Letting an agent rebuild my information diet

I’ve uninstalled Xiaohongshu and Douyin many times, and every time I’ve installed them again.

In June 2025 I unfollowed every business podcast on Xiaoyuzhou and wrote in my notes: “I don’t want utilitarianism to invade me and make me live in anxiety every day.” But a few days later, lying awake, I picked up my phone again, and what I saw was more of the same, mostly anxiety. This September and October, I wrote “the more I scroll, the more anxious I get” in my notes several more times.

These past few days I asked an agent to help me reorganise my information sources. Along the way I gradually realised: information sources decide not only what I see, but what I want, and even what I think the world is like.

The feed wants things for me

Take Xiaohongshu. I once tried to use it as a search tool. Its home page has a search box, but unlike early search engines, whose home page was just a clean search box, here the search bar is a thin strip at the top and the rest of the screen is packed with recommended posts. These days many search sites also put news and recommendations on their home pages; for the sake of traffic and ads, search itself takes a back seat. I meant to search, but the moment I saw those images my attention was pulled away and I forgot what I’d come to do. The headlines are worded to manufacture curiosity, the pictures are glossy, and what comes across is anxiety. Being steered by software like that makes me deeply uncomfortable. Platforms aren’t only after traffic, but also ads and other revenue, all part of their profit calculations.

Podcasts are the same. Many of the popular shows Xiaoyuzhou recommends aren’t about real progress in technology but about anxiety itself. Listening once in a while tells you something about the public mood; listen too much, and all that’s left is anxiety. I used to follow a few hosts I respected, and later found their new episodes said much the same as the old ones, with little that was new.

More worrying are people who resell second-hand information. I once followed a business news podcast. One episode turned to tech, and it sounded so unprofessional that I unfollowed on the spot. If they couldn’t get software and tech right, an industry I know, it was hard to trust the rest of what they said. That’s only my inference and may not be fair, but I doubt the quality of that kind of information. And with so many ads woven into the show, you have to listen with extra care.

The whole media industry, WeChat Official Accounts included, runs on a traffic logic that serves itself: making money from traffic and ads. Some will do anything for traffic, producing low-quality, sensational, anxiety-selling content. They treat information as a business rather than pursuing quality, and the result is content that all looks the same. Last year I read a memoir by a former new-media editor-in-chief, who likened new-media writers to “children imitating how grown-ups dress and talk.”¹ I wrote in the margin: that’s exactly what news podcasts are like.

It reminds me of a book I’m reading, René Girard’s Deceit, Desire, and the Novel. Girard argues that desire isn’t spontaneous but imitated: we often want something because we see others wanting it.² Internet media manufactures desire too. There are things I didn’t particularly want, until I saw them, and then the desire was stirred up, far stronger than before. Relationships are the same. Many shows discuss love over and over, telling you again and again how hard love is in this era, and it only makes you want it more. Listening, you start to feel disappointed in yourself, asking why you don’t have it, then envying others, and envy makes the desire stronger still. The more you listen, the more easily it’s aroused.

Over time you end up wrapped inside it. A communications course I took covered the “filter bubble”. The idea is contested: critics say people aren’t entirely passive in front of algorithms; supporters say a bubble doesn’t form all at once but changes how people get information bit by bit.³ I side with the latter, because I’ve been stuck in one myself: anxiety, relationships, quitting, job hunting, the same kind of content over and over, more anxious the more I read.

Information shapes us, and makes us think the world is the way it shows us. It reminds me of Plato’s allegory of the cave: prisoners who only ever see shadows on the cave wall take the shadows for reality.⁴ The feed is that wall.

Why consuming less isn’t enough

This isn’t my first attempt to sort out my sources. My only method used to be: consume less. Unfollow, uninstall, delete. At the end of last year I looked at myself through the Theory of Constraints (TOC) and found I’d been pushing hard on everything that wasn’t the bottleneck: reading, scrolling, watching videos, saving videos, with hundreds of hours of unheard podcasts piled up. My conclusion that day was “I should stop the input.”

Consuming less lasts a while, but it only tells me what I don’t want, not what I do. The freed-up time fills up with the feed again within days.

Other people’s solutions aren’t total abstinence either, but changing the entry point. Make Time suggests turning off notifications, replacing default pushes with active choices.⁵ I also feel every notification is fighting for my attention, and what I’ve seen leaves a trace in my mind. Today on Hacker News I read Vitalik Buterin’s “Digital minimalism”. He didn’t leave X, he just switched to a client without a recommendation algorithm, and a few months later he’d lost interest in the platform’s daily fights.⁶ A few days ago I read The Anthology of Balaji, which makes a similar point: media has been taken over by social media, and we should take it back into our own hands.

So the question isn’t how much to consume, but who decides the entry point.

Letting an agent be my information officer

Last Saturday I went to hear Cai Xin talk about her personal agent. She listens to lots of podcasts, discusses them with AI afterwards, and saves the summaries in her own knowledge base. She also has an “information officer” that gathers information from X and YouTube every day and reports back to her.⁷ Karpathy tried something similar: unhappy with YouTube’s recommendations, he had Claude pick out videos recommended in years of Hacker News comments and turned them into a YouTube-like site, with no account needed and preferences stored only in your browser.⁸

They’re doing the same thing: decide who to read first, then let AI fetch.

There’s a more fundamental reason to hand this to AI. When people see information they react emotionally, and every item takes up attention. AI’s attention is almost unlimited: it can process huge amounts of information, isn’t swayed by headlines, and doesn’t get more anxious the more it reads. So extracting information should go to AI, leaving our own limited attention for more important things.

Using AI for research has really meant handing over the attention I’d otherwise spend on research, so I’m less easily distracted too. That has a cost, of course. When I searched myself, I’d often stumble on something interesting along the way, something I’d wanted to learn about anyway; now those surprises are rarer. But a lot of what search turns up now comes from social media, full of emotional wording, so the quality is questionable to begin with. Weighing the two, I’d rather let AI filter first. That’s not to say nothing else is worth reading, but that my own attention should go more to classic, professional, authoritative things. Information others have already compiled is more like a literature review: let AI read it, no need to check it item by item myself. What really matters is where information comes from.

So I had an agent try it too.

Start from sources I already trust. I didn’t start from scratch. I first listed sources I already trusted: Hacker News, a podcast list Cai Xin recommended, and a few writers and shows she mentioned. The agent checks these sources regularly for updates and flags what it thinks I might be interested in.

Text only. What I really need is transcripts, so the agent can search them. The agent checked each show’s RSS feed; eight of them include transcripts directly, and for the rest it takes YouTube captions and turns them into transcripts. With the text, I don’t listen start to finish. I discuss it with AI directly: what this episode covers, what I didn’t know, what the key points are, which parts are worth hearing in full. I’m bound to miss some things this way, but it’s broadly fine. It’s a trade-off; otherwise I simply couldn’t keep up.

Let AI search. Besides the fixed sources, I also have AI search the web for things related to questions I’m currently thinking about. Today, for instance, Shanghai hosted the AI Engineer conference. I didn’t know in advance; I only found out from WeChat Moments. Having missed it, I had the agent search online: pick out the talks I’d be interested in, find what the speakers themselves have written and said, and check every night for new recordings. I used to only be able to read other people’s retellings, because looking everything up one by one took too long. Now the agent searches, and I just read and judge.

Email is a source too. My Gmail had so much junk that important course notices were buried. While cleaning up, I found that Lenny’s Newsletter had been publishing all along and had been discussing personal agents for a long time, and I’d never seen it. So I unsubscribed from marketing emails, kept Lenny’s newsletter, and filed it under its own “Reading” label.

The agent fetches, I decide

After this round, the agent does all the time-consuming work: checking sources for updates, tidying transcripts, searching the web for related content. But which sources to read often, which to drop, which episode I want to hear, those are my decisions. If I handed that step over too, I’d just have swapped one feed for another.

The hard part isn’t finding good content either, but keeping good content from turning into a backlog again. So my knowledge base only holds what I’ve actually listened to, and finished transcripts sit alongside my own thoughts.

One more thing: many podcasts discuss anxiety, but the more they discuss it, the less it’s solved. Rather than mulling things over and over, better to get on with your own work, find a goal, and move forward. It’s the same thing Carnegie says about worry: keep busy, and worry has nowhere to stay. Sorting out my sources is about making time to do things.

Who I want to become

Making time to do things, yes, but what to do, and who to become: information sources are quietly answering that too.

Recently I listened to a podcast episode about Eileen Chang. Professor Song Mingwei talked about how Chang wrote of a “fear of listening”: the voice of social trends is so loud that the individual’s voice gets drowned out. He said Chang believed in individualism all her life, and following your own choices means giving up a lot.⁹ Most of the voices in a feed are that voice of social trends: what to be anxious about, what to want, what kind of life counts as normal. I often get stuck between two forces myself: what society expects of us on one side, and who I want to become on the other, wearing myself out going back and forth.

Changing your sources means changing the voice you hear. That day listening to Cai Xin, my view suddenly opened up. For a while I’d been wondering: now that AI is here, what can I do, and where does a person’s value lie? Someone in the audience who writes for a living asked her what to do about how anxious writers are. She said people aren’t defined by their jobs; we aren’t born to work. We should find what we truly care about beyond work, then use AI to strengthen it, and become fuller people.⁷

It reminded me of what Romain Rolland wrote in Jean-Christophe: many people die at twenty or thirty, and from then on just repeat themselves.¹⁰ Seen that way, AI’s arrival is actually a good thing. It pushes everyone to rethink who they want to become, instead of being stuck in a job, living by a script society has written.

So sorting out my information sources is, in the end, about hearing my own voice again amid the voice of the trends. With the entry point in my own hands, I have room to think: who do I want to become?

I don’t plan to turn this into a product. My process won’t necessarily suit anyone else; everyone can have their own agent build their own. What I can share is experience, which is this post.