香港期货配资
Anthropic CEO 达里奥
编辑：米奇 来源：财经会议圈
7月24日，英伟达创始人黄仁勋刚开通X账号，首条动态就转发了25家科技企业的联署公开信《开放权重与美国AI领导力》，核心诉求直指美国政府：
不要限制开源AI模型。
这份名单几乎囊括了硅谷半壁江山：微软、Meta、谷歌、AMD、Hugging Face悉数在列，连一贯走闭源路线的OpenAI也表态“希望美国在开源和闭源上都能赢”，唯独坚持全闭源、主打Claude系列的Anthropic缺席。
次日，Anthropic工程师Julian Schrittwieser的一条嘲讽推文引爆舆论：他阴阳怪气地表示“期待CUDA和GPU驱动开源”，还调侃微软该把Windows、Office开源。
这一类比被吴恩达直接驳斥：
反对开源的本质是阻碍他人开放技术，而非企业自主选择闭源；且AI模型闭源属于基础设施权力垄断，和工具链生态封闭的CUDA完全不是一回事。该工程师被全网声讨三天，期间联署企业反而从25家激增至133家，OpenAI更补刀称“我们都签了，Anthropic一声不吭”。
舆论压力下，7月28日达里奥终于发声，《我们对开源权重模型的立场》（Our position on open-weights models）先是撇清“从未主张禁止开源”，随即抛出老生常谈的风险论：
担忧他国借开源获得军事优势、模型被滥用发动网络和生物攻击，最终落脚点仍是“限制高端芯片流向中国”“管控模型蒸馏”“强制安全风险检测”。
这套说辞早被行业看腻：
此前他任职OpenAI时就曾以“风险过高”为由扣下GPT-2，结果全量发布后预言的灾难无一发生；今年5月黄仁勋已公开点出其“上帝情结”——总以先知姿态替行业定调。
如今现实早已打了达里奥的脸：
Hugging Face遭遇AI攻击时，正是中国智谱的开源模型GLM 5.2成了防御利器；美国芯片限制越严，中国自研速度越快。当全行业都站在开源一边，达里奥手里只剩“反华”这一张磨花了的旧牌，可这牌从2018年用到现在，早已没人信了。
Anthropic CEO达里奥在行业共识下的孤立处境，这篇文章也点破了所谓“中国威胁论”已成其手中唯一的且无用的“废牌”。
文章英文原名：Our position on open-weights models
官网发布时间：2026-07-27（美西时间），
北京时间 7 月 28 日凌晨官网原文链接：
https://www.anthropic.com/news/our-position-on-open-weights-models
我们对开源权重模型的立场
2026 年 7 月 27 日作者：Dario Amodei，Anthropic 首席执行官
过去几天，围绕开源权重模型（尤其是来自中国的模型）产生了大量讨论。
有报道称，部分美国官员正在考虑禁止美国企业使用中国开源权重模型。作为回应，众多科技企业签署公开信支持开源权重模型，同时也有人指责 Anthropic 希望推动禁令，以此保护自身业务。
读过我过往文章的读者应当清楚：
我并不认为全面禁令是有效的手段。
但为消除一切误解，我在此明确表态：Anthropic 从未主张封禁开源权重模型。
不具备高危能力的开源权重模型属于一种公共产品。
除运行所需算力之外，它几乎没有额外成本，并且能够为企业、开发者、科研人员创造价值。
全面禁止开源权重模型既不是正确解决方案，也从来不是我们所倡导的政策。
尽管如此，我并不认同这份支持开源权重公开信中的两项观点：
开源权重模型必然更容易搭建安全防御体系；
AI 能力的广泛普及，对防御方的帮助一定会大于攻击者。
我认为在诸多高风险领域，现实情况大概率恰恰相反。
以生物恐怖风险举例：
攻击者可以利用随处可得的原料改造病原体，诱发大规模疫情；
即便在最优条件下，研发对应的防御手段往往需要数年时间（曲速行动是罕见特例）。这类问题应当依靠发布前严谨的实证测试得出结论，而非预先主观假定。
我认为政策应当重点应对先进人工智能带来两大类核心风险：
WQ国家研发出显著领先美国的 AI 系统，借此获得永久性military advantages or deepen domestic surveillance。在这一风险场景下，模型是否开放权重、美国企业是否使用海外开源模型，都不是核心矛盾。最大的风险往往来自built exclusively for military and intelligence purposes
强大 AI 系统被滥用，实施大规模网络攻击、设计危险生物制剂，或是因对齐失效产生有害行为。开源权重模型带来一类独特风险：权重一旦发布就无法撤回，平台无法追踪、限制使用者的行为。闭源 API 模型则可以持续部署防护机制、使用限制与行为日志。
以上结论不代表应当禁止开源权重模型，而是需要定向政策管控风险。
以下是我们主张的政策框架：
第一，对存在国家安全风险的主体，限制先进芯片与芯片制造设备的获取能力。前沿算力是训练先进模型（无论开源、闭源）的基础前提。管控算力源头，是从根源处理竞争风险，而非在模型分发的下游环节发力。
第二，定向打击产业级大规模模型蒸馏。已有可信消息显示，海外企业蒸馏美国前沿模型搭建竞品，却没有对等投入与合作。这种行为削弱美国实验室投入数十亿美元研发前沿模型的动力。我们支持设立定向法律规则与贸易框架，遏制无补偿的大规模模型蒸馏行为。
第三，所有具备足够能力的模型，无论开源权重还是闭源，都必须执行强制性安全测试。想要发挥效果，测试必须在公开发布前完成，安全标准应当随模型能力同步提升。安全测试并非万能解药，但可以建立标准化流程，在技术大范围扩散前识别潜在危险。
有人提出，Anthropic 的立场出于商业利己考量。
但所有 AI 实验室都会面临先进 AI 带来的基础安全风险。我们对于开源权重的观点，一脉相承于长期坚持的安全框架；即便 Anthropic 不存在商用 API 业务，我们也会维持相同立场。
开源权重的争论常常被简化成二元选择题：
要么封禁开源模型，要么完全无限制放开。
这是一种虚假两难。
我们支持能力有限的开源权重模型，同时支持出台政策缓解高能力开源模型的风险，拒绝一刀切禁令。
粗暴的全面禁令不仅无法解决核心风险，还会对美国创新造成巨大的次生伤害。
一、英文官网全文（原版无删减）
Our position on open-weights models
July 27, 2026A post by Dario Amodei, Anthropic CEO
Over the last few days there has been a lot of discussion about open-weights models, especially those from China. Reports suggest that some US officials are considering banning the use of Chinese open-weights models by US companies. In response, many tech companies have signed a letter supporting open-weights models, and some people have accused Anthropic of wanting a ban on open models to protect our business. Anyone who has read my past writing should know that I don’t regard such a ban as a useful measure, but let me state it clearly so that there is no confusion: Anthropic has never advocated for a ban on open-weights models.
Open-weights models that don’t contain dangerous capabilities are a public good. They don’t cost anything besides the compute required to run them, and they provide value to businesses, developers, and researchers. A blanket ban on open-weights models is neither the right solution nor something we have advocated.
That said, I disagree with two claims in the open letter supporting open weights:
That open-weights models necessarily make safety defenses easier to build.
That broader access to model capabilities will inevitably help defenders more than attackers.
I think the reverse is likely to be true in many high-stakes domains. Consider bioterror: an attacker can use widely available materials to engineer a pathogen capable of causing a large pandemic, while defending against such a pathogen can take years even under the best circumstances (Operation Warp Speed is a rare exception). Questions like this need to be settled by rigorous pre-release empirical testing rather than assumed in advance.
I see two major categories of risks from advanced AI that policy should address:
The risk that authoritarian states develop substantially more capable AI systems than the United States and use them to achieve permanent military advantages or deepen domestic surveillance. In this scenario, whether models are open-weight or whether US companies use foreign open models is not the core concern. The greatest risks often come from secret models built exclusively for military and intelligence purposes, not public open-weight releases.
The risk that powerful AI systems are misused to carry out large-scale cyberattacks, design dangerous biological agents, or act in harmful ways due to alignment failures. Open-weight models create a distinctive version of this risk: once weights are released, they cannot be recalled, and there is no built-in ability to track or restrict how they are used. Closed API models can implement ongoing guardrails, restrictions, and logging.
This does not mean open-weight models should be prohibited. It means we need targeted policies to manage the risks. Here is our preferred policy framework:First, restrict access to advanced chips and chip manufacturing equipment for actors that pose national security risks. Access to frontier compute is the foundational prerequisite for training advanced models, open or closed. Controls here address the root of the competitive risk, rather than focusing downstream on model distribution.
Second, take targeted action against industrial-scale model distillation. There have been credible reports of foreign firms distilling frontier US models to build competing systems without reciprocal investment or collaboration. This practice undermines the incentives for US labs to invest billions in frontier model research. We support targeted legal rules and trade frameworks to address uncompensated large-scale distillation.
Third, require mandatory safety testing for all sufficiently capable models, open-weight and closed-weight alike. To be effective, testing must happen before public release, and standards should scale with model capability. Safety testing is not a silver bullet, but it creates a structured process to identify hazards before they are widely accessible.
Some have suggested that Anthropic’s position is motivated by commercial self-interest. All labs face the same fundamental safety risks from advanced AI. Our views on open weights follow directly from our longstanding safety framework, and we would hold this position if Anthropic had no commercial API business.
The debate over open weights is often framed as a binary choice: ban open models or embrace fully unrestricted open release. That is a false dilemma. We support open-weights models with limited capabilities香港期货配资, and we support policies that mitigate the risks of highly capable open models without blanket prohibitions. The alternative — crude blanket bans — will fail to address core risks while creating massive unintended harm to US innovation.
文章为作者独立观点，不代表正规黄金期货配资_外盘原油期货无息配资_正规外盘期货开户观点
相关文章
