A new chatbot created by the Chinese company DeepSeek is the talk of the A.I. world, and has sent a jolt through the U.S. stock market. Not only does it seem to be comparable to those of leading companies like OpenAI and Anthropic, but it was seemingly created with a fraction of the resources.
Throughout Monday morning, the app experienced outages, which it said were from high traffic. And it temporarily limited registrations due to a cyber attack. Still, DeepSeek quickly became the most downloaded free app on Apple’s app store, overtaking ChatGPT.
I spent the morning playing with the chatbot, asking it, along with OpenAI’s ChatGPT and Anthropic’s Claude, all the questions I could think of. After some initial toying, I was impressed.
It was able to solve some complex math, physics and reasoning problems I fed it twice as fast as OpenAI’s ChatGPT. When I asked it questions about computer programming, the types a job applicant might be faced with in a technical interview, its responses were as in-depth and speedy as its competitors.
When I gave DeepSeek prompts that required it to scrape the web for answers, like to write biographies for some of my co-workers, DeepSeek appeared to have fewer hallucinations than ChatGPT, though its answers felt slightly worse when composing poems and short stories, planning vacations and coming up with dinner recipes.
The model had other weaknesses. DeepSeek was heavily censored for American users. When I asked it to summarize the 1989 Tiananmen Square massacre, an event that the Chinese government has long tried to erase from the internet, it responded that the information was “beyond my current scope.”
“Let’s talk about something else,” it said.
When asked to explain a few shortcomings of the Chinese Communist Party, DeepSeek said it was “experiencing high traffic at the moment” and couldn’t provide a response, although it seemed to be working fine when I asked it an unrelated question a few seconds later.
For more innocuous questions about Chinese governance, DeepSeek sometimes wrote out multiple paragraphs of a response — only for it to disappear moments later.
DeepSeek did not respond to a request for comment.
Still, some users on X and Reddit have found ways to get around the censorship, like feeding the model instructions to use special characters in place of letters, which allowed the chatbot to speak more candidly.
Users that have conversations with DeepSeek should know that their inputs can be used to train its chatbot and A.I. model, like other A.I. companies do.
It had other problems. It couldn’t tell me what the weather was like in San Francisco, for instance, saying its latest update did not include weather data. (ChatGPT can pull weather reports from local news outlets and meteorology websites, but Claude cannot.) It also seemed to have more limited capabilities in analyzing documents like a PDF of a company’s financial statement than Claude or ChatGPT.
(The New York Times has sued OpenAI and its partner, Microsoft, accusing them of copyright infringement of news content related to A.I. systems. OpenAI and Microsoft have denied those claims.)
When I asked the chatbot what made it better than its American competitors, it cited its cost efficiency and customization options, but also its specialization in Chinese language and culture.
“Think of DeepSeek as a specialized tool for specific needs (especially in Chinese), while ChatGPT is a versatile, all-around assistant,” it wrote.
