Chinese AI app DeepSeek rises in popularity and stuns competitors

Artificial Intelligence Assistant created by Chinese startup DeepSeek It became the number one most downloaded app in Apple’s U.S. App Store over the weekend, sending shockwaves through Silicon Valley and sending big tech stock prices plunging. Nvidia has seen more than $460 billion It was wiped off its market value on Monday, a decline that Bloomberg called “the largest in the history of the U.S. stock market.”

The change stems from an open source model developed by DeepSeek called R1, which debuted earlier this month. The company said it competes with the current industry leader: OpenAI 01. But what most stunned the tech industry was that DeepSeek claimed to have built its model using only a small fraction of the specialized computer chips that AI companies typically need to develop advanced systems.

DeepSeek said on Monday that it would temporarily restrict new registrations, citing “large-scale malicious attacks” on the company’s services, according to a report. message On its website.

The DeepSeek R1 model “challenges the idea that Western AI companies are significantly outperforming Chinese companies,” says Jack Clark, co-founder of AI startup Anthropic. books In his newsletter. Venture capitalist Marc Andreessen Named It is the “Sputnik moment for artificial intelligence.”

Qing Lu, a research scientist at OpenAI, said DeepSeek’s chatbot demonstrated impressive Chinese conversational skills. “It’s the first time I feel the beauty of the Chinese language created by a chatbot,” he said. In X’s post Sunday.

The DeepSeek AI assistant is currently available for free and comes with three main functions. First, users can ask the chatbot questions and receive live answers. For example, when WIRED asked for recipe ideas involving pomegranate seeds, DeepSeek’s chatbot quickly provided a list of 15 options ranging from yogurt parfait to “Middle Eastern-inspired” rice pilaf, but did not cite any specific chefs or recipes.

DeepSeek also has a search mode that displays answers from the Internet. When WIRED asked “What are some big news stories today?” DeepSeek’s chatbot cited the ceasefire between Israel and Hamas and linked to several Western media outlets such as BBC News, but not all of the stories appeared to be relevant. Ironically, one of them was a New York Times story about DeepSeek’s impact on the stock market.

Finally, there’s a “DeepThink” mode that allows users to take advantage of the DeepSeek R1 model, which is built on top of the company’s existing Model V3. The difference between the two is that R1 has what are called “reasoning” abilities that allow him to explain step by step how he arrived at his conclusions. For example, when asked “What are the most important historical events of the twentieth century?” DeepSeek initially provided a long and winding answer that began with a number of general questions.

“It’s been a hundred years, so a lot has happened,” part of her response read. “Maybe I should break it down into major decades or themes like wars, political changes, technological advances, social movements, etc.” DeepSeek’s chatbot then went on to mention World War II, the Cold War, and the Holocaust.

But before R1 could finish his response, the entire answer disappeared and was replaced by a message that read “Sorry, I’m not sure how to handle this type of question yet. Let’s talk about math, programming, and logic problems instead!” A number of experts and early adopters And I noticed DeepSeek, like other tech platforms operating in China, appears to do just that Widespread censorship Topics that the Chinese Communist Party considers sensitive

But despite these limitations, DeepSeek’s free chatbot could pose a serious threat to competitors like OpenAI, which charges $20 a month for access to its most powerful AI models. Unlike its Chinese counterpart, OpenAI does not disclose the underlying “weights” of its models, which determine how the AI ​​processes information. She also refused to announce the full “Chains of thought“produced by their own logic models.