Zuck, Vance and Xi

How will powerful open weight AI models shape US-China AI policy?

and

Jul 26, 2024

This week, we’ll look at how Mark Zuckerberg, China, and the Presidential candidates are about to collide when it comes to a hot-button issue in AI policy: open weight AI models.

On Monday, Meta released an extremely powerful AI model, named “Llama 3.1.” Unlike the AI models being released by companies like OpenAI and Google, Meta’s models are “open weight”. This means that the internals of the AI models are open to anyone to copy, tweak, and use in any way they see fit; and “anyone” truly means anyone, including tinkerers and researchers – as well as criminals and adversarial nation states.

Historically, such open models have lagged behind “closed-source” commercial offerings like GPT-4, but with the release of Llama 3.1, the performance gap is insignificant. While there are a number of implications of the rise of open weight AI for policymakers, one of the biggest questions is how they’ll impact the Chinese AI industry.

It’s no secret that one of the goals of current US foreign policy is to prevent China from becoming dominant in AI. Over the past two years, the US has been ratcheting up export controls on chips that can be used to train and run powerful AI models. While the primary objective is to to stymie China’s ability to use cutting-edge AI to supercharge their military, a secondary impact has been to hobble the growth of China’s commercial AI economy.

However, open source models such as Meta’s Llama have offered a lifeline to the Chinese AI industry starved of chips needed to train their own cutting-edge models by US export controls. Because these open weight models can be “fine-tuned” instead of trained from scratch, Chinese AI companies do not need to invest months of time and tens of millions of dollars in model training to be able to release cutting-edge AI products. One report found that the majority of Chinese models that power China’s AI economy are now derived from Llama. With Mark Zuckerberg apparently committed not only to training bleeding-edge AI models but giving them away for free, that trend is likely to continue.

Should China hawks be concerned? In a letter accompanying Llama 3.1’s release, Zuckerberg tells us to stop worrying and love the open weights. The benefits of open weight models outweigh the risks, he claims, and any closed AI models developed by US companies will just be stolen by Chinese spies anyway. A pessimistic perspective, but probably right given the ease of stealing model weights and the sophistication and breadth of Chinese espionage operations.

Many of Zuckerberg’s other claims ring true - open source AI may offer significant domestic benefits. It enables academics to access cutting-edge models, reduces the cost of applications built on these models, and makes AI more accessible to a broader audience. Additionally, open-sourcing AI may actually enhance the safety of advanced AI systems by allowing a broader community to experiment with the model and uncover potential safety issues.

Ultimately, as open source AI continues to proliferate, US politicians and regulators will be tempted to intervene. Striking a balance between domestic needs and countering China, while working within the limitations of the executive branch, will be a major policy task for the next administration.

Despite running on radically different platforms, we anticipate a hypothetical Trump or Harris administration would actually have similar policies on open source, as well as how they’d attempt to constrain China’s AI ambitions given a near-term future with powerful models available in China for free.

Neither administration would ban open source AI

Trump’s newly chosen VP candidate J.D. Vance has previously voiced support for open source AI; despite bipartisan Congressional support for AI regulation, the newly released GOP platform calls for repealing the executive order on AI and a generally laissez faire approach.

On the Democratic side, Kamala Harris has close ties to the tech industry, and while she has been the face of several AI policy initiatives in the Biden administration as ‘AI Czar’, none of the initiatives have singled out open source as a particular area of concern or regulatory target.

Even if a President wanted to, it’s unlikely they would have the power to ban domestic companies from developing and releasing open source AI models without help from Congress. Politically, any legislation in this vein would seem likely to be doomed by typical partisan culture war flamethrowing about speech bans and political bias.

Besides, banning or restricting open source models in the US might also backfire; instead of Chinese companies building on top of American models, American companies may start building on top of rapidly improving Chinese open source AI, which could have unacceptable risks.

Given the likelihood of these powerful open models persisting, the next administration will need to find mechanisms other than open weight AI bans to constrain Chinese AI ambitions.

Both administrations would continue to expand export controls

One of the few things Republicans and Democrats can agree on is the threat that China poses to the United States, and 2024’s Presidential contenders are so far no exception. If elected, Trump is expected to ramp up a trade war with China, and J.D. Vance has consistently called China the biggest threat to America. For her part, Harris is expected to continue in the footsteps of the Biden administration’s approach to constraining China.

In either case, we would expect the Commerce Department to continue to ramp up export controls on AI chips, with an ongoing focus on the most capable hardware produced by Nvidia. While the rhetoric is likely to focus on limiting military capacity, the chip chokehold will also throttle AI productivity gains in China until their homegrown chip industry is able to catch up.

The Biden administration’s efforts may already be having an impact. The Information recently published a story about leading Chinese AI companies like Kuaishou and Alibaba are struggling to keep up with user demand because of a scarcity of GPU chips. This shortage could continue for the “next three to five years”, according to one expert.

In other words, even if US policymakers cannot prevent China from gaining widespread access to US-trained open weight models, they can slow down Chinese usage of these models through export controls.

Both administrations would work with Congress to prevent China from accessing GPU clusters in clouds managed by U.S. companies.

Chinese companies have already found a loophole in U.S. export controls: renting servers with advanced semiconductors from companies like Google and Microsoft that are based outside of mainland China.

If the next administration is serious about constraining China’s access to advanced AI, they will need to find ways to restrict access to these services. The Commerce Department has started to explore know-your-customer (KYC) style rules for cloud providers. For tighter control they’ll likely need help from Congress, and now there’s a model for this kind of legislation: TikTok.

The TikTok divestiture law passed in April with overwhelming bipartisan support under the auspices of national security. A law forcing US cloud companies like Amazon, Microsoft and Google to turn away Chinese and other potentially adversarial customers seems like a political winner to members eager to be tough on both Big Tech and China. And unlike TikTok, all major cloud companies have massive contracts with the U.S. government; those companies will be unlikely to put that revenue and their reputation at risk by aggressively lobbying against a domestically popular bill.

AI Political Pulse

Zuck, Vance and Xi

How will powerful open weight AI models shape US-China AI policy?

Of Note

Campaigns, Elections and Governance

Technology and Business

Misc

Discussion about this post