China has introduced an affordable, open-source alternative to OpenAI’s ChatGPT, sparking excitement among scientists and concern in Silicon Valley.
The AI lab behind this leap, DeepSeek, launched its large language model called DeepSeek-V3 in late December 2024. They claim it was developed in just two months for about $5.58 million, which is much faster and cheaper than similar projects in the U.S.
Following this, they released an even more advanced model, DeepSeek-R1, on January 20. Early tests show that DeepSeek-V3 matches the abilities of OpenAI’s GPT-4o and Anthropic’s Claude Sonnet 3.5, and it even outperformed models like Meta’s Llama 3.1 and Alibaba’s Qwen2.5 in areas such as problem-solving and coding.
The new R1 model has also exceeded the performance of ChatGPT’s latest o1 model in many tests. This success, at a fraction of the cost compared to other models, has impressed AI experts. It raises questions about whether China’s AI developments could soon lead the field.
Satya Nadella, CEO of Microsoft and a partner with OpenAI, expressed concern over these advancements at the World Economic Forum in Davos, Switzerland. He stated that the progress coming from China should be taken very seriously.
AI systems learn by analyzing training data from human interactions. This allows them to produce results based on recognized patterns. DeepSeek’s models rely on extensive text data, similar to OpenAI’s GPT-3.5, which was trained on about 570GB of text, equivalent to around 300 billion words from various online sources.
DeepSeek models, like R1, incorporate reasoned thinking, allowing them to tackle complex problems more effectively. This feature is particularly attractive to scientists and engineers keen on integrating AI tools into their research.
Unlike ChatGPT’s model, DeepSeek is “open-weight,” which means while users can’t access the training data, they can modify the underlying algorithm. Additionally, it offers a more affordable solution, costing significantly less than its competitors.
The buzz around DeepSeek isn’t just due to performance; its low-cost approach is remarkable when compared to the tens of millions spent by other companies on their AI models. U.S. export controls limiting Chinese firms’ access to top-tier AI chips have compelled DeepSeek developers to create more efficient algorithms, allowing them to achieve strong results with less computing power. For example, while ChatGPT required around 10,000 Nvidia GPUs for its training, DeepSeek managed with only about 2,000.
It remains to be seen how these advancements will translate into practical applications in science and technology. Many experts and investors are closely monitoring the situation to understand the true impact of DeepSeek’s innovations.
