用 AI 搜文献,只搜出来十几篇,你就以为这个方向没人做?很可能不是没人做,是你的工具只连了几个免费来源。这一篇讲两种更靠谱的检索方式:用接入论文库的 Agent,或者自己给 Agent 接论文 API。附一段可以直接复制的提示词。
「古法手搓」和「AI 新世界」,应该彼此尊重
今天坐地铁去学校,路上突然想明白一件事:「古法手搓」和「AI 新世界」的人,应该彼此尊重。
每个人的通勤方式都不一样。有人开车,有人坐公交,有人骑车。大家选的不一样,但很少有人去评价别人怎么上班。
我妈 60 多岁了,到现在还习惯手洗衣服。我不手洗,直接扔洗衣机。这么多年,我们俩谁也没看不惯谁。
可到了学术圈,事情好像变了。用 AI 的人常常被「道德审判」,好像用了 AI,做的研究就低人一等。
前几天听播客,姚顺宇提到 99% 的代码都是 AI 写的。工业界这么积极地拥抱 AI,却很少听到有人说这是「编程不端」。
我觉得,写代码和写论文,道理是相通的。
我们评价一个 AI 产品,看的是体验好不好、有没有解决用户真实的需求,很少有人追问它背后的代码是不是 AI 写的。
评价一项学术研究也一样,应该看它有没有提出一个好问题,有没有得出有意义的发现。
A Nature journal paper ranks AI writing help on 5 levels.
Most researchers chase level 4: let AI write.
The most valuable is level 5: let AI find your paper's problems.
Which stage do you trust AI with least: outline, draft or edit?
Lately my students keep asking how to write a paper's introduction. I went back to the CARS model and publishers' guidelines, and shared the 5 AI prompts I actually use. The rule in every prompt: only the material I provide, no invented references.
作为老师,我想瞎聊一下为什么教授 们更加倾向年龄小的 PhD。
1. 很多年龄大的 PhD,他们已经有了家庭。读博期间的科研收入无法支撑他基本的家庭生活。可能很多人说,老师为什么不能多给博士生发点工资呢?有两个限制:第一,不是所有博导都用充足的科研经费的。第二,作为学生,他们的科研劳务在教育体系里也有一个上限的。 我所在学校,博士生每个月 1000-2000 已经算不错的了。
2. 很多年龄大的 PhD 并不是因为对科研感兴趣才读的博士。有一部分毕业很久的,可能是为了职称晋级,选择了读博。这部分 PhD 在读书期间很难有高的科研产出。
3. 年龄小的博士(主要是硕士毕业后直接读的)可塑性更强。他们更愿意按照导师的培养方式完成学校的任务。
I recently watched an interview with Fei-Fei Li. One line stayed with me: "They're talking too much about replacing humanity" — we should be talking far more about how AI can empower people, not just how it might replace us and our jobs.
I couldn't agree more, especially in academia.
Many colleagues around me focus mostly on AI's downsides. Will it replace us? Will students lose their creativity if they use it? Yet many of them have never really used AI tools in depth.
I'm a heavy AI user myself. Like Fei-Fei Li, what I feel most is how much AI empowers my teaching and research.
I'm not worried that AI will replace me, or that it will weaken my thinking. I'm the one asking the questions. I'm steering AI toward my goals.
I've never been against students using AI either. In fact, I tell them to ask AI about everything.
AI is a tool for efficiency. The real question is how we use it well, and how we spend the time it saves on the creative questions that only we can ask.
How do you use AI in your teaching or research?
#AI #HigherEducation #FeiFeiLi #research
AI 让我完成了从「古法」到新世界的蜕变
1. 在 YouTube 上看了一个李飞飞的采访,被一句话击中:我们应该更多地讨论如何赋能人类,而不是总在谈论如何取代人类、取代工作岗位。
我平时主要在两个圈子里:一个是学校,一个是 AI 创业者。两边的差别非常大。
在学校,很多老师对 AI 持悲观态度。就像李飞飞说的,大家讨论最多的是 AI 会不会取代人以及AI 的伦理问题。我观察到很多老师用 AI 用得比较浅,有的还没真正用过最好的模型或 Agent。
AI 创业者的圈子正好相反。大家每天都对 AI 很兴奋,也实实在在地被 AI 赋能着。
我很庆幸去年走出了舒适区,认识了更多人,学了更多 AI 技能,也看到了一个更多样、更丰富的世界。
AI 让我完成了一次从「古法」到新世界的蜕变。
2. 李飞飞在采访里还说:即便作为一名 AI 科学家,她也觉得自己基本跟不上 AI 的发展速度了。
有句玩笑话叫「只要你学得够慢,就不用学」。放在 AI 上,还真是这样。上周我翻了翻自己给学校老师做培训的课件,发现基本都过时了。去年讲 Cursor、v0、Bolt,今年讲用小龙虾赋能科研,现在这些都不用再学了。
3.纳瓦尔说过,人应该有三个爱好:一个让你赚钱,一个让你健康,一个让你聪明。我最近在培养第二个,让自己健康的那个。
4.今天偶然刷到一段视频,讲海伦·凯勒小时候怎样在莎莉文老师的帮助下学会语言。我感叹生命的顽强,也对莎莉文老师充满敬意:她成功的教会了一个又盲又聋的孩子使用语言!!!!
5.周三有朋友问我最近怎么样。我说挺好的,就是太忙了。年纪大了,精力有限,每天都在和自己和解,学着做减法。
他说:「刘邦 49 岁还在村口看狗打架,54 岁当了汉王。」
我说:好的,我也要当王。
6. 使用了 Grokbot 2 天后 ,我取消了Codex 的订阅。我开始给马斯卡送钱了。
7. 我的公众号不知道碰了微信的哪条红线,感觉已经被限流两周了。有在腾讯的朋友吗?
What makes an AI literature review actually usable?
I have two non-negotiables for any literature review.
1. Every reference must be real. I need to click a link and see the actual paper.
2. It can't read like Wikipedia. Listing findings one by one isn't a review. I want critical synthesis: where studies agree, where they conflict, and why.
Recently I tried a tool whose workflow matches how I actually write. It searches for papers first, then writes from them. Every citation links back to its source, and I can open any paper and read it in detail.
👉 Two things I especially liked:
➝ It's smart about screening. It knows which papers come from authoritative journals and keeps the core ones.
➝ I can edit the draft online, right where it's written.
My honest score: 75/100.
The sources check out, and it does compare findings across studies, but the critical judgment is still thinner than I'd want. A solid first draft, not a finished review.
🎯 Here's my process, in case it's useful:
1️⃣ Opened Yila and chose "Literature Review"
2️⃣ Described my topic: AI feedback on student writing, empirical studies since 2022, APA 7
3️⃣ It pulled 40 candidate papers from OpenAlex and PubMed
4️⃣ Screened them down to 30 core empirical studies
5️⃣ Wrote a review organized by themes, with references and DOIs
6️⃣ Downloaded the Word draft and the paper list. Every reference has a DOI, so I can open any paper and read it in full.
👉 What I'd still do myself: tighten the argument, add my own judgment on conflicting findings, and read the key papers in full.
Link if you'd like to try it on your own topic: yila.ai/
❄️ What are your non-negotiables for a literature review?
别给自己设限,你会学会任何事情
1. 周二,一个学生来找我聊天。他说上完我的课以后,对AI产生了很大的兴趣,还给我看了他在GitHub上找到的几本开源书,说准备开始学了。
我说:“不需要。”
学AI最大的门槛在于,它是一门实践的艺术。这一点,读了很多年书的同学往往很不适应。他们习惯用学高数、物理这些基础学科的方法,拿起一本书,从第一页往后看,看完就算学会了。但学AI的逻辑不一样,它靠的是「实践」。
2. 过去三年,AI给我带来最重要的变化,是「不再自我设限」。我突然发现,自己能做的事情越来越多。
3.过去一周的感觉是:在战斗,内心却十分平静。
4.最近看到邓亚萍和李小鹏的一个对谈,她说:“记住每一次好的感觉,然后一直重复。”
这和我最近打磨skill的逻辑一模一样:
做成一件事 → 把方法练出来 → 沉淀成skill → 效率提升10倍 → 用省下的时间探索新的事情 → 继续沉淀
5.我又重温了一遍小奇的夺冠专场,太喜欢他那个「中专生思维」的段子了。
大意是:他说中专生喜欢一件事,从来不在乎别人的眼光。他有个发小是木匠,原本养鸽子,给鸽子打了个笼子,结果打大了。发小依旧是「中专精神」:题答错了,就改题干。于是他改养孔雀了。
小奇说,自己在大城市待久了,有了“本科生思维”,见面问的第一个问题是:“你拿到许可了吗?”
发小说:“我要谁许可?孔雀同意了。”
他又问:“那你哪儿整来这么多孔雀?”
发小说:“快递发来的呗。”
我感觉这个「中专生思维」挺好的呀。
6. 我忘了在哪儿听过这样一个故事:一个读书人快要饿死了,他想到的不是赶紧挣钱买一碗饭,而是去研究怎么种稻子。(我时不时的提醒自己,不要做这样的读书人)
Today was our literature exam: students read assigned papers in advance, then defend them in front of faculty.
About 80% had memorized the papers. Only a few were truly thinking critically.
Two of them argued that one paper's ideas were outdated and it shouldn't be on the reading list anymore.
I love students like this. The point of reading literature isn't to remember it, it's to have the nerve to question it.
一个不懂科研的朋友用AI写了篇心理学论文,我照着他“简单粗暴”的方法,在传播学、医学、工学各写了一篇,请北大、华科、清华的博士打分。结论:四篇论文的形式和表达强于普通研究生,医学论文质量一般;但在已有研究上扩展,工科论文效果出奇地好。@bourneliu66
I gave my students a video-editing assignment.
The ones with editing experience started planning shoots and After Effects work.
The one with zero experience generated images with GPT, animated them in Kling, and cut it all with Codex. Done first.
Prior expertise can quietly cap how far you push AI. I keep reminding myself to unlearn.












