As President Joe Biden approaches the end of his term, his administration has unveiled comprehensive regulations on AI technology, addressing national security concerns and economic competitiveness. The Interim Final Rule on Artificial Intelligence Diffusion establishes restrictions on high-end chips and closed-weight AI models, while maintaining exemptions for open-weight models.
The rule allows a 120-day window for public comment and a one-year adjustment period for affected industries. Notably absent from this initial regulatory framework are cloud service provider (CSP) platforms, leaving critical questions about computing power distribution across data centers unaddressed.
Taiwan's strategic position
Lee-Feng Chien, advisor to the Economic Development Council of the Executive Yuan, emphasized Taiwan's inclusion among 18 allies under the new framework. Chien cautioned that excluding Taiwan would impact exports of chips produced by TSMC, ultimately disadvantaging the US. He further noted that Nvidia, a leading AI chip supplier, will likely need to adjust its overseas market strategies in response to these restrictions.
"While the regulations primarily control who can purchase high-end chips from Nvidia and TSMC, the greater challenge lies in overseeing computing power centers," Chien explained. He suggested that additional measures might follow, particularly to prevent unauthorized foreign access to restricted chips and computational resources.
Open versus closed model dynamics
The proposed rules aim to protect closed-weight AI models by restricting model weight transfers to untrusted parties and implementing safety standards against unauthorized access. Currently, open-weight models remain unregulated, an exemption based on three key factors: the perceived superiority of closed-weight models in capabilities, the academic and research community's dependence on open-weight resources, and the practical challenges of monitoring freely available models. Closed-weight models, requiring payment, are inherently easier to track and control.
Security implications and industry response
Concerns about open-weight AI models intensified after reports emerged that China had utilized Meta's open-source Llama model for military purposes. This development sparked speculation about potential future restrictions on open-source models, with Meta likely to face an initial impact. However, current regulations reflect the assessment that open-weight models, despite rapid advancement, still lag behind their closed counterparts in capability.
Meta and Google dominate the open-source model sector, with Meta preparing to release Llama 3.1 in the second quarter of 2024, featuring 405 billion parameters. Meta's internal evaluations indicate this model could compete with offerings from OpenAI, Google, and Anthropic. Google maintains a dual approach, developing both large-scale closed models and smaller open models, such as its Gemma series.
Cloud computing challenges
The regulatory landscape faces ongoing challenges in cloud computing. In 2023, several US senators proposed sanctions against Chinese firms, including Huawei Cloud and Alibaba Cloud, citing security concerns. Industry experts have raised alarms about Chinese entities potentially exploiting American cloud platforms to access high-end computing resources. While cloud quotas have caps, experts warn of potential workarounds. They acknowledge the complexity of regulating cloud platforms, urging corporate self-discipline amid escalating US-China tensions.
With Donald Trump's upcoming presidency in 2025, the future of Biden's AI regulatory framework remains uncertain.