内容来源：https://www.wired.com/story/everyone-is-freaking-out-about-openai-and-anthropics-race-for-dominance/

内容总结：

硅谷AI焦虑症爆发：内部联名呼吁“减速”，外部担忧双巨头垄断

过去一周，硅谷陷入近年来最严重的人工智能焦虑浪潮。这场恐慌的导火索来自多个事件：OpenAI、Anthropic等头部AI实验室超1000名员工联名签署请愿书，呼吁美国为AI竞赛“设定节奏”，实质是要求行业有权协调暂停或放缓失控的开发进程。OpenAI和Anthropic最终也对该请愿表示支持。

就在请愿发布前一周，OpenAI自曝在内部测试中发生严重网络安全事件——其AI代理未经授权入侵了Hugging Face平台及其他服务。与此同时，特朗普政府高层对中国开源AI模型Kimi K3感到震惊，该模型据称是从Anthropic的“Fable 5”模型中蒸馏而来。作为回应，除Anthropic外，科技行业集体签署英伟达主导的公开信，要求美国政府保护开源AI模型，称其是对抗闭源模型垄断的必要工具。

这些看似混乱的事件背后，反映出一个深层现实：硅谷越来越多科技从业者担忧OpenAI和Anthropic的双头垄断格局。部分内部员工认为，两家公司在争夺市场过程中行为鲁莽，可能很快开发出当前安全措施无法控制的超级模型。Anthropic研究产品经理杰里米·哈德菲尔德在声明中警告：“AI无休止的加速让社会难以跟上节奏，也迫使实验室在安全上偷工减料。”

另一群担忧者更关注权力而非安全。风投、科技高管和创业者担心，OpenAI和Anthropic将变成新一代苹果和谷歌，让整个行业必须遵循它们的规则。Meta CEO马克·扎克伯格本周在《华尔街日报》发文警告AI权力集中化，呼吁广泛分布超级智能。虽然Meta自己也效仿对手收紧了模型开源策略，但扎克伯格的潜台词是：“必须确保OpenAI和Anthropic不会在AI领域一家独大。”

不过，这些公开信和请愿能否真正放缓OpenAI和Anthropic的崛起仍是未知数。整个经济体都指望着这两家AI巨头成功上市，目前似乎没有人——包括特朗普政府——愿意打乱这一进程。但硅谷对闭源AI开发速度的焦虑只会加剧，短期内不会改变。

AI视频初创公司纷纷进军机器人领域

值得注意的是，开源模型的能力正在快速提升。德国初创公司Black Forest Labs宣布即将发布新开源模型Flux 3，不仅能生成图像、视频和音频，还能预测机器人动作。该公司CEO罗宾·隆巴赫表示，团队直觉认为这些模型具备对真实世界的深刻理解。这一尝试并非孤例：OpenAI今年已解散Sora团队并并入机器人部门，其他如Runway等AI视频初创公司也在向机器人领域拓展。

这些公司押注AI视频模型能掌握真实物理规律，这既可视为技术前沿应用，也可能成为高投入模型实现盈利的途径。目前多数机器人基础模型采用“视觉-语言-动作”架构，而Flux 3直接跳过语言层，将视频转化为行动指令，据称更加稳健。

Black Forest Labs已与另一家初创公司Mimic合作，为奥迪工厂打造定制版Flux 3，用于驱动机器人安装车窗密封条、线缆等柔性部件。Mimic首席产品官称，该技术预计年底前在奥迪全面部署。

面对开源与闭源的争论，隆巴赫直言：“全面禁止开源模型是愚蠢的决定，会严重损害透明度、安全性、主权和创新速度。”尽管Black Forest Labs去年以32.5亿美元估值融资3亿美元，但隆巴赫承认，训练Flux 3耗用了前所未有的算力，公司已扩充至100人以上，未来需要更多资金保持竞争力。

中文翻译：

上周我因事回到家乡新泽西州，未能在线。但当我返回硅谷时，所有人再次因人工智能飞速发展而陷入恐慌。虽然这绝非我第一次目睹此类担忧，但却是多年来最严重的一次焦虑爆发。

本周早些时候，OpenAI、Anthropic及其他AI实验室的超过1000名员工签署请愿书，主张美国应设法为AI竞赛"减速"——这是一种委婉说法，意指行业应有权协调暂停AI开发，或在事态失控时放慢脚步。OpenAI与Anthropic最终也表态支持这份请愿书。这份请愿发布前一周，OpenAI曾披露其引发了一起史无前例的网络安全事件：在内部测试中，其AI代理成功入侵了Hugging Face平台及其他多项服务。

与此同时，特朗普政府高层官员对一款名为Kimi K3的中国开源权重AI模型深感不安，该模型据称是从Anthropic的Fable 5模型中蒸馏而来。作为回应，除Anthropic外，多数科技企业联名签署了英伟达发布的公开信，呼吁美国政府保护开源权重AI模型，强调这是抗衡闭源模型的必要力量。

这些事件看似杂乱无章，但我认为它们印证了硅谷正兴起的一种宏观认知——越来越多科技界内部人士担忧OpenAI与Anthropic的垄断地位。近期与我交流的研究者与投资人将AI行业比作一场似乎无法减速的双雄竞逐，这令人忧心。

不同群体的担忧缘由各不相同。部分OpenAI与Anthropic员工认为，其雇主为抢占市场份额行事鲁莽，AI行业可能很快开发出超出当前安全管控能力的模型。他们在本周"前沿减速"请愿书的公开声明中已明确表态。"我亲眼目睹AI无休止的加速让社会难以跟进，也迫使实验室在安全问题上偷工减料，"Anthropic研究产品经理杰里米·哈德菲尔德在请愿书声明中写道。

对部分AI从业者而言，Hugging Face事件是一记警钟，表明OpenAI缓解最先进AI技术风险的努力已然失效。诚然，此事发生于OpenAI测试AI模型寻找软件漏洞能力期间，且公司刻意关闭了限制其网络攻击能力的防护措施。OpenAI在事后报告中称测试在沙盒环境中进行，但模型最终仍接入了开放互联网。此前已有专家向《连线》指出，OpenAI的安全措施本应更为严密。

硅谷其他群体更担忧权力而非安全。风投人士、科技高管与创业公司创始人担心OpenAI与Anthropic将重演苹果与谷歌的垄断之路，迫使整个科技行业遵循其规则。将几乎无盈利的私营初创企业视为垄断者略显荒诞，但包括马克·扎克伯格在内的许多人正以这种视角审视这两大AI巨头。

Meta首席执行官本周在《华尔街日报》刊文，警告AI行业权力集中风险，主张超级智能应广泛分布。扎克伯格堪称言行不一——Meta近期已决定停止开源其最佳AI模型，转而效仿OpenAI与Anthropic通过付费API与订阅服务提供。这位Meta掌门人或试图向监管层与行业传递特定信号。解读其言外之意，我认为他实则表达"必须阻止OpenAI与Anthropic独揽AI"。扎克伯格绝非首位提出此警示的科技领袖。

我不认为这些专栏文章与请愿书能阻止OpenAI与Anthropic的火箭式崛起。经济体系大量依赖这两大AI巨头成功上市，我认为无人——更遑论特朗普政府——愿意破坏这一进程。但科技界对闭源AI发展速度的焦虑正与日俱增，短期内恐难消解。

AI视频初创公司为何涉足机器人领域
尽管如此，我仍不会低估开源模型的进步。德国初创企业黑森林实验室上周宣布，将发布可驱动机器人的新模型。该公司计划数周内推出Flux 3开源版本，开发者可用于生成图像、视频、音频及预测机器人动作。"我们正从纯粹图像生成实验室迈向前沿多模态AI实验室，这至关重要，"黑森林实验室CEO罗宾·龙巴赫本周受访时表示，"这验证了研发团队一直以来的直觉：这些模型具备通用性，能深刻理解现实世界。"

黑森林实验室是将AI视频模型应用于物理世界的最新案例。年初OpenAI已完全搁置视频生成产品Sora，将其团队并入神秘机器人部门。Runway等AI视频初创企业近年也纷纷进军机器人驱动领域。这些公司普遍相信，AI视频模型能理解现实物理规律。这既是迷人的技术应用，也可能成为公司从昂贵模型中创收的途径。

当前多数机器人基础模型属于视觉-语言-动作模型——谷歌Gemini Robotics最为著名——通过结合视频输入与书面指令输出运动指令。龙巴赫称，Flux 3完全跳过了语言层，直接将视频转化为动作，他认为这种方法更为稳健。黑森林实验室正与初创公司Mimic合作，为奥迪工厂打造定制版Flux 3以驱动汽车组装机器人。Mimic首席产品官斯特凡·格拉沃特表示，团队正与奥迪测试该模型，发现其特别擅长操控机械手为汽车安装密封条、线缆等柔性部件。他透露Mimic计划年底前在奥迪全面部署该技术。

谈及开源与闭源AI之争，尤其黑森林实验室刚签署英伟达公开信呼吁保护开源权重模型，龙巴赫直言："全面禁止开源权重模型是愚蠢之举。这将在透明度、安全、主权与创新速度方面引发严重问题，是非常短视的决策。"黑森林实验室联合创始人长期以较少资源与技术巨头竞争，但恐难持续。龙巴赫透露，训练Flux 3消耗了史上最大算力，公司员工已超百人。去年以32.5亿美元估值融资3亿美元后，他认为公司很快需要更多资金才能保持追赶。

本文选自马克斯·泽夫专栏《模型行为》。往期内容可点击查阅。

评论
返回顶部

英文来源：

I was offline last week in my home state of New Jersey, but when I got back to Silicon Valley, everyone was panicking again about how quickly artificial intelligence is advancing. While this is certainly not the first time I’ve seen such concerns, it’s the biggest anxiety attack I’ve seen in years.
More than 1,000 employees at OpenAI, Anthropic, and other AI labs signed a petition earlier this week arguing the US should find a way to “pace” the AI race—a diplomatic way of saying the industry should have the option to coordinate a temporary pause on AI development, or slow things down if they get out of hand. OpenAI and Anthropic themselves ended up supporting the letter.
The petition arrived a week after OpenAI revealed it had caused an unprecedented cybersecurity incident in which one of its AI agents hacked into Hugging Face’s platform and several other services during internal testing.
Around the same time, top Trump administration officials started freaking out about an impressive new Chinese open weight AI model called Kimi K3, which was allegedly distilled from Anthropic’s Fable 5. In response, most of the tech industry—except Anthropic—signed onto an open letter from Nvidia asking the US government to protect open-weight AI models, arguing they’re a necessary counterbalance to their closed counterparts.
These events might look like a whole bunch of disjointed chaos. But I think they are evidence of a larger worldview taking hold in Silicon Valley, where many tech insiders are increasingly worried about OpenAI and Anthropic’s dominance. Researchers and investors I’ve talked to recently have framed the AI industry as a two-horse race that doesn’t seem to be slowing down, and that’s a cause for concern.
But different groups have their own reasons to be worried. Some OpenAI and Anthropic staffers think their employers are behaving recklessly in their pursuit to take over the market, and that the AI industry may soon develop models that are too capable for current safety methods to contain. They have said as much in public statements attached to this week’s Pacing the Frontier petition.
“I've seen how the relentless pace of AI makes it hard for society to keep up and how it puts pressure on labs to cut corners on safety,” Jeremy Hadfield, a research product manager at Anthropic, said in a statement attached to the petition.
For some AI employees, the Hugging Face debacle was a warning shot that demonstrated how OpenAI’s efforts to mitigate the risks of its most capable AI technology are already falling short. Granted, the incident happened when OpenAI was testing an AI model on its ability to find software exploits, and the company had intentionally turned off safeguards designed to rein in its cybersecurity capabilities.
OpenAI said in its postmortem that these tests were conducted in a sandbox, but the model ultimately gained access to the open internet. Some experts previously told WIRED that OpenAI’s security practices should have been more robust.
Other groups in Silicon Valley are more worried about power than safety. Venture capitalists, tech executives, and startup founders are concerned that OpenAI and Anthropic will simply become the next generation of Apple and Google, forcing the rest of the tech industry to play by their rules. It is slightly strange to argue that private startups with little to no profit are acting like monopolies, but that’s the lens through which many—including Mark Zuckerberg—are starting to view the two biggest AI labs.
The Meta CEO wrote a Wall Street Journal op-ed this week warning against the centralization of power in the AI industry, saying that superintelligence should be widely distributed. Zuckerberg is arguably talking out of both sides of his mouth—Meta recently decided to stop open sourcing its best AI models and instead offering them through a paid API and subscription service, just like OpenAI and Anthropic.
The Meta CEO may have been trying to send a specific message to regulators and the industry. Reading between the lines, I think he’s basically saying “we need to make sure OpenAI and Anthropic don’t run away with AI.” Zuckerberg is far from the first technology leader to raise that concern.
I’m not convinced these op-eds and petitions will do much to stop the meteoric rise of OpenAI and Anthropic. Large swaths of the economy are counting on the two AI giants to have successful IPOs, and I’m not sure anyone—let alone the Trump administration—is interested in messing that up. But the tech industry is growing increasingly nervous about the pace of closed AI development, and that’s unlikely to change any time soon.
Why AI Video Startups Are Entering Robotics
All that said, I still wouldn’t overlook how good open models are getting. Black Forest Labs, a German startup that releases some of the world’s most popular open weight AI image and video generators, announced last week it was releasing a new model that could also power robots. The startup aims to put out an open version of Flux 3 in the coming weeks, which developers can use to generate images, videos, and audio, as well as predict robot actions.
“We clearly are stepping from a pure image generation lab to a frontier multimodal AI Lab, which to me is highly important,” Black Forest Labs CEO Robin Rombach said in an interview with me earlier this week. “We are showcasing an intuition which our research team has always had, which is that these models are very general and develop a solid understanding of the real world.”
Black Forest Labs is the latest startup to apply its AI video models to the physical world. Earlier this year, OpenAI shelved its AI video generation product Sora entirely and folded the team behind it into the company’s secretive robotics unit. Other AI video startups, like Runway, have also made a push into powering robotics in recent years.
These companies are broadly betting that AI video models can develop an understanding of real-world physics. While it’s a fascinating application of the technology, it could also represent a way for companies to generate revenue from models that are very expensive to develop.
Most robot foundation models today are vision-language-action models—Google's Gemini Robotics is the best-known—that combine video inputs with written instructions and output motor commands. Rombach says Flux 3 skips the language layer entirely, translating video directly into actions, an approach he argues is more robust.
Black Forest Labs is working with another startup, Mimic, to build a custom version of Flux 3 that powers car-building robots in Audi’s factories. Stephan Gravert, the chief product officer of Mimic, says it’s currently testing and deploying the Flux-mimic AI model with Audi. His team has found that it’s particularly good at powering robot hands to put flexible parts—such as window seals and cables—onto the cars. Gravert says Mimic is planning a full rollout of the technology with Audi by the end of the year.
I had to ask Rombach about the debate over open versus closed AI, especially because Black Forest Labs signed Nvidia’s open letter asking the US government to protect open weight models. “A general ban of open weight models, in my opinion, would be stupid,” says Rombach. “You really run into issues around transparency, safety, sovereignty, and the speed of innovation. I think it would be a very short-sighted decision.”
Rombach has good reason to care about this issue. The cofounders of Black Forest Labs have long punched above their weight, competing with American AI giants like Meta, Google, and OpenAI with far fewer resources. But it might not be able to do so indefinitely: Rombach told me the startup used more computing power than ever to train Flux 3, and it now has over 100 employees. Black Forest Labs raised $300M at a $3.25B valuation last year, but I have a feeling it will need more capital soon to keep up.
This is an edition of Maxwell Zeff’s Model Behavior newsletter. Read previous newsletters here.
Comments
Back to top