美国或正误入AI竞赛歧途
The US may be running the wrong AI race
译文简介
中国对DeepSeek这类小型廉价模型的青睐，可能会被证明是更明智的选择
正文翻译
题图。
Surrounded by kick-boxing, piano-playing humanoid robots at a high-tech fair in Shenzhen last month, some tech influencers were asking: can the west catch up with China?
上个月在深圳一场高科技展会上，当被拳击、弹钢琴的人形机器人环绕时，一些科技界意见领袖不禁发问：西方能否赶上中国？
That question would have sounded absurd two decades ago, but it is anything but today. This week, the Australian Strategic Policy Institute published its latest critical technology tracker, covering high-impact global research in 74 areas. It found that China now leads in 66 of those technologies, in fields as varied as computer vision, quantum sensors and nuclear energy, with the US ahead in the other eight.
这个问题若在二十年前提出会显得荒谬，但如今情况已截然不同。本周，澳大利亚战略政策研究所发布了最新的关键技术追踪报告，涵盖74个领域的全球高影响力研究。报告发现，中国目前在计算机视觉、量子传感器和核能等不同领域的66项技术中处于领先地位，美国则在其余8项技术中保持优势。
ASPI’s researchers highlighted a familiar story across many technologies. An early and overwhelming US lead in research output in the first decade of this century has been surpassed by China’s persistent long-term investment in fundamental research. In 2005, China accounted for just 6 per cent of the world’s most highly cited research papers but that share had risen to 48 per cent this year. The comparable proportion of US publications fell from 43 per cent to 9 per cent. At a time when the US is defunding many federal science programmes, China is doing the opposite by “building the whole technology ecosystem”, says Jenny Wong-Leung, one of the report’s authors.
该研究所的研究人员强调了一个在许多技术领域都相似的趋势：本世纪头十年美国在研究产出上的早期压倒性领先优势，已被中国对基础研究的长期持续投资所超越。2005年，中国仅占全球最高被引研究论文的6%，但今年这一比例已升至48%。而美国论文的相应比例则从43%下降至9%。报告作者之一黄梁詹妮指出，在美国削减许多联邦科学项目资金之际，中国正反其道而行，通过“构建完整的技术生态系统”持续投入。
ASPI’s findings accord with Nature’s latest ranking of research institutions, tracking articles across 145 science journals. In terms of research output, nine of the world’s top 10 research institutions are Chinese with only Harvard University in the top tier. China is now mass manufacturing research; it truly has become a scientific superpower.
澳大利亚战略政策研究所的研究结果与《自然》杂志最新研究机构排名相符，该排名追踪了145种科学期刊的文章。就研究产出而言，全球前10名研究机构中有9家来自中国，仅哈佛大学位列第一梯队。中国正在实现科研的规模化生产；它确实已成为一个科学超级大国。
Published research, though, does not automatically translate into technological capability. Moreover, the location of research expertise does not always map with successful commercialisation of technology — as a long line of frustrated British scientists can attest.
然而，已发表的研究成果并不会自动转化为技术能力。此外，研究专长的所在地并不总是与技术成功商业化相匹配——这一点许多受挫的英国科学家可以作证。
However, a separate report from the Special Competitive Studies Project in the US earlier this year also highlighted the striking progress that China has made in adopting many frontier technologies. According to the SCSP’s staff assessment, the US is still leading in semiconductors, synthetic biology and quantum computing while China dominates in advanced batteries, 5G and commercial drones. But the most contested, and arguably most consequential, area is artificial intelligence.
不过，美国特别竞争研究项目今年早些时候发布的一份独立报告也强调了中国在采用许多前沿技术方面取得的显著进展。根据该机构工作人员的评估，美国仍在半导体、合成生物学和量子计算领域保持领先，而中国则在先进电池、5G和商用无人机领域占据主导地位。但竞争最激烈、且可以说最具决定性意义的领域是人工智能。
President Donald Trump has said that the US will do “whatever it takes” to lead the world in AI. And the big US tech companies, including OpenAI, Alphabet, Microsoft, Meta and Amazon, are making colossal investments to fulfil that ambition. OpenAI alone is planning to invest $400bn over the next few years to build out its Stargate data centres across the US. Last month, the Trump administration launched the Genesis Mission to boost the private AI sector by sharing the public data sets and computing resources of the country’s 17 national laboratories. “We’re essentially pitting our private capitalists against this nation state of China. The stakeholders here have two very different sets of resources, attributes, strengths and weaknesses,” says David Lin, a senior adviser to the SCSP.
美国总统特朗普曾表示，美国将“不惜一切代价”在人工智能领域引领世界。包括OpenAI、Alphabet、微软、Meta和亚马逊在内的美国科技巨头正为实现这一雄心投入巨额资金。仅OpenAI就计划未来数年投资4000亿美元，在全美建设其“星际之门”数据中心。上月，特朗普政府启动“创世纪计划”，通过共享美国17个国家实验室的公共数据集和计算资源来推动私营人工智能领域发展。“我们实质上是在让本国私人资本与中国这个国家主体进行较量。相关利益方拥有截然不同的资源禀赋、属性特征及优势短板，”美国特别竞争研究项目高级顾问林戴维表示。
But the US and China are also adopting very different approaches to adopting AI. The big US companies mostly favour massive, proprietary, “closed-weights” models, such as ChatGPT and Gemini, which may be best suited to achieving generalisable intelligence. By contrast, Chinese AI companies favour smaller, cheaper (and arguably less safe) “open-weights” models, such as DeepSeek and Alibaba’s Qwen, that can be more readily adapted by developers. In part, China is making a virtue of necessity because US export restrictions have denied it access to the state-of-the-art silicon chips needed to build the most powerful foundation models. But it also reflects China’s priority in rapidly diffusing the technology.
但美中两国在人工智能应用路径上存在显著差异。美国科技巨头大多青睐ChatGPT和Gemini这类庞大、专有、“闭源权重”的模型，这类架构可能最适于实现通用智能。相比之下，中国人工智能企业更倾向于采用DeepSeek和阿里通义千问等更轻量化、成本更低（且安全性存疑）的“开源权重”模型，这类模型更便于开发者进行适配改造。这种差异部分源于中国的现实条件——美国出口管制使其无法获得构建最强大基础模型所需的尖端芯片，但同时也反映出中国优先考虑技术快速普及的战略取向。
Michael Power, the former global strategist of the investment firm Ninety One, reckons the US is making a “catastrophic strategic error” in betting so heavily on giant closed AI models. “China’s model is turning out to be far more effective in terms of usable compute in the real world,” Power tells me, especially considering the country’s lower energy costs. Even Sam Altman, OpenAI’s chief executive, has expressed his personal concern that “we have been on the wrong side of history here”.
投资管理公司NinetyOne的前全球策略师迈克尔-鲍尔认为，美国在封闭式大型人工智能模型上投入如此巨大，是在犯“灾难性的战略错误”。鲍尔告诉我，尤其是在考虑到中国较低的能源成本的情况下，“中国的模式在现实世界中的可用计算能力方面，结果证明要有效得多”。就连OpenAI的首席执行官萨姆-奥特曼也表达了他个人的担忧，即“我们站在了历史错误的一边”。
A recent study by MIT and Hugging Face found that Chinese open models have now overtaken comparable US models in terms of global adoption. Many US companies, including Airbnb, have become fans of the “fast and cheap” Qwen. In this critical area too the question arises: can the west catch up with China?
麻省理工学院与美国机器学习公司Hugging Face的近期研究发现，中国开源模型在全球采用率方面现已超越同类美国模型。包括爱彼迎在内的许多美国企业已成为“快速且廉价”的通义千问模型的拥趸。在这一关键领域同样浮现出这样的疑问：西方能否追上中国的步伐？
评论翻译
很赞 ( 16 )
收藏
I dont know if the the US per se , or rather Sam Altman. It seems his strategy evolves around becoming "too big too fail" and has enthralled many actors to dance to his tune. Time to cut him lose , in my opinion
我也不知道这锅该美国背，还是该山姆-奥特曼自己扛。感觉他的套路就是玩“大而不能倒”，还忽悠了一堆人跟着他的节奏走。要我说，是时候把他踢出局了。
@Gudwin
I suppose what MSM1992 was trying to say is that by spreading the tentacles of massive debt into so many 3rd party balance sheets OpenAI strategy is becoming “too big to fail” indeed, so that in case of bubble bursting, the Fed will be required to bail them out because failing to do that would trigger the domino effect...
我猜@MSM1992想说的是，OpenAI这波操作把巨额债务的触手伸进这么多第三方资产负债表里，战略上确实把自己搞成了“大而不能倒”——万一泡沫炸了，美联储都得下场捞人，不然多米诺骨牌一倒，谁也别想跑……
@tdal1moe
Altman is just another grifter.
奥特曼就是个江湖骗子罢了。
@Maru Kun
A particularly talented grifter - who started off saying what he was doing was so important for the good of humanity that he would do it on a non profit basis. And now look at the situation. The Chinese have a much better claim to that given they are at least open sourcing their models.
这哥们儿堪称顶级画饼大师——起家时吹得天花乱坠，说自己的事业关乎人类福祉，甚至要搞非营利。现在看看这局面？要论格局还得看中国，至少人家模型都直接开源了。
@JS1958
My old boss in China always used to say you have to invest in fundamental and basic research even though you don't know if it will commercialise in a product.
我之前在中国的老板总说，就算不知道能不能转化成产品，也得砸钱搞基础研究。
He went further and mused, we employ thousands of PHDs from around the world for pure research. I read the title of their research, and I don't understand it. Should I smother them with process, and reporting and accountability, or just keep them fed and stay out of their way. You don't manage experts undertaking first of a kind research, you take care of them and hope in 10 or 20 years they produce something we can build into a product.
他还琢磨得更深，说我们雇了全球几千个博士搞纯研究。我连他们研究课题的名字都看不懂。我是该拿流程、汇报和考核把他们压得喘不过气，还是管好后勤别瞎指挥？对搞前沿探索的专家不该管束，得供养——盼着十几二十年后，他们能折腾出点能产品化的东西。
They key for me was "10 or 20 years" ... Its a long game where the money and hope taps need to be left on.
对我来说关键是那句“十几二十年”……这属于长期投资，钱和希望这两根水管都得一直开着。
@hb
The tragedy of the US tech industry is the best and brightest are working hard to sell consumers more useless rubbish. Ad revenue is the prime directive and talent is wasted on manufacturing addiction.
美国科技圈的悲哀就是：最顶尖的人才都在拼命给消费者推销更多没用的垃圾。广告收入成了最高纲领，人才全浪费在制造用户成瘾上了。
@Liszt
And a lot of the best US STEM grads go off to Wall Street to become quants.
而且很多美国顶尖的理工科毕业生都跑去华尔街当量化分析师了。
@American Person
This reminds me of the famous Bell research lab, one of the most productive sand pits in history. How was it so successful? Nobody ever let the bean counters get in the way of the boffins. That’s how.
这让我想起了著名的贝尔实验室，史上最高产的创新摇篮。为什么它能这么牛？因为从来不让那些只会算账的会计去干扰搞科研的天才。就这么简单。
@ctgarvey101
It's turning out a bit like windows (paid openai) v Linux.(Free Deep seek) I use cloud US anthropic to write code and train a deepseek local LLM... Which in turn can now write code after fine tuning... Deepseek is free,sufficient for a whole pile of tasks and can execute locally on the gpu. I think a lot of folks are implementing like this. So they are using the closed premium model to improve the open model so that they never have to pay for the closed model again .keeping with Linux v Windows....whom has the most active servers on the Internet? I still need to have an openai sub.. But the day to day stuff will be done on deepseek local or similar.
这发展得有点像Windows（付费的OpenAI）对Linux（免费的DeepSeek）。我用云端的美国anthropic公司的Claude来写代码和训练本地部署的DeepSeek模型……调教完的本地模型现在也能写代码了。DeepSeek免费，一堆任务都能搞定，还能在本地GPU上跑。感觉好多人都在这么搞。他们用付费闭源模型来调教开源模型，以后就再也不用给闭源模型交钱了。继续用Linux和Windows类比……网上活跃服务器最多的是哪个系统？我暂时还得续费OpenAI订阅，但日常活儿肯定都用本地DeepSeek或者同类工具解决了。
@Eclair
Did I read somewhere that the current US administration is cutting research subsidies to Harvard?
我是不是在哪儿看到说，美国现政府要砍哈佛的研究经费补贴来着？
@Wreck Center
Harvard and many other schools. Gutting national science grants. Shifting to vibes-based research at NIH. Immigration policies forcing out some academics and scaring others from coming.
哈佛等名校遭殃，国家科研经费被砍，国立卫生研究院开始搞“玄学式研究”。移民政策逼走学者，吓得别人不敢来。
The problem is that DJT just isn't really that bright.
问题在于，川普这人脑子是真不太灵光。
Just as he can blow up low-level civilians in drug boats while pardoning a drug kingpin, he can claim to do whatever it takes to win in AI vs China, but then gleefully burn down the crown jewel research machine that got us our lead in the first place.
就像他能炸毁毒贩小艇却赦免大毒枭一样，嘴上喊着要不惜一切代价在AI领域赢过中国，转身就笑嘻嘻地把自己领先的科研王牌给砸了。
Who knows if China or the US has the winning AI approach. We're led by a felon following the path to the greatest y=f(self-enrichment, retribution) so we'll see where this leads.
谁知道中美谁的AI路线能赢呢？我们这里这领头的是个罪犯，只管走那条最大化“自肥+报复”的函数曲线，结局如何走着瞧呗。
@Outliar
I do find it ironic that US companies have opted for the route of protecting and profiting off their AI's intellectual property, which is being trained on the collective IP of humanity (stolen). At least the Chinese have been consistent on this matter.
美国公司一边用全人类的集体智慧成果（说难听点就是偷来的）训练自家AI，一边又忙着搞知识产权保护来赚钱，这操作真是够讽刺的。至少在这件事上，中国人倒是始终如一。
@CaptainThunder
Having attempted to run a team building AI, I noted that American and British researchers tended to chase the “big problems” and are easily distracted by trying to solve largely academic challenges. Trying to get them to focus on creating a product - even the relatively simple use cases we had on the spec that would pay the bills - was a problem.
我之前试着带过AI团队，发现英美研究员总爱追那些“大问题”，动不动就被纯学术挑战带偏了。想让他们专注做能赚钱的产品——哪怕是需求文档里那些相对简单的用例——简直难如登天。
From a lot of the highly speculative use cases I see from Western devs, especially around generative AI, I’m getting the feeling that’s a general trend. Build the tech first, find the application later.
看多了西方开发者那些天马行空的用例（尤其是生成式AI方向的），感觉这都快成行业通病了。先闭门造车搞技术，产品落地随缘。
Whereas the Chinese approach described in this article - use fairly simple base models to build applications, and presumably evolve them to meet the use cases when they fall short - seems better both for stable, iterative development, and also helping dev teams keep the wolf from the door without quite such epic levels of cash burn.
相比之下，这篇文章里提到的中国模式——用现成的基础模型快速搭应用，不够用了再迭代升级——反而更靠谱。既能稳扎稳打推进项目，又能让团队在烧完史诗级预算前先吃上饭。
@Fair is foul, foul is fair
A few years ago, if you asked which country, US or China, is suppressing free speech critical of GENOCIDE and defunding universities, it would have been an obvious answer. No longer.
几年前，要是问你哪个国家在YA制关于种族MJ的批评言论、削减大学经费，美国还是中国，答案本来明摆着。现在情况可不一样了。
@GrayCrane
When I think about US this old joke comes to my mind: in the race to fly into space US invested millions and millions or dollars to create a pen that can write into space. The Russians used a pencil.
每次想到美国，我脑子里就蹦出那个老段子：为了太空竞赛，美国砸了几百万美元研发能在太空写字的笔，结果俄罗斯人直接掏出了铅笔。
@Madivan
Quite an astute observation. This is what happens when you have too much money to spend. It also highlights the issue that all this money only exists digitally. It can evaporate in the blx of an eye, so those that have access to it need to invest it before that happens. When it does, we the people are going to bail you out, no?
眼光够毒辣啊。这就是钱多得没处花的结果。而且这也暴露了一个问题：这些钱全是数字，说没就没，所以有钱人得赶在泡沫破灭前赶紧投资。等真崩盘了，最后还不是得我们公民来兜底，对吧？
@GoodfightGoodnight
And the Russians realised why the Americans stopped using pencils in space… graphite particles floating around caused circuits to short out, and were a fire hazard. This story still circulates decades later but it’s nonsense.
俄罗斯人才明白为什么美国人上太空不用铅笔了……石墨颗粒在舱内乱飘，容易造成电路短路，还容易引发火灾。这故事都传了几十年了还在流传，纯属瞎扯淡。
@Herethereeverywhere
China has already won. Better get used to it.
中国早就赢麻了，早点习惯吧。
@Another Risk Manager
Is that why it's economy is smaller? What has China invented the past few centuries? just curious
难道这就是它经济规模更小的原因吗？中国过去几百年发明了什么？纯好奇
@Sa88
At the minute the West has Nvidia chips and that's about it. Chinese are dominating in solar, batteries, EVs, nuclear, manufacturing - you name it. Even in defence - cheap Chinese jets downed French ones.
眼下西方也就剩英伟达芯片能拿得出手了。中国在太阳能、电池、电动车、核电、制造业这些领域简直杀疯了——要什么有什么。连军工都卷，便宜的中国战机都把法国货给打下来了。
As for what China invented. Why focus on the last few centuries ? What was the West doing for the last 1,500 years while China was inventing gunpowder, paper money, printing, compass, ocean faring vessels.
至于中国发明过什么？干嘛老盯着最近几百年？过去一千五百年西方在躺平的时候，中国早就搞出了火药、纸币、印刷术、指南针和远洋巨轮。
That focus on the last few centuries is blinding the West to China's rise as the technological powerhouse.
西方这种只盯着近代史的毛病，让他们完全看不清中国正在崛起成科技巨无霸的事实。
@Orcadian1
What’s also interesting is Ukraine - a European country - is now manufacturing, at least in terms of defence. From small seeds grow mighty trees when the chips are down.
还有个挺有意思的点是，乌克兰——一个欧洲国家——现在也开始搞制造业了，至少在国防领域是这样。这就是绝境之下小树苗也能长成参天大树。
@VS
"What has China invented ?" is the new "what have the romans ever done for us?"
“中国发明了什么？”简直跟“罗马人到底给咱们干过什么好事？”一个模子刻出来的。
@R0cketman
Actually ‘invented’ very little recently. It’s why they still bang on about gunpowder like we bang on about the ‘66 World Cup. Copied, scaled, lowered cost, often improved. Very little true invention. But that’s fine, all the ingredients are there to lead on this too. The time is coming.
说实话，最近真没搞出什么原创发明。所以他们才老把火药挂嘴边，就跟我们英国人天天念叨66年世界杯夺冠一个德行。基本都是抄袭、放大规模、压低成本，有时候还改良得不错。但真正的从零到一？少得可怜。不过也没什么，该有的底子都在，要引领这波也不是不行。时候快到了。
@Paul Thind
The insecurities about China advancing in tech now preoccupy the western mind. How sad is that?
西方现在满脑子都是对中国科技进步的焦虑，真是可悲。