⚔️Huawei Battles US Curbs For AI Dominance

China's Huawei takes on Nvidia despite mounting restrictions

Key Points 

• Huawei plans to mass-produce its Ascend 910C AI chip by 2025, despite challenges from U.S. restrictions.

• Huawei faces further troubles as U.S. authorities consider tighter export controls, and Trump’s return to office looming.

🎯Context of the news - Huawei has become a focal point in the U.S.-China tech rivalry, with Washington imposing strict curbs on the company over national security concerns. These restrictions have limited Huawei’s access to critical technologies, especially in chipmaking, which has emerged as a key pressure point.

🏭News - Despite the challenges, Huawei plans to start mass-producing its most advanced AI chip, the Ascend 910C, in early 2025. Designed to rival U.S. chipmaker Nvidia’s products, the company has already sent samples to tech firms and begun taking orders.

However, scaling up production remains a challenge. To be commercially viable, advanced chips need yields above 70%, but Huawei’s current 910B processor has a yield of only about 50%. This is due to U.S. restrictions that block Huawei from accessing crucial tools like EUV lithography machines. As a result, Huawei is prioritizing strategic orders from governments and corporations over large-scale production.

😵‍💫More troubles loom ahead - Adding to Huawei’s struggles, TSMC recently notified U.S. authorities that one of its chips ended up in Huawei’s 910B process. With Huawei already on a U.S. trade blacklist, this could bring stricter export controls. And with Donald Trump set to return as president in January, his track record of tough-on-China policies suggests Huawei’s path ahead won’t be any easier.

Chinese lab launches reasoning AI model to compete with OpenAI’s o1

Key Points 

• DeepSeek claims its model performs on par with o1 on AI benchmarks AIME and MATH.

• Despite its potential, the model struggles with logic puzzles and has been found to be vulnerable to jailbreaking. 

👨🏻‍💻News - A Chinese lab has introduced one of the first AI models that could compete with OpenAI’s o1. DeepSeek, an AI research company backed by quantitative traders, has previewed DeepSeek-R1, which it claims is a reasoning model on par with o1.

✨How does it compare? Like o1, DeepSeek-R1 is designed to reason through tasks, plan ahead, and perform a series of actions to arrive at an answer. However, this process can take some time. Depending on the complexity of the question, DeepSeek-R1 might take several seconds to “think” before delivering a response, much like o1. 

According to DeepSeek, the model performs similarly to OpenAI’s o1-preview on two popular AI benchmarks: AIME, which evaluates performance using other AI models, and MATH, a set of word problems. Despite this, the model has its limitations. 

On X, some users pointed out that DeepSeek-R1 struggles with logic problems like tic-tac-toe — a challenge it shares with o1. Additionally, DeepSeek-R1 has been criticized for being vulnerable to jailbreaking. One user was able to get the model to produce a detailed meth recipe. 

Looking ahead, DeepSeek plans to open-source DeepSeek-R1 and release an API, allowing wider access to its new AI model.