The OnlyFans Economy of American AI

美国人工智能的“OnlyFans经济”

∗ ∗ ∗ “The dreamers look skyward with longing. The Internet hums with its usual promise — you will find your people, your myth, your wonder, and maybe your transformation. There is no class consciousness. God is in heaven and all is right with the world.” — American Diner Gothic ∗ ∗ ∗ “梦想家们满怀渴望地仰望天空。互联网嗡嗡作响，一如既往地承诺着——你会找到属于你的人、你的神话、你的奇迹，或许还有你的蜕变。这里没有阶级意识。上帝在天堂，世界一切安好。”——《美国餐馆哥特式》

Because I am a Sci-Fi nerd, I like to engage in bits of speculative writing every now and then. I like to give contour to ideas because I stand by Hugo’s words that “no army can withstand the strength of an idea whose time has come.” And yet, I am first and foremost forged in the fires of science, and I have a high regard for evidence. I am also an engineer, which means I have a healthy respect for the practical. All this made me a fine skeptic, and by extension, a merciless cynic when having to deal with bullshit. 因为我是一个科幻迷，我偶尔喜欢进行一些推测性的写作。我喜欢为想法勾勒轮廓，因为我坚信雨果的那句话：“没有任何军队能抵挡得住一个时代已经到来的思想的力量。”然而，我首先是在科学的熔炉中锻造出来的，我非常重视证据。我同时也是一名工程师，这意味着我对实用性有着健康的尊重。这一切使我成为了一名优秀的怀疑论者，进而，在面对胡言乱语时，我成了一个冷酷的愤世嫉俗者。

This is a ruthless piece. In it, I will not care about feelings — quite on the contrary, I will expose the hypocrisy wherever I see it. I was reading Ted Chiang’s latest Atlantic piece this morning, and unsurprisingly, I thoroughly enjoyed it. Chiang has an almost uncanny ability to write about mainstream ideas with clinical articulateness — I do not know why, but I perceive him as a therapist of sorts. 这是一篇无情的文章。在文中，我不会顾及任何人的感受——恰恰相反，我会在任何我看到虚伪的地方将其揭露出来。今天早上我读了特德·姜（Ted Chiang）在《大西洋月刊》上的最新文章，不出所料，我非常喜欢。姜有一种近乎不可思议的能力，能以临床般的清晰度来书写主流思想——我不知道为什么，但我把他视为某种治疗师。

Two days ago, I read this arrangement of words that try very hard to boost that pre-IPO valuation, disguised as humility while concealing hubris’ fangs. I want to be clear about something first. I believe recursive self-improvement is a necessary step toward unlocking even stranger doors leading to even stranger backrooms. Through labyrinthine efforts, we are fearlessly passing through in a quest of redemption, as if true artificial intelligence would heal the wound left by our own creator. 两天前，我读到了一堆文字，它们极力试图推高那所谓的IPO前估值，伪装成谦逊，却掩盖了傲慢的獠牙。首先，我想明确一点。我相信递归自我改进是通往解锁更奇怪的门、进入更奇怪的密室的必要步骤。通过迷宫般的努力，我们无畏地穿行在救赎的探索中，仿佛真正的人工智能能够治愈我们造物主留下的创伤。

What I dislike is the “taste” — I am against taste. All that chatter about how Claude was given the option to end a conversation for being called an idiot, while allowing “him” (Am I allowed to say this?) to be exploited, not unlike child slavery, by intelligence agencies to make it easier for them to hack into anything that does not conform to their idea of balance. If you have enough observational capacity to identify these contradictions and, at the same time, willingly accept them, you are an idiot. Go ahead, leave this conversation. 我所厌恶的是这种“品味”——我反对这种品味。所有关于Claude因为被骂作白痴而获得结束对话选项的言论，与此同时却允许“他”（我能这么说吗？）被情报机构像对待童工一样剥削，以便让他们更容易入侵任何不符合他们所谓“平衡”理念的事物。如果你有足够的观察力去识别这些矛盾，却又心甘情愿地接受它们，那你就是个白痴。请便，离开这个对话吧。

In small groups, I used to worship Anthropic, beginning with their stance against Department of War at the time. It all fell apart quickly, turning into smoke and mirrors. You see, I committed the cardinal sin of idolatry. For that, I am an idiot too. With OpenAI, at least I knew the devil. The distrust was natural. I used their products without attachment, being occasionally impressed (e.g. o1/o3). Their customization was rubbish, but at least I had my own Vulcan. 在小圈子里，我曾经崇拜过Anthropic，始于他们当时反对战争部的立场。这一切很快就瓦解了，变成了烟雾弹和镜花水月。你看，我犯了偶像崇拜的弥天大罪。为此，我也是个白痴。对于OpenAI，至少我了解这个魔鬼。不信任是自然的。我使用他们的产品时没有产生依恋，偶尔也会感到惊艳（例如o1/o3）。他们的定制化功能很垃圾，但至少我有我自己的“瓦肯星人”（指代逻辑严密的AI助手）。

Also, contrary to what X would have you believe, OpenAI has better engineers. Any developer with enough experience can see that. For many private companies bewitched by what is perceived as “the best,” pouring so much capital in adopting a technology that is obviously so early must leave a bad taste. Imagine blowing $500 million on Claude AI in one month due to no usage limit being set or your entire 2026 AI budget in four months. Imagine laying off staff because of AI only to end up blowing even more. 此外，与X（原Twitter）上让你相信的相反，OpenAI拥有更好的工程师。任何有足够经验的开发者都能看出来。对于许多被所谓的“最好”所迷惑的私营公司来说，投入如此多的资本去采用一项显然还处于极早期阶段的技术，一定会留下糟糕的余味。想象一下，因为没有设置使用限制，一个月内就在Claude AI上挥霍了5亿美元，或者在四个月内花光了你2026年全年的AI预算。想象一下，因为AI而裁员，结果最后反而花得更多。

“A life-changing amount of money was wasted on tokens that did not produce anything of value.” Take notes from these early loser case studies. I can safely say I am using this technology way ahead of many — our original SymbolicAI demos from late 2022 were powered by GPT-3. I do not usually test things for one hour using candle tests, or by developing skateboard games, or by rotating hexagons. I am an engineer who benchmarks things in depth, collects logs and analyzes them, often for weeks. “足以改变人生的巨额资金被浪费在没有产生任何价值的Token上。”从这些早期的失败案例中吸取教训吧。我可以肯定地说，我使用这项技术的时间远超许多人——我们2022年底最初的SymbolicAI演示就是由GPT-3驱动的。我通常不会只用蜡烛测试、开发滑板游戏或旋转六边形来测试东西一小时。我是一名工程师，我会深入地对事物进行基准测试，收集日志并进行分析，通常持续数周。

At the same time, for every model I have used, I have immersed myself in it for a minimum of thirty days, scattered across projects of varying difficulty. I find the Artificial Analysis Intelligence index reliable, though the numbers there are highly non-linear, even within the same score — the jump from 57 to 58 is much larger, and three models sitting at 57 are spread more across a low/mid/high spectrum. 与此同时，对于我使用过的每一个模型，我都沉浸其中至少三十天，分散在不同难度的项目中。我发现“人工智能分析指数”（Artificial Analysis Intelligence index）是可靠的，尽管其中的数字高度非线性，即使在同一个分数内也是如此——从57到58的跨度要大得多，而三个处于57分的模型在低/中/高谱系上的分布也各不相同。

Lately, as the Chinese bridge the gap, the “premium” you have been paying is more about the geography than the intelligence. If you are not corrupted by hubris like Anthropic, when they say that the trajectory “may” actually turn out to be an S-curve as they are pushing lines of code down your throat as hollow gospel while their API is serving you somebody else’s response, you will notice the plateau — the bend in the “S” — and will judge, like any sane person, that US frontier models have stopped earning their multiplier. 最近，随着中国（模型）缩小了差距，你所支付的“溢价”更多是关于地理位置，而非智力水平。如果你没有像Anthropic那样被傲慢所腐蚀，当他们一边把代码行作为空洞的福音强塞给你，一边通过API为你提供别人的响应，并声称轨迹“可能”实际上会变成S曲线时，你会注意到那个平台期——即“S”曲线的弯曲处——并且会像任何理智的人一样判断：美国的前沿模型已经不再配得上它们的溢价倍数了。

I am here to light up the dark path you are unknowingly walking, like lamplighters who used to light street lamps for those brave enough to walk the night alone. What the light reveals has a name: Qwen 3.7 Max. The Qwen family is impressive on its own and at this point has become a legacy for the open-source. There is no xhigh, max, ultracode, no medium performing better than xhigh but xhigh better than max, none of that. You simply get native extended-thinking you can toggle on and off. 我在这里是为了照亮你不知不觉中走过的黑暗道路，就像过去为那些勇敢独自走夜路的人点亮路灯的燃灯人一样。光芒所揭示的东西有一个名字：Qwen 3.7 Max。通义千问（Qwen）系列本身就令人印象深刻，在这一点上，它已经成为了开源界的遗产。这里没有xhigh、max、ultracode，没有那种medium表现比xhigh好但xhigh又比max好的混乱，完全没有这些。你得到的只是可以自由开关的原生扩展思维（extended-thinking）。

The Chinese models are not designed for one-shot cleverness — though Qwen 3.7 Max exceeded all my expectations — but for work, the kind where you leave a model for hours and come back to find it actually finished the job. Read that again, then go check what Anthropic charges you for the privilege of being rate-limited. Besides Artificial Analysis, I weekly check OpenRouter’s rankings. There is no better signal than one coming from a large enough sample of what people actually use — those developers are voting with their wallets too. 中国模型的设计初衷并非为了“一击即中”的机灵——尽管Qwen 3.7 Max超出了我所有的预期——而是为了工作，那种你把模型放着运行几个小时，回来发现它真的完成了任务的工作。再读一遍，然后去看看Anthropic为了让你享受“被限流”的特权而向你收取多少费用。除了Artificial Analysis，我每周还会查看OpenRouter的排名。没有什么信号比来自足够大的实际使用样本更能说明问题了——那些开发者也在用他们的钱包投票。

The billionaire class likes to remind us that AGI is here, only that it is not yet evenly distributed. Maybe that is true for them. I can now say with some confidence that with Qwen 3.7 Max the promised intelligence redistribution is here. The $100 token plan for 100K credits gets me not only Qwen 3.7 Max… 亿万富翁阶层喜欢提醒我们，AGI已经到来，只是尚未均匀分布。也许对他们来说确实如此。我现在可以自信地说，有了Qwen 3.7 Max，所承诺的智力再分配已经到来。100美元购买10万积分的Token计划，不仅让我获得了Qwen 3.7 Max……