Where value goes when building is free 当“造东西”变免费，价值去了哪
AI is pushing the cost of writing software toward zero. It’s tempting to read that as “developers are finished”. I read it the way an economist would: when the cost of one input collapses, value doesn’t disappear. It moves to whatever has just become the bottleneck. The interesting question isn’t whether building gets cheap, but what becomes scarce once it does. AI 正在把写软件的成本压向零。很容易把这理解成“程序员要完了”。我更愿意用经济学的眼光看：一项投入的成本一旦大幅下降，价值不会消失，而是转移到新的瓶颈上。值得问的不是“做软件会不会变便宜”，而是它变便宜之后，什么会变得稀缺。
2026-07-16 · updated 2026-08-08 · first published on boheastill.com最初发表于 boheastill.com
When a cost goes to zero, value moves to the complements
This is an old pattern. When one part of a system gets cheap, demand and value shift to the parts that now constrain it. When building is plentiful, three things become scarce. Judgment: of everything that can be built, what should be, and for whom. Verification: once generating an answer is free, knowing whether it’s correct is the hard part. And the last mile: turning a capability into something that runs, and is trusted, in a specific and messy real situation. Cheaper code isn’t the end of the story. Value is moving out of output and into judgment and verification.
The thing AI structurally cannot do: be accountable
One complement matters more than the rest, and a model can’t supply it however good it gets. A machine that moves, and can damage a part or hurt a person, needs someone who can be held responsible for it: a name on the work, someone who carries the reputational and legal weight and can be looked in the eye when it matters. A model can’t be accountable. It can’t sign off, can’t be liable, and can’t stand behind a decision when a regulator, an insurer or an injured person asks who is answerable. The more capable the automation, the more that accountable person is worth.
Some areas commoditize more slowly, for structural reasons
Not all software loses value at the same rate. In safety-critical and regulated work, anywhere a mistake has physical or irreversible consequences, the things that slow projects down elsewhere become a moat: liability, regulation, insurance, and the hard requirement that a responsible person stand behind the system. That friction means prices fall slowly, because cheap and fast was never the whole requirement; provably safe, with someone answerable, was. The fields where verification and accountability dominate are, not by coincidence, the ones least threatened by cheap generation.
What this means for how I work
I'm not betting that I write code better than a model. That's a losing bet. I'm betting on the layer above the code: taking a powerful capability and being the accountable, trusted person to whom it's safe to hand a system with real consequences. That's why the way I work leans on things that hold their value when code is abundant (proof over claims, built-in verification, bounded consequences, honest scope), and why I'm drawn to problems where being right and being answerable matter more than being fast. When building is free, that's where the value goes.
成本归零时，价值迁移到互补品
这是个老规律。系统里的一部分变便宜了，需求和价值就会转到现在卡脖子的那些部分。当“做出来”不再稀缺，有三样东西会变得稀缺。判断：在所有能做的东西里，哪些真该做、为谁做。验证：生成答案几乎不花钱之后，判断它对不对才是难点。还有最后一公里：把一项能力变成在具体、复杂的真实环境里能跑、能被信任的东西。代码变便宜不是故事的结局，价值正在从“产出”转向“判断和验证”。
AI 结构上做不到的那件事：被问责
有一样东西比其他都重要，而且模型再强也给不了。一台会动、可能损坏工件或伤到人的机器，需要有人能为它承担责任：作品上有署名，承担声誉和法律上的分量，出了事能站出来面对。模型没法被追责。它不能签字确认，不能承担法律责任，当监管机构、保险公司或受伤的人问“谁来负责”时，它也站不出来。自动化越强，这个能负责的人就越有价值。
有些领域贬值得更慢，原因在结构上
不是所有软件都以同样的速度贬值。在安全关键和受监管的领域，也就是出错会造成物理伤害或无法挽回后果的地方，那些在别处拖慢项目的因素，反而成了护城河：法律责任、监管、保险，以及必须有一个负责人为系统担保的硬性要求。这些阻力让价格降得很慢，因为客户要的从来不只是又便宜又快，而是能证明安全、并且有人负责。验证和问责最重要的领域，恰好也是最不怕“生成变便宜”的领域，这不是巧合。
这对我怎么做事意味着什么
我不是在赌自己写代码比模型强，那注定会输。我赌的是代码之上的那一层：把一项强大的能力接过来，做那个可以被追责、被信任，让人放心把有实际后果的系统交给他的人。所以我的工作方式押在那些代码再多也不会贬值的东西上（用证据而不是说法、内建验证、后果有界、范围说实话），也因此我更喜欢那些“做对”和“有人负责”比“做得快”更重要的问题。做软件不再值钱的时候，价值就流到了这里。