Are We All Artists Now?

我们现在都是艺术家了吗？

Here’s something I’ve been thinking about for a while now — a mental model I keep coming back to whenever I try to understand how anyone actually gets good at something. I think mastering a skill happens in two phases, and I’ve always just thought of them as the science of something and the art of something. 这是我思考已久的一个问题——每当我试图理解一个人如何真正精通某项技能时，我总会回到这个思维模型。我认为掌握一项技能分为两个阶段，我一直将其简单地称为“科学”阶段和“艺术”阶段。

The science is where you’re learning something — a craft, a trade, a discipline, a sport, whatever — and you start by inheriting the rules. Somebody already went through the trial and error. They figured out what works, wrote it down, handed it to you so you don’t have to relearn the whole thing from scratch. You follow it, you absorb the pattern, you do it the way it’s supposed to be done, because that’s what’s been laid out for you. And honestly, that’s a fine way to get competent fast. Nothing wrong with it. “科学”阶段是你学习某样东西的过程——无论是手艺、行业、学科还是运动——你从继承规则开始。前人已经历过试错，他们找出了行之有效的方法，将其记录下来并传授给你，这样你就不必从零开始重新摸索。你遵循规则，吸收模式，按照既定的方式去做，因为这就是为你铺好的路。老实说，这是快速获得胜任能力的好方法，这并没有什么错。

Then, at some point, if you keep going long enough, something shifts. You stop asking “what’s the rule” and start asking “why does this rule even exist, and does it still apply here?” You start bending things, breaking things, going back and reexamining stuff you used to just take for granted. Not randomly — that’s important — but with this kind of earned judgment about which parts of the rulebook are actually load-bearing and which parts are just habit nobody’s bothered to question in a while. That’s the art. 然后，在某个时刻，如果你坚持得足够久，情况就会发生转变。你不再问“规则是什么”，而是开始问“为什么会有这条规则，它在这里还适用吗？”你开始尝试变通、打破常规，回头重新审视那些你曾经视为理所当然的事物。这并非随机为之——这一点很重要——而是基于一种习得的判断力，去辨别规则手册中哪些部分是真正的“承重墙”，而哪些部分只是无人质疑的陈规陋习。这就是“艺术”。

Turns out I didn’t come up with this myself, which, fine. Once I started poking at it I found it maps pretty closely onto existing thinking about skill acquisition — the Dreyfus model of skill mastery describes something similar, moving from novice rule-following (the science) toward this expert-level improvisation past the rules (the art). So I’m not being nearly as clever as I thought I was. But I’m going to keep using my own language for it anyway, because I think it’s the more useful frame for what I actually want to talk about here: “art” and “science” as names for two tiers of any skill. Not fine art. Not literal science. Just rule-following versus rule-transcending. 事实证明，这并非我的原创，但这没关系。当我深入研究时，发现它与现有的技能习得理论非常吻合——德雷福斯技能习得模型（Dreyfus model）描述了类似的过程，即从新手遵循规则（科学）向专家级的超越规则即兴发挥（艺术）过渡。所以我并没有自己想象中那么聪明。但我还是会继续使用我自己的术语，因为我认为对于我接下来要讨论的内容，这是一个更有用的框架：“艺术”和“科学”作为任何技能的两个层级。不是指纯艺术，也不是指字面意义上的科学，仅仅是指“遵循规则”与“超越规则”。

So that’s the mental model. What I want to spend this piece doing is pushing on it a bit — specifically, why I think generative AI is stuck pretty firmly in the science tier, why that might be making the art tier more valuable rather than less, and where I think that actually leaves people. 这就是那个思维模型。我想在这篇文章中进一步探讨它——具体来说，为什么我认为生成式 AI 牢牢地困在“科学”层级，为什么这反而可能让“艺术”层级变得更有价值，以及我认为这会将人类置于何种境地。

The tell

破绽

Here’s something I’d bet almost everyone reading this has felt: you read something, or look at something, and some part of you just goes — yep, that’s AI. You can’t always say exactly why. It’s a little too smooth. A little too resolved. Technically fine, but somehow thin. I don’t think that reaction is some mystical sixth sense. I actually think it might be a pretty accurate read of what’s happening underneath. 我敢打赌，几乎每个读到这里的人都有过这种感觉：当你阅读或观看某样东西时，内心深处会产生一种直觉——“没错，这就是 AI 做的”。你未必能说出确切原因。它太圆滑了，太完美了。技术上没问题，但总觉得缺乏深度。我不认为这种反应是什么神秘的第六感，我实际上认为这可能是对底层逻辑相当准确的解读。

Generative models are trained to predict the most statistically likely continuation given everything they’ve seen. That process pulls outputs toward the center of the training distribution — smoothed, average, safe. Human work, even mediocre human work, tends to carry more irregularity in it: specific choices, small “wrong” decisions that somehow turn out right, texture that comes from one particular person’s particular constraints and mood and history on that day. If AI output tends to read like an average and human output tends to read like a point of view, maybe that difference is genuinely detectable — not just something we’re imagining. 生成式模型通过训练，旨在根据已有的所有数据预测统计学上最可能的后续内容。这个过程将输出结果拉向训练分布的中心——平滑、平均、安全。人类的作品，即使是平庸之作，往往也带有更多的不规则性：特定的选择、一些看似“错误”却歪打正着的决定，以及源于某个人在特定当天的特定限制、情绪和经历所带来的质感。如果 AI 的输出倾向于“平均值”，而人类的输出倾向于“观点”，那么这种差异或许确实是可以察觉的——而不仅仅是我们的想象。

“But models keep getting better” — sure, but at what, exactly?

“但模型一直在进步”——没错，但究竟是在哪方面进步？

The obvious pushback here: models keep improving, constantly. Doesn’t that just erode this whole argument over time? I don’t think it does, actually, and I think it’s worth walking through why models get better, because as far as I can tell, none of the reasons actually touch the thing I’m talking about: 这里显而易见的反对意见是：模型在不断进步。这难道不会随着时间的推移削弱我的论点吗？事实上，我不这么认为。我认为有必要梳理一下模型为何会进步，因为据我所知，这些原因都没有触及我所讨论的核心：

• Scale — more parameters, more data. This just gives a model a bigger map of the territory it was trained on, and more resolution once it’s there. 规模——更多的参数，更多的数据。这只是给了模型一张更大的训练领域地图，并在该领域内提供了更高的分辨率。
• Architecture and training method improvements — these make navigating that map more efficient, less wasted effort on noise. 架构和训练方法的改进——这些改进使得导航地图的效率更高，减少了在噪音上的无效投入。
• Post-training and fine-tuning — this one’s closer, honestly. It’s sculpting which parts of the existing map the model gets pulled toward by default. That’s a real form of discernment. But it’s discernment done by humans, in advance, on the model — not something the model works out itself, in the moment, for a case nobody already judged. 训练后处理与微调——老实说，这更接近本质。它是在塑造模型默认倾向于地图的哪些部分。这确实是一种辨别力，但这是人类预先对模型进行的辨别，而不是模型在当下、针对一个无人评判过的案例时自己得出的结论。
• Inference-time stuff — longer context, more reasoning steps. This just gives the model more time to search the map it already has before it answers. 推理时技术——更长的上下文，更多的推理步骤。这只是给了模型在回答前更多的时间去搜索它已有的地图。

Bigger map. Better-curated map. More time to search the map. That’s basically every lever I can name, and every one of them is about navigating the existing space better. None of them is “redraw the map because this part of it is wrong.” Which isn’t a knock on the technology, it’s just not what any of these methods are built to do. And honestly, I think that absence is the actual evidence here — not just something I’m asserting because it happens to support what I already wanted to believe. 更大的地图，更精细的地图，更多搜索地图的时间。这就是我能列出的所有杠杆，它们每一个都是为了更好地在现有空间内导航。没有一个是“重绘地图，因为这部分错了”。这并不是对技术的贬低，只是这些方法的设计初衷并非如此。老实说，我认为这种缺失正是关键证据所在——这不仅仅是我为了支持自己的观点而做出的断言。

What’s actually missing

真正缺失的是什么

Here’s where I want to push back on my own instinct a bit. The naive version of “art versus science” makes it sound like creativity is just rule-breaking — like the more you depart from convention, the more creative you are. But I don’t think that’s really true. Randomly ignoring constraints doesn’t produce insight, it produces noise. A musician who “breaks” a harmonic rule isn’t just playing random notes — they’re breaking exactly the one rule that’s stopped being load-bearing, while keeping everything else intact enough that the thing still holds together as music. So maybe the sharper version of the claim is this: creativity isn’t measured by distance from the rules, it’s measured by how precisely you can tell which constraints are still structural and which ones are just calcified habit. That’s a much harder thing to do than either blindly obeying or blindly departing, and I think that’s the actual content of “transcending the rules”. 在这里，我想对自己的一些直觉进行反思。“艺术与科学”的简单版本听起来像是创造力仅仅是打破规则——仿佛你越背离传统，就越有创造力。但我认为事实并非如此。随机无视约束并不能产生洞察力，只会产生噪音。一位“打破”和声规则的音乐家并不是在随意弹奏音符——他们打破的恰恰是那条不再具有承重意义的规则，同时保持其他部分完整，使作品依然作为音乐而成立。所以，这个论点更精准的表述可能是：创造力不是由与规则的距离来衡量的，而是由你能多精确地辨别哪些约束仍是结构性的，哪些只是僵化的习惯来衡量的。这比盲目服从或盲目背离要困难得多，我认为这才是“超越规则”的真正内涵。