Xiluluke's Weekend Letter Vol. 13
- No.
- 200
- Date
- Read
- 15 min
- Words
- 3,298
Today is the gaokao, the national college entrance exam. I wonder if any of my followers are taking it this year? Wishing you all a steady performance.
I'm in Fushun County in Zigong right now. When I woke up this morning I heard rain pouring outside. I opened the curtain a little and saw a bus carrying exam takers driving this way.
I don't know why, but it always seems to rain during the gaokao. In my head there's always this picture of students heading to the exam halls on a rainy day. When I got to Fushun yesterday it was unbearably hot, and today the temperature dropped eight degrees all at once. I don't like rain, but I have to admit that slightly milder weather is better for thinking. So the students should do better in weather like this!
I worked all day and got back to the hotel at six. I'd been planning to watch today's WE vs. TES match in the afternoon, but by the time I got to the hotel, the second game was already half over. I watched for a bit, and WE had already lost two games!
Civilization
Chatting with Boss Quan, I found that my idea of civilization is different from a lot of people's.
I think civilization is the opposite of barbarism. "Barbaric" sounds like a pointed word, but it actually fits a lot of situations. On a big scale, someone with violent tendencies who always wants to threaten others with violence is barbaric. On a small scale, someone stubborn who always forces others to bend to their will is barbaric too.
When it comes to civilization, many people might think "reason" is its key word, because reason drives us to understand things more systematically and structurally, and lots of rational thinking can grow from there. But I think civilization isn't something one word can describe. Civilization is a system.
If I had to pick one key word for civilization, I'd say "benevolence." What comes to mind when I think of benevolence? I think of what Liu Bei said: "Do not fail to do good because it is small, and do not do evil because it is small."1 That's also a reflection of personal character and cultivation.
Then there's anger. Everyone gets angry. But Lu Xun said: "When the weak are angry, they draw the knife against the weaker. When the strong are angry, they draw the knife against the stronger."2
At the start of the New Culture Movement, Chen Duxiu championed "Mr. De" and "Mr. Sai," that is, "Democracy" and "Science."3 These were among the good things in Western culture, and they were brought into China. These days, lots of things are supposed to combine Chinese and Western ways, since so much of Western culture has come into China. I think the original point of combining them was to take the other side's strengths to make up for our own weaknesses. But when it comes down to individuals, it can turn into looking only at the other side's weaknesses and leaning on our own strengths.
Take "elite culture" and "meritocracy," for example. Both are very common in the Western world. They have their strengths, but also their flaws, and some people only see the former and ignore the latter. Meritocrats, for instance, always credit their success to their own hard work and deny every other factor.
Lots of things are like this. Something originally neutral gets packaged as civilization and treated as gospel, when in fact it twists what civilization really means.
Time
It's gaokao season again. Looking back, my high school years were the happiest years of my life. I never imagined it's already been six years.
As the old sayings go, "Time is a galloping horse under the whip, the days and months are fallen petals on flowing water," and "Flowers bloom again, but no one is young twice."4 Most of the time I don't feel time flying by. Only at certain moments, when something suddenly wakes up an old memory, do I start to sigh.
Like now, when I come across cartoons I watched all the time as a kid on streaming, the scenes from back then come back vividly.
As a kid at my grandma's house, in the middle of summer, I'd put a chair in front of her bedroom door with a little fan in front of me, cool air blowing, and my favorite shows on TV.
My favorite show as a kid was Pleasant Goat and Big Big Wolf, but there were so, so, so many cartoons that stuck with me: Fruity Robo, Happy Heroes, SpongeBob SquarePants, Beast Armor, Lolo's Adventures, Little Chef Fugui and on and on. They had twisty, exciting plots, and they really taught you something too.
Now when I see compilations of these cartoons on my phone, I can't help thinking that those lazy afternoons were more than ten years ago.
Or when I hear a song I used to listen to a lot as a teenager.
There's one song I don't even like that much, but whenever I hear it, I think of more than ten years ago: "Yi Xiao Qing Cheng." I was in middle school then and thought it was so romantic (mostly because of the drama Love O2O it went with, though the lead actor has since been banned). So now whenever I hear it, I think of summer, breezes, school and youth.
Speaking of youth. I used to see this line online all the time: "You can't have youth and the feeling of youth at the same time."5 Back then it didn't move me at all. Now I'm 24, and it suddenly hits me that my youth seems to be gone.
I don't have a clear idea of what youth is. I guess it's those days when carefree and anxious were all mixed up. They weren't my most creative days, but they really were the most comfortable time of my life. Looking back now, it seemed like I thought about a lot back then, but there wasn't much in my head. The happiness was real, though. I still miss playing cards and Legends of the Three Kingdoms with my roommates in college.
Since I started working, each year seems to go faster than the last. The good thing is that this year I've slowly started to feel life is real, I've found a direction, and I can feel my creativity moving forward.
Reading
This week I opened two books, Poor Charlie's Almanack6 and Google's AI Brain: Hassabis. I've put the first one aside for now and I'm mostly reading the second. I'm a bit over four chapters in and have a few things to say, so I'll write them here.
First, AI, which has only really come into public view in the last ten years, was actually being researched way back in the last century. Hassabis was one of the first people of our time to set out to work on artificial intelligence. What blew me away is that in an era when CPU and GPU technology wasn't yet powerful, these AI scientists could see AI's wide future and influence decades ahead of their time.
These scientists were already in their thirties when they finished their PhDs. Choosing at that age to devote your life to a project whose scientific and business prospects were both unclear is something ordinary people can hardly do, just like very few ordinary people can hold a single stock for thirty years.
Reading this book hasn't just given me a more concrete picture of how AI developed. It also introduced me to a lot of people, and the one who struck me most was Geoffrey Hinton.
Why? He doesn't have much direct connection with Hassabis. But when Hassabis was hiring scientists for his company DeepMind, most of them were Hinton's students.
That made me curious, so I went and looked him up.
Geoffrey Hinton. He just won the Nobel Prize in Physics in 2024 and is called the "godfather of deep learning" and the "godfather of AI." Among his honors, even the Turing Award hardly stands out.7
His great-great-grandfather was George Boole, the founder of Boolean algebra, which laid the foundation for the binary logic of modern computers.8
His great-grandfather, Sir George Everest, led the trigonometric survey of the Indian subcontinent, and Mount Everest is named after him.9
All I can say is, I've wandered into the house of the gods.
But shock aside, the most interesting part of what I've read in this book is this: researching AI in this era means taking part in the era itself. These AI tech giants, whether Hassabis or Hinton's many students active in AI, are all in their prime. They still have immeasurable creativity and potential waiting to be unleashed.
Maybe it's because I've always read timeless classics that I assumed everyone I looked up to was someone long gone. Studying and learning about AI has introduced me to a group of people living in the same era as me. It's a strange and wonderful feeling.
"Computers will soon be able to do anything the brain can do."10
Vibe Coding
This has almost become a regular in my weekly, but I can't help it. I do it nearly every day.
Good news this week: the verification-code check on my Codex got lifted, so I can use it again. But that didn't make me all that happy, because for me it just means one more account to work with, with roughly 60 million tokens a week. Claude alone gives me nowhere near enough tokens.
All week I'd been looking for a model to replace Codex and hadn't found one yet, and then Codex unblocked me first, haha.
On the coding side this week, I've got two things to share.
The first is an idea. I think in vibe coding, the number of tokens can bluntly be equated with productivity. The more tokens, the more productive you are. Paying more money is really just buying more tokens.
The second is a coding workflow I learned this week.
When I used to code, sometimes I'd talk through a lot with the AI in one go and have it build everything at once. The result was that what came out was wrong all over the place.
So sometimes I'd talk with the AI point by point instead, and fix things point by point. The downside was that I had to sit at the computer chatting with it all day, making decisions nonstop, and at the end of the day I had no idea what I'd actually done.
The first way does too much at once to keep an eye on. The second is too scattered to manage.
But thinking it over, neither way is wrong. The method is the problem.
When I talked with the AI about a huge piece of work, my problem was that I didn't talk through my needs, my ideas and the result I really wanted in detail. I just gave one long description. It looked like a lot, but a real big project needs its requirements broken apart and discussed piece by piece. You can't just bundle it all up and say it in one breath.
- Get clear on your core needs
- Every small point needs a clear, specific answer
That's the core of a big task. Now I do it really well with a Skill called "Grill Me." The core of this Skill is simple: it keeps asking you questions, and when it thinks things can wrap up, it stops.
Once it stops, it starts writing the spec. After the spec, it writes the execution plan, and finally it implements it.
The key point here is to get the AI to run longer tasks in one go as much as possible, and the longer the task runs, the longer the plan should be. A couple of days ago, for example, I had a plan of about 2,000 lines. It took three full rounds of five-hour token limits to get through... The method is great, but for those of us on the $20 subscription, it's a bit of a stretch.
- Let the AI understand your core needs
- Let the AI ask follow-up questions to draw out the details
- Let the AI write the spec itself
- Let the AI write the plan itself
- Let the AI carry it out itself
- Let the AI review it itself, over several rounds
- Let the AI commit it itself
Of course, this is already the simplified version, just a quick workflow I've figured out over the past few days. But even just this has sent the quality of my work soaring. After three or four big tasks so far, I've had very few small bugs to fix. I can clearly feel the quality of the tasks shooting up.
Also, when I'm discussing a project with the AI, I use the best model with Max turned on and go into detail. Usually after forty or fifty minutes of discussion, a round's tokens are about used up, and I have to wait for the next refresh before it can start the task. But for the actual work I've switched to low instead of Max. It saves tokens. The detailed execution plan is already written, after all, so there's no need to run it on Max.
Book of Truth
Finally, let me talk about the product I've been working on lately, Book of Truth.
A fairly complete version was actually ready on Monday, but in my full vision for the product, it's only about half done.
I got the AI writing flow working end to end and debugged it, then sent it to a friend to use first, since he'd already been writing a novel with my Markdown version of Book of Truth.
Over the last few days I've been working on the Book of Truth system again, and during this stretch of development my ideas and thinking have been shifting a little.
First, I'm not worried at all about AI's prose. What AI writes now is already better than lots of beginner writers. So from the start, the core of the Book of Truth system has been about using a system to constrain and manage AI's writing. That is, building a precise system around a novel so the AI writes carefully inside it.
When I said I'd gotten the writing flow working end to end, I meant I've proven out a writing system that can actually be used.
But that's not enough. Writing is one thing, but the foundation of AI writing is a detailed outline. The outline system is a part I care about a lot, and I've put a lot of thought into designing it.
Right now the Book of Truth outline system is designed like this:
- Whole-book outline
- Volume outlines
- Story stages
- Story units
- Chapter outlines
Five layers, each building on the last, and story stages are the design I'm proudest of. Lots of AI writing systems jump straight from the volume outline to chapter outlines, or the better ones have five-chapter or ten-chapter outlines. But I think those designs have no soul.
Story stages are my current idea. A volume's story can be broken into many story stages, big and small. The big and small part is the key. Some story stages might be only a few chapters, while others might run dozens or even a hundred. Mix big and small and the plot gets real ups and downs.
The unit outlines under the story stages give finer control over each stage's pacing.
That's my current idea, but I haven't finished a novel from scratch with this system yet. It's still in the testing phase.
Still, designing all these layers of the outline system is, at heart, a way of managing the AI in fine detail.
During development, I've come up with a whole new, different idea. My thought is: in the early part of a novel, set up a basic framework for the world and the characters and develop it a bit, and then I wonder, could the AI work out where the story goes on its own?
Because I think that sometimes, managing a book this finely from start to finish might work well, but it can also make the story feel rigid and lifeless.
The soul of AI web novel writing is the AI. If people keep restricting it and only let it write from what they hand it, you lose part of the point of AI.
I think the strength of large models comes down to one word: "extrapolation." They can think almost like humans, and their knowledge is in no way weaker than ours.
So we can hand the content to it, let it write on its own and improvise, and only give broad, general constraints.
In the early part of a story, give it a fixed world, the matching characters and part of the plot, then hand all of those elements over to the AI. Let it judge from the characters' personalities where their stories go next, and let it work out the story on its own. I think that would be really interesting.
I'm working on some concrete designs for this possibility now, but I'm still feeling my way.
Footnotes
-
From Liu Bei's deathbed edict to his son Liu Shan, which reads "Do not do evil because it is small, and do not fail to do good because it is small." See Pei Songzhi's annotation to the "Biography of the First Ruler" in the Book of Shu of Records of the Three Kingdoms, citing the Collected Works of Zhuge Liang. ↩
-
Lu Xun's original wording is "When the brave are angry, they draw the blade against the stronger; when cowards are angry, they draw the blade against the weaker," from "Random Thoughts" in Huagai Ji (1925). ↩
-
"Mr. De" (Democracy) and "Mr. Sai" (Science) were put forward by Chen Duxiu during the New Culture Movement, most fully in "A Reply to the Charges Against This Magazine," published in New Youth in 1919. ↩
-
Both are folk sayings from Zengguang Xianwen, an old primer for children. The more common form of the first is "Time flies like an arrow, the days and months like a weaver's shuttle." "Flowers bloom again, but no one is young twice" also appears in Yuan dynasty drama (Guan Hanqing's The Injustice to Dou E, among others). ↩
-
This is a saying that circulates widely online and is often attributed to Camus, George Bernard Shaw and others, but none of these has a verifiable source and the author is unknown. Best treated as a maxim of doubtful origin. ↩
-
Poor Charlie's Almanack is a collection of the sayings and speeches of Charlie Munger, Warren Buffett's longtime partner and vice chairman of Berkshire Hathaway, edited by Peter Kaufman. ↩
-
Geoffrey Hinton shared the Turing Award with Yoshua Bengio and Yann LeCun in 2018, and in 2024 shared the Nobel Prize in Physics with John Hopfield for foundational contributions to artificial neural networks and machine learning. ↩
-
George Boole (1815–1864) was an English mathematician and logician who founded Boolean algebra, laying the groundwork for the binary logic of modern digital circuits and computers. He was indeed Hinton's great-great-grandfather. ↩
-
Sir George Everest (1790–1866) was a British surveyor and Surveyor General of India who led the Great Trigonometrical Survey across the Indian subcontinent. Mount Everest is named after him. To be precise, his relation to Hinton is more indirect: Hinton's great-great-grandmother, Mary Everest Boole (George Boole's wife), was George Everest's niece. So Everest wasn't Hinton's "great-grandfather" but an uncle of his great-great-grandmother. ↩
-
A quotation from the book. For the original source and speaker, refer to Google's AI Brain: Hassabis itself. ↩