One Week Later : My AI Reading List - W1
A Deep Dive into Evals and Model Reasoning.
Hi there, Rosa here! Welcome to the very first issue of my Substack, "Rambling About AI," and the launch of my first column series: One Week Later. This series is created for a slogan: News worth reading only after the noise fades.
哈喽，我是Rosa！欢迎来到我的Substack《Rambling About AI》创刊号，我的第一个专栏系列“One Week Later”也正式上线啦！这个系列的灵感来自一句话：喧嚣过后，方见真章。
Each week, I'll dive into the articles or podcast I've been reading/listening in the AI world and handpick the most practical ones to share with you in a reading list. I generally pick articles that are at least a week old. This way, I wish to help you find some of the hidden gems you might have missed in the whirlwind of the past week.
之后每周，我都会花一些时间深入在AI领域阅读文章和播客，并精选出其中最实用的几篇，做成阅读清单分享给你。我会尽量选择一周之前或者更早的文章，帮助你回看一些可能遗漏的高价值作品。
We're kicking things off with a catch-up issue for week one, but you can expect fresh updates every Sunday or Monday from now on.
第一周是补发的内容，之后你就可以期待每周日或周一的新鲜推送了！
Full disclosure: I use AI to help write this blog in both Chinese and English—it's a big part of how I'm practicing both languages. Thanks for your patience if some phrasing comes out a bit awkward!
友情提醒：我在AI的帮助下创作中英双语博客，以此帮助我学习语言。如果这对你造成了不便，还请见谅！
🌟 First up: AI Evaluation! Here are a few super practical guides and case studies.
🌟 先来看看AI评估，这里有几篇实用的技术案例和指南。
1️⃣ 《Eval Driven System Design - From Prototype to Production》
This is an awesome case study from OpenAI and Fractional. It's a hands-on guide that shows you how to build a production-level AI automation system around the Evals tool, kicking those tedious manual processes to the curb.
OpenAI和Fractional联手做的一个超赞案例，手把手教你如何用Evals工具为核心，搭建一套生产级的AI自动化应用，把那些繁琐的人工流程给替代掉。
想快速了解OpenAI Evals？官方介绍在这：
Want a quick intro to OpenAI Evals? Here's the official doc:
2️⃣ 《Mastering AI Evals: A Complete Guide for PMs》
A top-notch newsletter for any Product Manager out there! It gives you a complete rundown on how to design a solid evaluation framework for your AI product.
这篇Newsletter质量超高，非常适合产品经理！它完整而细致地介绍了如何给AI产品设计一套靠谱的评估体系。
Plus, don't miss its sister article on finding the right metrics: "Evaluating AI Products: How to Find The Right Metrics"
另外，也别错过它的姐妹篇，专门讲怎么找对评估指标：《Evaluating AI Products: How to Find The Right Metrics》
3️⃣ 《A statistical approach to model evaluations》
For those who want to get a bit more technical, check out this piece from Anthropic. It dives deep into the statistical methods for AI evals and even comes with a full paper. This one's for you.
想搞点硬核的？Anthropic去年底发的这篇文章，深入聊了聊AI评估里的统计学方法，还附上了论文，适合深入阅读。
🌟 Some Research About Model Reasoning
Next, let's talk model reasoning—and what better time, with Apple's (hilariously debunked) new paper making waves.
再来借着苹果那篇引发争议的论文，我们聊聊模型的推理能力。
4️⃣ 《The Illusion of Thinking》&《The Illusion of the Illusion of Thinking》
So, Apple dropped a new paper with some pretty bold claims, arguing that current LLMs have basically no real reasoning power and just crash and burn when faced with complex tasks.
It got debunked almost immediately , and Anthropic's Claude fired back—by using AI to publish its own paper titled "The Illusion of the Illusion of Thinking" just to clap back at Apple.
苹果的最新论文，观点挺大胆的，说现在的LLM根本没啥真正的推理能力，一碰复杂的任务就崩溃。结果嘛，光速翻车；Claude甚至刚刚由AI发布了一篇论文《The Illusion of the Illusion of Thinking》为了反驳苹果的观点。
5️⃣ 《Tracing the thoughts of a large language model》
Ever wonder how LLMs actually "think"? Anthropic dives deep into reverse-engineering their internal thought processes.
想知道大模型是怎么“想”问题的吗？Anthropic这篇给你揭秘！他们逆向工程了大模型的内部工作原理来追踪它的“思维”。
What's cooler? The accompanying "Attribution Graphs" tool they built is open-source, so you can even play around with it:
更酷的是，他们还把用来追踪思维的工具“归因图”开源了：
6️⃣ 《Reasoning models don't always say what they think》
Another mind-bender from Anthropic.
They found that what a model says it's thinking during its Chain-of-Thought (CoT) process might not be what it's actually thinking. So yeah, that CoT? Not always the full story.
另一篇来自Anthropic的“LLM读心术”研究。
他们发现模型进行思维链（CoT）推理时，“说的”和“想的”可能不是一回事儿。思维链读起来挺有趣的，并不能完全信任。
7️⃣ 《Detecting misbehavior in frontier reasoning models》
AI supervising AI? You bet.
OpenAI walks through how they're using one LLM to monitor the "thought process" of another frontier model to catch misbehavior like reward hacking.
用AI监督AI是可行的吗？OpenAI就介绍了这么个方法，用一个大模型来“监视”另一个前沿模型的“思考过程”，揪出那些不老实的行为（比如奖励欺骗）。
好啦，以上就是本周的全部“干货”！你对哪个话题最感兴趣，或者有啥想看的？在评论区给我留言吧！咱们下周日或周一再见！
And that's a wrap for this week!
I'd love to hear what you thought. Which article caught your eye? Got any topics you want me to dive into? Drop a comment below! Catch you next Sunday or Monday!