DeepSeek: Everything You Need to Know About the AI Chatbot App

DeepSeek has recently gained significant popularity.

The Chinese AI lab, DeepSeek, made headlines this week as its chatbot app climbed to the top of the charts on both the Apple App Store and Google Play. The AI models developed by DeepSeek utilize efficient training methods, prompting analysts on Wall Street and technology experts alike to speculate on the United States' ability to retain its dominance in the AI industry and the sustainability of demand for AI chips.

So, how did DeepSeek rise to global fame so rapidly?

DeepSeek’s Trader Origins

DeepSeek is backed by High-Flyer Capital Management, a quantitative hedge fund in China that employs AI technology to assist in trading decisions.

Liang Wenfeng, a notable AI enthusiast, co-founded High-Flyer in 2015. His journey into trading began while studying at Zhejiang University, which ultimately led to the establishment of High-Flyer Capital Management as a hedge fund focused on AI algorithms in 2019.

In 2023, High-Flyer initiated DeepSeek as a research lab aimed at developing AI tools, distinct from its financial endeavors. It later evolved into an independent company, sharing the same name.

From the outset, DeepSeek created its own data centers for training models. However, like many AI firms in China, DeepSeek faced challenges due to U.S. export restrictions on hardware. For its more recent models, the company relied on the Nvidia H800 chips, which are less powerful than the H100 chips available to U.S. companies.

DeepSeek's team is noted for its youth, with aggressive recruitment of PhD students specializing in AI from leading Chinese universities. They also employ individuals from varied backgrounds to help the technology understand a broader range of subjects, as reported by The New York Times.

DeepSeek’s Strong Models

In November 2023, DeepSeek introduced its first lineup of models: DeepSeek Coder, DeepSeek LLM, and DeepSeek Chat. However, it was the launch of the next-generation DeepSeek-V2 models in spring that caught the attention of the AI sector.

The DeepSeek-V2 series, a versatile system for analyzing text and images, excelled in various AI benchmarks and was significantly more cost-effective than other competing models at that time. This performance pressured DeepSeek's local competitors, including ByteDance and Alibaba, to reduce their pricing and, in some cases, offer their models for free.

The launch of DeepSeek-V3 in December 2024 further increased the company's visibility.

Internal testing at DeepSeek indicated that V3 surpasses both readily available models like Meta’s Llama and exclusive models accessible only via API, such as OpenAI’s GPT-4. DeepSeek also introduced the R1 reasoning model in January, claiming it matches the performance of OpenAI’s o1 model on critical benchmarks.

The R1 model is designed to fact-check itself, allowing it to avoid common errors typically associated with other models. Reasoning models like R1 usually take longer—ranging from seconds to minutes—compared to standard models but provide greater reliability in areas like physics, science, and math.

However, there are restrictions related to R1, DeepSeek V3, and other models due to their Chinese origin. They must pass assessments from China’s internet regulator to ensure their outputs reflect “core socialist values.” Consequently, R1 does not respond to queries regarding sensitive topics, such as Tiananmen Square or Taiwan's independence.

A Disruptive Approach

DeepSeek's business strategy is not entirely transparent. The company prices its offerings significantly below market prices and also provides some for free. Notably, it has opted not to pursue additional investments, despite substantial interest from venture capitalists.

The company credits breakthroughs in efficiency for its dramatic cost advantages. Yet, some experts question the accuracy of the figures DeepSeek has provided.

Regardless, developers have embraced DeepSeek’s models. While not open source in the traditional sense, they are available under licenses that allow for commercial utilization. According to Clem Delangue, CEO of Hugging Face—a platform hosting DeepSeek’s models—over 500 derivative models based on R1 have collectively been downloaded 2.5 million times.

DeepSeek’s competition with larger, established companies has been described as revolutionary yet exaggerated. Their rise was linked to an 18% drop in Nvidia's stock price in January and prompted responses from high-profile figures like OpenAI’s Sam Altman. Reports indicate that U.S. Commerce Department officials advised staff against using DeepSeek products on government devices.

DeepSeek is now available on Microsoft's Azure AI Foundry, which consolidates AI services for businesses. When discussing DeepSeek's influence on AI budget allocation at Meta, CEO Mark Zuckerberg highlighted the strategic importance of AI infrastructure spending. In March, OpenAI characterized DeepSeek as “state-subsidized” and advised that the U.S. should consider restricting its models.

During Nvidia’s fourth-quarter earnings update, CEO Jensen Huang lauded DeepSeek's innovation, stating that their reasoning models significantly require more computing power.

On the flip side, several companies and certain countries, including South Korea and New York State, have prohibited the use of DeepSeek due to apprehensions about its origins.

The future trajectory of DeepSeek remains uncertain. Enhanced models are anticipated, but the U.S. government appears to be increasing its scrutiny concerning perceived foreign threats. Reports have suggested a potential U.S. ban on DeepSeek products for government use.

This article was originally published on January 28, 2025, and will be regularly updated.

AI, technology, market