内容来源：https://aibusiness.com/ai-policy/enterprises-shaky-spot-amid-calls-ai-slowdown

内容总结：

关于是否应放缓人工智能发展的争论，正给企业带来新的不确定性。一方面，Anthropic和OpenAI等前沿AI实验室呼吁全球放缓前沿模型研发、并推动美国联邦层面加强监管；另一方面，中国企业推出的开源模型以更低成本快速追赶，令市场对开放权重模型的未来监管前景产生疑虑。分析人士指出，OpenAI和Anthropic的监管诉求，表面上是为社会利益，实则可能意在借政府之力维护自身在AI竞赛中的领先地位，尤其是在两家公司即将IPO之际。面对不确定性，企业倾向于为“确定性”买单，即便前沿模型成本更高。专家建议，企业应将AI治理视为一项持续运营能力，积极通过早期访问计划贴近前沿技术，尤其需加强网络安全准备，以应对AI压缩攻防周期带来的挑战。

中文翻译：

由谷歌云赞助

选择你的第一批生成式AI用例

要开始使用生成式AI，首先应聚焦于能够改善人类信息体验的领域。

随着企业对更便宜、不受监管的中国开源模型的需求不断增长，要求放缓AI发展的呼声也随之而来。

美国国内要求对AI进行联邦监管的呼声日益高涨，加之围绕是否需要全球AI放缓的争论，给企业带来了不确定性——它们不确定这些限制是否会阻碍其使用开放权重模型，尤其是来自中国的模型。

周一，中国外交部发言人郭嘉昆将前沿AI实验室Anthropic和OpenAI近期要求放缓AI发展的呼吁称为“散布恐慌、对抗和恶性竞争”。郭嘉昆此番表态之前，Anthropic首席执行官兼联合创始人达里奥·阿莫代伊于上周末发表了一篇文章，呼吁全球AI行业放缓前沿模型能力的推进步伐，并呼吁民主国家与前沿AI公司进行协调。

阿莫代伊的文章发表四天前，OpenAI首席全球事务官撰写了一份声明，请求制定联邦AI政策。该声明发布的同一天，一位曾供职于Anthropic的AI研究员宣布辞职，称OpenAI和Anthropic都没有以负责任的方式行事。该研究员还声称，这项技术可能导致人类灭绝。这些最新进展之前，Anthropic、谷歌、Meta和OpenAI的员工于7月发出呼吁，敦促美国政府支持国际治理，以把控AI发展节奏。

面对所有这些要求联邦政府采取行动的呼声，企业发现自己处于不确定的境地，这让本就在某些方面不稳定的技术更加难以预测。

“企业不得不在选择AI平台、供应商、技能、架构和运营模式的同时，清楚地知道底层技术可能在这些投资完全成熟之前就发生变化，”RPA2AI Research创始人卡什亚普·孔佩拉表示。“这场摊牌式辩论一次性增添了多项新的不确定性。”

这些担忧、疑虑和悬而未决的问题包括：前沿研发是会继续保持目前这种狂飙突进的速度，还是会放缓；开放模型是否会被限制；以及某些重要的AI能力是否仍将处于私人控制之下。

然而，Futurum Group分析师戴维·尼科尔森表示，目前的一些担忧可能恰恰是OpenAI和Anthropic想要的，因为企业正在意识到，尽管前沿模型性能令人印象深刻，但其成本高于开放模型，而且企业可以用更小、开放权重以及完全开源的AI模型，以更低的成本完成更多工作。

“随着时间一天天过去，真正需要前沿模型最尖端能力才能完成的工作负载越来越少，”尼科尔森说，并指出Anthropic和OpenAI即将进行备受期待且利润丰厚的IPO。

“毫无疑问，前沿模型的最尖端能力令人惊叹，但当你开始核算你所支付的费用并审视手头的任务时，会发现有更经济高效的方式来完成OpenAI和Anthropic向市场提供的东西，”他继续说道。

随着OpenAI和Anthropic面临来自以开放模型为主的开发者越来越大的价格压力，呼吁联邦AI监管似乎不再那么像是为了社会利益，而更像是为了保住它们在AI竞赛中的领先地位。

“为了维护它们在这个领域的双头垄断地位，它们愿意让政府来监管自己，”尼科尔森说，“它们所要求的是联邦政府介入，保护它们免受中国竞争的冲击。”

阿里巴巴、月之暗面和智谱等中国AI厂商近期的进步意味着，前沿模型制造商已经意识到，开放模型几乎与它们发布的模型一样好，但成本却低得多。

“它们正受到大量涌现的开放权重模型的围攻，”Informa TechTarget旗下Omdia的分析师连杰·苏表示。“这是它们在IPO前夕面临的商业风险。”

他补充说，中国厂商还认为，呼吁监管是美国独有的现象，因为中国已经在实施一些法规。例如，中国近期出台了规定，防止未成年人通过与AI聊天机器人互动而形成关系。

“也许是为了IPO，也许放缓节奏符合（Anthropic的）利益，但这仍然是一件非常重要的事情，因为与中国相比，美国从根本上缺乏那种AI治理监管，”苏说。他补充说，Anthropic也看到了来自开源的风险，因为开源——一种去中心化且高度公开的技术框架——意味着更少的治理。

不过，尼科尔森表示，对企业而言，对监管是否会延伸到开源AI技术以及涉及中国模型的地缘政治紧张局势的疑虑，意味着它们可能会屈服并无论如何都使用前沿模型，以便对冲未来可能出现的监管风险。

“（企业）愿意为确定性买单，”他说。

孔佩拉表示，当企业应对走向监管的不确定性时，最好的方法仍然是主动出击。

“治理必须成为一种运营能力，而不是一项打勾式的例行公事，”他说。他补充说，企业不能把AI治理当作一份只是定期翻阅的政策文件。

“对于能够访问数据、编写代码、使用工具、做出决策并自主行动的系统来说，这是不够的，”孔佩拉补充道。他说，企业应优先考虑早期访问和预览计划，以更接近前沿能力，尤其是在AI网络安全方面。

“网络备战成为核心，”孔佩拉继续说道。“随着智能体系统获得更强的攻击和防御能力，安全团队需要假定AI将压缩攻击和响应两个周期。”

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
Calls for an AI slowdown come as enterprise demand is building for cheaper and unregulated open source models from China.
The mounting calls for federal regulation of AI in the U.S. and debate over the need for a global AI slowdown have introduced uncertainty for enterprises unsure whether restrictions will hinder their use of open-weight models, especially those from China.
On Monday, Chinese foreign ministry spokesperson Guo Jiakun called the recent requests for an AI slowdown from frontier AI labs Anthropic and OpenAI "fear mongering, confrontation and vicious competition." Jiakun's comment comes after Anthropic's CEO and co-founder, Dario Amodei, published an essay over the weekend calling for the global AI industry to slow the advancement of frontier model capabilities and for democratic countries to coordinate with frontier AI companies.
Amodei's essay came four days after OpenAI's chief global affairs officer authored a statement asking for a federal AI policy. That statement was published the same day that an AI researcher who worked for Anthropic quit, saying that neither OpenAI nor Anthropic acts responsibly. The researcher also claimed that the technology could lead to human extinction. The recent developments followed a July call from employees at Anthropic, Google, Meta and OpenAI urging the U.S. government to back international governance to pace AI development.
With all the calls for the federal government to act, enterprises find themselves in an uncertain position, adding even more unpredictability to a technology that is already unstable in some ways.
"Enterprises are having to choose AI platforms, vendors, skills, architectures and operating models while knowing that the underlying technology may change before those investments fully mature," said Kashyap Kompella, RPA2AI Research founder. "The showdown debate adds several new uncertainties at once."
Those fears, doubts and open questions include whether frontier development will continue at its current torrid pace or slow, open models become restricted, and some important AI capabilities remain privately controlled.
However, some of the fears at this point could be exactly what OpenAI and Anthropic want, because enterprises are realizing that despite the impressive performance of frontier models, they're more expensive than open models and that enterprises can do more with less with smaller and open-weight and fully open source AI models, said David Nicholson, an analyst at Futurum Group.
"Every day that goes by, fewer and fewer workloads legitimately need what the leading edge of frontier models can deliver," Nicholson said, noting that Anthropic and OpenAI are nearing highly anticipated and lucrative IPOs.
"There is no question that the leading edge of frontier models is amazing, but when you start pricing the cost of what you're paying for and looking at the task at hand, there are more economically efficient ways to do what OpenAI and Anthropic are both offering to the market," he continued.
With OpenAI and Anthropic under increasing price pressure from developers of mostly open models, the appeals for federal AI regulation appear less about doing what is good for society and more about preserving their lead in the AI race.
"In the interest of preserving their duopoly in this space, they are willing to let the government regulate them," Nicholson said, "What they're asking for is for the federal government to step in and protect them from Chinese competition."
The recent advancements by Chinese AI vendors such as Alibaba, Moonshot AI and Z.ai, among others, mean that frontier model makers have realized that open models are nearly as good as the models they release, but available at far lower cost.
"They are under siege by a lot of the emergence of all these open weight models," Lian Jye Su, an analyst at Omdia, a division of Informa TechTarget, said. "It is a business risk that they are facing at the verge of the IPO."
He added that Chinese vendors have also argued that the call for regulation is unique to the U.S., since China is already enacting some regulations. For instance, China recently introduced rules to keep minors from forming relationships with AI chatbots.
"Maybe it's for the IPO, maybe it's in [Anthropic's] interest to slow things down, but it's still a very important thing because the U.S. fundamentally doesn't have that AI governance regulation as compared to China," Su said. He added that Anthropic also sees a risk from open source because open source -- a decentralized and highly public technology framework -- entails less governance.
For enterprises, though, doubt about whether regulation will extend to open source AI technology and geopolitical tensions involving Chinese models means they could cave in and use frontier models regardless of the cost, so that they can hedge themselves against the future possibility of regulation, Nicholson said.
"[Enterprises are] willing to pay for certainty," he said.
While enterprises deal with uncertainty about the move toward regulation, the best approach remains proactive, Kompella said.
"Governance has to become an operating capability, not a checkbox exercise," he said. He added that enterprises can't treat AI governance as a policy document that they only look at periodically.
"That is inadequate for systems that can access data, write code, use tools, make decisions and act autonomously," Kompella added. He said that enterprises should prioritize early access and preview programs to get closer to frontier capabilities, particularly with regard to AI cybersecurity.
"Cyber preparedness becomes central," Kompella continued. "As agentic systems gain stronger offensive and defensive capabilities, security teams need to assume that AI will compress both attack and response cycles."