A puzzle一道谜题
2026-06-30
An old verdict
Back when I was doing my master's, one of my professors told me neural networks were "outdated." We were talking about them, and he waved the whole idea off. Sure, he said, a neural net can fit anything, predict anything, near-perfectly. But you can't explain how. Nobody can tell you what actually happens inside the thing. And if you can't explain it, you can't write a good paper about it. So: a dead end.
And yet, here we are.
Watching the line move
I've followed the GPT releases closely — 1, 2, 3, all the way to 5 — and I kept grading each one against the only ruler I really trust: how a software engineer grows.
The early models were junior. Then junior turned into mid-level. Mid turned into senior. And now, with a bit of guidance, I'll watch one execute at something close to principal level. Every time I drew a line and said that's the ceiling, the next model — Fable, lately — stepped over it like the line wasn't there.
I've gone from "GenAI isn't really possible" to "...maybe." I used to think I understood the whole core of it: internet-scale data, fast GPUs, a big model predicting the next token. I thought I knew its limits. I don't say that anymore. I keep telling people it's slowing down — and I keep not seeing it actually slow down.
But here's the part I still hold onto. At its core, it's an aggregation of human knowledge. I haven't seen any real evidence it can step outside that boundary and make something genuinely new.
The same puzzle, in my own house
Which brings me back to Homie, and especially the agent fleet.
The puzzle is this: an LLM mimics human knowledge and predicts the best next answer or action. The fleet does the same thing, one layer up. And because it does, it runs into the exact problems a human team runs into.
Multiple tasks in flight. Builders and reviewers working in parallel. And at the end of the day, everyone congests on the same chokepoint — merging PRs, fighting the same merge conflicts a real developer curses at. Coordination breaks down. Each builder, each architect, is heads-down on its own task, under-communicating, occasionally stepping right on another's toes.
I built a fleet of agents and it inherited the headaches of a human org. Same puzzle, smaller machine.
Does it have to be shaped like us?
It feels like building a humanoid robot.
One argument for the human shape is simple: every facility, every tool, every doorway and staircase was built for human bodies. So a human-shaped bot reuses all of it. That's a real advantage.
But step back and ask the task-centered question, and the human shape stops looking obvious. If you only care about the task, two legs and ten fingers might be a strange, inefficient way to get it done.
A hundred steps back
Now go further back. The way I interact with a coding agent — it mimics human behavior, learned from a giant pile of human-written code. And every modern programming language was designed for humans to read and write, not for machines.
So is this even the right way to do it? A machine should speak machine language, shouldn't it? Let it talk in whatever form is actually native to it.
And then the counter-argument arrives, quiet and a little annoying: at the end of the day, we're still building all of this for humans.
So which is it?
I don't have the answer. Just a thought. A puzzle. Mostly me, talking to myself.
一句旧判词
读硕士那会儿,有位教授跟我说,神经网络已经"过时"了。当时我们正聊到这个话题,他一摆手,把整套想法都否了。他说,没错,神经网络什么都能拟合、什么都能预测,几乎分毫不差。可你说不清它是怎么做到的。没人讲得明白那玩意儿内部到底发生了什么。既然解释不了,你也就写不出一篇像样的论文。所以——死路一条。
可如今呢,我们走到了这一步。
看着那条线一点点挪动
我一直在密切关注 GPT 的每一代——1、2、3,一路到 5——而我衡量它们,用的始终是我唯一真正信得过的那把尺子:一个软件工程师是怎么成长起来的。
早期的模型是初级水平。后来初级长成了中级,中级又熬成了资深。到了现在,只要稍加引导,我能眼看着它干出接近首席工程师的活儿。每一次我画下一条线,说这就是天花板了,下一代模型——最近是 Fable——都像那条线根本不存在似的,一步跨了过去。
我从"GenAI 其实没戏"变成了"……也许有戏"。我曾以为自己摸透了它的核心:互联网规模的数据、飞快的 GPU、一个庞大的模型在预测下一个 token。我以为我清楚它的极限在哪。这话我现在不说了。我一遍遍跟人讲它在放缓——可我一遍遍没看见它真的慢下来。
但有一点我至今仍然认定。说到底,它是人类知识的聚合。我还没看到任何真凭实据,能证明它可以跨出那道边界,造出真正崭新的东西。
同一道难题,搬进了我自家屋里
这就把我带回了 Homie,尤其是那支 agent 队伍。
难题是这样的:一个 LLM 模仿人类的知识,预测最优的下一句回答、下一个动作。而这支队伍做的是同一件事,只是高了一层。也正因为如此,它撞上的,恰恰是一个人类团队会撞上的那些问题。
多个任务同时在跑。builder 和 reviewer 并行干活。可到头来,所有人都堵在同一个瓶颈上——合并 PR,跟真实开发者咒骂的那种 merge conflict 死磕。协作开始崩盘。每个 builder、每个 architect,都埋头扎在自己那摊事里,彼此沟通不够,时不时还正好踩到对方脚上。
我搭起了一支 agent 队伍,它却把一个人类组织的头疼病全继承了过来。同一道难题,只是换了台更小的机器。
它非得长成我们这副模样吗?
这感觉就像在造一台人形机器人。
支持人形的一个理由很简单:每一处设施、每一件工具、每一道门、每一段楼梯,都是按人的身体造的。所以一台人形机器人能把这一切照单复用。这是实打实的好处。
可退一步,换个以任务为中心的问法,人形这件事就没那么理所当然了。如果你只在乎把活干成,那两条腿加十根手指,也许是一种奇怪又低效的干法。
退后一百步
再往回退。我跟编程 agent 打交道的方式——它模仿的是人类行为,学自一大堆人写的代码。而每一门现代编程语言,都是为人的读写设计的,不是为机器。
那么,这条路本身就走对了吗?机器理应说机器的语言,不是吗?让它用真正属于它自己的那种形式去表达。
接着,那个反驳声悄悄冒了出来,带着点恼人的意味:说到底,我们造这一切,终归还是为了人。
那到底哪个对?
我没有答案。只有一个念头。一道难题。多半,就是我在自言自语罢了。