正如当时所写的那样，这一切如今都已发生。不久前，我们曾指出，鉴于人工智能在代码分析方面表现出色，克隆大型应用程序将成为可能。然后我们会身处何地？就是这里，正是如此。一个勤奋的个人已经发布了针对模仿Claude的微软Office和Adobe Creative Suite程序的源代码和二进制文件，令人失望的是，他没有将后者的部分内容命名为Fauxtoshop。
As it was written, so it has come to pass. A while back, we suggested that given how good AI was at code analysis, it would become possible to clone large applications. And then where would we be? Here, that’s where. An industrious individual has released source code and binaries for Claude-copied Microsoft Office and Adobe Creative Suite programs, disappointingly not calling elements of the latter Fauxtoshop.
它们处于包含大量错误的不完整预alpha阶段，但具备强大的功能，这证明了当人工智能被赋予一项零创造力、穿着写有“这就是概念验证的样子”和“意料之外的后果”字样T恤的任务时，可以具有多么疯狂的生产力。他们的衣橱里或许还可以添上一首《Karma Police》。显而易见，微软和Adobe正面临着来自其竞争对手所描述的“洁净室重新实现”的某种竞争。
They’re in pre-alpha states of incomplete bugginess, but with substantial functionality, which is proof of how insanely productive AI can be given a task needing zero creativity and wearing "This is what a proof of concept looks like" and "Unintended Consquesences" T-shirts. They could also add Karma Police to their wardrobe. It’s clear that Microsoft and Adobe are facing some competition from what their rival describes as "clean-room reimplementations."
目前也不清楚是否有任何法律禁止这样做。你无法对功能申请版权，代码是用Rust语言生成的，并声称是在洁净室条件下完成的，虽然声称这些作品在很大程度上属于衍生作品可能会有一些依据，但这里存在一个大问题。那些依靠历史上最大规模知识产权盗窃来训练人工智能的公司，将需要无与伦比的法律技巧来声称不应允许人工智能玷污它们自己的知识产权。
It’s also unclear that there is any law against it. You can’t copyright functionality, the code has been generated in Rust with claimed clean-room conditions, and while there may be something in claiming that the works are substantially derivative, there’s a big problem here. Companies that sell AI trained on the biggest theft of IP in history are going to need unmatched legal dexterity to claim that AI shouldn’t be allowed to sully their own IP.
厚颜无耻过去曾被定义为一个因谋杀父母而受审的孩子，以自己是孤儿为由乞求宽恕。我们将不得不更新这个定义。代码克隆后的世界将会非常有趣。实际上，一切都将变成开源，这切断了那些曾假定用户会被锁定的巨头们的大量收入来源。从头构建替代方案的动力可能会消失，无论是对开源还是闭源而言。转向纯订阅制的云托管产品——在那里没有人能将他们盗取AI的提示词打在二进制文件上——似乎是一种显而易见的应对之策，但人工智能可以控制和抓取网页界面，而这正是它所需要的全部。这是一场人类无法获胜的军备竞赛。
Chutzpah used to be defined as a child on trial for murdering their parents begging for mercy on the grounds that they’re an orphan. We’ll have to update that. The post-codeclone world is going to be interesting. Everything becomes effectively open source, which closes down substantial revenue streams for big companies who had assumed lock-in. The motivation to build ab-initio alternatives may go away, in open source as well as closed. The push to subscription-only cloud-hosted products, where nobody can get their thieving AI prompts on binaries, seems an obvious counter move, but AI can control and scrape a web interface, which is all it needs. It's an arms race that won’t be won by humans.
代码克隆只是人工智能极具破坏性、以至于威胁到其创造者的最新案例。另一个例子则是微软与OpenAI在内部邮件中承认的厄运循环：通过抽走支付内容费用的广告收入，人工智能扼断了它赖以依赖的数据。这就像是一家把业务建立在他人知识产权之上的公司在进行森林砍伐。最大的潜在故障点是资金。鉴于北京高度重视将人工智能发展为全球影响力载体，在面对更便宜或免费的中国模型时，收入将如何维持是一个令人担忧的问题。
Code cloning is just the latest case of AI being so disruptive it threatens its creators. Another is the doom loop that Microsoft-OpenAI acknowledged in internal emails: by siphoning away the ad revenue that pays for content, AI chokes off the data it depends on. It's deforestation by companies that built their business on other people's IP. The biggest potential point of failure is the money. How revenues will hold up against cheaper or free Chinese models is a concern, given the high importance assigned by Beijing to develop AI as a global influence vector.
这些模型本身效率很高，部分原因是对高端芯片的出口管制以及旧数据中心出现某种程度的过剩。限制获取技术成为一种演进驱动因素，而对那些实施限制的人来说结果不佳，这已经不是第一次了。一个更大的问题是：资金将从何而来，以建设支撑人工智能用来证明其增长合理性的各项声明所需的础设施？Gartner表示，目前的投资回报尚不足以推动这种增长，而一个更引人入胜的SemiAnalysis播客则预测，到2030年总资本支出将达到11万亿美元，并质问这些资金将从何而来。
The models themselves are very efficient, in part because of export regulations on high-end silicon and something of a glut of older datacenters. This wouldn’t be the first time restricting access to technology is an evolutionary driver with a poor outcome for those doing the restricting. A much bigger issue is: where will the money come from to build the infrastructure that is needed to support the claims AI is using to justify its growth. Gartner says that there’s not enough return on investment to drive that growth, while a much more engaging SemiAnalysis podcast predicts total capex by 2030 of $11 trillion and asks where it’s going to come from.
如果你对理解人工智能感兴趣却又没有关注SemiAnalysis，那你就是对理解人工智能不感兴趣，虽然该机构一贯保持绝对的乐观，但也极其消息灵通。在这种情况下，人工智能的投资市场结构被剖析得淋漓尽致。这非常诡异。新云（Neocloud）由那些对资金有着贪婪渴望但没有投资级评级的初创公司组成。因此，英伟达为它们提供后盾，保证以固定价格租赁未售出的产能，时间长到足以确保足够的收入来吸引其他投资。
If you’re interested in understanding AI and aren't following SemiAnalysis, you’re not interested in understanding AI, and while the outfit is relentlessly optimistic it’s also wildly well-informed. In this case, the structure of the investment market for AI is picked apart. It is very weird. Neocloud is made up of startups with a voracious appetite for cash, but no investment grade status. Thus, Nvidia backstops them, guaranteeing to rent unsold capacity for a fixed price for long enough to guarantee enough revenue to attract other investment.
其他资金雄厚的科技巨头也在做类似的事情。这种后盾支持，也被更讽刺地称为“资产负债表即服务”，实际上将更多风险转移给了投资级公司。它是一个日益重要的资本支出赋能工具，但正如SemiAnalysis所说，它必然是有限的，特别是如果人工智能本身已经在企业内部炸开了一个大洞的话。
Other cash-rich big tech companies do similar things. This backstopping, also known more wryly as "balance sheet as a service," effectively shifts more risk to investment-rated companies. It's an increasingly important capex enabler, but as SemiAnalysis says, it’s necessarily limited, especially if AI itself has blown a hole in the enterprise.
或者以旧计算资源的残值为例。通常情况下，数据中心芯片的账面价值和创收能力会在五年左右的时间里贬值到某个名义价值。这种情况在这里并没有发生，这个年份的芯片突然产生了与它们全新时一样多甚至更多的收入。这就是一个习惯了快速淘汰的供应受限市场的独特性质。但这当然是件好事吧？不。在这样一个由奇异且难以建模的因素所定义的市场中，这是另一个奇异且难以建模的因素。
Or take the residual value of the older compute. Normally, the book value and revenue capabilities of datacenter chips depreciates over five years or so to some nominal value. That’s not happening here, where chips of that vintage are suddenly producing the same or more revenue than they did when they were new. Such are the peculiarities of a supply-constrained market used to rapid obsolescence. Surely a good thing, though? No. It’s another strange and hard-to-model factor in a market defined by strange and hard-to-model factors.
甚至不要提通用人工智能（AGI）。或者说那些危险得无法使用的模型。或者是模型崩溃。许多投资者不喜欢这种奇异且难以建模的事物，其中也包括庞大的机构资金池。有些模型可以将风险厌恶型的投资者与投机性投资融合在一起，但那些还记得近二十年前摧毁全球经济的聪明模型的人并非都已经不在人世了。
Don’t even mention AGI. Or models too dangerous to use. Or model collapse. Many investors do not like the strange and hard to model, including huge pools of institutional funds. There are models that can blend in risk-averse investors to speculative investments, but those who remember the clever models that blew up the global economy nearly twenty years ago are not yet all dead.
关于从哪里弄到足够的资金来让AI如承诺般运作这个问题，简而言之就是：没人知道。如果资本支出没有上限，会发生什么？参见前一个答案。在考虑人工智能时，脑海中浮现出动画片段或许是一种幼稚的顽疾，但也无可奈何。代理式AI？《幻想曲》中作为魔法师学徒的米老鼠。科技巨头以最快速度推出AI？正在锯自己坐着的树枝的威利狼。AI投资市场？《非凡数字马戏团》中凯恩的任何一次任务。无论我们愿意与否，我们都只能随波逐流。®
The tl;dr of the answer to where do you get enough money to make AI work as promised, is: nobody knows. What happens if the capex has no cap to ex? See previous answer. It may be an infantile affliction to have animation clips come to mind when considering AI, but it can’t be helped. Agentic AI? Mickey Mouse as the sorcerer's apprentice in Fantasia. Big Tech pushing out AI as fast as it can? Wile E Coyote sawing through the branch he’s sitting on. AI investment market? Any of Caine’s quests in the Amazing Digital Circus. We’re all along for the ride, whether we like it or not. ®