Last week in Beijing, Pangzi dragged me to an AI meetup. The room was wall-to-wall young people, and the two of us "old-timers" stuck out badly. Once the talk got going, one young man said: "So AVL Code is you guys? How hard is this AI-coding game — no way you pull it off without a thousand GPUs." Before either of us could answer, it dawned on him: "I get it — you must be that post-70s bunch who were started young on this stuff, the ones with a soft spot for it." Pangzi and I traded looks — in these kids' eyes, does every decision we post-70s old codgers make come down to "sentiment"? Then again, his read was interesting. But do I even count as having been started young?
I first touched a computer in junior high, but I didn't really start playing until the computer room at Jixi No.1 High School, where a couple dozen Laser machines glowed green. Pangzi truly was started young — from the computer room at Changchun No.1 Experimental Primary School to formal mechanical-typewriter touch-typing drills at the Normal University's affiliated high school; and yet in high school he was still on Basic, while I was already writing Apple assembly.
1. When Resources Went From Scarce to Abundant
Scarcity, and everyone striving on their own — that was the true backdrop of that era. Pangzi and I really were beneficiaries of the "computer literacy must start with the children" wave. The country wasn't rich then, but it still worked hard to put computers into some schools — though the mainstay model was actually the China Education Computer; the ones that could run to an Apple were the vanishingly few "provincial key schools." Scarcity makes whoever touches the resource cherish it to death. The administrator of my high school's computer room happened to be my aunt's eldest son, which put me in the front row: during the day there were classes and you couldn't grab a seat even after they ended; but every evening, once the school had emptied out, I could let myself in with his key — unlimited time, unlimited joy — and start to "play" with the machines.
Sometimes I think: when resources are scarce, people cherish them more, and it patterns the choices of a life. Among my peers, some became rich, some became officials, some have already retired or even become grandfathers — and Pangzi and I are still writing code. That is the result of being patterned by that era, we who touched computers earlier than most of our peers. Today a computer is no longer an unimaginable luxury for a Chinese household; there are still families who can't afford one, but even a single phone connects you to the world. Maybe asking Doubao or DeepSeek is already the most effective way for a kid in the mountains to make up for a shortage of teachers. Of course, every era has its own scarce resource — for us, it's compute. Our public copy says we trained the model on just a thousandth of the data, but anyone paying attention can guess the truth: our compute can only support that much. A 1.8B binary model is the ceiling of what we can do at our current compute. But there is so much to do in the AI era — the other day I saw Beijing roll out a policy supporting one-person companies (OPCs). The truth is, once you start moving, anyone who wants to launch a dream in the digital world can always find a place and the resources for it.
2. When Knowledge Is Within Reach
The computer knowledge of that "started-young" era was scarcer and more of a luxury still. By the time I reached university, the computer room was all x86 clones, and programming naturally shifted to Turbo C. The library at the Qiqihar Institute of Light Industry held next to nothing (it was really just a reading room). I remember one boxed three-volume set of Turbo C — the exact titles are a blur now, genuine or pirated I couldn't tell — but it was the only copy in the whole place. My roommate and I were both into programming, so we founded the school's student computer association together, one of us chairman and the other vice-chairman; and to keep that set close for the long haul, we pulled in a third person and borrowed and returned it in relay — since each volume had a fixed loan period, we would arrange to check out the next one the moment we returned the last, keeping the set circulating in our dorm until we had read it to pieces. It may look laughable to young people today, who might even sneer at how shabby we were; but in that era, for someone who genuinely wanted to learn programming, reference material was exactly that scarce. Books were not just the sacred vessel of knowledge — they could even decide which road you took.
Back then, Pangzi at the Harbin Institute of Technology (HIT) was using Quick C, which almost nobody had heard of, rather than the Turbo C everyone knew. The reason was pure chance: passing the campus press one day, he found a pile of books on clearance and picked up a cheap four-volume Quick C set from Hope Publishing. Afterward he could not find the software anywhere; in the end it was a high-school classmate of mine who tracked it down for him — they were both running HIT's student computer association, and that fellow asked around until he dug up Quick C version 2.5. Looking back, the scarcity of knowledge breeds contingency, and I can't help imagining: if what Pangzi bought that day had not been Quick C but, say, "The Thirty-Six Stratagems of Business Warfare," would Antiy's road have been smoother? But then, maybe we would never have met.
Today, every question gets a thorough, lucid answer from a large model; knowledge is there for the taking, explained far more clearly than any textbook, and patient enough to be asked a hundred times over. In an age when knowledge can be drawn on at almost zero cost, at will, mastering knowledge is no longer the prerequisite for most work — this is the age of the genuine doer, the age of "learning by doing." It is also why AVL Code's very first blog said we want to be Don Quixote, not Hamlet.
3. What Muscle Memory Means
The other day, Kimi K3 launched, paying tribute to Windows XP by recreating it as a web page, and I sent it to Pangzi. He said: I'll make a "tribute" to you, too. So with AVL Code + Kimi he generated a set of nostalgic "tribute" pages — web operating demos of the PDP-11, the China Education Computer, the Apple, the Sun-3/280, and a PC clone (DOS) (at cs.avlcode.cn). He told A-Guan: "This series is a tribute to Lao Zhang, and this Apple is the very one from his high school." But I know what he was really up to — showing off that he had used more old machines than I had, especially HIT's Sun-3/280. In truth, he used the Sun because the Sun room was cheap: you could not game on it, so few people went — scarcity again. And the PDP-11 in the tribute he never touched himself; as a kid he simply found his father's PDP-11 assembly listings and stared at them like holy scripture for a few days.
I even won a bet over it. In 2001, when we were developing together at a civilian-products institute on some base in Wuhan, I bet the others that Pangzi was bound to mispronounce the English word "address," reading it as add-ress. Sure enough, I won. That was my confident analysis — helped along, of course, by his already terrible English: a kid who, before ever studying English, first learned that ADD was the "addition" instruction will, once he finally does study English, instinctively break "address" into ADD-Ress. That is a kind of "muscle" memory. But everyone's memory differs: I trust the flexibility of code more, because there is boundless possibility in it, while Pangzi leans harder on the rigor and formalism of data structures. Yet all memory has to be adapted to the future — adapted to this era in which natural language can be executed.
Plenty of people ask me: isn't the programmer the profession most easily replaced? Now that the large models are here, aren't you dejected? The program you used to write in a month — can't AI now finish it in a few hours? Of course it can, obviously. But honestly we are delighted — after all, there is no psychological burden in cursing out an AI at midnight.
So what exactly is the muscle memory that "started-young" era left us? Perhaps it is the limit of a single 360KB low-density floppy, or the constraint of 640KB of memory. To save memory, when we built Chinese-language software we would even skip the full font, prying out just the bitmaps of the Chinese characters the menus actually used; and for any program of real size, you always had to obsess over how to hang the data off into expanded memory.
These, perhaps, are the true muscle memory of our "started-young generation": you have a little bit of resource, but you know resources are forever a luxury. So we never see a shortage of resources as an obstacle, nor a blind spot in our knowledge as a limit — "as long as your resolve doesn't slip, there are always more ways than there are difficulties" — because this is exactly how we came up. So while everyone talks ten-thousand- and hundred-thousand-GPU clusters, it does not stop us from training a domain-specific Landi binary model on just twenty cards and a thousandth of the data; nor does it keep us from carrying AVL Code forward with at least one update a day.
We slap the screen and call to the Landi model: giddy up, little donkey — quicken your step!
The nostalgic tribute pages mentioned here live in The Time Machine Room, made with respect by the Antiy AVL Code team.
AVL Code — the AVL security engine, with intelligence at your side. From the Antiy Landi team.