Pichai says AI feeds back into R&D, and the next year will accelerate due to RSI
Google CEO Sundar Pichai said on the earnings call that AI is helping improve training and release speed, and he expects the next twelve months to keep accelerating due to RSI. The author believes this is not an official announcement of recursive self-improvement, but AI feeding back into the R&D pipeline has already been written into the investor communication minutes. The real change is that the engineering iteration cycle has been compressed; the next stage is a contest over who can hand the R&D pipeline over to AI faster.
View original postThe A-share market has been really bad recently, and my mood is complicated. The biggest realization I've had recently is this: the biggest misunderstanding of our generation of stock investors is mistakenly believing that the nation's fate equals stock prices. The latest issue of The Economist also said that China's stock market is a rare exception, and raised a very piercing soul-searching question: "If this country is winning the future, why are its stock market investors so depressed?" Many of us, when looking at the macro narrative, are always used to directly equating "industry won" with "investment made money." Over the past few years, in new energy, solar power, power batteries, smart hardware, and open-source AI, whether judged by production capacity scale, speed of going global, or technological iteration, China's industrial chain has demonstrated what can only be called terrifying dominance globally. But on the K-line charts of the secondary market, major global stock markets are taking turns hitting record highs, while this stock market alone has become a painfully rare global exception. What this contrast tears open is a hardcore financial common sense lesson that all stock investors and entrepreneurs must make up for: A country's industrial competitiveness and stock investors' return on capital (ROE) have never been the same set of accounts at the fundamental level. The underlying design of Western capital markets is a cash machine extremely biased toward shareholder returns. Even when US tech giants make money, they take large amounts of cash to buy back and cancel shares and pay dividends to the secondary market, prioritizing maintaining high return on equity and stock price bubbles. What they pursue is maintaining high profit margins, building high patent walls, and using capital premiums to reward those who hold shares. But the underlying logic of China's industrial narrative is an "industrial expansion machine" that pushes total factor productivity to the extreme. Once any industry is proven to have prospects, massive industrial capital, local policies, and startup teams immediately rush in; what everyone competes over is not who pays shareholders more dividends, but who can break through costs at the fastest speed, drive prices to the limit, and directly squeeze global competitors out of the market. This extreme cost-effectiveness and overcapacity have won the country global market share and won consumers extremely cheap goods, but what is left for the listed companies themselves is razor-thin profit margins and endless capital expenditure. The missions of stock markets in different systems have been diametrically opposite from day one: one is an asset pool for residents to accumulate wealth and enjoy capital compounding, the other is, to a large extent, a capital tool for supplying blood to the real economy, sharing innovation risks, and financing heavy industrial assets at low cost. When you put in your hard-earned savings, you think you are sharing the dividends of industrial upgrading, but in fact every cent of yours, as high-intensity R&D consumables and capacity subsidies, is silently being thrown into a super furnace that is reshaping the global industrial map. Only when you understand this huge structural misalignment will you understand where that depression comes from: the world is being reshaped by Made in China, but the contributors who created the manufacturing miracle have to endure the longest decoupling in the capital market.
View original
最近A股行情很不好,心情复杂,最近最大的一个感触是:我们这代股民最大的误解,是误以为国运等于股价。 《经济学人》最新一期文章也说中国股市是个罕见例外,提出了一个非常扎心刺骨的灵魂拷问: “如果这个国家正在赢得未来,那为什么它的股市投资者们却如此沮丧?” 咱们很多人看宏观叙事,总习惯把“产业打赢了”直接等同于“投资赚到钱了”。 过去几年,新能源、光伏、动力电池、智能硬件、AI 开源,无论从产能规模、出海速度还是技术迭代来看,中国产业链在全球都打出了堪称恐怖的统治力。 但在二级市场的 K 线图上,全球主要股市都在轮番刷出历史新高,唯独这里的股票市场成了一个全球罕见的痛苦例外。 这种反差撕开的,是所有股民和创业者必须补上的一堂硬核金融常识: 国家的产业竞争力,和股民的资本回报率(ROE),在底层从来不是同一套账本。 西方资本市场的底层设计,是一台极致偏向股东回报的提款机。 美股科技巨头哪怕赚了钱,也要把大笔现金拿去回购注销股票、分红给二级市场,优先维持高净资产收益率和股价泡沫, 它们追求的是维持高利润率、筑高专利墙,用资本溢价去奖励持股的人。 而中国产业叙事的底层逻辑,是一台把全要素生产率推向极致的“工业扩张机器”。 任何一个行业一旦被证明有前景,海量的产业资本、地方政策和创业团队就会瞬间涌入; 大家拼的不是谁给股东多分红,而是谁能以最快的速度把成本打穿、把价格卷到极限,把全球竞争对手直接挤出市场。 这种极致的性价比和产能过剩,给国家赢得了全球市场份额,给消费者赢得了极致便宜的商品,但唯独留给上市公司本身的,是薄如刀片的利润率和永无止境的资本开支。 股市在不同体系里的使命从第一天起就截然相反: 一个是拿来给居民沉淀财富、享受资本复利的资产池, 另一个在很大程度上,是承担为实体经济输血、分担创新风险、以低成本进行产业重资产融资的资本工具。 当你把辛苦积攒的积蓄投进去,你以为自己是在分享产业升级的红利, 实际上你的每一分钱,都在作为高强度的研发耗材和产能补贴,被无声砸进了重构全球工业版图的超级熔炉里。 看懂了这个巨大的结构性错位,你才会明白那种沮丧到底从何而来: 世界正在被中国制造重塑,但制造奇迹的功臣们,却不得不承受资本市场上最漫长的一场脱钩。
Body · AI translation
A single remark from Google CEO Sundar Pichai at the just-concluded earnings call instantly set the tech world ablaze: We've reached a stage where AI is helping us improve the speed of training and release. I expect the next twelve months to keep accelerating due to RSI. Many people's first reaction was sheer bewilderment: Is Google effectively making an official announcement of Recursive Self-Improvement?
I think equating a single executive's earnings-call remark directly with a technological singularity is most likely overreading it.
But in this kind of investor communication, which has always been meticulous about every detail, the fact that AI is now feeding back into training and release speed is something genuinely written into the minutes.
After all, what used to bottleneck large-model R&D was often not the ceiling of compute power, but the physical limits of human engineers cleaning dirty data, manually tuning parameters, and writing training scaffolding.
Now the giants are playing a different game—pouring the inference compute of the previous generation of mature models back into the R&D pipeline for the next generation of frontier models.
From automatically distilling high-quality synthetic corpora to automated code debugging and cluster scheduling, the entire R&D process is being restructured into a closed-loop, self-reinforcing flywheel.
What is truly spine-chilling about this is not some ethereal self-awareness at all, but the exponential compression of engineering iteration cycles.
When a company's model not only sells tokens externally but also ceaselessly optimizes its own next-generation model day and night, the generational gap in R&D efficiency will instantly widen.
So rather than rushing to argue online about whether the singularity has arrived, it is better to see the cruelest reality right now: The life-or-death line for large models in the next stage is no longer simply a contest of how many cards have been stacked, but of who can hand over their engineering R&D pipeline entirely to AI faster.