The U.S. government is moving forward with plans to keep AI software out of Chinese hands. The Commerce Department is considering banning the export of closed-source AI models. In doing so, it misses out on open-source options that cannot be contained in any way.
With the plans, revealed by Reuters, sources speak of “guardrails” that keep advanced AI models contained. This involves only the limited distribution of LLMs like OpenAI’s GPT-4 or Google Gemini, not their functionality. Currently, AI vendors can sell their technology without any restrictions to anyone, including the Chinese government.
The Biden administration fears the use of these models by China to wage cyber warfare. In that regard, the move seems mostly preventive. Future LLMs may give new impetus to such a use case, but currently the use of GenAI tech is limited to writing more convincing phishing emails than before.
The Llama 3 problem
Central to the plans are parties such as OpenAI, Google and Anthropic. These companies are leading AI players and keep their most advanced models in-house. The LLMs can only be deployed through proprietary platforms or with APIs; running GPT-4 locally, for example, is impossible.
That is not true of Meta’s Llama models, version 3 of which appeared recently. These open-source LLMs have been considered the state of the art for the restriction-free deployment of GenAI since the publishing of Llama 1 early last year. Any developer can start working with the various Llama models themselves, which they can train, fine-tune or fundamentally modify as desired. That doesn’t mean the LLMs are inferior to OpenAI/Google/Anthropic offerings, if benchmarks are to be believed.
Read more: Meta unveils powerful open-source model Llama 3 and chatbot Meta AI
Where does the boundary lie?
The question is how the possible U.S. plans deal with this open-source issue. For some time, several IT companies have been denouncing the potential impact on open-source that regulations bring. For example, the Cyber Resilience Act threatened to curtail contributions to open-source because developers would suddenly be held accountable for code that someone else deploys uninvited for their own purposes. That danger now seems to have passed.
Regardless, the Biden administration wishes to put AI technology along a yardstick. AI models can be scored for intelligence through a variety of benchmarks, so a specific score could determine whether an LLM may be shared with China. This is reminiscent of existing U.S. legislation that blocks AI hardware from export above a certain level and places restrictions on ASML machines.
Those previous export restrictions appear to have been partially successful. For example, Chinese parties are still producing remarkably advanced chips with old machines, although their precise quality is debatable. Consequently, the very latest ASML machines cannot be made to work without the expertise from the Dutch chip company. The same is not true of Nvidia’s GPUs, which enable AI models such as GPT-4 and Gemini. China has access to AI chips it really shouldn’t have had under the restrictions, it recently revealed. If it can source them from the U.S., Europe or elsewhere, there’s nothing that stops these GPUs from being used to deploy or develop advanced AI solutions.
Again, it is unclear how successful America’s export limits against China are. Open-source AI development will presumably remain a thorn in the side of Washington’s desire to shield AI from Beijing. China itself has long been working to reduce its dependence on the West as well. That move was already predicted by ex-CEO of ASML Peter Wennink: blocking exports to China only motivates that country to take a technological leap forward itself, he suggested.
Also read: China blocks Windows, Intel and AMD for government use