The US is losing its lead in AI | The Vergecast

以下是内容的中文翻译： 由David Pierce主持的《The VergeCast》播客节目，深入探讨了中美之间日益激烈的AI竞赛，这场竞赛已成为一个重要的科技和政策议题。历史上，谷歌、Anthropic和OpenAI等硅谷巨头在前沿AI开发领域占据主导地位。然而，中国正迅速追赶，DeepSeq、月之暗面（Moonshot）和阿里巴巴等公司最近推出了“前沿模型”，挑战了美国的主导地位。 The Verge的资深AI记者Hayden Field指出，长期以来关于中国AI能力“落后六个月”的说法，现在被认为是最好的情况，实际差距可能更短，甚至在日常模型方面已经持平。资深政策记者Lauren Feiner强调，中美科技/AI竞赛是华盛顿经久不衰的话题，其核心目标是“击败中国”。 文中强调的一个核心担忧是“蒸馏”（distillation），即较小或较新的AI模型向更大、更成熟的模型学习。Hayden将此解释为AI领域的“快速致富方案”，能够以更少的资源实现快速发展。例如，Anthropic声称，中国公司DeepSeq、月之暗面（Moonshot AI）和Minimax利用欺诈性创建的账户，与其Claude模型进行了总计1600万次交互，以提升自身的AI能力。这种做法使中国企业能够迅速缩小技术差距。 Lauren指出，DeepSeq在一年半前出现，削弱了美国通过限制获取先进芯片来减缓中国AI进步的策略。DeepSeq证明了即使使用技术含量较低的硬件，也能以更高的效率开发出强大的模型，这迫使人们重新评估出口管制的有效性。 主持人探讨了如果中国AI超越美国可能带来的影响。Hayden指出，担忧主要围绕数据隐私问题和中国获得更大的全球影响力，前OpenAI员工也公开表达了类似担忧。Pierce将此与TikTok辩论相提并论，当时人们担心的是中国获得经济和文化软实力。Lauren进一步阐述道，她设想了一种情景：如果某款中国AI产品无疑更胜一筹，其他国家可能不得不将其整合到关键系统（如医疗保健或国防）中，这将带来地缘政治和监控风险，类似于过去对华为电信基础设施的担忧。 对中国获胜的担忧也深刻影响着美国AI公司。Hayden解释说，这些公司利用这一论调，主张监管机构给予“更长的缰绳”（即更宽松的监管），声称更少的监管将使他们能够更快创新并保持领先地位。这与政府在AI竞赛中“速度致胜”的口号不谋而合，无意中赋予了科技公司巨大的权力。然而，Pierce指出了其中的悖论：一些AI领导者一方面主张设立监管机构以推卸责任，另一方面又要求放松管制以“击败中国”。 展望中期选举，Lauren观察到，“击败中国”与“保护儿童”一样，仍然是国会中一个有力的政治论点。如果某个重大事件预示中国AI占据主导地位，这一论调可能会获得更多关注，从而可能改变政策制定者对国内AI监管的立场。 Hayden证实，美国AI实验室确实“害怕”中国获胜，这种恐惧源于中国公司，部分通过蒸馏技术，展现出的出人意料的效率和成本效益。这种生存恐惧，加上上市和吸引投资者的目标，促使他们主张优先考虑速度和创新的政策。 最终，播客总结道，美国内部的混乱——监管辩论、法律挑战和行业内讧——可能无意中使中国受益。这一认识正促使来自DeepMind、OpenAI和Anthropic等公司的美国AI领导者考虑采取更统一的监管方法，他们担心持续的混乱将让中国取得领先。

The VergeCast, hosted by David Pierce, dives into the escalating AI race between the US and China, a contest that has become both a significant tech and policy story. Historically, Silicon Valley giants like Google, Anthropic, and OpenAI have dominated frontier AI development. However, China is rapidly catching up, with companies like DeepSeq, Moonshot, and Alibaba recently unveiling "frontier models" that challenge US supremacy. Hayden Field, The Verge's senior AI reporter, notes that the long-held "six months behind" notion for China's AI capabilities is now considered a best-case scenario, potentially much shorter, or even at parity for everyday models. Lauren Feiner, the senior policy reporter, emphasizes that the US-China tech/AI race is a perennial topic in Washington, with the overarching goal of "beating China." A central concern highlighted is "distillation," where smaller or newer AI models learn from larger, more established ones. Hayden explains this as a "get rich quick scheme" for AI, allowing rapid development with fewer resources. Anthropic, for instance, claims that Chinese companies DeepSeq, Moonshot AI, and Minimax collectively generated 16 million exchanges with its Claude model using fraudulently created accounts to improve their own AI. This practice allows Chinese firms to quickly close the technology gap. Lauren points out that DeepSeq's emergence a year and a half ago undermined the US strategy of slowing China's AI progress by restricting access to advanced chips. DeepSeq demonstrated that powerful models could be developed with less advanced hardware and greater efficiency, forcing a re-evaluation of export control effectiveness. The hosts ponder the ramifications if China's AI overtakes the US. Hayden suggests fears revolve around data privacy issues and China gaining more global power, a sentiment echoed by former OpenAI employees who have publicly expressed their concerns. Pierce draws parallels to the TikTok debate, where the fear was China gaining economic and cultural soft power. Lauren extends this, envisioning a scenario where if a Chinese AI product is indisputably superior, other nations might be compelled to integrate it into critical systems (like healthcare or defense), posing geopolitical and surveillance risks, similar to past concerns over Huawei's telecom infrastructure. The fear of China winning also profoundly influences US AI companies. Hayden explains that these companies leverage this narrative to argue for a "longer leash" from regulators, asserting that less oversight will allow them to innovate faster and maintain the lead. This aligns with the government's "speed wins" mantra in the AI race, inadvertently granting tech companies significant power. However, Pierce notes the paradox: some AI leaders simultaneously advocate for regulatory watchdogs to absolve themselves of responsibility while also demanding deregulation to "beat China." Looking ahead to the midterms, Lauren observes that "beating China" remains a potent political argument in Congress, alongside "protecting children." This narrative might gain further traction if a significant event signals China's AI dominance, potentially shifting policymakers' stance on domestic AI regulation. Hayden confirms that US AI labs are genuinely "terrified" of China winning, driven by the unexpected efficiency and cost-effectiveness shown by Chinese firms, partly through distillation. This existential fear, combined with the goal of going public and attracting investors, pushes them to advocate for policies that prioritize speed and innovation. Ultimately, the podcast concludes that the internal chaos within the US—regulatory debates, legal challenges, and industry infighting—could inadvertently benefit China. This realization is prompting US AI leaders from DeepMind, OpenAI, and Anthropic to consider a more unified approach to regulation, fearing that continued disarray will allow China to pull ahead.