An AI assistant launched by the Chinese startup DeepSeek has quickly ascended to become the top-downloaded app in Apple’s US App Store over the weekend, creating waves throughout Silicon Valley and triggering a sharp decline in major tech stock prices. Nvidia suffered a staggering loss of over $460 billion in its market value on Monday, a drop that Bloomberg deemed as “the largest in the history of the US stock market.”
This upheaval is attributed to an open-source model created by DeepSeek, named R1, which made its debut earlier this month. The company asserts that its model competes directly with the prevailing industry frontrunner: OpenAI’s 01. However, what astonished the tech community most was DeepSeek’s claim of constructing its model using only a fraction of the specialized computer chips typically required by AI firms to develop leading-edge systems.
On Monday, DeepSeek announced that it would be temporarily halting new registrations due to “large-scale malicious attacks” targeting its services, as stated in a message on its website.
The emergence of DeepSeek’s R1 model “challenges the assumption that Western AI companies retain a substantial advantage over their Chinese counterparts,” said Jack Clark, co-founder of the AI startup Anthropic, in his newsletter. Venture capitalist Marc Andreessen referred to it as “AI’s Sputnik moment.”
Cheng Lu, a research scientist at OpenAI, praised DeepSeek’s chatbot for its remarkable proficiency in Chinese conversational language. “It’s the first occasion where I can appreciate the beauty of the Chinese language generated by a chatbot,” he commented in a post on X over the weekend.
Currently, DeepSeek’s AI assistant is available at no cost and boasts three primary features. Firstly, users can query the chatbot and receive direct responses. For instance, when WIRED asked for recipe suggestions involving pomegranate seeds, DeepSeek’s chatbot swiftly delivered a list of 15 options, ranging from yogurt parfaits to a “Middle Eastern-inspired” rice pilaf, though it did not reference specific chefs or recipes.
Secondly, DeepSeek’s app includes a search mode that retrieves information from the internet. When WIRED inquired, “What are some important news stories today?”, the chatbot cited the Israel-Hamas ceasefire and linked to various Western news outlets, such as BBC News, but not all stories appeared relevant. Ironically, one of the stories referenced was from the New York Times discussing DeepSeek’s effect on the stock market.
Finally, there’s a “DeepThink” mode that enables users to leverage DeepSeek’s R1 model, which is based on the company’s existing V3 model. The key distinction between the two is that R1 possesses so-called “reasoning” capabilities that allow it to elucidate step by step how it arrived at its conclusions. For example, when queried about “the most important historical events of the 20th century,” DeepSeek initially provided a lengthy and wandering reply, commencing with broad inquiries.
“That’s a hundred years, so there’s a lot that happened,” read a portion of its response. “I should probably categorize it by decades or major themes such as wars, political shifts, technological progress, social movements, etc.” DeepSeek’s chatbot subsequently referenced World War II, the Cold War, and the Holocaust.
However, before R1 could finalize its answer, the entire response vanished, and a message replaced it, stating “Sorry, I’m not sure how to approach this type of question yet. Let’s talk about math, coding, and logic problems instead!” Numerous experts and early users have pointed out that DeepSeek, akin to other tech platforms operating in China, appears to heavily censor topics viewed as sensitive by the Chinese Communist Party.
Despite these restrictions, DeepSeek’s complimentary chatbot could present a formidable challenge to rivals like OpenAI, which charges $20 monthly for access to its most advanced AI models. In contrast, OpenAI does not reveal the underlying “weights” of its models, which dictate how the AI interprets information, and has chosen not to publicly disclose the full “chains of thought” generated by its reasoning systems.
