Editor's Note: After Fable briefly emerged and then disappeared, some power users began to realize that the changes in AI models are no longer easily explained by just scores, parameters, or tool invocation. This article, starting from the user experience brought by Fable, discusses a broader realization: the cutting-edge AI competition is transitioning from "who can train a stronger model" to "leading models aiding in training the next generation of leading models," entering a self-accelerating phase.
The author believes that Anthropics's Mythos and its derivative models may signify that this threshold has already been crossed. The truly important change is not just the advancement in product forms such as Claude Code and Codex, but that the model itself is beginning to be part of the R&D system, driving internal development efficiency and accelerating the next round of model iterations. Within this framework, the years 2023 to 2026 may be a critical window for countries to enter the cutting-edge AI competition. As computing power, chip exports, model capabilities, and talent further concentrate, the difficulty of catching up for latecomers will sharply rise.
The most incisive part of the article is the redefinition of "sovereign AI": whether a country truly possesses AI capabilities in the future will not depend on whether it has an application or model shell labeled with the national flag, but on whether it has controllable computing power, talent, training systems, and cutting-edge models. If critical systems are merely a layer of packaging on top of models from top U.S. or Chinese labs, once model access is cut off, the impact may no longer be just the discontinuation of a product but a systemic paralysis across the economic, educational, medical, military, and infrastructural levels.
This is also the core warning of this article: AI is not just a mere coding tool, nor is it a battle of copyright, art, or mathematical ability; it is a general-purpose technology reshaping national capabilities, industrial order, and civilizational infrastructure. The brief appearance and disappearance of Fable may have been just a minor event, but it has allowed more people to foresee the true form of the cutting-edge AI race: the window is closing, computing power is becoming a strategic resource, and those who do not participate in the construction may only become spectators in the next stage.
Below is the original text:
If you ever used Fable during its availability, you would know that its uniqueness cannot be captured in benchmark tests. I often release benchmarks because many people care about these things, but for a long time, they have meant almost nothing to me. The only thing I care about is a benchmark that cannot be measured in numbers: the shape of a model's mind. How deeply can it perceive users, infer intent; how far can it think and iterate on existing information; how "alive" it makes one feel. Fable was precisely different and special in this aspect. It made me feel like I was back in 2023. From my timeline, it wasn't just me who had this experience. As soon as it was disabled, many people's reactions were like their wings had been broken.
Since November of last year, we've been accelerating on a steeper trajectory. For some of us, it has been as clear as the tolling of a bell. Over the past five months, this sensation has only intensified. After using Fable, I am more certain than ever that the shift we are feeling is not only due to advancements in tools like Claude Code and Codex, but also from the unleashed capabilities of Mythos following its training completion in early February. Mythos will birth many "offspring," and its internal development efficiency boost has truly altered the landscape of this competition.
I'm not saying the competition is over. It's not. Other major labs will also train equally powerful models — I believe some already have. Ultimately, they will decipher the magic Anthropic injected into Mythos and replicate it. However, for some, the competition is already over. The frontier has now become an acceleration system: leading models will help produce the next generation of leading models. Many have been predicting for years that we would eventually cross this threshold. Now, that threshold has been crossed.
In 2023, the leaked Series C financing roadshow deck from Anthropic accurately foretold what is happening now. The deck stated: "We believe that the company training the most powerful 2025/2026 models will lead well beyond anyone else in subsequent cycles." In many ways, they were remarkably prescient, and I believe we are currently living in the moment they described back then. If you want your country to be at the frontier, I think the window is only three years: from February 2023 to February 2026. That's the window. And now, it has closed.
Elon seized this window and, through an almost Herculean effort, approached the frontier in just over two short years — 26 months, to be precise. Hardly anyone else with sufficient capital has seriously tried, including many nation-states. European countries have done worse than nothing: they've spent a lot of time erecting barriers that prevent those who might have tried. Now, as leaders are finally glimpsing the contours of the future, some of these barriers are hastily being dismantled. But it's too late.
Seeing the future too late is sometimes worse than not seeing it at all because you have to live through the illusion of what might have been. European leaders could have built truly sovereign frontier models. They had the time, the resources, and could have attracted talent. But by not doing so, they have ensnared themselves in reliance on systems from others: systems that do not belong to them, are beyond their control, cannot be fully audited by them, and may not always be granted for their use — these systems could even be revoked in an instant due to a whimsical decision. They have utterly failed their own people. In my estimation, they have been doing so for decades, but this failure may be among the most catastrophic.
The sudden disappearance of Fable and Mythos has awakened many. Today, there are voices everywhere demanding the construction of sovereign AI projects. However, the ships have already sailed. Many ships have already sailed. The issue is not only whether Mythos-level models can be obtained to kickstart the training of the next generation of models. Hardware constraints will also become increasingly severe. We are entering a new phase: the computational power required to reach this level of capability will itself become a strategic resource.
In the next stage of the competition, chip exports will be allocated, rationed, politicized, and increasingly seen as a tool of national security. The Biden administration has previously outlined this with the almost forgotten "dispersion rules." Those rules even set limits on chip exports to allies. The nuclear weapon analogy has been overused, but soon, state-of-the-art computational power will indeed be treated like uranium: monitored, licensed, tracked, guarded, and kept within U.S. borders. Those in control of this computational power will use all necessary means to maintain this situation and prevent other countries from establishing their own complete tech stacks. Initially, there will be political pressure, the use of preferential agreements for access to their models, assurances and commitments, some treaties guaranteeing other nations will never be suddenly cut off—while only providing just enough computational power to let other countries say to themselves and their citizens: we are also part of it, we are also "contributing." If these don't work, they will resort to more forceful methods. Of course, they might not even have to, as the premise of these scenarios is that other countries are genuinely trying to join this race, and currently, they show no sign of doing so.
The adoption of this technology is still in its very early stages. It won't stay this way forever. Imagine a future nation: its economy, institutions, schools, military, hospitals, and infrastructure all deeply intertwined with a "national-level" AI model that it takes pride in. But in reality, this model is just a thin shell layered on top of a leading U.S. or Chinese lab model. Then imagine this model, like this week's Fable, suddenly subject to an embargo. The effect would seem like an airstrike. The result would be a catastrophic paralysis. This is the power that most countries have handed over to others through inaction.
There will be no second chances unless there is espionage or intentional leakage. Open-sourcing will partially alleviate this issue and is significant in some ways, but relying solely on open-source will not save those countries lacking computational power, talent, and proprietary cutting-edge models. By 2030, training computational power is likely to undergo three orders of magnitude leaps, and these capabilities will increasingly concentrate in the U.S. and China. The ability to train beyond a certain threshold will only belong to the leaders; the models needed for the next generation of models will also be in their hands. What's worse, as open-source models approach Mythos-level capabilities, even surpassing this level—which is likely to begin later this year— they will face regulation and global suppression. I say this not with joy, but it is inevitable. Capabilities will reach a tipping point, and even without an "almost out of control" event, the public will demand regulation. But there's no need to wait for the public to reach that point because the tolerance threshold of the powerful is much lower. They will coordinate to drive a global regulatory crackdown and will use cutting-edge, increasingly powerful closed-source lab models to help them do so.
Whatever your original timeline was for AI development, even just using Fable for a day should significantly accelerate that timeline. Over the course of this upcoming year, that trajectory will become increasingly clear. But for most, what it reveals will only be the depth of missed opportunities and the enormity of what they have squandered. Global leaders should have seen in 2022 what many have already seen—what we have been talking and writing about almost continuously since: the most consequential technology in human history. It has nothing to do with coding, copyrights, art, or math. It is about civilization, society, and the transformation of humanity itself. It is a miracle. And if you have not been involved in building it, then from now on, you will only be a spectator.
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
