The embargo on advanced chips imposed by the US on China has proven to be a blessing in disguise for AI development there, finds Satyen K. Bordoloi.
Filmmaking changed forever in the first half of June 2024. On the 10th of that month, Kuaishou Technology announced Kling—a free text-to-video creation AI program. Though OpenAI had announced and teased Sora on February 15th, most of the videos they showcased were enhanced by VFX, and it wouldn’t be released for another 10 months. People only began to create stunning AI videos via Kling and DreamMachine, launched by Luma on June 12th.
Whenever such a new tool came out, I told my friends in Bollywood to try it. Six months later, at the end of December, after Sora finally released nearly a year after its announcement, I asked them again which of the many AI video-making tools was the best. The answer was unanimous: Kling. How extraordinary this is can only be gauged when we consider how the US tried to stifle China’s AI advances.
US-China Trade Wars: The trade war between the US and China has been like a game of ping pong, with each side volleying tariffs and restrictions back and forth. Under Trump, between 2018 and 2020, the US imposed tariffs on over $360 billion worth of Chinese goods, citing unfair trade practices and intellectual property theft.
China retaliated with tariffs on $110 billion of US products. The Biden administration continued and even expanded some measures, targeting high-tech sectors like AI and semiconductors, while China banned exports of rare earth metals.
The Chip Embargo and Its Objectives: The Biden administration’s decision to impose a chip embargo on China was designed to hinder the country’s ability to develop advanced AI systems by denying access to high-performance chips. We will not delve into whether Western fears of a belligerent China justify these sanctions. Instead, we will discuss how this was a blessing in disguise for China and what it can teach us about AI.
The embargo—first highlighted by the Huawei affair and recently by TP-Link fears—naturally presented significant challenges, but surprisingly spurred innovative strategies and adaptations within the Chinese tech sector.
Chinese companies responded by developing new methods to extract more value from weaker chips, focusing on smaller, more specialised AI models, and investing heavily in local chip manufacturing. Hence, despite the restrictions, Chinese AI companies managed to sustain and advance their AI capabilities, often through creative workarounds and a shift in focus from hardware to software and model efficiency.
The Chinese Strategy to Circumvent the Embargo was multipronged.
Local Chip Manufacturing: One of the primary strategies employed is the development of homegrown AI chip suppliers. Companies like Huawei, Baidu, Alibaba, and Tencent made strides in producing their own AI chips. While the West has its NVIDIA chips like the A100, China now has Huawei’s Ascend 910B and Baidu’s Kunlun Gen 2, both using the 7nm process node technology.
While these domestically produced chips may lag behind their Western counterparts in terms of performance and stability, they are increasingly proving to be viable alternatives. Huawei and Baidu’s chips are seen as competitors to Nvidia’s AI chips.
Efficient Code and Edge Models: The embargo forced Chinese AI companies to focus on other strategies, like developing more efficient code and smaller, specialised AI models. Think of this approach as the difference between a bulked-up gym enthusiast and a well-trained martial artist. Having the brute power of advanced chips is like a gym enthusiast with massive muscles.
Still, it is the martial artist, with precise techniques developed through efficient training, who emerges victorious in a street fight with a worthy opponent. The US has the brute power of advanced chips, but the Chinese figuring out the precise movement of martial artists to do more with what they have could prove the eventual winner.
In AI terms, this means that instead of relying solely on powerful chips, companies are optimising their models to require fewer resources. Professor Winston Ma, a law professor at New York University, recently noted that “the coming year is the year of small models.”
Instead of LLMs (Large Language Models), SMLs (Small Language Models) will see significant adoption and development due to their efficiency, cost-effectiveness, and ability to operate on all devices, making them suitable for a wider range of applications over LLMs. With less training data and speed, they’ll allow for quicker response times.
Another aspect is enhancing engineering capabilities and algorithms to compensate for the unavailability of advanced chips. By improving software and model training techniques, Chinese companies like Alibaba and Tencent achieve high performance even with less advanced hardware. Zhang Ping’an, a senior Huawei executive, said it best when he advocated that the mindset of relying solely on the most advanced AI chips needs to be abandoned in favour of innovative engineering and algorithmic advancements.
Flexibility over bulky muscles and agility over brute force seems to have become the new Chinese mantra for AI, which even Western AI companies realise is the better approach.
Cloud Services and Global Workarounds: Another tactic Chinese companies employ is renting cloud services located in the United States or other regions unaffected by the embargo. This allows them to access advanced computing resources without directly violating the restrictions.
Although the US government has not yet addressed this loophole effectively, it remains a viable option for Chinese companies to leverage global cloud infrastructure.
Supervised, Unsupervised, and Semi-Supervised Learning: Instead of the one-size-fits-all mentality of the West, AI models in China are being trained using different techniques such as supervised learning, unsupervised learning, and semi-supervised learning. Supervised learning is when models are trained on labelled data, which is crucial for tasks like object recognition and sentiment analysis.
Unsupervised learning uses unlabeled data to find patterns and structures useful for clustering and anomaly detection. Semi-supervised learning combines both, leveraging a small amount of labelled data and a large amount of unlabeled data to improve overall performance.
Open-Source AI Model Development: China has also been forced to rely on open-source AI model development. This has helped in China’s AI advancements because such models offer transparency, customizability, and flexibility, allowing continuous improvement and refinements. It enables the incorporation of the latest advancements in AI research rapidly and at scale while fostering a collaborative environment that drives future innovations.
National Strategies and Investments: China’s AI strategy is guided by key documents such as the New Generation Artificial Intelligence Development Plan (AIDP) from 2017 and the Made in China 2025 initiative launched in 2015. Both emphasise the importance of AI for enhancing national competitiveness and security and outline goals for achieving world-leading levels in AI technology while reducing dependence on foreign technologies.
The Made in China 2025 document aimed to transform China’s manufacturing sector from producing low-cost, low-quality goods to becoming a global leader in high-tech innovative products, especially in key industries like AI, robotics, aerospace, and advanced manufacturing. The number of patents coming out of China in each of these sectors in the last five years proves that the Chinese have succeeded.
This success is due to significant investments in AI research and development by the Chinese government, with regional governments pledging billions of yuan to foster local AI industries. Take the city of Beijing, which invested heavily in developing its AI ecosystem, focusing on areas such as autonomous vehicles and smart city technologies, emerging as an AI development hub for China and the world.
In the Three Body novel trilogy, writer Cixin Liu wrote about how an alien species tries to stop Earth’s progress by preventing the advancement of key technologies. The US tried to do something similar to China. In Liu’s novel, the Earth-alien conflict wipes out our solar system. One can only hope that the US-China trade war, leading to attempts to halt China’s AI advancements, will not lead to the same outcome.
In case you missed:
- A Manhattan Project for AI? Here’s Why That’s Missing the Point
- You’ll Never Guess What’s Inside NVIDIA’s Latest AI Breakthrough
- What are Text-to-Video Models in AI and How They are Changing the World
- AIoT Explained: The Intersection of AI and the Internet of Things
- AI vs. Metaverse & Crypto: Has AI hype lived up to expectations
- AI Taken for Granted: Has the World Reached the Point of AI Fatigue?
- Rethinking AI Research: Shifting Focus Towards Human-Level Intelligence
- The Rise of Personal AI Assistants: Jarvis to ‘Agent Smith’
- OpenAI’s Secret Project Strawberry Points to Last AI Hurdle: Reasoning
- Google’s Willow Quantum Chip: Separating Reality from Hype
