The US is genuinely worried about China gaining access to open-source models.
Recently, Vinod Khosla expressed his views on the global AI race: “We are in an AI war with China, and whoever develops the best models will dominate the world economically, socially, and politically.”
Khosla isn’t alone in this thought. In a recent instance, US officials were also worried about China potentially using AI for malicious purposes, such as disrupting elections, launching cyberattacks, or even developing bioweapons.
In the same vein, Khosla said that while open source is good for VCs and innovation, it’s bad for national security. Many people took this out of context.
For instance, Stanford HAI co-director and professor Fei-Fei Li responded, saying, “Open-source projects helped propel science and technology in public and private sectors, benefiting developers, researchers, students, entrepreneurs, businesses, etc. – Human Genome, Linux, ImageNet, PyTorch, and much more!”
Similarly, Meta AI chief Yann LeCun weighed in, saying, “He is genuinely worried about China getting its hands on it. That worry is misguided.”
LeCun expressed his support for open-source projects. He said Chinese AI scientists and engineers are quite talented and very much able to “fast follow” the West and innovate themselves: “There are lots of good ideas from Chinese publications that make the community advance.”
“The Chinese government is even more worried about a lack of control of AI technology than their counterparts in liberal democracies,” LeCun said.
Additionally, LeCun added that in a future where every citizen’s digital diet is mediated by AI assistants, people’s knowledge and opinions about politics, history, value systems, etc., will be clearly affected.
Something he believes is becoming an increasingly large red flag for the Chinese government.
Open-sourcing AI assistants is beneficial when considering the potential impact on authoritarian regimes and the dissemination of knowledge. AI assistants are rapidly becoming repositories of human knowledge.
Once they can be downloaded and run locally, the control exerted by authoritarian governments over the information received by their citizens will be significantly diminished.
In other words, the Great Firewall of China would become ineffective in such a scenario.
LeCun believes that a future in which everyone has access to a variety of AI assistants with diverse expertise, language abilities, cultures, value systems, political opinions, and interpretations of history is the future we want.
That future can only come about through open-source AI platforms enabling a large selection of diversely fine-tuned systems.
This is not the first time. A few months ago, Khosla and Marc Andreessen had a debate on the open-sourcing of AI models, sparked by Elon Musk’s lawsuit against OpenAI. Andreessen advocated for open-source AI as a way to promote transparency and prevent monopolisation by Big Tech.
At the time, Khosla again supported closed AI development, arguing that, like nuclear weapons, open sourcing poses a risk to national security. He said, “It’s totally OK not to share the science.”
China is Leading the AI Race
Last year, Alibaba developed Qwen-72B, one of the largest models, and released its latest open-source model, Qwen1.5-72B, which surpassed Claude 2.1 and GPT-3.5 Turbo on several benchmarks. Notably, Qwen is also used in building large multimodal models (LMMs) and other generative AI-related projects.
Apart from Qwen, Tencent released ‘Hunyuan,’ an LLM for enterprise usage, in September last year. This move marked a significant advancement for China as companies in the country strive to establish themselves as leaders in the technology industry, specifically in the generative AI field.
Tencent vice president Jie Jiang highlighted the competitive landscape, stating that over 130 LLMs had surfaced in China by July.
In December, DeepSeek, a company based in China, aimed to “unravel the mystery of AGI with curiosity.” It open-sourced its DeepSeek LLM, a 67-billion parameter model meticulously trained from scratch on a dataset consisting of 2 trillion tokens in both English and Chinese, clearly hinting at its bid to go global. This model outperformed Llama 2, Claude 2, and Grok-1 in various metrics.
Moreover, Kai-Fu Lee’s AI startup open-sourced its foundational LLM called Yi-34B, which outperformed Llama 2 on various key metrics. Lee said that he wanted to provide an alternative to Meta’s Llama 2, which had been “the gold standard and a big contribution to the open-source community.”
With the release of these models, the fear of China rising up against US open-source models is not unfounded.
However, gaining prominence in open-source models, US-based Abacus AI has developed Smaug-72B, a model that tops the leaderboard with an average score of 80, outperforming Mistral. Smaug-72B is a fine-tuned version of Qwen-72B.
This case of Abacus AI’s Smaug model shows that researchers are increasingly interested in using open-source models.
Additionally, speaking at the AI Expo for National Competitiveness, ex-Google CEO Eric Schmidt said, “In the case of AI, we are well ahead two or three years, probably, of China, which in my world is an eternity.”
However, he also mentioned that the next twenty years would focus on national competitiveness overall, attributed to China’s emphasis on dominating specific industries.
However, Schmidt remained optimistic, saying, “We’re the likely winner if we don’t screw it up.” Specifically, Schmidt pointed out that premature regulation could hold the US back, contrary to Khosla’s assumptions.
