2026年6月15日 · 9 分钟阅读

The Window Has Closed

The Window Has Closed Author : Andrew Curran (@AndrewCurran ) — William Armstrong Professor of the Humanities, Wesleyan University Date : 2026 06 15 0...

The Window Has Closed

Author: Andrew Curran (@AndrewCurran_) — William Armstrong Professor of the Humanities, Wesleyan University Date: 2026-06-15 09:31 AM (ET) Tags: #AI #SovereignAI #Fable #Mythos #Anthropic #Geopolitics #RSI #ComputeAsUranium #Inevitabilism Source: https://x.com/AndrewCurran_/status/2066332670817456584 Article: https://x.com/i/article/2066289802295779328

原文

If you used Fable while it was available, you know it is special in ways that will not show up on benchmarks. I post benchmarks all the time because they matter to many people, but for a long time they have not mattered to me at all. I only care about a benchmark that is not measured in numbers. It is the shape of a model's mind: how deeply it can perceive the user and infer intent, how far it can think and iterate upon what it has been given, how alive the model feels. Fable is different and special in exactly this way. It makes me feel like I'm back in 2023. I see from my timeline that I am not the only one who has experienced this. As soon as it was disabled, many people reacted as if their wings had been torn off.

Since November, we have been on a steeper trajectory. To some of us, it was as clear as a bell being rung. That feeling has steadily increased over the last five months. After using Fable, it now seems clearer than ever to me that the shift we felt was driven not only by tool advances in Claude Code and Codex, but by Mythos emerging from its training run in early February. Mythos will have many children, and the resulting boost to internal development has changed the shape of the race.

I am not saying the race is over. It is not. The other big labs will train models just as capable — I'm sure some already are. Eventually they will divine the magic Anthropic put into Mythos and replicate it. But the race is over for some people. The frontier is now an accelerating system in which the leading models will help produce the next leading models. That we would reach this threshold has been predicted by many people for years. It has now been crossed.

In 2023, Anthropic's leaked Series C pitch deck predicted exactly what is now happening. It said, "We believe that companies that train the best 2025/26 models will be too far ahead for anyone to catch up in subsequent cycles." They have been remarkably prescient in many ways, and I think we are now standing inside the moment that they described. If you wanted your nation to be at the frontier, I think you had three years to do it: February 2023 to February 2026. That was the window. And it is now closed.

Elon seized upon that window and, through a Herculean effort, approached the frontier in just over two years — 26 months, to be exact. Almost no one else with the required capital made a serious attempt, including many nation-states. European states did worse than nothing: they spent much of that time erecting barriers to anyone who might have tried. Some of those barriers are now being frantically torn down as leaders finally begin to see the shape of things to come. But it is too late.

Seeing the future too late can be worse than not seeing it at all, because you are forced to experience the vision of what could have been.

European leaders could have built truly sovereign frontier models. They had the time and the resources, they could have recruited the talent. By failing to do so, they have consigned themselves to dependence on systems they do not own, do not control, cannot fully inspect, and may not always be allowed to use — systems that can be instantly removed on a whim. They have utterly failed their people. By my estimation, they have been doing this for decades, but this failure may be the most catastrophic of all.

The sudden disappearance of Fable and Mythos woke a lot of people up. Calls for sovereign AI projects are everywhere today. But again, the ship has sailed. Many ships have. The problem is not just having access to Mythos-level models to bootstrap the next generation. There will increasingly be hardware constraints as well. We are now entering the period in which the compute required to reach that level will itself become a strategic resource.

During the next phase of the race, chip exports will be rationed, narrowed, politicized, and increasingly treated as instruments of national security. The Biden administration already sketched this out with their now long-forgotten diffusion rules. They had chip export caps even for allies. The nuke analogy is overused, but soon the most advanced compute actually will be treated like uranium: monitored, licensed, tracked, guarded, and kept within America's borders. The people in control of it will use all means necessary to keep it that way and to prevent others from building their own stack. It will start with political pressure, with sweetheart deals on using their models, with assurances and agreements, with treaties that guarantee to nations that they will never suddenly be cut off — by providing just enough compute that other nations can tell themselves and their citizens that they are part of this, that they are "contributing." And if all of that doesn't work, they will use harder methods. They may not have to do any of this, as these scenarios would require other nations to actually make a serious attempt to join the race, which they show absolutely no signs of doing.

The adoption of this technology is still in its infancy. It won't always be this way. Imagine a future nation whose economy, institutions, schools, military, hospitals, and infrastructure are all intertwined with a proud "national" AI model that is in fact a thin wrapper around one of the leading American or Chinese labs models. Then imagine that model being embargoed in an instant, as Fable was this week. The effect would look like an airstrike. And the result would be catastrophic paralysis. This is the power most nations have now handed to others through their inaction.

Barring espionage or a deliberate leak, there will be no second chance. Open source will mitigate some of this, and it will matter enormously in some respects, but open source alone will not save states that lack compute, talent, and frontier-grade models of their own. There will probably be three more orders-of-magnitude jumps in training compute before 2030, and that capacity will be increasingly concentrated in America and China. The ability to train beyond a certain threshold will belong only to the leaders, along with the models needed to train the models that come next. And even worse, as open source models approach Mythos-level capabilities — and beyond — which will probably start to happen later this year, they will face regulation and a global crackdown. I take no pleasure in saying this, but it is inevitable. Capabilities will reach a point where the public will demand regulation, even without a near miss. But reaching that point won't be necessary, because the people in power will have a much lower tolerance level for this than the public. They will coordinate to enact a global regulatory crackdown, and they will use the increasingly powerful closed lab models at the frontier to help them do it.

Whatever your timelines are, using Fable for even a day should probably have shortened them. This trajectory will become more apparent to everyone over the rest of this year, but for most people it will only reveal the depth of the missed opportunity and the scale of what they squandered. Global leaders should have seen what many of us saw in 2022, and what we have done little but talk and write about ever since: this is the most important technology in the history of humanity. It is not about coding. It is not about copyright. It is not about art or math. It is about the transformation of civilization, society, and humanity itself. It is a Wonder. And if you are not already part of building it, you will now only be a spectator.

关键论点速览

Fable 不是 benchmark 级的领先 —— 真正差距在 "shape of a model's mind"：感知深度、推断意图、思考迭代、感觉"alive"。无可复制，只能等"很多 Mythos 的孩子"。
三年窗口（Feb 2023 → Feb 2026）已闭 —— 想自建前沿模型的国家窗口期。Anthropic 2023 Series C pitch deck 早有预言："训练出 2025/26 最佳模型的公司将领先到无法追赶"。
前沿 = 加速系统 —— 领先模型帮助训练下一代领先模型（RSI 阈值已跨越）。其他大厂会追上，但国家级玩家已无机会。Elon/xAI 26 个月追上前沿是孤例。
算力即铀矿 —— 下阶段 chip exports 会被 rationed、licensed、tracked、guarded、kept within America's borders。Biden diffusion rules 是预演。先甜头（条约/sweetheart deals 给刚好够的算力让别国自欺"参与"），不行就用更硬手段。
未来图景 = airstrike —— 想象一个国家经济/机构/军队/医院都依赖"国家 AI"（实为美/中前沿模型的薄壳）。一旦 embargo，"效果像 airstrike"，灾难性瘫痪。
开源不够 —— 救不了缺 compute/talent/前沿模型的国家。2030 前还会有 3 个数量级的训练 compute 跃升，集中在美中。开源模型接近 Mythos 级别（今年晚些开始）会面临全球监管打压。
结尾 —— 这是人类历史上最重要的技术。"It is a Wonder." 不在场即永远只是观众。

写作风格

地质学笔调（geological key）：strata 已沉、fault 已闭、bell 已响。Substack《The Reflexive Machine》评其为"AI 时代最美的写作之一"，但"以所有必然论都错的方式错了"——把当下结构误当自然事实。

主要批评（The Reflexive Machine, Substack）

Curran 把层搞反了：决定性的不是 compute（底层），而是 legal-jurisdictional layer（法律-管辖层）——Fable 事件中真正按停一切的，是一封周五下午的信件
"Ownership was not a shield. It was not even friction." —— 拥有整栈的 Anthropic 仍被自己的政府按停
引 Benjamin Bratton《The Stack》：行星计算是垂直结构、国家按水平领土统治，两者错位。"Foreign national" 是领土分类，无法切非领土的层，只能整层拔除
真正的主权 ≠ 拥有整列垂直栈，而是："能从可达层组装多少、拉走一层能多快重组、多了解别人手按在哪层"
结尾反驳："The window was never the point. The architecture is."

数据与反响

互动：1M+ 浏览、2.6K 点赞、356 RT、2.8K 收藏
大V 转发/引述：Robert Scoble、Siqi Chen、Derya Unutmaz、Scott Leibrand、Jeffrey Emanuel
Derya Unutmaz：不认同"窗口已闭"的悲观，认为其他国家仍有机会

作者背景

Andrew S. Curran，Wesleyan University William Armstrong 讲席教授（人文学），18 世纪/启蒙运动学者，专注种族思想史。著作《The Anatomy of Blackness》《Darwin's Seed》。同时活跃于 X 撰写 AI 时代评论。