Chinese AI startup DeepSeek recently released its latest large language model, DeepSeek-V3-0324, launching a challenge to leading American AI companies such as OpenAI and Anthropic with its comprehensively upgraded technical architecture. This leap forward not only showcases China's ambition in the field of AI but also elevates the Sino-U.S. AI competition to a new height.

01. Founder of AI and former president of Google China, Kai-Fu Lee, stated that DeepSeek has significantly narrowed the technological gap with American leaders like OpenAI through algorithmic innovation and efficient use of domestic Hardware. This progress indicates that China is only three months behind the USA in core AI technology and has even taken the lead in certain areas. In an interview with Reuters, Kai-Fu Lee said:

Previously, I believed the gap was six to nine months and that we were completely behind. Now, I think that in certain core technology areas we are only three months behind, but in some specific fields, we have already achieved a lead.

Earlier this year, DeepSeek released an AI inference model trained on lower performance chips, attracting international attention. The company claims that this model only utilized Hardware resources valued at 6 million USD, distributed across 2000 NVIDIA H800 chips. In contrast, American companies like OpenAI and Meta poured billions of dollars into similar projects.

Benchmark tests released this week on the AI platform Hugging Face show that DeepSeek's latest model, DeepSeek-V3-0324, exhibits competitiveness in inference and coding. This model features an advanced "chain-of-thought" visualization capability, originally developed by OpenAI but not made available to users.

Compared to its predecessor, the V3 version has achieved significant improvements in the following dimensions:

• Inference capability: The efficiency of solving complex logical problems improved by 40% through a new training architecture.

• Code generation: In the auto-completion tests of programming languages like Python, the accuracy reached 92%, approaching the level of GPT-4.

• Cost advantage: Training was completed with 2,000 NVIDIA H800 chips valued at only 6 million dollars, costing just 1/20 of similar projects in the USA.

"This is no longer a race; we are defining the new future of AI," said the head of DeepSeek technology. The model has been opened to developers globally, with its open-source strategy directly targeting Meta's Llama series, forming a dual competitive advantage of "high performance + low cost."

Since the release of version V1 in December 2023, DeepSeek has maintained an astonishing iteration speed:

• 2023.12: Launched the basic version V1 model.

• 2024.01: Released the optimized R1 model focusing on enterprise scenarios.

• 2024.03: The current V3 version achieves a generational breakthrough in technology.

This "quarterly revolution" type of update frequency breaks the industry norm of a six-month to one-year upgrade cycle, forcing Western giants to reassess the evolution speed of AI in China.

Kai-fu Lee pointed out: "DeepSeek is able to achieve chain of thought functionality through new reinforcement learning methods, which indicates they are catching up with the USA, learning quickly, and may even be more innovative." This development challenges the view that US semiconductor sanctions hinder the progress of AI technology in China. Lee described the sanctions as a "double-edged sword," creating obstacles in the short term but also forcing Chinese companies to innovate under constraints.

The rise of DeepSeek has raised concerns in Silicon Valley and Washington. The rapid progress and efficient resource utilization of this startup starkly contrast with the massive investments of US giants in datacenters and dedicated chips.

Silicon Valley companies have raised their vigilance, with Anthropic listing China's AI technology as the "greatest strategic threat" in its latest financing documents. Meanwhile, the Capital Markets have started to adjust their layouts, with investment Institutions like Sequoia Capital establishing special Funds to increase investment in local AI projects. In terms of commercial applications, DeepSeek's technology has been successfully implemented in BYD's smart factories and CM BANK's financial risk control system.

Consulting agency TechInsight predicts that by 2025, the share of China's AI models in the global open-source market will increase from the current 15% to 35%, with DeepSeek expected to become a representative Chinese enterprise in this field.

Data released by QuestMobile shows that in the month following the launch of the DeepSeek APP, the number of active users surged to over 0.18 billion. The Doubao APP also successfully surpassed 100 million, while Tencent Yuanbao and Nano AI Search, bolstered by the powerful support of the DeepSeek large model, also stood out, entering the industry's top 5.

Edit/New