Xiluluke's Weekend Letter Vol. 07

Most of our lives are just ordinary habits running along, and how well they run decides who you are. — William James

---

In a blink, another week is almost over.

Looking back, what did I spend my time on this week?

• Researching and using AI
• Planning my column
• Playing Roco Kingdom
• Cardio
• Walking

Researching and Using AI

Research

This week I tried one model from abroad and two Chinese ones.

ChatGPT Image-2

This is an image generation model that just came out, and it's amazing at making images. Yet Liang Mouyin and I both took it off the beaten path and used it to design our blogs.

MiMo-V2.5-Pro

One of the Chinese models, from Xiaomi. I paid ¥30 for its Token Plan Lite, and I've already used up almost all of the plan's quota.

DeepSeek-V4-Pro/Flash

DeepSeek's new models came out this week, and I'm already using them.

Using Third-Party APIs in Claude

I heard the Claude desktop app now lets you run Claude Code with third-party APIs. I was thrilled, so I dug into it right away.

Before, if you wanted to plug a third-party API into Claude Code, you could only do it in the terminal. I don't really like the terminal, for a simple reason: I think an app fits how we use things today better. It looks nice and it's easy to use.

Most importantly, coding in the app makes it easy to manage past tasks. I like keeping tasks for different directions in their own chat threads.

Practice

ChatGPT Image-2

After two days of the free quota, I signed up for ChatGPT Plus and used this model to redesign my blog from top to bottom. Compared with a week ago, my blog is a whole new thing.

From using it, the model has a few big strengths:

• First, it's smart. It understands what I say and produces exactly the image I need.
• Second, it's fast. Send it the requirements or a rough draft, and it returns an image in very little time, usually with no second round of edits.
• Third, the quota is generous. I don't know where to check ChatGPT's token quota, but in recent use, at no more intensity than I use Claude Code, I've never hit a five-hour limit. Does ChatGPT just not have a five-hour limit?

But it still has some problems:

• First, in chat it calls the image model for no reason. For example, I upload an image and say clearly in the prompt that I want help writing a prompt and not to call the image model, and it still just generates an image. That leaves me speechless.
• Second, it tends to freeze when you edit an image a second time. Sometimes an image has small flaws, like a cup with two handles. When I select the extra handle with the brush and ask it to remove it, there's a chance it freezes. Luckily, when I go back to the screen after it freezes, I still get the result I wanted.
• Third, every time you re-edit an image, it creates a separate chat, so I end up with a huge pile of chats. I don't think I've ever seen an AI company add chat folders or anything like that. Does that go against some logic? Or do AI companies just want users to delete or archive their chats?

MiMo-V2.5-Pro

Honestly, I don't even know what to write as a usage log for this model, because it all went by so fast.

I don't know how many tokens the credits in the Lite plan add up to, but I used them all up without really doing any projects.

If you asked me now what exactly I did with it, I couldn't even piece together one complete workflow.

• In my "Materialist Dialectics Q&A Tool," I used it to build an image-sharing feature and design a half-finished image template.
• I had it look through my codebase and delete some junk code.

That's it. Nothing more.

My impression is that it burns through tokens far too fast. I did pay ¥30, after all.

If I'd done the same work with Sonnet, I might not even have touched the five-hour limit. With MiMo, I've already used up a month's quota.

More importantly, I think its ability is just average. I'm not a professional programmer, but I've used Sonnet, Opus, Codex and others, including the new Chinese DeepSeek V4, and they all do better work than MiMo.

DeepSeek-V4-Pro/Flash

I was genuinely happy to see DeepSeek release new models, because DeepSeek's tokens are famous in the industry for being cheap and plentiful.

More importantly, my "Materialist Dialectics Q&A" project is built on DeepSeek's API. Before, I could only use the DeepSeek Chat model for conversations, and the output was okay. But when I saw DeepSeek V4's performance report and found it was every bit as capable as Claude's Sonnet, I was even happier.

So I immediately tried switching my API model to V4 Pro. But after half an hour of testing, I found my product isn't a good fit for the Pro model.

The Q&A tool runs on a web page, and V4 Pro not only uses a lot of tokens but also thinks for a long time. Ask it a longer question and it often hits the timeout, which makes for a terrible experience. So in the end I switched to V4 Flash.

Compared with Pro, Flash is much faster, and its answers are much better than DeepSeek Chat's. I'm really excited about it.

On top of that, I plugged the V4 Pro model into Claude Code and have been coding with it for a while. It's pleasantly surprised me. The report's claim that it's "no weaker than Sonnet" holds up completely. It gets work done reliably and fixes bugs fast.

So good.

Even better, DeepSeek is running a promo right now with 75% off tokens. Bless them, that's practically free!

Planning My Column

This week I ran into a problem while writing my column.

The piece I wrote this week is called "What Kind of Problem Is Subjectivism?" After finishing it, I found myself stuck in the mud.

I analyzed the nature and roots of subjectivism in depth, concluded that subjectivism is a problem rooted in epistemology, and compared it with materialist epistemology.

By the end, I realized that for readers, what I'd written was just a castle in the air.

Because to understand these errors systematically, you first need a materialist epistemology and worldview. And apart from the systematic analysis, what I'd put out so far didn't present any of those basics.

So, though I'd planned to keep going with a long treatise, I pulled back and left it at a light touch.

Over the next month or two, I plan to focus on breaking down two essays, "On Practice" and "On Contradiction." Whenever anyone has asked me how to read the Selected Works, I've always told them to start with these two. So in the whole column, they're the bridge between what came before and what comes after.

I don't know how well I'll do, but I have to do it, and I should.

Playing Roco Kingdom

It might seem odd to put this here, but I really did spend some time this week playing Roco Kingdom.

I was a loyal player of the Roco Kingdom browser game as a kid, but I stopped once I got to middle school. Now there's a mobile version, and as someone who plays for nostalgia, I came back.

This week I did nothing in it but hunt for shiny pets. I got shiny versions of Aurora Thousand-Beast, Big-Ear Hood, Crescent Snow Bear and Fallen Star Rabbit, and picked up a shiny Beise along the way.

But honestly, it felt pretty empty. I don't know what I got out of it.

That's my attitude to games. Very conflicted. I can't not play at all, but I can't really focus on playing either.

Thinking about it, I don't like story in games. If there's a skip button, I always skip. If there isn't, I mash the space bar.

Honestly, a game uses its story to tie together all the models, art and so on. Without the story, what's left is pure gameplay: catching pets, battling, exploring. But none of that really interests me.

My favorite games used to be roguelikes and building games. The first gives you that great feeling of mowing down enemies. With the second, I'd get hooked on the thrill of logical planning and couldn't pull myself away.

In the end, I play far less than I used to. I think the only problem with games is that they take too much time.

Right now Xiaoheihe shows a little over 300 hours for me, while my friend Jiahua has 3,000. Honestly, I don't even think that's a lot for him, because I once met someone with thousands of hours in a single game.

I can't even imagine.

Cardio

After resting five days last week, I'd recovered quite a bit. This week I started getting back into it, but I still didn't go to the gym. I just rode the bike trainer at home every day.

After five days of riding, I found my aerobic fitness has slipped a lot, far more than a little.

Holding 140 W for an hour already gets my heart rate to 150, and on one or two days I ran out of steam after only about 50 minutes.

It's unbelievable. How did I get so weak? Last year I could hold 200 W for a full hour.

It's exactly because it's so unbelievable that I haven't gone back to the gym yet. I'm sticking with the trainer for a while to see if a few more days of training improve things.

Walking

This week I started deliberately setting aside time to walk after dinner.

Honestly, I'm cosplaying Steve Jobs. I never had a walking habit before. Only after reading Steve Jobs did I learn that Jobs loved walking and talking with people, because great ideas often come up while you walk.

Later I read in a Dan Koe piece that he thinks of walking as a kind of meditation:

If you really can't write, try doing what Jobs or Tarantino did: stop, go for a walk, do something else. Your brain keeps working in the background, and inspiration often strikes when you least expect it. That's the so-called "aha!" moment. Share those moments, and someone will always benefit.

— Dan Koe

I haven't felt that flash of inspiration yet, but so far, getting out for a walk doesn't seem like a bad thing.

• I have the time. If I didn't spend it walking, I'd basically spend it scrolling short videos or playing games. Using it for a walk is actually great.
• I work out a lot most of the time, but I still don't burn enough calories. An hour's walk is an easy way to move more.

So why not?

Now every evening a little after seven, I head out for a walk. While walking, I try hard not to look at my phone. I put on headphones, listen to music, and pay attention to the people passing by, the music, the buildings, the plants, the wind.

An hour goes by fast. I can only manage a little over 4 km. Afterward, even without any new ideas, my body feels really good.

Walking just to get inspired would be a bit too calculating anyway.

The only pity is that it would be nice to have someone to talk to while walking.