Researcher in Causal Inference
我最近在听一个抖音上的AI音乐账号。它做R&B翻唱，做得特别好——节奏处理、转音、和弦行进，技术上挑不出毛病。巧的是，陶喆最近也出了一张R&B翻唱专辑。如果你不告诉我哪个是AI做的，我可能会选AI的那首，然后说：”陶喆这张专辑做得真好。”
I’ve been listening to an AI music account on Douyin recently. It does R&B covers—really well. The rhythmic treatment, vocal runs, chord progressions—technically flawless. Coincidentally, David Tao recently released an R&B cover album. If you didn’t tell me which was AI-made, I might pick the AI track and say: “David Tao really outdid himself on this album.”
但一旦你告诉我那是AI做的，一些东西就变了。我开始注意到风格的统一化——每首曲子之间确实有细微差异，技巧也不错，但我就是没有耐心听下去了。
But the moment you tell me it’s AI-made, something shifts. I start noticing a uniformity of style—there are subtle differences between tracks, the technique is solid, but I just lose patience.
这个心理反应让我想了很久。
This psychological reaction stayed with me for a while.
第一反应是：也许我只是同情陶喆，觉得”人类要被AI取代了好可怜”，所以主观上偏向真人。但我仔细想了想，不完全是这样。
My first thought: maybe I just feel sorry for David Tao—”poor human about to be replaced by AI”—so I’m subjectively biased toward the real person. But thinking more carefully, that’s not quite it.
让我失去耐心的不是”这是机器做的，所以不好”。而是：AI做的每一首歌，都是在”已知的好听”这个空间里做最优解。它的每一个决策都是概率上的安全选择。听三首觉得惊艳，听三十首就麻了——因为新鲜感耗尽了。
What makes me lose patience isn’t “it’s machine-made, therefore bad.” It’s that every AI-generated song is an optimal solution within the space of “known good sounds.” Every decision it makes is a probabilistically safe choice. Three tracks feel impressive; thirty tracks feel numbing—because the novelty is exhausted.
而陶喆——或者任何一个真正有意思的音乐人——给我的感觉是不一样的。不是因为他的技术更好（可能真没有），而是因为他的选择里有不安全的跳跃。
David Tao—or any genuinely interesting musician—feels different. Not because his technique is better (it might genuinely not be), but because his choices contain unsafe leaps.
我试着用一个概念来理解这种跳跃：创造性赌注。
I tried to understand this leap through a concept: the creative bet.
当年陶喆把台语和R&B混在一起做了《Susan说》。没有任何数据支持”台语+R&B=好听”这个假设。没有先例，没有市场验证，没有算法会推荐这个组合。这纯粹是一个人，基于自己在台北长大、听美国R&B、同时泡在闽南语文化里的生命经验，做出的一个赌注。
Years ago, David Tao mixed Taiwanese dialect with R&B in “Susan Said.” No data supported the hypothesis that “Taiwanese + R&B = good.” No precedent, no market validation, no algorithm would recommend this combination. It was purely one person, drawing on the life experience of growing up in Taipei, listening to American R&B, while steeped in Hokkien culture, making a bet.
周杰伦把中国风和嘻哈混在一起，也是同样的事情。事后来看这些都是”当然会成功”的，但事前没有任何”当然”可言。
Jay Chou mixing Chinese traditional elements with hip-hop was the same kind of thing. In hindsight these all seem “obviously destined to succeed,” but beforehand there was nothing “obvious” about it.
AI能不能做这种跳跃？理论上可以——你可以让它随机组合风格。但这里有一个AI目前解决不了的问题。
Can AI make this kind of leap? Theoretically yes—you could have it randomly combine styles. But there’s a problem AI currently can’t solve.
创造性赌注的第一次尝试，几乎一定是不完美的。这是关键。
The first attempt at a creative bet is almost certainly imperfect. This is key.
如果台语R&B的第一首歌反响平平，这意味着什么？两种可能：这个方向根本不行。或者：这个方向是对的，但你还没找到正确的打开方式。
If the first Taiwanese R&B song gets a lukewarm response, what does that mean? Two possibilities: this direction simply doesn’t work. Or: the direction is right, but you haven’t found the right way to open it.
人类创作者能做的，是在失败的反馈中区分这两种情况。也许是：这首歌没打动人，但我在现场唱的时候，副歌那里有几秒观众的表情变了。所以方向可能是对的，只是主歌的进入方式不对。台语的部分应该更晚出现、更自然地融入，而不是一上来就给。
What human creators can do is distinguish between these two cases within the feedback of failure. Maybe: this song didn’t move people, but when I sang it live, during the chorus, for a few seconds, the audience’s expressions changed. So the direction might be right, but the verse’s entry is wrong. The Taiwanese dialect part should appear later, blend in more naturally, instead of hitting immediately.
AI面对同样的反馈，处理方式是：台语+R&B → 负面反馈 → 降低这个组合的权重。它把反馈当作结果。而人把反馈当作线索。
Facing the same feedback, AI processes it as: Taiwanese + R&B → negative feedback → lower this combination’s weight. It treats feedback as a verdict. Humans treat feedback as a clue.
这种”在失败中读出线索而非结论”的能力，需要两个AI没有的东西：身处社会现场的感知力——不是数据层面的”好评率43%”，而是一张张脸、一个眼神、一次犹豫的鼓掌；以及在负面反馈中保持信念的能力——相信自己的直觉，即使短期数据说你错了。
This ability to “read clues rather than conclusions from failure” requires two things AI doesn’t have: the perceptive power of being physically present in social reality—not data-level “43% approval rate,” but face after face, a shift in someone’s eyes, a hesitant round of applause; and the ability to maintain conviction amid negative feedback—trusting your intuition even when short-term data says you’re wrong.
AI没有信念。它只有概率。
AI has no conviction. It only has probabilities.
最近有一个有趣的案例：在一档节目里，AI写的”秋元康风格”歌词在粉丝投票中比秋元康本人写的更受欢迎。
A recent interesting case: on a TV show, AI-written “Akimoto Yasushi-style” lyrics were preferred by fans over lyrics actually written by Akimoto Yasushi himself.
这说明什么？在秋元康已经确立的风格空间里，AI已经比他本人更”秋元康”了。AI学会了他的用词习惯、情感节奏、叙事结构，然后把这些特征推到了极致。就像AI做的寿司：如果”最好的寿司”被定义为小野二郎的风格，那AI在那个定义上可以做到超越小野二郎。
What does this mean? Within the style space Akimoto already established, AI is already more “Akimoto” than Akimoto himself. AI learned his word choices, emotional rhythms, narrative structures, then pushed these features to their extreme. Like AI-made sushi: if “the best sushi” is defined as Jiro Ono’s style, AI can surpass Jiro Ono within that definition.
但定义本身是人建立的。
But the definition itself was built by a human.
那秋元康该怎么办？他的核心能力从来不是”写好歌词”——那只是技术输出。他的核心能力是读懂日本社会在某个时刻的集体情绪，然后找到一个形式来承载它。AKB48的”可以去见面的偶像”概念，总选举制度，毕业制度——这些不是”写词”层面的事情，是社会情绪的翻译和文化系统的架构。
So what should Akimoto do? His core ability was never “writing good lyrics”—that’s just technical output. His core ability is reading the collective mood of Japanese society at a given moment, then finding a form to carry it. AKB48’s “idols you can meet” concept, the general election system, the graduation system—these aren’t “lyric-writing” level things. They’re translations of social mood and architectures of cultural systems.
AI在歌词层面打败了他，但AI不知道2025年的日本年轻人需要什么样的新的文化形式。这不是一个数据分析问题，是一个创造性赌注。
AI beat him at the lyric level, but AI doesn’t know what new cultural form Japan’s young people need in 2025. This isn’t a data analysis problem. It’s a creative bet.
围棋可能是AI”统治人类”最完整的已完成案例。AlphaGo在2016年击败李世石的时候，很多人说”围棋完了”。
Go is perhaps the most complete case of AI “dominating humans.” When AlphaGo defeated Lee Sedol in 2016, many declared “Go is finished.”
但围棋没有完。人们仍然看柯洁下棋，不是因为柯洁比AI强——所有人都知道他不是——而是因为看一个有限的、会疲劳的、有情感的存在在极端条件下做选择，这件事本身有价值。他的某一步棋是冒险还是求稳？他知不知道对手的意图？时间压力下他会不会崩？这些问题只在人类对局中才有意义。AI下棋没有勇气可言，因为它不会恐惧。
But Go didn’t end. People still watch Ke Jie play, not because Ke Jie is stronger than AI—everyone knows he isn’t—but because watching a finite, fatigue-prone, emotional being make choices under extreme conditions has value in itself. Was that move risky or conservative? Does he sense his opponent’s intention? Will he crack under time pressure? These questions only matter in human games. AI plays without courage, because it knows no fear.
这和人跑不过汽车是一回事。没有人会办一场”人类 vs 汽车”的100米比赛，但奥运会100米决赛依然是全世界收视最高的体育赛事之一。赛道转移了：不再是”谁最强”，而是”人能做到什么”。
This is the same as humans not outrunning cars. Nobody would stage a “human vs. car” 100-meter race, but the Olympic 100-meter final remains one of the world’s most-watched sporting events. The track has shifted: no longer “who’s strongest,” but “what can humans achieve?”
电子竞技可能比围棋更有韧性。游戏版本不断更新，规则不断变化，而且是不完全信息博弈——你看不到对手在做什么。这里面有一种创造性的诡计：故意做一个看起来很蠢的操作来诱骗对手，需要理解对方作为一个人会怎么反应。这种心理博弈是人对人赛事独有的魅力。
Esports may be even more resilient than Go. Game versions constantly update, rules keep changing, and it’s an incomplete information game—you can’t see what your opponent is doing. There’s a kind of creative trickery: deliberately making a seemingly stupid play to bait your opponent, requiring understanding of how another person would react. This psychological game is the unique charm of human-vs-human competition.
不是所有领域都有”功能转变”的优雅叙事。有些领域，AI的作用就是暴露谁一直在裸泳。
Not every field has an elegant “functional transformation” narrative. In some fields, AI simply exposes who’s been swimming naked.
我在哈佛医学院见过很多这样的情况：研究助理跑统计模型，但不理解模型的假设和原理。有了数据，只要你能提出问题，GPT就能做出来——根本不需要那些不求甚解的中间人。
I’ve seen plenty of this at Harvard Medical School: research assistants running statistical models without understanding the models’ assumptions or principles. With data, as long as you can formulate the question, GPT can do it—there’s no need for those middlemen who never sought deeper understanding.
这里AI的角色不是推动什么优雅的转变，就是淘汰。把那些本来就不应该存在于一个健康学术体系里的工作清掉。
Here AI’s role isn’t driving some elegant transformation. It’s simply elimination. Clearing away work that shouldn’t have existed in a healthy academic system in the first place.
但这里有一个微妙的变化：“求甚解”的标准本身在提高。 以前你会跑Cox回归就算入门了。以后入门可能意味着：你能告诉我，为什么在这个具体的临床问题里，proportional hazards假设是不合理的？应该用什么替代？什么时候average hazard比restricted mean survival time更适合作为终点？这些不是计算问题，是判断问题——和寿司师傅、音乐人、棋手面对的，本质上是同一类问题。
But there’s a subtle shift: the standard for “seeking deeper understanding” itself is rising. Before, knowing how to run a Cox regression counted as entry-level. In the future, entry-level might mean: can you tell me why, in this specific clinical problem, the proportional hazards assumption is unreasonable? What should replace it? When is average hazard more appropriate than restricted mean survival time as an endpoint? These aren’t computation problems—they’re judgment problems. Fundamentally the same kind of problem facing sushi chefs, musicians, and Go players.
回到最初的问题。很多科幻作品预言AI会”统治人类”——天网觉醒、虚拟现实囚笼、中央网络独裁。
Back to the original question. Many sci-fi works predicted AI would “dominate humanity”—Skynet awakening, virtual reality prisons, central network dictatorships.
实际发生了什么？书店没有消亡——美国独立书店数量在过去五年增长了70%，TikTok上的#BookTok反而把人引回了线下书店。黑胶唱片没有消亡——2024年美国黑胶销售额达到14亿美元，而且50%的买家甚至没有唱片机。这些”应该被数字化消灭”的东西不但活着，还以一种新的形态繁荣了。
What actually happened? Bookstores didn’t die—the number of independent bookstores in the US grew 70% over the past five years, and TikTok’s #BookTok actually drove people back to physical stores. Vinyl records didn’t die—US vinyl sales reached $1.4 billion in 2024, and 50% of buyers don’t even own a turntable. These things “destined to be killed by digitization” are not only alive but thriving in new forms.
科幻弄错的地方在于：它们预测了戏剧性的、突然的、外部强加的灾难——网络觉醒、集体奴役。而实际发生的是渐进的、可以调整的、有时候甚至是反直觉的共生。
Where sci-fi went wrong: it predicted dramatic, sudden, externally imposed catastrophes—network awakening, collective enslavement. What actually happened is gradual, adjustable, sometimes even counterintuitive coexistence.
把这些放在一起看：
Putting these together:
自动点餐机取代了收银员的功能，但酒吧的bartender——那个跟你聊天、帮你调整心情、喝多了劝你回家的人——反而更有价值了。AI做的R&B翻唱在技术上无可挑剔，但陶喆的价值在于他能做出AI从概率上不会探索的方向。围棋AI比人类强得多，但柯洁的对局仍然有人看，因为人在极端条件下的选择本身就是一种表演。研究助理的机械操作被GPT取代了，但”在这个临床问题里该用什么统计方法”这种判断反而更重要了。
Self-ordering machines replaced cashiers’ functions, but the bartender—who chats with you, adjusts your mood, tells you to go home when you’ve had too much—becomes even more valuable. AI R&B covers are technically impeccable, but David Tao’s value lies in his ability to explore directions AI wouldn’t probabilistically attempt. Go AI is far stronger than humans, but Ke Jie’s games still draw audiences, because human choices under extreme conditions are themselves a performance. Research assistants’ mechanical operations got replaced by GPT, but the judgment of “what statistical method should we use for this clinical question” becomes even more important.
AI消灭的都是可以被还原为算法的人类活动。留下来的都是需要一个具体的人、在具体的处境中、做出带有个人判断的选择的活动。
What AI eliminates is human activity that can be reduced to algorithms. What remains is activity that requires a specific person, in a specific situation, making choices that carry personal judgment.
但这个模式有一个不太舒服的推论：大多数人大多数时候做的事情，其实是可以被还原为算法的。
But this pattern has an uncomfortable corollary: most of what most people do most of the time can, in fact, be reduced to algorithms.
不是每个音乐人都是陶喆或周杰伦。不是每个厨师都能开创新风格。不是每个棋手都有让人着迷的棋风。不是每个研究者都在”求甚解”。中间层——有技术但没有强烈个人风格的人——他们原来靠”比普通人做得好”来生存。现在AI也比普通人做得好了。
Not every musician is David Tao or Jay Chou. Not every chef can pioneer a new style. Not every Go player has a captivating playing style. Not every researcher is “seeking deeper understanding.” The middle tier—people with technique but no strong personal style—used to survive by “being better than average.” Now AI is also better than average.
这就是书店的故事里那段2000-2015年的低谷期。调整过来不意味着没有代价。Borders倒闭了，大量独立书店关门了，很多人失去了工作。”书店最终复兴了”这句话对那些在低谷期失业的店员来说，可能不是什么安慰。
This is the 2000-2015 trough in the bookstore story. Adjusting doesn’t mean there’s no cost. Borders went bankrupt, masses of indie bookstores closed, many people lost their jobs. “Bookstores eventually revived” probably isn’t much comfort to the clerks who lost their jobs during the trough.
AI在任何”已定义的优秀”上都会超过人类。但”定义什么是优秀”这件事，到目前为止，仍然只有人能做。
AI will surpass humans in any “already-defined excellence.” But “defining what excellence is”—so far—only humans can do that.
当AI在旧定义里做到了天花板，人的唯一出路就是创造一个新天花板。然后AI会追上来，人再创造下一个。这个循环，也许就是人和AI长期共存的基本节奏。
When AI hits the ceiling within an old definition, the only way out for humans is to create a new ceiling. Then AI catches up, and humans create the next one. This cycle may be the fundamental rhythm of long-term human-AI coexistence.
这意味着创造性赌注不再是少数天才的特权，而是所有人的生存技能。不是每个人都需要成为周杰伦，但每个人都需要问自己：我做的事情里面，有多少是AI能直接替代的？剩下的那些——那些需要我在具体处境中做出不安全的判断的部分——那才是我存在的理由。
This means creative bets are no longer the privilege of rare geniuses, but a survival skill for everyone. Not everyone needs to become Jay Chou, but everyone needs to ask: how much of what I do can AI directly replace? What remains—those parts requiring me to make unsafe judgments in specific situations—that’s the reason for my existence.
而且，和我在上一篇文章里写过的一样：学了之后，继续问自己——这个判断，是我真的做出来的，还是我被框架说服了？在AI时代，这个问题变得更尖锐：这个选择，是我真的在创造，还是我只是在概率上的安全区里做优化？
And, as I wrote in my previous post: after learning, keep asking yourself—did I truly make this judgment, or was I convinced by a framework? In the AI era, this question becomes sharper: am I truly creating with this choice, or just optimizing within a probabilistically safe zone?
这个问题，值得一直问下去。
This question is worth asking forever.
写这篇文章时，AI翻唱账号仍然在日更。陶喆的新专辑评价褒贬不一。我还是会两个都听。但我注意到：AI的歌我是在跑步时当背景音乐听的，陶喆的歌我是坐下来、戴上耳机、认真听的。也许这个区别本身，就已经是某种答案了。
Writing this, the AI cover account still posts daily. David Tao’s new album reviews are mixed. I still listen to both. But I notice: I listen to the AI tracks as background music while running; David Tao’s tracks I listen to sitting down, headphones on, paying attention. Perhaps this difference itself is already some kind of answer.