Xiluluke's Weekend Letter Vol. 12

Right now I'm in Mianzhu, Sichuan, on a work trip here for the first time. I didn't expect a free bus from the high-speed rail station into town. That's a pretty nice perk for residents. But when I got to the bus stop, the bus was already packed, the kind of packed where there's no way you're squeezing on.

The bus was full, but the driver wasn't leaving. I figured it had to wait for its scheduled time, so I checked, and sure enough, it didn't leave until 18:10. At 18:10 it did pull out. But then I saw the next bus wasn't until 19:40. So I checked the train schedule, and sure enough, a train was arriving at Mianzhu South at 19:20.

A free bus is great, but there's only one bus at a time. How many people can squeeze onto one bus?

Short on time, I ended up taking a cab to the hotel and started writing this.

Life has been really full lately. All of a sudden this weekly has been going for three months and hit its 12th issue. It's gone from fumbling around at the start to the fairly practiced shape it has now. Simply put, it's my weekly journal.

I skipped posting several days this week. Not because I didn't want to, but because I've been vibe coding my new project and pouring almost all my time and energy into it. By six or seven every evening, my brain won't let me write another word.

So it makes me treasure the chance to write this weekly even more, because it's one of the few times left when I can sink in and write.

Liking Specific Things

Yesterday I saw my brother looking at car models. I rarely see him interested in anything else. He showed me a Lamborghini model, though I've forgotten what it's called.

I asked him what he liked about the car. He said it looked cool. I asked where it was cool, and he couldn't say. I asked if he thought it was cool overall, and he said yes.

I said I think all Lamborghinis look the same. He said they definitely don't. I asked how they're different, and again he couldn't say.

I've found this is actually pretty common. Lots of people feel they like something, but when you dig in, they don't know what exactly. Take the Lamborghini's exposed engine. My brother thought the design was cool but didn't know why. I think it's simple: ordinary cars don't show off their engines.

And what engine it is, how much horsepower, how many cylinders, he didn't know either.

It's not that liking something means you have to know every detail. But I hope the things we like are specific, not some abstract presence.

"Cool," for example, is a very abstract word. Cool can be specific too: the exposed engine I mentioned, or high-tech butterfly doors, or a sleek fastback body. Those are all specific ways of saying it.

Complimenting someone works the same way. Just saying "you're amazing" sounds like flattery. But if you say "you're really good at XX" or "you nailed such-and-such," it lands completely differently.

What matters is saying things specifically, whether it's the nouns or the verbs.

It's the same for people and for things.

And I think when someone loves something, they always go deep into the specifics. Take Ultraman. To me, all the Ultramen look the same too. But to someone who can rattle off every Ultraman, each one has its own unique, specific features.

Or take the new bike I got over Chinese New Year, a Cannondale LAB71. In China, the brand and model are a bit niche compared with foreign brands like Trek and Specialized. For the same budget I could have bought a Trek or a Specialized, but I still went with Cannondale. Why? Because I like Cannondale's less rounded design, which makes it look more elegant, and because I love this green color. Just those two things, nothing else.

If we're talking cool, Trek and Specialized are cool too, each in its own way. But in my eyes, the Cannondale is the one that stands out.

So like something specific. Maybe it's only in the specifics that you find what you really love.

Occam's Razor

Recently, listening to Meng Yan's podcast, I heard Li Jigang bring up a theory.

Occam's razor comes down to one line: entities should not be multiplied beyond necessity.

Its core idea is: when several theories can all explain the same phenomenon, prefer the one with fewer assumptions and a simpler structure.

It doesn't say the simple explanation is always right. It says don't bring in extra assumptions when you don't need them. If your computer suddenly can't get online, the Occam's razor approach is to check the router, the cable and the network settings first, not to start by suspecting a hacker.

Once I learned about this theory, I found it really interesting. It reminded me of something I read in How to Take Smart Notes.

Niklas Luhmann, who created the Zettelkasten method, published dozens of books across many fields in his lifetime and was incredibly prolific. When people asked him how he managed to be so productive and so good, Luhmann just said he took notes every day.

Most people can't make sense of that. Take notes every day? Who can't do something that simple? You become a writer just by taking notes every day? That's how it goes. People have a hard time believing a simple method can solve a big problem. They always feel a big problem needs a big method, a Big Idea, to solve it. But it's often the other way around.

Just like lots of people never think of walking and running as nothing more than stepping forward with one leg, then the other. Alternate your two legs and you can walk anywhere you want. It looks like a long road, but really you're just repeating one alternating motion over and over.

Occam's razor teaches more than that. It's not just for explaining problems. It works in lots of other areas too.

I'm glad I came across this theory recently, because I'm vibe coding a big project right now: the AI-based long-form novel writing and management project I introduced on my WeChat account a few days ago.

The project already had a demo at the end of April, but it wasn't a proper app. On May 16 I first wanted to turn it into an app, and as of today something fairly usable has landed.

But of those ten-plus days, the project was stuck in place for about ten. The reason was that my thinking kept circling around things that didn't need to be there. What really got me on track was one day suddenly realizing that thinking about all this was pointless. What mattered was getting the truly core parts of the software working. Once I held onto that, I slowly got back on track.

The idea here is: don't think about lots of complicated stuff yet. Get the most essential, core part of the software right first, and deal with everything else later.

Then I came across Occam's razor, and it came at exactly the right time.

With that razor in hand, over the last few days I found lots of things in my software that had no reason to exist, so I went wild cutting them. After I'd cut them, not only did it not affect using it, the whole app became cleaner, with basically no noise.

I'm starting to feel the project is heading in a good direction.

Vibe Coding

There's basically no way I'm not writing about this. I looked through what I've posted lately, and almost all of it has to do with AI, because I've been using AI way too much.

Before, my vibe coding mostly revolved around my blog. If my experience used to stop at the blog, now I've been through building a whole piece of software.

I can't write a line of code or read a line of code, and I really did build software of my own.

That would have been impossible in any era before. But in the age of AI, I actually did it. And it shipped.

This week my main work was fixing bugs and adding a really important outline system feature.

Only after fixing bugs myself do I get why there are so many jokes online about programmers who look like old men but are actually twenty. It really fries your brain. For the last few days I've dealt with at least ten bugs a day. And fixing a bug isn't the end of it. Often touching one thing affects everything.

Some bugs don't get fixed on the first try. They take several attempts and several different models.

Some bugs get fixed, but the fix spawns a few new problems that need fixing too.

So even with AI's help, I spend all day, morning to night, fixing bugs. No exaggeration.

Luckily, things always get better bit by bit. When I sent the first version of the app to people to try, it took over twenty minutes just to get to the main screen. Seriously hilarious.

But after these few days of testing and nonstop fixes, it's basically fine overall now.

So right now my mind is mostly on designing the outline system. Only once that's complete will Book of Truth be more or less finished.

Actually, there are quite a few AI novel projects online right now. Xingyue Writing was really popular a while back. But after trying it for a day, my verdict is it's a garbage project, nowhere near my Book of Truth.

Xingyue Writing and lots of open-source long-form AI novel tools online have similar problems: they're designed to be too complicated. The interface is full of random tools, there are lots of random features, and the onboarding is a mess. You go in and fumble around for half an hour without knowing how to start writing a novel. What good is a tool like that?

Honestly, I'm really optimistic about Book of Truth right now, because I have some original ideas in how I've imagined it that I haven't seen anywhere online yet.

Now for the system design. The AI long-form novel tools online today are basically all multi-agent designs. So far I haven't gone multi-agent, but judging by the trend, multi-agent seems to be where projects like this end up.

But the key is how the agents in the system should be designed. That's where the systems really differ.

I'm still slowly feeling my way on this. I actually don't want to add too much to the software, because I've found that sometimes when you put too many constraints on AI, what it writes is total garbage. On the other hand, if you just give it an outline and some simple style constraints, it can write really well.

How much you constrain the AI is the key here. Without oversight, AI will definitely go off the rails. But too much oversight and it feels stifled. You need to find the balance, both by being clever in the design and by testing and validating through lots of practice.

A quick aside: I'm planning to start a community for AI long-form novel writing and run this project as a long-term thing. If you're interested in writing long-form novels with AI, get in touch.

:::info That said, I haven't officially promoted this product or community online yet. I've only posted about it to leave a footprint on the internet. :::

My current view on AI web novel writing is that even the cheapest AI models write well enough to use. So the problem isn't the prose at all. It's how to use AI effectively and manage the writing of a novel.

And Book of Truth is the answer I'm putting on the table.

What's fun about vibe coding is that there's no real technical barrier, and it gives you a real sense of creation. Turning the ideas in your head into reality is the best way to feel that.

Vibe coding also does a great job of curbing my appetite. I basically sit at the computer for hours at a time, and at mealtimes I just make a simple fat-loss meal. If anything, I've gotten more disciplined, haha.

Final Words

That's it for this week's issue. It's not exactly fun stuff, but I had a blast writing it.

Happy June, everyone.

I've stopped treating daily posting as a task. When I can write, I write. When I can't, I don't post. Whatever happens, happens. I'd also like to write some more everyday-life stuff.

We'll see.