionality: move from simple to complex, or from complex to simple.
Locality: change one feature at a time; avoid overly large jumps.

  • Lesson: not only can you make simple data more complex—you can also make complex data simpler.
  • Lesson: maintain an interpolation mindset. Simplicity vs. complexity is not binary—it is continuous and multidimensional.

Physics of AI is experiment- and phenomenon-driven, prioritizing intuition and breadth—touching many sides of the elephant and assembling a coherent picture.

  • 很多HBR，哈佛商学院评论，还有麻省理工科技评论的文章，前者我更多会打印PDF来收集管理阅读，后者的英文原版有强相关性的我才会PDF阅读，中文版一般是新闻类型的比较多，所以网页浏览即可。主要是相关性或者觉得意义较为重要的，我需要记录随笔，所以才需要打印PDF记录和留存。btw，很多网页直接打印PDF排版很不好，总会有页眉页脚的遮挡，纽约客BBC，尤其是经济学人的网页，TMD，还比较幸运，淘到一个好用的chrome拓展——PrintFriendly，可以排版的非常美观，甚至还贴心的免费加入了AI summary的功能。

Amanda Askell

  • 上次说Dario的妹妹，Daniela 或许也该出一本书。
    • @readswithravi
      • Daniela Amodei on the underrated power of curiosity:

“The ability to be curious, to learn across a lot of disciplines and to have a strong foundation of wanting to have impact, regardless of the area that you're working in, I think that is an underrated quality.”

  • 然后看一篇文章：Why big AI labs are hiring so many philosophers 就是Anthropic的Amanda Askell领导的宪法AI相关的。
    • 义务论 后果论，苏格拉底在给一堆机器人讲哲学。？？？ 我为什么用的量词是一堆？ 感觉有点不礼貌，甚至没有像对待动物一样…… 是物品，而且感觉人类在于动物之上，动物于物品之上？
    • 文章intro比较了文科哲学生就业形式好于CS毕业的学生，尤其是AI更需要哲学。 但是imo，确实连哲学慢慢因为用于AI也要慢慢卷学历，certification，但是两个专业学生的行业不同 选项不同，CS更多自由职业，startup，文科专业性门槛不高所以相互可以兼容，岗位动态。
    • AI宪法这些在Dario的传记和Anthropic的post中介绍有很多了，包括Lex的podcast里。

New Season - Blue moon 🔵🌙

08月23日星期日
曼城
曼城
21:00
CST

伯恩茅斯

I'm free to be
whatever I
Whatever I choose
and I'll sing the blues
if l want

And New season !!!!!!!!!!!!

to start over again

夏日，周末，黄昏时候跑步，向着西边方向，一直向西，还有两个绝佳的album，Oasis和linkin park，There'r hopes, days, today, tomorrow

ff7a3580497b00cc2e2c7c250d8baa04.png

You cannot be interested and inconsistent.
Thanks for being an insider till the end!
Till next, stay safe and stay hydrated!