我的征尘是星辰大海。。。
The dirt and dust from my pilgrimage forms oceans of stars...
-------当记忆的篇章变得零碎,当追忆的图片变得模糊,我们只能求助于数字存储的永恒的回忆
作者:黄教授
手机视频列表
别被AGI神话吓住了
视频
音频
原始脚本
别被 AGI 神话吓住,中美 AI 竞争从来不是一战定乾坤。 当下关于 AI 竞争,最主流的焦虑莫过于美国只要先做出 AGI,通用人工智能,就能对中国形成降维打击。 中国现在靠性价比抢占中低端 AI 市场,不过是重蹈造不如买的芯片老路,迟早会被卡脖子,最终满盘皆输。 但一组刚刚出炉的真实数据,彻底打破了这个焦虑。 2026年3月,全球最大 AI 模型 API 聚合平台 OPPO Rotor 的数据显示,中国大模型周调用量已连续3周超越美国。 最新一周达7,359万亿 token,是美国的2.3倍,占全球总调用量的36%。 更关键的是,平台上近半数用户是美国开发者。 这场胜利是在全球市场的客场,实打实打出来的,绝非本土市场的自娱自乐。 这场格局反转的背后,藏着中美 AI 竞争最核心的真相。 它从来不是曼哈顿工程式、一战定乾坤的和竞赛,而是一场全链条能力的长跑。 硅谷最聪明的大脑们并非全知全能,他们正在陷入一场为技术论的战略性误判。 先戳破两个最核心的认知幻象,所有关于 AGI 降维打击的焦虑本质都源于两个被刻意混淆的概念,我们必须先把它讲透。 第一个幻象是把弱 AGI 和强 AGI 画上了等号。 我们当下能遇见,中美正在全力竞争的是弱 AGI ,它是能力更强的通用大模型,能完成绝大多数人类的智力工作,具备跨场景迁移能力。 但依然依赖人类提供的目标、数据、硬件与能源,无法突破物理规则的约束,落地窗口期在3~5年。 而大家恐惧的降维打击,来自科幻级别的强 AGI ,它具备完全的自主意识与自我迭代能力,能独立设定目标,突破人类认知边界,甚至自主重构供应链,一旦实现确实会改变文明格局。 但它的落地时间,哪怕是硅谷最激进的从业者,也不敢给出明确的时间表。 就像我们永远说不准可控核聚变何时能商用。 把强 AGI 的生存及恐惧套到弱 AGI 的产业竞争上,本质是用神话制造焦虑,完全不符合现实逻辑。 第二个幻象是把 AGI 当成了原子弹,以为造出来就能锁定胜局。 1945年美国先造出原子弹,立刻形成了不可逆的核威慑,直接定了二战的终局。 但这套逻辑在弱 AGI 的竞争里完全不成立。 核心有三个无法逾越的区别。 原子弹造出来就赢了,AGI 造出来只是起跑。 原子弹不需要用户,不需要迭代,三颗就能改变战争走向。 但大模型的能力提升,60%以上来自全球用户的万亿次调用与场景反馈。 没有持续的真实场景交互,再强的模型也会快速停滞。 中国拥有全球最大的 AI 调用量、最完整的制造业数字化场景,这些是 AGI 迭代的核心燃料,美国根本锁不住。 原子弹的核心资源能垄断, AGI 却不行。 铀矿、重水、顶尖核物理学家都是稀缺的、可被垄断的资源。 但 AGI 依赖的电力,中国占全球发电量的1/3。 算力产能,中国智能算力增速全球第一。 10亿级用户基数,中国都占据绝对优势,根本无法被封锁。 原子弹不需要落地,AGI 离开落地就是玩具。 原子弹的价值是威慑,不用给普通人用。 但 AGI 的商业价值与产业影响力,90%来自落地到各行各业的日常场景。 IDC 数据显示,2026年全球 AI 推理算力占比已突破70%。 其中90%的 token 需求都来自文档处理、代码开发、智能体调度这些日常场景,而非实验室里的顶尖科研,这正是中国的核心优势所在。 硅谷精英的战略性误判,高估了技术,低估了生态。 硅谷聚集了全球最聪明的大脑,他们当然不是傻子。 他们的判断对了一半,AGI 确实是下一代人类文明的核心基础设施。 谁先实现顶尖突破,谁就能掌握下一代技术革命的话语权,这一点毫无争议。 但他们陷入了三个致命的认知盲区,这很可能让他们在这场长跑中逐渐失去优势。 第一,他们把 AGI 当成了核武器,却忘了它本质是电力。 电力改变世界,靠的不是第一台实验室里的发电机,而是覆盖全球的电网,家家户户的电器,全产业链的配套。 AGI 也是一样,没有规模化的落地场景,没有海量的中低端算力支撑,没有全行业的适配应用。 再强的模型也只是实验室里的演示品,无法形成真实的产业影响力与战略威慑力。 第二,他们高估了模型创造模型的闭环能力,忽略了物理世界的硬约束。 很多人以为只要 AGI 成熟了,就能用模型设计模型,用高端算力制造中低端算力,形成自我强化的正循环,彻底碾压对手。 但他们忘了,AGI 哪怕能设计出世界上最先进的芯片,也造不出光刻机、晶圆厂,离不开工厂、工人、电力与原材料,这些物理世界的壁垒不是靠智能就能凭空突破的,而中国恰恰在这些领域拥有不可替代的全产业链优势。 第三,他们严重低估了中低端场景对智能进化的核心作用,陷入了为技术论的陷阱。 他们以为 AGI 的突破只靠顶尖实验室里的科学家与高端算力,却忘了人类的智能不是靠几个天才在实验室里想出来的,而是靠几十万年的生存实践,海量的场景交互,一步步进化出来的。 AGI 也是一样,亿万次的日常调用、海量的场景反馈才是它持续进化的核心燃料。 而硅谷的精英们长期身处高端研发的闭环里,恰恰忽略了这个最朴素的规律。 中国的正确路径不是甘心做中低端。 而是农村包围城市。 很多人说中国做中低端 AI 市场,就是甘心接受美国做高端研发,中国做低端落地的分工,迟早会被卡脖子。 这个说法完全误解了中国的战略逻辑。 核心路径从来不是放弃高端,只做低端,而是以中低端为基本盘,用规模化优势反哺高端突破,形成全链条的闭环能力。 这是一套典型的农村包围城市的长期战略,分为三个核心步骤。 第一,牢牢守住基本盘,掌控全球 AI 的基础设施命脉。 占全球90%需求的中低端场景是我们的核心优势所在。 我们要做的就是用极致的性价比、全链条的成本优势,垄断全球中低端算力市场,成为真正的世界 Token 工厂。 这不仅能带来持续的营收,更能拿到海量的场景数据,积累丰富的工程化经验。 这些都是反哺高端突破的核心燃料。 第二,坚持以战养战,用基本盘的积累持续突破卡脖子环节。 我们绝对不能走,造不如买的老路。 必须用中低端市场赚来的钱,积累的经验,持续投入顶级通用模型、高端芯片、 HBM 先进封装等核心环节。 现在的真实情况是,中国顶尖模型与美国的差距已经从2023年的10~12个月缩小到了4~7个月,不是代差,而是版本差。 华为昇腾、智谱 GLM 等已经构建了完全自主可控的高端模型与算力供应链,哪怕完全脱钩,也能持续闭环迭代,根本不会出现卡脖子到完全无法发展的情况。 第三,用全链条能力构建真正的战略平衡。 最终的中美 AI 格局不会是一方碾压另一方,而是形成美国有高端技术优势,中国有全链条落地优势的相互依赖格局。 美国离开中国,AGI 无法规模化落地,失去持续迭代的核心燃料。 中国离开美国,也能靠自己的全链条能力闭环发展,不会被降维打击。 这种相互制衡的格局才是真正的战略安全。 最终的结论,中美 AI 竞争的终局,从来不是谁先做出 AGI 谁就赢了的零和博弈,而是一场全链条能力的长跑。 硅谷的 AGI 神话,本质是高估了单点技术突破的碾压性能力,低估了产业生态、规模化场景与全产业链配套的核心价值。 而中国的机会,从来不是跟着美国的节奏,砸钱去追一个单点的 AGI 突破,而是守住自己的基本盘,用全链条的优势,在长跑中逐步缩小差距,最终实现真正的战略自主。 AI 改变世界从来不是靠实验室里的几个天才模型,而是能让每一个普通人、每一家中小企业都用得起、用得稳的普惠算力体系。 而这恰恰是中国最大的机会。
修正脚本
别被 AGI 神话吓住,中美 AI 竞争从来不是一战定乾坤。 当下关于 AI 竞争,最主流的焦虑莫过于美国只要先做出 AGI,通用人工智能,就能对中国形成降维打击。 中国现在靠性价比抢占中低端 AI 市场,不过是重蹈造不如买的芯片老路,迟早会被卡脖子,最终满盘皆输。 但一组刚刚出炉的真实数据,彻底打破了这个焦虑。 2026年3月,全球最大 AI 模型 API 聚合平台 OPPO Rotor 的数据显示,中国大模型周调用量已连续3周超越美国。 最新一周达7,359万亿 token,是美国的2.3倍,占全球总调用量的36%。 更关键的是,平台上近半数用户是美国开发者。 这场胜利是在全球市场的客场,实打实打出来的,绝非本土市场的自娱自乐。 这场格局反转的背后,藏着中美 AI 竞争最核心的真相。 它从来不是曼哈顿工程式、一战定乾坤的核竞赛,而是一场全链条能力的长跑。 硅谷最聪明的大脑们并非全知全能,他们正在陷入一场唯技术论的战略性误判。 先戳破两个最核心的认知幻象,所有关于 AGI 降维打击的焦虑本质都源于两个被刻意混淆的概念,我们必须先把它讲透。 第一个幻象是把弱 AGI 和强 AGI 画上了等号。 我们当下能预见,中美正在全力竞争的是弱 AGI ,它是能力更强的通用大模型,能完成绝大多数人类的智力工作,具备跨场景迁移能力。 但依然依赖人类提供的目标、数据、硬件与能源,无法突破物理规则的约束,落地窗口期在3~5年。 而大家恐惧的降维打击,来自科幻级别的强 AGI ,它具备完全的自主意识与自我迭代能力,能独立设定目标,突破人类认知边界,甚至自主重构供应链,一旦实现确实会改变文明格局。 但它的落地时间,哪怕是硅谷最激进的从业者,也不敢给出明确的时间表。 就像我们永远说不准可控核聚变何时能商用。 把强 AGI 的生存级恐惧套到弱 AGI 的产业竞争上,本质是用神话制造焦虑,完全不符合现实逻辑。 第二个幻象是把 AGI 当成了原子弹,以为造出来就能锁定胜局。 1945年美国先造出原子弹,立刻形成了不可逆的核威慑,直接定了二战的终局。 但这套逻辑在弱 AGI 的竞争里完全不成立。 核心有三个无法逾越的区别。 原子弹造出来就赢了,AGI 造出来只是起跑。 原子弹不需要用户,不需要迭代,三颗就能改变战争走向。 但大模型的能力提升,60%以上来自全球用户的万亿次调用与场景反馈。 没有持续的真实场景交互,再强的模型也会快速停滞。 中国拥有全球最大的 AI 调用量、最完整的制造业数字化场景,这些是 AGI 迭代的核心燃料,美国根本锁不住。 原子弹的核心资源能垄断, AGI 却不行。 铀矿、重水、顶尖核物理学家都是稀缺的、可被垄断的资源。 但 AGI 依赖的电力,中国占全球发电量的1/3。 算力产能,中国智能算力增速全球第一。 10亿级用户基数,中国都占据绝对优势,根本无法被封锁。 原子弹不需要落地,AGI 离开落地就是玩具。 原子弹的价值是威慑,不用给普通人用。 但 AGI 的商业价值与产业影响力,90%来自落地到各行各业的日常场景。 IDC 数据显示,2026年全球 AI 推理算力占比已突破70%。 其中90%的 token 需求都来自文档处理、代码开发、智能体调度这些日常场景,而非实验室里的顶尖科研,这正是中国的核心优势所在。 硅谷精英的战略性误判,高估了技术,低估了生态。 硅谷聚集了全球最聪明的大脑,他们当然不是傻子。 他们的判断对了一半,AGI 确实是下一代人类文明的核心基础设施。 谁先实现顶尖突破,谁就能掌握下一代技术革命的话语权,这一点毫无争议。 但他们陷入了三个致命的认知盲区,这很可能让他们在这场长跑中逐渐失去优势。 第一,他们把 AGI 当成了核武器,却忘了它本质是电力。 电力改变世界,靠的不是第一台实验室里的发电机,而是覆盖全球的电网,家家户户的电器,全产业链的配套。 AGI 也是一样,没有规模化的落地场景,没有海量的中低端算力支撑,没有全行业的适配应用。 再强的模型也只是实验室里的演示品,无法形成真实的产业影响力与战略威慑力。 第二,他们高估了模型创造模型的闭环能力,忽略了物理世界的硬约束。 很多人以为只要 AGI 成熟了,就能用模型设计模型,用高端算力制造中低端算力,形成自我强化的正循环,彻底碾压对手。 但他们忘了,AGI 哪怕能设计出世界上最先进的芯片,也造不出光刻机、晶圆厂,离不开工厂、工人、电力与原材料,这些物理世界的壁垒不是靠智能就能凭空突破的,而中国恰恰在这些领域拥有不可替代的全产业链优势。 第三,他们严重低估了中低端场景对智能进化的核心作用,陷入了唯技术论的陷阱。 他们以为 AGI 的突破只靠顶尖实验室里的科学家与高端算力,却忘了人类的智能不是靠几个天才在实验室里想出来的,而是靠几十万年的生存实践,海量的场景交互,一步步进化出来的。 AGI 也是一样,亿万次的日常调用、海量的场景反馈才是它持续进化的核心燃料。 而硅谷的精英们长期身处高端研发的闭环里,恰恰忽略了这个最朴素的规律。 中国的正确路径不是甘心做中低端。 而是农村包围城市。 很多人说中国做中低端 AI 市场,就是甘心接受美国做高端研发,中国做低端落地的分工,迟早会被卡脖子。 这个说法完全误解了中国的战略逻辑。 核心路径从来不是放弃高端,只做低端,而是以中低端为基本盘,用规模化优势反哺高端突破,形成全链条的闭环能力。 这是一套典型的农村包围城市的长期战略,分为三个核心步骤。 第一,牢牢守住基本盘,掌控全球 AI 的基础设施命脉。 占全球90%需求的中低端场景是我们的核心优势所在。 我们要做的就是用极致的性价比、全链条的成本优势,垄断全球中低端算力市场,成为真正的世界 Token 工厂。 这不仅能带来持续的营收,更能拿到海量的场景数据,积累丰富的工程化经验。 这些都是反哺高端突破的核心燃料。 第二,坚持以战养战,用基本盘的积累持续突破卡脖子环节。 我们绝对不能走造不如买的老路。 必须用中低端市场赚来的钱,积累的经验,持续投入顶级通用模型、高端芯片、 HBM 先进封装等核心环节。 现在的真实情况是,中国顶尖模型与美国的差距已经从2023年的10~12个月缩小到了4~7个月,不是代差,而是版本差。 华为昇腾、智谱 GLM 等已经构建了完全自主可控的高端模型与算力供应链,哪怕完全脱钩,也能持续闭环迭代,根本不会出现卡脖子到完全无法发展的情况。 第三,用全链条能力构建真正的战略平衡。 最终的中美 AI 格局不会是一方碾压另一方,而是形成美国有高端技术优势,中国有全链条落地优势的相互依赖格局。 美国离开中国,AGI 无法规模化落地,失去持续迭代的核心燃料。 中国离开美国,也能靠自己的全链条能力闭环发展,不会被降维打击。 这种相互制衡的格局才是真正的战略安全。 最终的结论,中美 AI 竞争的终局,从来不是谁先做出 AGI 谁就赢了的零和博弈,而是一场全链条能力的长跑。 硅谷的 AGI 神话,本质是高估了单点技术突破的碾压性能力,低估了产业生态、规模化场景与全产业链配套的核心价值。 而中国的机会,从来不是跟着美国的节奏,砸钱去追一个单点的 AGI 突破,而是守住自己的基本盘,用全链条的优势,在长跑中逐步缩小差距,最终实现真正的战略自主。 AI 改变世界从来不是靠实验室里的几个天才模型,而是能让每一个普通人、每一家中小企业都用得起、用得稳的普惠算力体系。 而这恰恰是中国最大的机会。
英文翻译
Don’t be intimidated by the myth of AGI. The AI competition between China and the US has never been decided in a single battle. Currently, the most prevalent anxiety about AI competition is that if the US achieves AGI (Artificial General Intelligence) first, it will deliver a dimensionality-reduction strike against China. Some argue that China’s current strategy of capturing the mid-to-low-end AI market through cost-effectiveness is merely repeating the old chip industry mistake of “buying rather than building”—that it will eventually be choked off and lose entirely. But a fresh set of real data has completely shattered this anxiety. In March 2026, data from OPPO Rotor, the world’s largest API aggregation platform for AI models, showed that China’s weekly large model invocation volume had surpassed the US for three consecutive weeks. The latest week reached 7,359 trillion tokens—2.3 times that of the US, accounting for 36% of global total invocations. More critically, nearly half of the platform’s users are US developers. This victory was won in the global market, on the road, not a self-congratulatory celebration in the domestic market. Behind this reversal of the landscape lies the most crucial truth of US-China AI competition. It has never been a Manhattan Project-style nuclear race that decides everything in one go, but rather a long-distance race of full-chain capabilities. The brightest minds in Silicon Valley are not omniscient; they are falling into a strategic miscalculation of techno-supremacy. Let’s first debunk two core cognitive illusions. All the anxiety about a dimensionality-reduction strike from AGI fundamentally stems from two deliberately conflated concepts. We must clarify them thoroughly. The first illusion is equating weak AGI with strong AGI. What China and the US are now fiercely competing for is weak AGI—a more capable general-purpose large model that can complete most intellectual tasks of humans and possesses cross-scenario transferability. However, it still relies on human-provided goals, data, hardware, and energy, and cannot break the constraints of physical laws. The deployment window is around 3–5 years. The dreaded dimensionality-reduction strike comes from sci-fi-level strong AGI, which possesses full autonomous consciousness and self-iterative capability, can independently set goals, break through the boundaries of human cognition, and even autonomously restructure supply chains. If realized, it would indeed change the course of civilization. But even the most aggressive practitioners in Silicon Valley dare not give a specific timeline for its deployment—just as we can never predict when controlled nuclear fusion will be commercially viable. Applying the existential fear of strong AGI to the industrial competition of weak AGI is essentially manufacturing anxiety through myth, completely inconsistent with real-world logic. The second illusion is treating AGI like an atomic bomb, believing that once built, it secures victory. In 1945, the US built the atomic bomb first, immediately establishing irreversible nuclear deterrence that directly ended WWII. But this logic does not apply to weak AGI competition at all. There are three insurmountable differences. First, building the atomic bomb meant winning; building AGI is only the starting line. The atomic bomb required no users or iteration—three bombs changed the course of war. However, more than 60% of a large model’s capability improvement comes from trillions of user invocations and scenario feedback globally. Without continuous real-world interaction, even the strongest model will quickly stagnate. China has the world’s largest AI invocation volume and the most complete digitalized manufacturing scenarios—these are core fuel for AGI iteration that the US simply cannot lock down. Second, the atomic bomb’s core resources could be monopolized; AGI’s cannot. Uranium ore, heavy water, and top nuclear physicists are scarce and monopolizable. But for AGI, China accounts for one-third of global electricity generation. China’s intelligent computing capacity growth rate is the fastest in the world. With a user base of over a billion, China holds an absolute advantage that cannot be blockaded. Third, the atomic bomb doesn’t need deployment; AGI without deployment is just a toy. The atomic bomb’s value is deterrence—it doesn’t need to be used by ordinary people. But 90% of AGI’s commercial value and industrial impact come from its application in everyday scenarios across industries. IDC data shows that by 2026, global AI inference computing power will account for over 70% of the total. Among that, 90% of token demand comes from routine scenarios like document processing, code development, and agent scheduling—not top-tier scientific research in labs. This is precisely China’s core advantage. Silicon Valley elites’ strategic miscalculation: overestimating technology, underestimating ecology. Silicon Valley gathers the world’s brightest minds—they are certainly not stupid. Their judgment is half correct: AGI is indeed the core infrastructure for the next generation of human civilization. Whoever achieves the top breakthrough first will hold the discourse power in the next technological revolution—this is uncontroversial. But they have fallen into three fatal cognitive blind spots, which may gradually erode their advantage in this long race. First, they treat AGI as a nuclear weapon but forget it is essentially electricity. Electricity changed the world not by the first generator in a lab, but by a global power grid, household appliances, and the entire industrial chain. The same goes for AGI: without large-scale deployment scenarios, massive mid-to-low-end computing support, and industry-wide adapted applications, even the strongest model is just a lab demo—unable to form real industrial impact or strategic deterrence. Second, they overestimate the closed-loop capability of models creating models, ignoring the hard constraints of the physical world. Many believe that once AGI matures, it can design models with models, use high-end computing to manufacture mid-to-low-end computing, forming a self-reinforcing virtuous cycle to crush opponents. But they forget that even if AGI could design the world’s most advanced chips, it cannot build lithography machines or wafer fabs—it still depends on factories, workers, electricity, and raw materials. These physical-world barriers cannot be broken through intelligence alone, and China’s irreplaceable full-chain industrial advantages lie exactly in these areas. Third, they severely underestimate the core role of mid-to-low-end scenarios in intelligent evolution, falling into the trap of techno-supremacy. They assume AGI breakthroughs rely solely on top scientists in elite labs and high-end computing, forgetting that human intelligence evolved not through a few geniuses in labs, but through hundreds of thousands of years of survival practice and massive scenario interactions—step by step. The same is true for AGI: billions of daily invocations and massive scenario feedback are the core fuel for its continuous evolution. Silicon Valley elites, long immersed in the closed loop of high-end R&D, precisely overlook this most basic rule. China’s correct path is not to be content with the mid-to-low end, but to “surround the cities from the countryside.” Many say that China focusing on the mid-to-low-end AI market means willingly accepting a division of labor where the US does high-end R&D and China does low-end deployment—that it will eventually be choked off. This view completely misunderstands China’s strategic logic. The core path has never been to abandon high-end and only do low-end, but to take the mid-to-low end as the foundation, using scale advantages to feed back high-end breakthroughs, forming a full-chain closed-loop capability. This is a classic long-term strategy of “surrounding the cities from the countryside,” consisting of three steps. First, firmly hold the basic market, controlling the lifeblood of global AI infrastructure. Mid-to-low-end scenarios accounting for 90% of global demand are our core advantage. What we need to do is to use extreme cost-effectiveness and full-chain cost advantages to monopolize the global mid-to-low-end computing market, becoming the world’s true Token factory. This not only brings sustained revenue but also massive scenario data and rich engineering experience—core fuel to support high-end breakthroughs. Second, sustain the war by fighting: use the accumulation from the basic market to continuously break through bottlenecks. We must absolutely not repeat the old path of “buying rather than building.” We must use the money earned from the mid-to-low-end market and the accumulated experience to continuously invest in core areas such as top-tier general models, high-end chips, HBM advanced packaging, etc. The current reality is that the gap between China’s top models and the US has narrowed from 10–12 months in 2023 to 4–7 months—not a generation gap but a version gap. Huawei’s Ascend, Zhipu’s GLM, etc., have already built fully autonomous high-end models and computing supply chains. Even with complete decoupling, they can continue closed-loop iteration and will not be completely stuck. Third, use full-chain capabilities to build true strategic balance. The final US-China AI landscape will not be one side crushing the other, but a mutually dependent pattern where the US has high-end technology advantages and China has full-chain deployment advantages. Without China, the US cannot scale AGI deployment and loses core fuel for continuous iteration. Without the US, China can still develop in a closed loop with its own full-chain capabilities, immune to dimensionality-reduction strikes. This mutual-checking pattern is the real strategic security. The final conclusion: The endgame of US-China AI competition has never been a zero-sum game where whoever builds AGI first wins. It is a long-distance race of full-chain capabilities. The Silicon Valley AGI myth essentially overestimates the crushing power of single-point technological breakthroughs and underestimates the core value of industrial ecology, scaled scenarios, and full-chain supporting industries. And China’s opportunity is never to follow the US rhythm and throw money at chasing a single-point AGI breakthrough, but to hold its own foundation, use its full-chain advantages to gradually close the gap in the long race, and ultimately achieve true strategic autonomy. AI changes the world not by a few genius models in labs, but by an inclusive computing system that every ordinary person and every SME can afford and use reliably. And that is precisely China’s greatest opportunity.
back to top