美国酸成柠檬精？咨询机构炒作 “中国开源 AI 威胁论”，网友吐槽：双标玩得真溜！
正文翻译
BEIjinG, March 23 (Reuters) - The dominance of China's open-source artificial intelligence is creating a "self-reinforcing competitive advantage", allowing it to challenge U.S. rivals despite restricted access to advanced AI chips, a U.S. congressional advisory body said on Monday.
北京，3月23日（路透社）——美国国会下属的美中经济与安全审查委员会（USCC）周一表示，尽管获取先进人工智能芯片受到限制，但中国开源人工智能的主导地位正形成“自我强化的竞争优势”，使其有能力挑战美国竞争对手。
Driven by their cheaper cost, Chinese large language models from firms including Alibaba (9988.HK), Moonshot and MiniMax (0100.HK) now dominate worldwide usage rankings on platforms like HuggingFace and OpenRouter.
阿里巴巴（港股代码：9988.HK）、月之暗面 Moonshot、稀宇科技 MiniMax（港股代码：0100.HK）等中国企业推出的大语言模型，凭借更低的成本优势，目前已在 HuggingFace、OpenRouter 等平台的全球使用量排行榜中占据主导地位。
Beijing's push to deploy AI throughout a wide range of sectors to upgrade its manufacturing base, factories, logistics networks and robotics is generating real-world data that feeds back into model improvement, the report said.
报告称，中国推动在众多行业部署人工智能，以升级其制造业基础、工厂、物流网络和机器人技术，这一举措正产生真实数据，反过来为模型优化提供支持。
"This open ecosystem enables China to innovate close to the frontier despite significant compute constraints," the U.S.-China Economic and Security Review Commission wrote in a report published on Monday.
“尽管面临严重的算力限制，但这种开放生态系统使中国能够在接近技术前沿的领域进行创新，”美中经济与安全审查委员会在周一发布的报告中写道。
"Chinese labs have narrowed performance gaps with top Western large language models," it added.
该委员会补充道：“中国实验室已缩小了与西方顶尖大型语言模型的性能差距。”
U.S. lawmakers have imposed successive rounds of export restrictions on China since 2022, banning them from acquiring the most advanced AI chips, though Washington approved exports of Nvidia's (NVDA.O) second-most advanced chip in December.
自2022年以来，美国立法者已对中国实施了多轮出口限制，禁止中国获取最先进的人工智能芯片，但美国政府在12月批准了英伟达（NVDA.O）第二梯队高端芯片的对华出口。
U.S. companies including ChatGPT developer OpenAI and Anthropic, creator of Claude, as well as traditional tech giants have, meanwhile, invested billions of dollars to remain at the forefront of the new technology.
与此同时，包括ChatGPT开发商OpenAI、Claude开发商Anthropic以及传统科技巨头在内的美国企业，已投入数十亿美元以保持在这项新技术领域的领先地位。
But their position could be under threat.
但它们的地位可能正受到威胁。
"Open model proliferation creates alternative pathways to AI leadership," the report said.
报告称：“开源模型的普及为人工智能领导地位开辟了替代路径。”
China POISED TO CAPITALISE ON SHIFT TO EMBODIED AI
中国有望借人工智能向具身化转型之机抢占先机
Some estimates suggest that around 80% of U.S. AI startups now use Chinese open-source AI models.
部分估算显示，目前约80%的美国人工智能初创企业在使用中国的开源人工智能模型。
DeepSeek's groundbreaking R1 model launched last year quickly overtook ChatGPT as the most downloaded model on the U.S. App Store.
深度求索（DeepSeek）去年推出的突破性R1模型迅速超越ChatGPT，成为美国App Store下载量最高的模型。
And Alibaba's Qwen family of models has surpassed Meta's (META.O) Llama in global cumulative downloads, according to HuggingFace.
HuggingFace数据显示，阿里巴巴的通义千问（Qwen）系列模型的全球累计下载量已超过Meta（META.O）的 Llama 模型。
As AI's frontiers shift from large language models towards agentic AI and physical, or embodied, AI, China may be better positioned to capitalise on its mass data collection efforts to boost development of humanoid robots, autonomous driving software or even dual-purpose technologies, the report said.
报告称，随着人工智能的前沿领域从大型语言模型转向智能体人工智能和具身智能（实体 AI），中国可能更有优势利用其大规模数据收集优势，推动人形机器人、自动驾驶软件乃至军民两用技术的发展。
"There's a bit of a deployment gap in the embodied AI space between the U.S. and China. That's something that over time compounds itself ... We're starting to see that compounding now," Michael Kuiken, the commission's vice-chair, told Reuters in an interview.
该委员会副主席迈克尔·奎肯在接受路透社采访时表示：“在具身人工智能领域，美中之间存在一定的部署上的差距。这种差距会随着时间推移不断扩大……我们现在已经开始看到这种扩大的趋势。”
The commission is also watching how China uses AI in sectors like biotech, quantum computing and advanced materials, he added.
他补充道，该委员会还在关注中国在生物技术、量子计算和先进材料等领域对人工智能的应用情况。
Beijing has designated embodied AI as a core future strategic industry, and many leading Chinese humanoid robotics firms plan public listings this year.
中国已将具身人工智能列为未来核心战略产业，多家中国领先的人形机器人企业计划今年上市。
Despite warnings from some Western research organisations of the potential security risks of an over-reliance on Chinese open-source AI models and their political bias towards Chinese government positions, many companies are adopting them anyway.
尽管部分西方研究机构警告称，过度依赖中国开源人工智能模型存在潜在安全风险，且这些模型可能对中国政府立场存在政治偏向，但许多企业仍在采用它们。
Siemens (SIEGn.DE) CEO Roland Busch said earlier on Monday that there were "no disadvantages" to using Chinese open-source AI to train the German company's AI models specialised for industrial automation, citing their cost advantage and ease of customising parameters.
德国西门子公司（SIEGn.DE）首席执行官罗兰·布希周一早些时候表示，使用中国开源人工智能训练该公司专为工业自动化设计的人工智能模型“没有劣势”，并指出中国开源模型具有成本优势且参数易于定制。
BEIjinG, March 23 (Reuters) - The dominance of China's open-source artificial intelligence is creating a "self-reinforcing competitive advantage", allowing it to challenge U.S. rivals despite restricted access to advanced AI chips, a U.S. congressional advisory body said on Monday.
北京，3月23日（路透社）——美国国会下属的美中经济与安全审查委员会（USCC）周一表示，尽管获取先进人工智能芯片受到限制，但中国开源人工智能的主导地位正形成“自我强化的竞争优势”，使其有能力挑战美国竞争对手。
Driven by their cheaper cost, Chinese large language models from firms including Alibaba (9988.HK), Moonshot and MiniMax (0100.HK) now dominate worldwide usage rankings on platforms like HuggingFace and OpenRouter.
阿里巴巴（港股代码：9988.HK）、月之暗面 Moonshot、稀宇科技 MiniMax（港股代码：0100.HK）等中国企业推出的大语言模型，凭借更低的成本优势，目前已在 HuggingFace、OpenRouter 等平台的全球使用量排行榜中占据主导地位。
Beijing's push to deploy AI throughout a wide range of sectors to upgrade its manufacturing base, factories, logistics networks and robotics is generating real-world data that feeds back into model improvement, the report said.
报告称，中国推动在众多行业部署人工智能，以升级其制造业基础、工厂、物流网络和机器人技术，这一举措正产生真实数据，反过来为模型优化提供支持。
"This open ecosystem enables China to innovate close to the frontier despite significant compute constraints," the U.S.-China Economic and Security Review Commission wrote in a report published on Monday.
“尽管面临严重的算力限制，但这种开放生态系统使中国能够在接近技术前沿的领域进行创新，”美中经济与安全审查委员会在周一发布的报告中写道。
"Chinese labs have narrowed performance gaps with top Western large language models," it added.
该委员会补充道：“中国实验室已缩小了与西方顶尖大型语言模型的性能差距。”
U.S. lawmakers have imposed successive rounds of export restrictions on China since 2022, banning them from acquiring the most advanced AI chips, though Washington approved exports of Nvidia's (NVDA.O) second-most advanced chip in December.
自2022年以来，美国立法者已对中国实施了多轮出口限制，禁止中国获取最先进的人工智能芯片，但美国政府在12月批准了英伟达（NVDA.O）第二梯队高端芯片的对华出口。
U.S. companies including ChatGPT developer OpenAI and Anthropic, creator of Claude, as well as traditional tech giants have, meanwhile, invested billions of dollars to remain at the forefront of the new technology.
与此同时，包括ChatGPT开发商OpenAI、Claude开发商Anthropic以及传统科技巨头在内的美国企业，已投入数十亿美元以保持在这项新技术领域的领先地位。
But their position could be under threat.
但它们的地位可能正受到威胁。
"Open model proliferation creates alternative pathways to AI leadership," the report said.
报告称：“开源模型的普及为人工智能领导地位开辟了替代路径。”
China POISED TO CAPITALISE ON SHIFT TO EMBODIED AI
中国有望借人工智能向具身化转型之机抢占先机
Some estimates suggest that around 80% of U.S. AI startups now use Chinese open-source AI models.
部分估算显示，目前约80%的美国人工智能初创企业在使用中国的开源人工智能模型。
DeepSeek's groundbreaking R1 model launched last year quickly overtook ChatGPT as the most downloaded model on the U.S. App Store.
深度求索（DeepSeek）去年推出的突破性R1模型迅速超越ChatGPT，成为美国App Store下载量最高的模型。
And Alibaba's Qwen family of models has surpassed Meta's (META.O) Llama in global cumulative downloads, according to HuggingFace.
HuggingFace数据显示，阿里巴巴的通义千问（Qwen）系列模型的全球累计下载量已超过Meta（META.O）的 Llama 模型。
As AI's frontiers shift from large language models towards agentic AI and physical, or embodied, AI, China may be better positioned to capitalise on its mass data collection efforts to boost development of humanoid robots, autonomous driving software or even dual-purpose technologies, the report said.
报告称，随着人工智能的前沿领域从大型语言模型转向智能体人工智能和具身智能（实体 AI），中国可能更有优势利用其大规模数据收集优势，推动人形机器人、自动驾驶软件乃至军民两用技术的发展。
"There's a bit of a deployment gap in the embodied AI space between the U.S. and China. That's something that over time compounds itself ... We're starting to see that compounding now," Michael Kuiken, the commission's vice-chair, told Reuters in an interview.
该委员会副主席迈克尔·奎肯在接受路透社采访时表示：“在具身人工智能领域，美中之间存在一定的部署上的差距。这种差距会随着时间推移不断扩大……我们现在已经开始看到这种扩大的趋势。”
The commission is also watching how China uses AI in sectors like biotech, quantum computing and advanced materials, he added.
他补充道，该委员会还在关注中国在生物技术、量子计算和先进材料等领域对人工智能的应用情况。
Beijing has designated embodied AI as a core future strategic industry, and many leading Chinese humanoid robotics firms plan public listings this year.
中国已将具身人工智能列为未来核心战略产业，多家中国领先的人形机器人企业计划今年上市。
Despite warnings from some Western research organisations of the potential security risks of an over-reliance on Chinese open-source AI models and their political bias towards Chinese government positions, many companies are adopting them anyway.
尽管部分西方研究机构警告称，过度依赖中国开源人工智能模型存在潜在安全风险，且这些模型可能对中国政府立场存在政治偏向，但许多企业仍在采用它们。
Siemens (SIEGn.DE) CEO Roland Busch said earlier on Monday that there were "no disadvantages" to using Chinese open-source AI to train the German company's AI models specialised for industrial automation, citing their cost advantage and ease of customising parameters.
德国西门子公司（SIEGn.DE）首席执行官罗兰·布希周一早些时候表示，使用中国开源人工智能训练该公司专为工业自动化设计的人工智能模型“没有劣势”，并指出中国开源模型具有成本优势且参数易于定制。
评论翻译
@sentencevillefonny
The Software as a Service frenzy has been hurting my pockets for a decade. Can’t buy, only rent…and everything comes coupled with ads, bloat and the useful features paywalled behind the tier right above the one you just paid for. I’m all for any alternatives, especially open source.
软件即服务（SaaS）的热潮已经让我钱包受损十年了。只能租不能买…… 而且所有服务都附带广告、冗余功能，有用的功能还藏在你刚付费的套餐之上的更高层级里，得再花钱解锁。我完全支持任何替代方案，尤其是开源的。
@Agitated-Ad-504
Currently experiencing that hellscape at work. Upper management decided they will ban all AI except Grok. Almost certain because it was contracted. Absolutely fucking brain dead. (I’m in Aerospace) that is supposed to be somehow restricted in what data it saves (or at least that's what they claim). Don't know if that's the case with Grok, but MS Copilot has a version specifically for enterprise use. Banning AI at workplaces is just a smart business decision because if you don't, some idiot will paste confidential information (or worse, personal data) into ChatGPT to ask it for help.
我现在正经历着那种地狱般的工作环境。高层决定禁止除 Grok 之外的所有 AI，几乎可以肯定是因为签了合同，简直蠢透了。（我在航空航天领域工作）据说这款 AI 在数据保存方面有某种限制（至少他们是这么宣称的）。不知道 Grok 是不是这样，但微软 Copilot 有专门针对企业用户的版本。不过职场禁止 AI 也算是个明智的商业决策 —— 不然真会有傻瓜把机密信息（甚至更糟的个人数据）粘贴到 ChatGPT 里寻求帮助。
@CtrlAltEvil
This is why I love Linux. There’s an open source alternative for practically everything. Plus none of the bloat.
这就是我喜欢 Linux 的原因。几乎所有东西都有开源替代方案，而且完全没有冗余垃圾功能。
@PauseLost2137
There’s an open source alternative for practically everything. cries in Adobe
几乎所有东西都有开源替代方案。在 Adobe的付费墙前哭泣
@therealowlman
If something good comes out of AI, the democratization of writing software is one of them.
如果人工智能能带来什么好事，软件编写的民主化绝对是其中之一。
@ChainsawBologna
That will be short-lived until they figure out how to put gates in to prevent you from writing software, and raise the price of tokens so high that you can't afford to spend them on writing software because you need them to hunt for affordable food. But the bubble is about to pop anyway. AI fatigue popping up all over. Compact portable models that run on-device for specific tasks is the real future. The monolithic datacenter model is failing faster than it can get going. No idea why the Tech Bros really thought people would be willing to spend $100/mon to chat with a graphics card in Altoona, Iowa to get work done.
这种民主化状态不会持续太久 —— 直到他们想出办法设置门槛阻止你编写软件，再把代币价格抬到天上去，让你根本没钱用来写软件，因为你得靠这些钱买得起食物。但 AI 泡沫终究要破裂的，现在到处都能感受到 AI 疲劳。能在设备上运行、针对特定任务的小型便携模型才是真正的未来，那种庞大的单体数据中心模式的失败速度比它发展的速度还快。真搞不懂那些科技大佬为啥会觉得，人们愿意每月花 100 美元，跟爱荷华州阿尔图纳市的一块显卡聊天来完成工作。
@swagonflyyyy
Open Source models are getting smaller and better over time, but to run the best ones you still need pretty expensive hardware, as those models are the only ones good for coding. The rest can automate simple tasks or be used for simple analysis or entertainment purposes, but anything more than that you need either the cloud or lots of hardware for large open source models. And if you are serious about running them locally, you need to switch to Linux. Period. Anything less than that will only hold you back.
开源模型正变得越来越小、越来越好，但要运行最顶尖的开源模型，仍然需要相当昂贵的硬件 —— 因为只有这些模型才适合编程。其他的开源模型只能自动化简单任务，或者用于简单的分析和娱乐，但如果要做更复杂的事，大型开源模型要么需要云端支持，要么需要大量硬件。而且如果你真的想在本地运行它们，就得切换到 Linux 系统，就是这么简单。低于这个标准的设备只会拖累你。
@CoastingUphill
Yes that’s why China keeps releasing those models
没错，这就是中国一直在发布这些开源模型的原因。
@CanadianPropagandist
Good! This push by Chinese AI companies to distribute very capable models under some extremely generous licensing (MIT / Apache) is one of the few things that will save AI tech from being completely gated by the same techbros that want to dominate entirely.
太好了！中国 AI 公司推动以极其慷慨的授权（如 MIT/Apache 协议）分发功能强大的模型，这是少数能避免 AI 技术被那些想完全垄断的科技大佬彻底封锁的举措之一。
@YumariiWolf
I had no idea 80% of AI startups in the US are using Chinese models, or that deepseeks R1 overtook chatGPT in US downloads on the App Store last year. Seems pretty plausible they could close the gap and overtake US companies.
我完全不知道美国 80% 的 AI 初创公司都在使用中国的模型，也不知道去年 Deepseek R1 在 App Store 的美国下载量居然超过了 ChatGPT。他们很有可能缩小差距，甚至超越美国公司，这看起来合情合理。
@ElvisDumbledore
"The people in charge" use anything they can to make you need them (heroin, oil, proprietary software, etc.). It's all about certain people keeping themselves at the top of the social hierarchy.
“掌权者” 会用尽一切手段让你依赖他们（比如海洛因、石油、专有软件等等）。这一切都是为了让某些人牢牢占据社会阶层的顶端。
@Swimming_Cover_9686
US AI lead is dumb throwing insane amounts of money at a problem that more compute really can't solve. By the time they plan on their investments paying out their hardware will be obsolete. It mathematically cannot work.
美国所谓的 AI 领先地位简直是蠢得可笑 —— 在一个更多计算能力根本无法解决的问题上砸下巨额资金。等他们盼着投资回本的时候，所用的硬件早就过时了。从数学角度来看，这根本行不通。
@Rich_Housing971
The reason they're against AI is precisely because they're closed-source models completely controlled by corporations only to extract more wealth from people. They are pro open source AI because those can be used by scientists and the average person.
人们反对某些 AI，恰恰是因为它们是闭源模型，完全被企业控制，目的只是从普通人身上榨取更多财富。而人们支持开源 AI，是因为这些 AI 能被科学家和普通人自由使用。
@blksilksheetz
its innovation when it’s cheaper and gets the job done for much less, hope this helps.
当一项技术更便宜、能以更低成本完成工作时，这就是创新 —— 希望这点能帮你理清思路。
@Fickle_Competition33
Reddit is funny, they go all rage against AI, but then when placed as US vs China AI, all of a sudden it becomes a good thing... The ideological relativism is strong here.
Reddit 真是太有意思了 —— 平时对 AI 骂声一片，但一旦把话题变成 “美国 AI vs 中国 AI”，AI 突然间就成了好事…… 这里的意识形态相对主义真是根深蒂固。
@Signal_Flight_7262
Love to see it
很乐见这种情况。
@Neither-Phone-7264
so that means that the US needs to start making open models to beat them!!! right????
所以这意味着美国得开始做开源模型才能打败他们啊！！！对吧？？？
@Material_Policy6327
“We will now pass a law making open source ai illegal!”
“我们现在要通过一项法律，把开源 AI 定为非法！”
@SGmoze
"I'm going to open-source Chinese models even harder now"
“现在我要更卖力地推广中国开源模型了。”
@progfix
Why not just all open source software!
为什么不干脆让所有软件都开源呢！
@IngwiePhoenix
"And it will be tremendous! We will show those bad people how we do it over here and how it's really done, then they will maybe reconsider and be nicer." Read that in a particular voice. :p
“这将会超棒的！我们要让那些坏人看看我们是怎么做事的，看看真正的正确做法，到时候他们可能会重新考虑，变得友善一点。”模仿特定语气读完这段话 :p）
@NeverLookBothWays
Nah we’re like Russia…eventually we’ll lose the brain share. We’re already behind on a lot of fronts, and where we could have accepted and learned from Chinese tech, we have instead siloed off and regressed, kicking out a lot of incredibly talented minds in the process. Meanwhile China is building a solid foundation that offsets energy costs with renewables. They are entering a Bronze Age while we’re still embracing stone.
不，我们现在就像俄罗斯…… 最终会失去人才优势的。我们在很多领域已经落后了，本可以接受并向中国技术学习，结果却自我孤立、倒退，还在这个过程中赶走了很多极具才华的人。与此同时，中国正在搭建坚实的基础，用可再生能源抵消能源成本。他们都要进入青铜时代了，我们却还在抱守石器时代的思维。
@broknbottle
To be honest, the US is carrying a lot of dead weight right now, like the Baby Boomers who refuse to let go of power, would rather die in office than see the next generation succeed.
说实话，美国现在背负着很多累赘，比如那些拒绝放权的婴儿潮一代，他们宁愿死在权力岗位上，也不愿看到下一代成功。
@CertainMiddle2382
Being one foot behind the lead, is much much much easier than being in the lead. As you know what you are trying is already working.
落后领先者一步，比保持领先要容易得多得多。因为你知道自己正在尝试的东西已经被验证是可行的。
@EffectiveCeilingFan
"Lead"? Pff. US is getting crushed in open weights, not even a competition. Absolute peak vibecoding performance is not the only metric. The Chinese stuff is cheaper, too. Not to mention, all I hear about recently is how dysfunctional Opus, GPT-5.4, and Gemini 3.1 Pro have been.
“领先”？呸。美国在开源模型领域被碾压得毫无还手之力，连像样的竞争都算不上。所谓的 “氛围编码” 根本不是唯一的衡量标准，而且中国的模型还更便宜。更别提最近我听到的全是 Opus、GPT-5.4 和 Gemini 3.1 Pro 有多拉胯的吐槽了。
@blackkettle
The real problem is that the US administration has absolutely no clue - none, not even a shadow of a - not even a concept of a plan about what to do. The admin is staunchly anti-intellectual from the top all the way down. There is no chance of 'competing'. None. Their response will most likely be 'tariffs' or something equally idiotic and when that immediately fails, they'll instead just choose to 'ban' the foreign models in the name of 'national security'. Decline upon decline.
真正的问题是，美国政府根本毫无头绪 —— 完全没有，连一丝一毫的计划概念都没有。从高层到基层，这个政府都坚决反智，根本没有 “竞争” 的可能，一点都没有。他们的回应很可能是 “加征关税” 之类同样愚蠢的操作，等这些措施立刻失败后，就会以 “国家安全” 为由 “禁止” 外国模型。真是一路衰败到底。
@sine120
Yeah, GPT-OSS held its ground for longer than most open models do, so I'll give it credit, but OpenAI was pretty much the only US company putting out something more than a toy. Credit to Llama for dominating for as long as it did, and other players like Granite and Phi were interesting to play with, but they were toys, in reality. I work in aerospace and have friends who work in other sensitive areas, they have need for local models to run internally on air-gapped networks, they all turn to Chinese models. Last year it was Deepseek and then GLM, this year there's a whole wealth of great options. Frontier models are still the best, don't get me wrong. I wouldn't trust open models, as good as they are, to my real work yet, but the gap went from feeling like frontier models were "twice as smart" last year to "maybe 10% better".
没错，GPT-OSS 比大多数开源模型坚持的时间要长，这点我得承认，但 OpenAI 几乎是唯一一家美国公司能拿出点超越 “玩具级” 的产品。Llama 能主导这么久确实值得肯定，Granite 和 Phi 这类产品玩起来也挺有意思，但实际上它们都只是玩具而已。我在航空航天行业工作，有朋友在其他敏感领域，他们需要能在内部隔离网络上运行的本地模型，现在全都会选中国模型。去年是 Deepseek 和 GLM，今年更是有一大堆优质选择。别误会，前沿闭源模型依然是最棒的，我还不会把开源模型（虽然已经很棒了）用在真正的工作中，但差距已经从去年感觉 “前沿模型聪明一倍” 缩小到了 “可能只比开源模型好 10%”。
@PureSignalLove
It's flipping for me right now. It's to the point where I don't trust frontier models due to their inconsistency. Opus 4.6 is great when it works, but when it goes lazy/stupid, it's less than useless.
我现在的看法已经反转了。因为前沿模型太不稳定，我已经不信任它们了。Opus 4.6 能用的时候确实厉害，但一旦变得懒散或犯蠢，就比没用还糟。
@sine120（回复 @PureSignalLove）
The nature of my work, I don't use them to write a ton of production code for me yet, just because investigating and reviewing is 10x more important than writing, and I needed as much usable context as I could get, but I've been doing side projects with Gemini CLI and do notice on days it just gets lazy or stupid. I suspect as inference capacity gets saturated, they deploy quantized models or somehow tune behavior to stop being so independent. Ultimately, local models will be the only way to guarantee you know what you're going to get. Once I don't get laughed out of the room for pitching we buy 1TB of RAM or VRAM, then I'll get serious about switching.
就我的工作性质而言，我还不会用它们写大量生产代码 —— 因为调研和审核比编写重要 10 倍，而且我需要尽可能多的可用上下文。但我一直在用 Gemini CLI 做副业项目，确实发现有些时候它会变得懒散或犯蠢。我怀疑是随着推理算力饱和，他们部署了量化模型，或者通过某种方式调整了行为，让它不再那么独立。归根结底，本地模型是唯一能保证你 “知道自己会得到什么结果” 的方式。等我提议买 1TB 内存或显存不会被人嘲笑的时候，我就会认真考虑切换到本地模型了。
@CavulusDeCavulei（回复 @sine120）
I don't work in a field as cool as yours and I am still a junior/mid engineer, but I tried to advocate to buy a gpu and use local models and I got basically laughed on because it's all on cloud now. I feel you.
我不在你这么酷的领域工作，还只是个初级 / 中级工程师，但我之前试着提议买显卡用本地模型，结果基本上被嘲笑了 —— 因为现在大家都用云端。我懂你的感受。
@Zc5Gwu
I think price is insanely good for open models, moreso even than a lot of frontier models. Look at rebench, Step-3.5-Flash $0.14 per problem at 59% versus Claude Sonnet 4.6 $1.02 per problem at 60.7%. 10x price difference for negligible accuracy difference.
我觉得开源模型的价格真的划算到离谱，甚至比很多前沿模型还划算。看看 Rebench 的数据：Step-3.5-Flash 每题 0.14 美元，准确率 59%；而 Claude Sonnet 4.6 每题 1.02 美元，准确率才 60.7%。价格差 10 倍，准确率却没差多少。
@sine120（回复 @Zc5Gwu）
For the price, open models are definitely best, I just need approved vendors for my job, and most open providers are not approvable.
单看价格，开源模型绝对是最优选择，但我的工作要求必须用经过批准的供应商，而大多数开源模型提供商都通不过审批。
@jld1532
Good thing we're getting a push for NFS funded American open weight models.
幸好现在有人在推动由 NFS（正确缩写NSF美国国家科学基金会） 资助的美国开源模型，算是个好消息。
@jld1532
Kimi K2.5 is legitimately good enough for me to use in my work. Not as a full human replacement by a long shot, but it's good enough that the thought of paying for AI is laughable. My university hosts it on the cluster free for all faculty and students. As soon as enough people realize this, for-profit AI is dead.
Kimi K2.5 对我来说真的足够好用，完全能用到工作中。虽然远远不能替代人类，但已经好到 “花钱买 AI 服务” 这种想法都显得可笑。我的大学在集群上部署了它，所有师生都能免费使用。一旦足够多的人意识到这一点，营利性 AI 就完了。
@autoencoder
In my non-agentic testing, Qwen 3 ripped it to shreds. It would go loopy thinking, burning tokens uselessly to me. It did not care about any thinking directives, it just thought excessively.
在我的非智能体测试里，Qwen 3 直接把 GPT-OSS 吊打。GPT-OSS 经常陷入无意义狂想，纯纯浪费 token，根本不听指令，就是没完没了瞎想。
@Fastpas123
Agreed, the us is struggling, Gemini was incredible back in 2.5 days but 3.1 is a broken joke.
同意，美国现在确实很挣扎。Gemini 2.5 版本的时候还挺厉害，但 3.1 版本简直是个烂笑话。
@procgen
5.4 in codex has been an absolute workhorse for me lately. Haven't found anything that comes close to replacing it.
最近 Codex 里的 GPT-5.4 简直是我的主力工具，还没找到能接近替代它的东西。
@Free-Combination-773
US is not crushed in open weights competition, it just didn't show up.
美国在开源模型竞争中不是被碾压了，只是根本没参与而已。
@TheMericanIdiot
It’s open source… you know the same way US built its tech sector…. The stupid have taken over.
这可是开源啊…… 美国当年就是靠这个建立起科技产业的啊…… 现在蠢货们掌权了。
@Lissanro
The issue is keeping things closed may give few months advantage over competitors but overall slows the research down. OpenAI likely wouldn't even exist if there was no the Attention is All You Need paper. DeepSeek and their release that was also accompanied by the detailed research paper about its architecture and training... That's what lays down the foundation to further build upon. Currently I prefer to run mostly Kimi K2.5 on my rig, but it would not even exist if DeepSeek did not share their research and architecture. It seems even large companies prefer it... For example Cursor AI picked Kimi K2.5 as a base model for their Composer 2. But then again, what else is there to pick in the larger size besides a few other top Chinese models? Obviously that's a rhetorical question.
问题在于，闭源可能会让你比竞争对手多几个月的优势，但总体上会拖慢整个研究的进度。如果没有《Attention is All You Need》那篇论文，OpenAI 可能根本就不会存在。而 DeepSeek 发布模型时，还附带了关于其架构和训练过程的详细研究论文…… 这才是为后续发展奠定基础的做法。目前我更喜欢在自己的设备上运行 Kimi K2.5，但如果 DeepSeek 没有分享他们的研究和架构，Kimi K2.5 可能也不会存在。看来连大公司都青睐中国开源模型…… 比如 Cursor AI 就选了 Kimi K2.5 作为他们 Composer 2 的基础模型。不过话说回来，除了其他几个顶尖的中国模型，大参数规模的开源模型还有别的选择吗？显然这还用问吗？
@mrdevlar（回复 @Lissanro）
Qwen3.5-35B is an excellent model and can do 90% of what I want. I get an additional 9% with clever prompting. It really depends on how intricate your use cases are and how willing you are to spend the time to put in guidelines and guardrails. So a single 3090 is fine for me.
Qwen3.5-35B 是个超棒的模型，能满足我 90% 的需求。通过巧妙的提示词，我还能再挖掘 9% 的潜力。这真的取决于你的用例有多复杂，以及你愿意花多少时间制定使用指南和限制条件。所以对我来说，单块 3090 显卡就够用了。
@Chicagoj1563
People are going to use what serves their needs most. If that is a low cost open model, then that is what people will use. The USA seems to think everyone will pay through the nose for ai.
人们会用最符合自己需求的东西。如果低成本的开源模型能满足需求，那大家就会用它。美国似乎觉得每个人都会为 AI 花大价钱。
@ea_man
Personally, I donate to people who release open source products and improve the entire ecosystem. I'm glad OpenAI or Entropic are pushing the boundaries, but no, I won't give them money.
我个人会给那些发布开源产品、改善整个生态系统的人捐款。我很高兴 OpenAI 或 Entropic 在推动技术边界，但我不会给他们钱。
@Guinness
They’re not wrong, they’re just assholes. China is beating the pants off of us now. The video generation model is incredibly impressive. Their open models are impressive enough. I'm not Anthropic, but I can use Kimi K2.5 for 95% of work and switch when I need to do a tough problem or refactor something. Anthropic is the only company left with an actual lead on China. And that’s the point. They don’t want the United States dominating yet another major industry. Which is why I found it rather confusing that they’ve stopped the release of their video generation model. I would think they’d want to decimate American cinema and media companies. So yeah, they’re not wrong. But also fuck them, open source is the way. Linux lead to an explosion of industries and technologies. Your entire digital life is 98% driven by the Linux kernel. They’re just pissed they’re not going to make money off of the next big thing.
他们没说错，但就是一群混蛋。中国现在把我们打得落花流水。他们的视频生成模型让人印象极其深刻，开源模型也已经足够惊艳了。我不是 Anthropic 的用户，但我能用 Kimi K2.5 完成 95% 的工作，只有遇到难题或需要重构的时候才切换其他工具。Anthropic 是唯一一家在 AI 领域还能对中国保持实际领先的美国公司。这才是关键 —— 他们不想让美国再主导另一个主要产业。这也是我觉得困惑的地方：他们居然停止了视频生成模型的发布，我还以为他们要靠这个吊打美国影视和媒体行业呢。所以是啊，他们没说错中国在 AI 领域的威胁，但开源才是正道。Linux 曾引领了产业和技术的爆炸式发展，你整个数字生活 98% 都由 Linux 内核驱动。他们只是因为没法从下一个大热门里赚钱而气急败坏而已。
@Global_Estimate7021
IMHO there's plenty of reasons to think the US is already cooked when talking about AI. 1. Massive AI acceptance gap. China 87% vs. 32% in the US. 2. Chinese local govs and companies pushing AI literacy to the public (bottom up) vs. US where it's being unsuccessfully implemented on companies first (top down). 3. China beats US+UK+EU in AI research volume. At the end of the day what matters for services like AI isn't how fast or strong your model is, but if people actually use it.
依我看，谈到 AI，有很多理由认为美国已经完了。1. 巨大的 AI 接受度差距：中国 87%，美国只有 32%。2. 中国地方政府和企业在向公众推广 AI 素养（自下而上），而美国是先在企业推行（自上而下），但完全失败了。3. 中国的 AI 研究产出量超过了美国 + 英国 + 欧盟的总和。归根结底，像 AI 这样的服务，重要的不是你的模型有多快、多强，而是人们是否真的会用它。
@JockY（回复 @Global_Estimate7021）
It's almost like there are advantages to opening up revolutionary technology to the people instead of hoarding it for the rich corporations.
把革命性技术向公众开放，而不是囤积给富裕企业，这似乎确实有优势啊。
@mrdevlar（回复 @JockY）
Remember when OpenAI spent those few months pretending they were a non-profit social good? Yeah, it's weird to realize that currently China is more humanistic than the United States on this topic.
还记得 OpenAI 那几个月假装自己是非营利性社会公益组织吗？是啊，现在意识到中国在这个话题上比美国更有人文精神，真是挺奇怪的。
@cutebluedragongirl
They fear what they cannot control.
他们害怕自己无法控制的东西。
@Box_Robot0
I like how an authoritarian country is doing more to contribute to AI freedom than whatever we have here.
真有意思，一个被贴上 “威权国家” 标签的国家，在推动 AI 自由方面居然比我们这里做得还多。
@Zc5Gwu（回复 @Box_Robot0）
It may not be authoritarian but it may be an oligarchy.
美国可能不是威权主义，但可能是寡头政治。
@bnolsen
I think China is doing this to try to damage the obvious false AI market that US companies have been over hyping.
我觉得中国这么做是为了打击美国公司过度炒作的虚假 AI 市场。
@Cuddlyaxe
I mean at the end of the day it's all strategic. China's edge is in hardware while America's is in software. The Chinese care more about ensuring proprietary American models don't achieve dominance rather than having their own super model everyone is paying API fees for. After all, in a world where America's software advantage is neutralized but Chinese hardware superiority remains, China wins the AI war. It's not that different from Meta's strategy releasing Llama as an open model in the first place they knew they couldn’t compete with OpenAI or Google in terms of quality, so they released it open to stop them from establishing dominance. Except this time it's on a national level.
归根结底，这一切都是战略性的。中国的优势在硬件，而美国的优势在软件。中国人更关心的是阻止美国专有模型占据主导地位，而不是拥有一个所有人都要付 API 费用的超级模型。毕竟，在一个美国软件优势被抵消、但中国硬件优势依然存在的世界里，中国就赢了 AI 战争。这和 Meta 最初发布开源模型 Llama 的策略没什么不同 —— 他们知道自己在质量上无法与 OpenAI 或 Google 竞争，所以就开源发布来阻止对方建立主导地位。只不过这次是在国家层面上操作而已。
@I_pretend_2_know
It isn't only that. There is also the "not-sustainable" financial side of the American way of doing AI. The American companies are playing a dangerous bet. They're throwing gigantic mountains of money into building their supermodels and reselling the services at a loss. Sooner or later, they'll need to jack up their prices. And it won't be just double the price of their plans. It will be more like 100 dollars per month for Claude/Gemini. And I seriously doubt that many people will be willing to pay that much. Americans are in a bubble. It will pop, and the pop will hurt.
不仅如此，美国搞 AI 的路子，在商业上根本不可持续。美国公司正在下一场危险的赌注：他们投入巨额资金打造超级模型，却亏本出售服务。迟早他们得涨价，而且不会只是套餐价格翻倍 ——Claude 或 Gemini 的月费可能会涨到 100 美元左右。我严重怀疑有多少人愿意付这么多钱。美国人活在泡沫里，泡沫迟早会破裂，而且破裂的时候会很疼。
@JockY（回复 @I_pretend_2_know）
Chinese actions have pushed the industry toward a services game by commoditizing the most valuable parts of American big AI and instead building cheap APIs and clever high-value services on top. OpenAI and Anthropic can't compete with that business model. They need to make the frontier models and win the services game, all while maximizing returns for investors, which seems almost impossible. Very soon 99% of people will be able to completely ignore OpenAI and Anthropic because cheap services will be so good that the expensive SOTA frontier models simply aren't needed. And then the pop will hurt them, but everyone else will be like "meh". This could have been avoided. The non-profit OpenAI could have been open and non-profit, dedicated to improving life for humanity. Instead they got greedy and chased the dolla dolla bills. Fuck 'em. Let them reap what they sow.
中国这波直接把美国 AI 最核心的技术打成了白菜价，直接把行业拖入服务竞争 —— 转而在其上搭建廉价 API 和高价值智能服务。OpenAI 和 Anthropic 根本无法与这种商业模式竞争：他们既要打造前沿模型，又要赢得服务市场，同时还要给投资人赚够钱，这几乎是不可能的。很快，99% 的人将完全无视 OpenAI 和 Anthropic，因为廉价服务会变得非常好，以至于昂贵的最先进前沿模型根本没必要存在。到时候泡沫破裂只会伤害到他们，其他人则会觉得 “无所谓”。这本来是可以避免的：非营利性质的 OpenAI 本可以保持开源和非营利，致力于改善人类生活，但他们却变得贪婪，一心追名逐利。去他们的，让他们自食其果。
@Specialist-Heat-6414
The 'threat' framing is a bit backwards. Open source models being good is just... good? Like Qwen and DeepSeek improved every local setup I run, full stop. The US advisory body warning about open source AI is basically asking: how do we maintain an advantage in a world where the tools are free and shared? The answer isn't restricting model weights -- it's building better infrastructure and applications on top. Closing off open weights would just accelerate the fork.
这种 “威胁” 的说法有点本末倒置。开源模型变好用，本身就是一件好事啊？比如 Qwen 和 DeepSeek 让我所有的本地设备配置都变得更棒，就这么简单。美国咨询机构警告开源 AI，本质上是在问：在工具免费共享的世界里，我们该如何保持优势？答案不是限制模型权重，而是在其上搭建更好的基础设施和应用。关闭开源模型只会加速技术分流，让大家更倾向于用中国开源模型。
@Specialist-Heat-6414
The advisory body is fighting the last war. Open source dominance is not a threat you counter with export controls or closed-source advantages. The model weights are already out. DeepSeek is already downloaded on servers everywhere. The actual question is whether the US builds the infrastructure layer on top of open models faster than China does. Tooling, deployment, integration, trust frxworks. That is still wide open and that is where the real lead gets built or lost.
这个咨询机构是在打一场过时的战争。开源主导地位不是靠出口管制或闭源优势就能对抗的威胁 —— 模型权重已经公开、遍地都是了，DeepSeek 已经被下载到世界各地的服务器上。真正的问题是，美国能否比中国更快地在开源模型之上搭建基础设施层：开发工具、部署方案、系统集成、安全信任体系。这一领域仍然是开放的，也是真正的领先优势所在 —— 要么在这里建立优势，要么在这里失去优势。
@shaneucf
US Big Tech even stopped publishing papers.
美国大型科技公司甚至已经停止发表研究论文了。
The Software as a Service frenzy has been hurting my pockets for a decade. Can’t buy, only rent…and everything comes coupled with ads, bloat and the useful features paywalled behind the tier right above the one you just paid for. I’m all for any alternatives, especially open source.
软件即服务（SaaS）的热潮已经让我钱包受损十年了。只能租不能买…… 而且所有服务都附带广告、冗余功能，有用的功能还藏在你刚付费的套餐之上的更高层级里，得再花钱解锁。我完全支持任何替代方案，尤其是开源的。
@Agitated-Ad-504
Currently experiencing that hellscape at work. Upper management decided they will ban all AI except Grok. Almost certain because it was contracted. Absolutely fucking brain dead. (I’m in Aerospace) that is supposed to be somehow restricted in what data it saves (or at least that's what they claim). Don't know if that's the case with Grok, but MS Copilot has a version specifically for enterprise use. Banning AI at workplaces is just a smart business decision because if you don't, some idiot will paste confidential information (or worse, personal data) into ChatGPT to ask it for help.
我现在正经历着那种地狱般的工作环境。高层决定禁止除 Grok 之外的所有 AI，几乎可以肯定是因为签了合同，简直蠢透了。（我在航空航天领域工作）据说这款 AI 在数据保存方面有某种限制（至少他们是这么宣称的）。不知道 Grok 是不是这样，但微软 Copilot 有专门针对企业用户的版本。不过职场禁止 AI 也算是个明智的商业决策 —— 不然真会有傻瓜把机密信息（甚至更糟的个人数据）粘贴到 ChatGPT 里寻求帮助。
@CtrlAltEvil
This is why I love Linux. There’s an open source alternative for practically everything. Plus none of the bloat.
这就是我喜欢 Linux 的原因。几乎所有东西都有开源替代方案，而且完全没有冗余垃圾功能。
@PauseLost2137
There’s an open source alternative for practically everything. cries in Adobe
几乎所有东西都有开源替代方案。在 Adobe的付费墙前哭泣
@therealowlman
If something good comes out of AI, the democratization of writing software is one of them.
如果人工智能能带来什么好事，软件编写的民主化绝对是其中之一。
@ChainsawBologna
That will be short-lived until they figure out how to put gates in to prevent you from writing software, and raise the price of tokens so high that you can't afford to spend them on writing software because you need them to hunt for affordable food. But the bubble is about to pop anyway. AI fatigue popping up all over. Compact portable models that run on-device for specific tasks is the real future. The monolithic datacenter model is failing faster than it can get going. No idea why the Tech Bros really thought people would be willing to spend $100/mon to chat with a graphics card in Altoona, Iowa to get work done.
这种民主化状态不会持续太久 —— 直到他们想出办法设置门槛阻止你编写软件，再把代币价格抬到天上去，让你根本没钱用来写软件，因为你得靠这些钱买得起食物。但 AI 泡沫终究要破裂的，现在到处都能感受到 AI 疲劳。能在设备上运行、针对特定任务的小型便携模型才是真正的未来，那种庞大的单体数据中心模式的失败速度比它发展的速度还快。真搞不懂那些科技大佬为啥会觉得，人们愿意每月花 100 美元，跟爱荷华州阿尔图纳市的一块显卡聊天来完成工作。
@swagonflyyyy
Open Source models are getting smaller and better over time, but to run the best ones you still need pretty expensive hardware, as those models are the only ones good for coding. The rest can automate simple tasks or be used for simple analysis or entertainment purposes, but anything more than that you need either the cloud or lots of hardware for large open source models. And if you are serious about running them locally, you need to switch to Linux. Period. Anything less than that will only hold you back.
开源模型正变得越来越小、越来越好，但要运行最顶尖的开源模型，仍然需要相当昂贵的硬件 —— 因为只有这些模型才适合编程。其他的开源模型只能自动化简单任务，或者用于简单的分析和娱乐，但如果要做更复杂的事，大型开源模型要么需要云端支持，要么需要大量硬件。而且如果你真的想在本地运行它们，就得切换到 Linux 系统，就是这么简单。低于这个标准的设备只会拖累你。
@CoastingUphill
Yes that’s why China keeps releasing those models
没错，这就是中国一直在发布这些开源模型的原因。
@CanadianPropagandist
Good! This push by Chinese AI companies to distribute very capable models under some extremely generous licensing (MIT / Apache) is one of the few things that will save AI tech from being completely gated by the same techbros that want to dominate entirely.
太好了！中国 AI 公司推动以极其慷慨的授权（如 MIT/Apache 协议）分发功能强大的模型，这是少数能避免 AI 技术被那些想完全垄断的科技大佬彻底封锁的举措之一。
@YumariiWolf
I had no idea 80% of AI startups in the US are using Chinese models, or that deepseeks R1 overtook chatGPT in US downloads on the App Store last year. Seems pretty plausible they could close the gap and overtake US companies.
我完全不知道美国 80% 的 AI 初创公司都在使用中国的模型，也不知道去年 Deepseek R1 在 App Store 的美国下载量居然超过了 ChatGPT。他们很有可能缩小差距，甚至超越美国公司，这看起来合情合理。
@ElvisDumbledore
"The people in charge" use anything they can to make you need them (heroin, oil, proprietary software, etc.). It's all about certain people keeping themselves at the top of the social hierarchy.
“掌权者” 会用尽一切手段让你依赖他们（比如海洛因、石油、专有软件等等）。这一切都是为了让某些人牢牢占据社会阶层的顶端。
@Swimming_Cover_9686
US AI lead is dumb throwing insane amounts of money at a problem that more compute really can't solve. By the time they plan on their investments paying out their hardware will be obsolete. It mathematically cannot work.
美国所谓的 AI 领先地位简直是蠢得可笑 —— 在一个更多计算能力根本无法解决的问题上砸下巨额资金。等他们盼着投资回本的时候，所用的硬件早就过时了。从数学角度来看，这根本行不通。
@Rich_Housing971
The reason they're against AI is precisely because they're closed-source models completely controlled by corporations only to extract more wealth from people. They are pro open source AI because those can be used by scientists and the average person.
人们反对某些 AI，恰恰是因为它们是闭源模型，完全被企业控制，目的只是从普通人身上榨取更多财富。而人们支持开源 AI，是因为这些 AI 能被科学家和普通人自由使用。
@blksilksheetz
its innovation when it’s cheaper and gets the job done for much less, hope this helps.
当一项技术更便宜、能以更低成本完成工作时，这就是创新 —— 希望这点能帮你理清思路。
@Fickle_Competition33
Reddit is funny, they go all rage against AI, but then when placed as US vs China AI, all of a sudden it becomes a good thing... The ideological relativism is strong here.
Reddit 真是太有意思了 —— 平时对 AI 骂声一片，但一旦把话题变成 “美国 AI vs 中国 AI”，AI 突然间就成了好事…… 这里的意识形态相对主义真是根深蒂固。
@Signal_Flight_7262
Love to see it
很乐见这种情况。
@Neither-Phone-7264
so that means that the US needs to start making open models to beat them!!! right????
所以这意味着美国得开始做开源模型才能打败他们啊！！！对吧？？？
@Material_Policy6327
“We will now pass a law making open source ai illegal!”
“我们现在要通过一项法律，把开源 AI 定为非法！”
@SGmoze
"I'm going to open-source Chinese models even harder now"
“现在我要更卖力地推广中国开源模型了。”
@progfix
Why not just all open source software!
为什么不干脆让所有软件都开源呢！
@IngwiePhoenix
"And it will be tremendous! We will show those bad people how we do it over here and how it's really done, then they will maybe reconsider and be nicer." Read that in a particular voice. :p
“这将会超棒的！我们要让那些坏人看看我们是怎么做事的，看看真正的正确做法，到时候他们可能会重新考虑，变得友善一点。”模仿特定语气读完这段话 :p）
@NeverLookBothWays
Nah we’re like Russia…eventually we’ll lose the brain share. We’re already behind on a lot of fronts, and where we could have accepted and learned from Chinese tech, we have instead siloed off and regressed, kicking out a lot of incredibly talented minds in the process. Meanwhile China is building a solid foundation that offsets energy costs with renewables. They are entering a Bronze Age while we’re still embracing stone.
不，我们现在就像俄罗斯…… 最终会失去人才优势的。我们在很多领域已经落后了，本可以接受并向中国技术学习，结果却自我孤立、倒退，还在这个过程中赶走了很多极具才华的人。与此同时，中国正在搭建坚实的基础，用可再生能源抵消能源成本。他们都要进入青铜时代了，我们却还在抱守石器时代的思维。
@broknbottle
To be honest, the US is carrying a lot of dead weight right now, like the Baby Boomers who refuse to let go of power, would rather die in office than see the next generation succeed.
说实话，美国现在背负着很多累赘，比如那些拒绝放权的婴儿潮一代，他们宁愿死在权力岗位上，也不愿看到下一代成功。
@CertainMiddle2382
Being one foot behind the lead, is much much much easier than being in the lead. As you know what you are trying is already working.
落后领先者一步，比保持领先要容易得多得多。因为你知道自己正在尝试的东西已经被验证是可行的。
@EffectiveCeilingFan
"Lead"? Pff. US is getting crushed in open weights, not even a competition. Absolute peak vibecoding performance is not the only metric. The Chinese stuff is cheaper, too. Not to mention, all I hear about recently is how dysfunctional Opus, GPT-5.4, and Gemini 3.1 Pro have been.
“领先”？呸。美国在开源模型领域被碾压得毫无还手之力，连像样的竞争都算不上。所谓的 “氛围编码” 根本不是唯一的衡量标准，而且中国的模型还更便宜。更别提最近我听到的全是 Opus、GPT-5.4 和 Gemini 3.1 Pro 有多拉胯的吐槽了。
@blackkettle
The real problem is that the US administration has absolutely no clue - none, not even a shadow of a - not even a concept of a plan about what to do. The admin is staunchly anti-intellectual from the top all the way down. There is no chance of 'competing'. None. Their response will most likely be 'tariffs' or something equally idiotic and when that immediately fails, they'll instead just choose to 'ban' the foreign models in the name of 'national security'. Decline upon decline.
真正的问题是，美国政府根本毫无头绪 —— 完全没有，连一丝一毫的计划概念都没有。从高层到基层，这个政府都坚决反智，根本没有 “竞争” 的可能，一点都没有。他们的回应很可能是 “加征关税” 之类同样愚蠢的操作，等这些措施立刻失败后，就会以 “国家安全” 为由 “禁止” 外国模型。真是一路衰败到底。
@sine120
Yeah, GPT-OSS held its ground for longer than most open models do, so I'll give it credit, but OpenAI was pretty much the only US company putting out something more than a toy. Credit to Llama for dominating for as long as it did, and other players like Granite and Phi were interesting to play with, but they were toys, in reality. I work in aerospace and have friends who work in other sensitive areas, they have need for local models to run internally on air-gapped networks, they all turn to Chinese models. Last year it was Deepseek and then GLM, this year there's a whole wealth of great options. Frontier models are still the best, don't get me wrong. I wouldn't trust open models, as good as they are, to my real work yet, but the gap went from feeling like frontier models were "twice as smart" last year to "maybe 10% better".
没错，GPT-OSS 比大多数开源模型坚持的时间要长，这点我得承认，但 OpenAI 几乎是唯一一家美国公司能拿出点超越 “玩具级” 的产品。Llama 能主导这么久确实值得肯定，Granite 和 Phi 这类产品玩起来也挺有意思，但实际上它们都只是玩具而已。我在航空航天行业工作，有朋友在其他敏感领域，他们需要能在内部隔离网络上运行的本地模型，现在全都会选中国模型。去年是 Deepseek 和 GLM，今年更是有一大堆优质选择。别误会，前沿闭源模型依然是最棒的，我还不会把开源模型（虽然已经很棒了）用在真正的工作中，但差距已经从去年感觉 “前沿模型聪明一倍” 缩小到了 “可能只比开源模型好 10%”。
@PureSignalLove
It's flipping for me right now. It's to the point where I don't trust frontier models due to their inconsistency. Opus 4.6 is great when it works, but when it goes lazy/stupid, it's less than useless.
我现在的看法已经反转了。因为前沿模型太不稳定，我已经不信任它们了。Opus 4.6 能用的时候确实厉害，但一旦变得懒散或犯蠢，就比没用还糟。
@sine120（回复 @PureSignalLove）
The nature of my work, I don't use them to write a ton of production code for me yet, just because investigating and reviewing is 10x more important than writing, and I needed as much usable context as I could get, but I've been doing side projects with Gemini CLI and do notice on days it just gets lazy or stupid. I suspect as inference capacity gets saturated, they deploy quantized models or somehow tune behavior to stop being so independent. Ultimately, local models will be the only way to guarantee you know what you're going to get. Once I don't get laughed out of the room for pitching we buy 1TB of RAM or VRAM, then I'll get serious about switching.
就我的工作性质而言，我还不会用它们写大量生产代码 —— 因为调研和审核比编写重要 10 倍，而且我需要尽可能多的可用上下文。但我一直在用 Gemini CLI 做副业项目，确实发现有些时候它会变得懒散或犯蠢。我怀疑是随着推理算力饱和，他们部署了量化模型，或者通过某种方式调整了行为，让它不再那么独立。归根结底，本地模型是唯一能保证你 “知道自己会得到什么结果” 的方式。等我提议买 1TB 内存或显存不会被人嘲笑的时候，我就会认真考虑切换到本地模型了。
@CavulusDeCavulei（回复 @sine120）
I don't work in a field as cool as yours and I am still a junior/mid engineer, but I tried to advocate to buy a gpu and use local models and I got basically laughed on because it's all on cloud now. I feel you.
我不在你这么酷的领域工作，还只是个初级 / 中级工程师，但我之前试着提议买显卡用本地模型，结果基本上被嘲笑了 —— 因为现在大家都用云端。我懂你的感受。
@Zc5Gwu
I think price is insanely good for open models, moreso even than a lot of frontier models. Look at rebench, Step-3.5-Flash $0.14 per problem at 59% versus Claude Sonnet 4.6 $1.02 per problem at 60.7%. 10x price difference for negligible accuracy difference.
我觉得开源模型的价格真的划算到离谱，甚至比很多前沿模型还划算。看看 Rebench 的数据：Step-3.5-Flash 每题 0.14 美元，准确率 59%；而 Claude Sonnet 4.6 每题 1.02 美元，准确率才 60.7%。价格差 10 倍，准确率却没差多少。
@sine120（回复 @Zc5Gwu）
For the price, open models are definitely best, I just need approved vendors for my job, and most open providers are not approvable.
单看价格，开源模型绝对是最优选择，但我的工作要求必须用经过批准的供应商，而大多数开源模型提供商都通不过审批。
@jld1532
Good thing we're getting a push for NFS funded American open weight models.
幸好现在有人在推动由 NFS（正确缩写NSF美国国家科学基金会） 资助的美国开源模型，算是个好消息。
@jld1532
Kimi K2.5 is legitimately good enough for me to use in my work. Not as a full human replacement by a long shot, but it's good enough that the thought of paying for AI is laughable. My university hosts it on the cluster free for all faculty and students. As soon as enough people realize this, for-profit AI is dead.
Kimi K2.5 对我来说真的足够好用，完全能用到工作中。虽然远远不能替代人类，但已经好到 “花钱买 AI 服务” 这种想法都显得可笑。我的大学在集群上部署了它，所有师生都能免费使用。一旦足够多的人意识到这一点，营利性 AI 就完了。
@autoencoder
In my non-agentic testing, Qwen 3 ripped it to shreds. It would go loopy thinking, burning tokens uselessly to me. It did not care about any thinking directives, it just thought excessively.
在我的非智能体测试里，Qwen 3 直接把 GPT-OSS 吊打。GPT-OSS 经常陷入无意义狂想，纯纯浪费 token，根本不听指令，就是没完没了瞎想。
@Fastpas123
Agreed, the us is struggling, Gemini was incredible back in 2.5 days but 3.1 is a broken joke.
同意，美国现在确实很挣扎。Gemini 2.5 版本的时候还挺厉害，但 3.1 版本简直是个烂笑话。
@procgen
5.4 in codex has been an absolute workhorse for me lately. Haven't found anything that comes close to replacing it.
最近 Codex 里的 GPT-5.4 简直是我的主力工具，还没找到能接近替代它的东西。
@Free-Combination-773
US is not crushed in open weights competition, it just didn't show up.
美国在开源模型竞争中不是被碾压了，只是根本没参与而已。
@TheMericanIdiot
It’s open source… you know the same way US built its tech sector…. The stupid have taken over.
这可是开源啊…… 美国当年就是靠这个建立起科技产业的啊…… 现在蠢货们掌权了。
@Lissanro
The issue is keeping things closed may give few months advantage over competitors but overall slows the research down. OpenAI likely wouldn't even exist if there was no the Attention is All You Need paper. DeepSeek and their release that was also accompanied by the detailed research paper about its architecture and training... That's what lays down the foundation to further build upon. Currently I prefer to run mostly Kimi K2.5 on my rig, but it would not even exist if DeepSeek did not share their research and architecture. It seems even large companies prefer it... For example Cursor AI picked Kimi K2.5 as a base model for their Composer 2. But then again, what else is there to pick in the larger size besides a few other top Chinese models? Obviously that's a rhetorical question.
问题在于，闭源可能会让你比竞争对手多几个月的优势，但总体上会拖慢整个研究的进度。如果没有《Attention is All You Need》那篇论文，OpenAI 可能根本就不会存在。而 DeepSeek 发布模型时，还附带了关于其架构和训练过程的详细研究论文…… 这才是为后续发展奠定基础的做法。目前我更喜欢在自己的设备上运行 Kimi K2.5，但如果 DeepSeek 没有分享他们的研究和架构，Kimi K2.5 可能也不会存在。看来连大公司都青睐中国开源模型…… 比如 Cursor AI 就选了 Kimi K2.5 作为他们 Composer 2 的基础模型。不过话说回来，除了其他几个顶尖的中国模型，大参数规模的开源模型还有别的选择吗？显然这还用问吗？
@mrdevlar（回复 @Lissanro）
Qwen3.5-35B is an excellent model and can do 90% of what I want. I get an additional 9% with clever prompting. It really depends on how intricate your use cases are and how willing you are to spend the time to put in guidelines and guardrails. So a single 3090 is fine for me.
Qwen3.5-35B 是个超棒的模型，能满足我 90% 的需求。通过巧妙的提示词，我还能再挖掘 9% 的潜力。这真的取决于你的用例有多复杂，以及你愿意花多少时间制定使用指南和限制条件。所以对我来说，单块 3090 显卡就够用了。
@Chicagoj1563
People are going to use what serves their needs most. If that is a low cost open model, then that is what people will use. The USA seems to think everyone will pay through the nose for ai.
人们会用最符合自己需求的东西。如果低成本的开源模型能满足需求，那大家就会用它。美国似乎觉得每个人都会为 AI 花大价钱。
@ea_man
Personally, I donate to people who release open source products and improve the entire ecosystem. I'm glad OpenAI or Entropic are pushing the boundaries, but no, I won't give them money.
我个人会给那些发布开源产品、改善整个生态系统的人捐款。我很高兴 OpenAI 或 Entropic 在推动技术边界，但我不会给他们钱。
@Guinness
They’re not wrong, they’re just assholes. China is beating the pants off of us now. The video generation model is incredibly impressive. Their open models are impressive enough. I'm not Anthropic, but I can use Kimi K2.5 for 95% of work and switch when I need to do a tough problem or refactor something. Anthropic is the only company left with an actual lead on China. And that’s the point. They don’t want the United States dominating yet another major industry. Which is why I found it rather confusing that they’ve stopped the release of their video generation model. I would think they’d want to decimate American cinema and media companies. So yeah, they’re not wrong. But also fuck them, open source is the way. Linux lead to an explosion of industries and technologies. Your entire digital life is 98% driven by the Linux kernel. They’re just pissed they’re not going to make money off of the next big thing.
他们没说错，但就是一群混蛋。中国现在把我们打得落花流水。他们的视频生成模型让人印象极其深刻，开源模型也已经足够惊艳了。我不是 Anthropic 的用户，但我能用 Kimi K2.5 完成 95% 的工作，只有遇到难题或需要重构的时候才切换其他工具。Anthropic 是唯一一家在 AI 领域还能对中国保持实际领先的美国公司。这才是关键 —— 他们不想让美国再主导另一个主要产业。这也是我觉得困惑的地方：他们居然停止了视频生成模型的发布，我还以为他们要靠这个吊打美国影视和媒体行业呢。所以是啊，他们没说错中国在 AI 领域的威胁，但开源才是正道。Linux 曾引领了产业和技术的爆炸式发展，你整个数字生活 98% 都由 Linux 内核驱动。他们只是因为没法从下一个大热门里赚钱而气急败坏而已。
@Global_Estimate7021
IMHO there's plenty of reasons to think the US is already cooked when talking about AI. 1. Massive AI acceptance gap. China 87% vs. 32% in the US. 2. Chinese local govs and companies pushing AI literacy to the public (bottom up) vs. US where it's being unsuccessfully implemented on companies first (top down). 3. China beats US+UK+EU in AI research volume. At the end of the day what matters for services like AI isn't how fast or strong your model is, but if people actually use it.
依我看，谈到 AI，有很多理由认为美国已经完了。1. 巨大的 AI 接受度差距：中国 87%，美国只有 32%。2. 中国地方政府和企业在向公众推广 AI 素养（自下而上），而美国是先在企业推行（自上而下），但完全失败了。3. 中国的 AI 研究产出量超过了美国 + 英国 + 欧盟的总和。归根结底，像 AI 这样的服务，重要的不是你的模型有多快、多强，而是人们是否真的会用它。
@JockY（回复 @Global_Estimate7021）
It's almost like there are advantages to opening up revolutionary technology to the people instead of hoarding it for the rich corporations.
把革命性技术向公众开放，而不是囤积给富裕企业，这似乎确实有优势啊。
@mrdevlar（回复 @JockY）
Remember when OpenAI spent those few months pretending they were a non-profit social good? Yeah, it's weird to realize that currently China is more humanistic than the United States on this topic.
还记得 OpenAI 那几个月假装自己是非营利性社会公益组织吗？是啊，现在意识到中国在这个话题上比美国更有人文精神，真是挺奇怪的。
@cutebluedragongirl
They fear what they cannot control.
他们害怕自己无法控制的东西。
@Box_Robot0
I like how an authoritarian country is doing more to contribute to AI freedom than whatever we have here.
真有意思，一个被贴上 “威权国家” 标签的国家，在推动 AI 自由方面居然比我们这里做得还多。
@Zc5Gwu（回复 @Box_Robot0）
It may not be authoritarian but it may be an oligarchy.
美国可能不是威权主义，但可能是寡头政治。
@bnolsen
I think China is doing this to try to damage the obvious false AI market that US companies have been over hyping.
我觉得中国这么做是为了打击美国公司过度炒作的虚假 AI 市场。
@Cuddlyaxe
I mean at the end of the day it's all strategic. China's edge is in hardware while America's is in software. The Chinese care more about ensuring proprietary American models don't achieve dominance rather than having their own super model everyone is paying API fees for. After all, in a world where America's software advantage is neutralized but Chinese hardware superiority remains, China wins the AI war. It's not that different from Meta's strategy releasing Llama as an open model in the first place they knew they couldn’t compete with OpenAI or Google in terms of quality, so they released it open to stop them from establishing dominance. Except this time it's on a national level.
归根结底，这一切都是战略性的。中国的优势在硬件，而美国的优势在软件。中国人更关心的是阻止美国专有模型占据主导地位，而不是拥有一个所有人都要付 API 费用的超级模型。毕竟，在一个美国软件优势被抵消、但中国硬件优势依然存在的世界里，中国就赢了 AI 战争。这和 Meta 最初发布开源模型 Llama 的策略没什么不同 —— 他们知道自己在质量上无法与 OpenAI 或 Google 竞争，所以就开源发布来阻止对方建立主导地位。只不过这次是在国家层面上操作而已。
@I_pretend_2_know
It isn't only that. There is also the "not-sustainable" financial side of the American way of doing AI. The American companies are playing a dangerous bet. They're throwing gigantic mountains of money into building their supermodels and reselling the services at a loss. Sooner or later, they'll need to jack up their prices. And it won't be just double the price of their plans. It will be more like 100 dollars per month for Claude/Gemini. And I seriously doubt that many people will be willing to pay that much. Americans are in a bubble. It will pop, and the pop will hurt.
不仅如此，美国搞 AI 的路子，在商业上根本不可持续。美国公司正在下一场危险的赌注：他们投入巨额资金打造超级模型，却亏本出售服务。迟早他们得涨价，而且不会只是套餐价格翻倍 ——Claude 或 Gemini 的月费可能会涨到 100 美元左右。我严重怀疑有多少人愿意付这么多钱。美国人活在泡沫里，泡沫迟早会破裂，而且破裂的时候会很疼。
@JockY（回复 @I_pretend_2_know）
Chinese actions have pushed the industry toward a services game by commoditizing the most valuable parts of American big AI and instead building cheap APIs and clever high-value services on top. OpenAI and Anthropic can't compete with that business model. They need to make the frontier models and win the services game, all while maximizing returns for investors, which seems almost impossible. Very soon 99% of people will be able to completely ignore OpenAI and Anthropic because cheap services will be so good that the expensive SOTA frontier models simply aren't needed. And then the pop will hurt them, but everyone else will be like "meh". This could have been avoided. The non-profit OpenAI could have been open and non-profit, dedicated to improving life for humanity. Instead they got greedy and chased the dolla dolla bills. Fuck 'em. Let them reap what they sow.
中国这波直接把美国 AI 最核心的技术打成了白菜价，直接把行业拖入服务竞争 —— 转而在其上搭建廉价 API 和高价值智能服务。OpenAI 和 Anthropic 根本无法与这种商业模式竞争：他们既要打造前沿模型，又要赢得服务市场，同时还要给投资人赚够钱，这几乎是不可能的。很快，99% 的人将完全无视 OpenAI 和 Anthropic，因为廉价服务会变得非常好，以至于昂贵的最先进前沿模型根本没必要存在。到时候泡沫破裂只会伤害到他们，其他人则会觉得 “无所谓”。这本来是可以避免的：非营利性质的 OpenAI 本可以保持开源和非营利，致力于改善人类生活，但他们却变得贪婪，一心追名逐利。去他们的，让他们自食其果。
@Specialist-Heat-6414
The 'threat' framing is a bit backwards. Open source models being good is just... good? Like Qwen and DeepSeek improved every local setup I run, full stop. The US advisory body warning about open source AI is basically asking: how do we maintain an advantage in a world where the tools are free and shared? The answer isn't restricting model weights -- it's building better infrastructure and applications on top. Closing off open weights would just accelerate the fork.
这种 “威胁” 的说法有点本末倒置。开源模型变好用，本身就是一件好事啊？比如 Qwen 和 DeepSeek 让我所有的本地设备配置都变得更棒，就这么简单。美国咨询机构警告开源 AI，本质上是在问：在工具免费共享的世界里，我们该如何保持优势？答案不是限制模型权重，而是在其上搭建更好的基础设施和应用。关闭开源模型只会加速技术分流，让大家更倾向于用中国开源模型。
@Specialist-Heat-6414
The advisory body is fighting the last war. Open source dominance is not a threat you counter with export controls or closed-source advantages. The model weights are already out. DeepSeek is already downloaded on servers everywhere. The actual question is whether the US builds the infrastructure layer on top of open models faster than China does. Tooling, deployment, integration, trust frxworks. That is still wide open and that is where the real lead gets built or lost.
这个咨询机构是在打一场过时的战争。开源主导地位不是靠出口管制或闭源优势就能对抗的威胁 —— 模型权重已经公开、遍地都是了，DeepSeek 已经被下载到世界各地的服务器上。真正的问题是，美国能否比中国更快地在开源模型之上搭建基础设施层：开发工具、部署方案、系统集成、安全信任体系。这一领域仍然是开放的，也是真正的领先优势所在 —— 要么在这里建立优势，要么在这里失去优势。
@shaneucf
US Big Tech even stopped publishing papers.
美国大型科技公司甚至已经停止发表研究论文了。
