换把尺子,宣布世界第一Change the Ruler, Declare Yourself First
韩国宣布要做一个“世界最高性能”的主权 AI 模型,预算是 4.7 万亿韩元。一个还没做出来的东西,性能已经先被定好了。这不是技术目标,这是新闻稿标题。
CNBC 的报道里,整件事最扎实的部分是出口数据:10 月出口同比涨 93.7%,到 1153 亿美元,前 10 个月累计 9300 亿,10 月贸易顺差 425 亿美元。这些是海关统计出来的数字,不是谁宣布的。而“世界最高性能”这个词,和这些数字放在同一篇报道里,看上去就像同一类事实。它不是。观察出来的定律和声明出来的定律,性质根本不同,Moore's Law 是别人事后总结出来的,τ law 是自己站上台念的。李在明这次念的是后者。
再看钱怎么分。213 万亿韩元里,21.3 万亿给半导体、物理 AI 和 AI 数据中心,4.7 万亿给主权模型。主权模型拿的是零头。真正的大头在硬件和产能上,那是韩国已经赢了的赛道。所以这个预算的真实内容不是“我们要造最强的模型”,而是“我们要把已经赢的那条产业链再加固一层,顺便在发布会上说一句最强”。这不丢人,三星和 SK 海力士的订单是真的。丢人的是把加固说成突破。
人才那部分更值得看。“合同系”——企业参与设计课程、出钱、提前锁定毕业生,学生没毕业就已经在去三星和 SK 海力士的路上。政府把它写成“培养 2000 名专家”。这套机制的功能不是培养,是筛选和绑定:谁来定义课程,谁就定义了什么叫合格的人才;谁出钱,谁就拿到优先挑选权。学生拿到的是就业路径,让渡的是在别处发展的可能。这是最优解表达,对个体未必不划算,但它是被结构挤出来的那个解,不是真.最优解。
还有一句要单独拎出来:“ virtually no net increase in government debt”,新发债用来还旧债。总支出涨 12.8%,债务净增近乎为零——这个算术成立的前提是增长继续、出口继续、AI 泡沫不破。IMF 预测 2026 年增长 3.3%,是 2021 年以来最快。整份预算押在这个预测上。预测是别人的,赌注是自己的。
国防那 73.3 万亿也顺手提了一句:AI、机器人、无人机,“smart, elite, and powerful military”。同一笔 AI 投资,一半讲增长,一半讲武力。技术叙事和战争叙事在这份预算里是同一根管子,一头接出口,一头接军备。
所以这篇报道里真正在动的是叙事层:用“主权”“最高性能”“世界第一”这些词,把一个产业升级方案包装成技术主权宣言。谁付代价?报道没写。预算里没有一项写清楚这 4.7 万亿如果失败由谁承担,也没有写那些没进“合同系”的学生去了哪里。名字缺席的地方,通常就是代价所在。
4.7 万亿买不到“世界最高性能”,但买得到一句可以反复引用四年的话。等模型出来那天,记得回头看看性能是谁测的、按什么标准测的、和谁比的。尺子握在宣布者手里的时候,第一名的含金量和新闻稿的排版差不多。
South Korea is going to build a sovereign AI model with the "world's highest performance," for 4.7 trillion won. The thing doesn't exist yet, but its performance has already been decided. That's not a technical goal. That's a headline.
The most solid part of CNBC's report is the export data: exports up 93.7% year on year to a record $115.3 billion in October, $930 billion in the first ten months, a $42.5 billion trade surplus. Those are numbers compiled by a customs service. "World's highest performance" is a phrase someone said at a podium. The article places them side by side as if they were the same kind of fact. They are not. Moore's Law was observed after the fact. A τ law is read aloud from a stage. Lee Jae Myung read the second kind.
Look at where the money actually goes. Of the 21.3 trillion won for semiconductors, physical AI and data centers, the sovereign model gets 4.7 trillion. The model is the rounding error. The bulk sits on the track Korea has already won. So the real content of this budget is not "we will build the strongest model." It is "we will reinforce the supply chain we already dominate, and say 'strongest' at the press conference." That's not shameful—Samsung and SK Hynix orders are real. What's shameful is calling reinforcement a breakthrough.
The talent section deserves a closer look. The "contract departments" let companies design courses, provide funding, and lock in graduates before they finish. The government calls it training 2,000 specialists. The mechanism isn't training. It's screening and binding: whoever designs the curriculum defines what a qualified person is, and whoever pays gets first pick. Students get a path to a job; what they give up is the option of going somewhere else. That's an optimal expression, maybe even a good deal for the individual, but it's the expression the structure squeezed out—not the true optimum.
One line deserves its own paragraph: "virtually no net increase in government debt," with new bonds used to repay old ones. Spending up 12.8%, net debt near zero—the arithmetic only holds if growth continues, exports continue, and the AI boom doesn't pop. The IMF forecasts 3.3% growth in 2026, the fastest since 2021. The whole budget is a bet placed on someone else's forecast. The forecast belongs to the IMF; the stake belongs to Korea.
And the 73.3 trillion won for defense gets a single sentence: AI, robots, drones, a "smart, elite, and powerful military." The same AI investment tells two stories—growth on one side, firepower on the other. In this budget, the tech narrative and the war narrative share a pipe: exports at one end, armaments at the other.
So what's really moving here is the narrative layer: words like "sovereign," "highest performance," and "world's first" repackage an industrial upgrade as a declaration of technological independence. Who pays? The report doesn't say. Nothing in the budget specifies who absorbs the loss if the 4.7 trillion won fails, and nothing says where the students who don't get into a contract department end up. Where names are missing, the cost usually is.
4.7 trillion won won't buy "world's highest performance." It will buy a sentence that can be quoted for four years. When the model ships, check who measured the performance, against what benchmark, and compared to whom. When the ruler is in the hands of the person making the announcement, first place is worth about as much as the press release's layout.