The US government is working on plans to keep AI software out of Chinese hands. The Commerce Department is considering banning the export of closed-source AI models. This means that it lacks the open-source options that cannot be restricted in any way.
With the plans, revealed by Reuters, the sources speak of “guardrails” that keep advanced AI models in check. This only concerns the limited distribution of LLMs such as OpenAI’s GPT-4 or Google Gemini, not their functionality. Currently, AI vendors can sell their technology to anyone, including the Chinese government, without any restrictions.
The Biden administration fears the use of these models to wage cyber wars. In that respect, the step seems mainly preventive. Future LLMs may give new impetus to cyber attacks, but currently the use of GenAI technology 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 used via their own platforms or with APIs; for example, running GPT-4 locally is impossible.
This does not apply to Meta’s Llama models, version 3 of which recently appeared. Since the opening of Llama 1 early last year, these open-source LLMs have been considered the state of the art for the free use of GenAI. Any developer can get started with the various Llama models, which can be trained, fine-tuned or fundamentally adjusted as desired. That does not mean that the LLMs are significantly inferior to the offerings of OpenAI/Google/Anthropic, if the benchmarks are to be believed.
Read more: Meta unveils powerful open-source model Llama 3 and chatbot Meta AI
Where is the border?
The question is how any American plans will deal with this open source problem. Various IT companies have long criticized the possible impact on open source that regulations may have. For example, the Cyber Resilience Act threatened to limit contributions to open source because developers are suddenly held responsible for code that someone else uses unsolicited for their own purposes. That danger now seems to have passed.
In any case, the Biden administration wants to put AI technology to the test. AI models can be scored for intelligence via all kinds of benchmarks, so a specific score could determine whether an LLM can be shared with China. This is reminiscent of existing US legislation that blocks AI hardware from export above a certain level and places restrictions on ASML machines.
These previous export restrictions appear to be partially successful. For example, Chinese parties still produce remarkably advanced chips with old machines, although their precise quality is questionable. The latest ASML machines cannot be started without the expertise of the Veldhoven company. The same isn’t true for Nvidia’s GPUs, which power AI models like GPT-4 and Gemini. China has access to AI chips that it should not have had under the restrictions, it recently emerged. It cannot be ruled out that fast AI hardware will also be obtained via third parties in the US, Europe or elsewhere.
Once again it is shadowy how successful America’s technology battle against China is. Open-source AI development will likely remain a thorn in the side of Washington’s desire to shield AI from Beijing. China itself has been working for some time to reduce its dependence on the West. Ex-CEO of ASML Peter Wennink already predicted this step: blocking exports to China only motivates that country to make a technological leap forward itself.
Also read: China bans Windows, Intel and AMD from government use