Artificial intelligence (AI) is undergoing a radical transformation with the emergence of DeepSeek, a Chinese startup founded in 2023 that has captured global attention with its innovative open-source models. Its latest creations, DeepSeek R1 y DeepSeek V3They have generated a significant impact on the technology industry, challenging giants like OpenAI and jeopardizing the dominance of US companies in the AI sector.
The impact of DeepSeek on the technology industry
The fact that DeepSeek has developed highly competitive AI models, but with lower resource consumption, has provoked a reaction in the markets. Nvidia, one of the leading semiconductor and AI hardware companies, suffered a drop of 17% in its shares, which represented a loss of 600 one billion dollarsThis fall caused it to lose its title as the world's most valuable company, surpassed by Apple y MicrosoftOther technology companies were also affected, with Alphabet falling a 4% and Microsoft a 2%.
This uncertainty in the markets was compounded by the fall in the index Nasdaq by 3.77%, along with losses in the futures of Dow Jones (-1.04%) and the S & P 500 (-2.29%)The volatility reflects fears that Chinese AI could displace US solutions in terms of efficiency and cost.
Technological restrictions and advances of DeepSeek
Despite restrictions imposed by the United States DeepSeek has made significant progress in selling advanced chips to Chinese companies without access to cutting-edge hardware, according to the journalist. Holger ZschaepitzDeepSeek's ability to develop efficient, low-cost models calls into question the enormous infrastructure investments that Western companies have made in this sector.
What is DeepSeek?
DeepSeek was founded in 2023 by Liang Wenfeng, who also directs High-Flyer Capital Management, the only company funding the startup. Its first model was DeepSeek Coder, an open-source model specializing in programming, launched in November 2023.
The DeepSeek team is made up of young graduates from the best universities in Chinawith a clear focus on technological innovation. By not depending on external investors, the company has been able to develop its technology without financial pressures, allowing it to work in long-term AI projects.
However, DeepSeek is subject to internet regulation in Chinawhich means that its AI does not respond to sensitive topics according to the criteria of the Chinese government. For example, when asked about the Tiananmen Square protests of 1989She avoids answering, while she does provide information about the Spanish Civil War.
DeepSeek R1: Advanced Reasoning AI
El November 20th 2023DeepSeek presented a preview of DeepSeek R1, as well as the January 20, 2024 He launched it as an open-source model. It is a advanced reasoning modeldesigned for complex logic and math tasks, surpassing in certain parameters GPT-4o by OpenAI.
Key Features of DeepSeek R1
- Superior performance in key metrics, such as:
- AIME (efficiency in AI models).
- MATH-500 (solving mathematical problems).
- SWE-bench Verified (programming tasks).
- Self-verification capability, thanks to its multi-step reasoning.
- Moderate response speed, but greater accuracy, due to its focus on advance planning and task execution.
- 671 billion parameters, with lighter versions of 1,500 billion to 70 billion, adapted for different devices, from laptops to high-performance servers.
- Open source on Hugging Face, under license MIT without commercial restrictions.
DeepSeek V3: AI Mixture-of-Experts
Days after the launch of R1, DeepSeek presented DeepSeek V3, language model based on the Mixture-of-Experts (MoE) architectureThis technology divides tasks into different specialized modules, optimizing performance by using only the experts needed for each query.
Key features of DeepSeek V3
- MoE Architecture: optimizes processing by dividing learning among multiple specialized experts.
- 671 billion parameters, with 37 billion activated per token.
- Advanced handling of text-based tasks, including:
- Coding.
- Traducción.
- Content writing and creation.
- It surpasses models like:
- Llama 3.1 405B.
- Claudia 3.5.
- GPT-4o in various parameters.
- Training cost of $5.5 million, significantly lower than that of OpenAI's GPT-4 ($80 million).
- Flexible license, which allows its download, modification and commercial use in multiple applications.
- Available: DeepSeek website, GitHub and its official application.
The rise of DeepSeek in the world of AI
DeepSeek has shown that it is possible challenging the AI giants with highly efficient and accessible open-source models. Their focus on advanced reasoning, resource optimization, and development without commercial restrictions This could mark a change in the industry, offering alternatives to proprietary business models like OpenAI.
As AI continues to evolve, DeepSeek's ability to compete without access to advanced hardware It raises questions about the future of AI development and the efficiency of the West's multi-billion dollar investments. With models like R1 and V3This Chinese startup could become a A key reference point in the next generation of artificial intelligence.
Original article by María Bastero | January 27, 2025 published in Marketing4eCommerce
