A newly released chatbot by a Chinese startup has momentarily taken the lead over OpenAI’s ChatGPT at the top of Apple’s US App Store.
This app is entirely free to use, and DeepSeek’s R1 model boasts capabilities that rival OpenAI’s o1 “reasoning” model. Unlike OpenAI’s offering, which is subject to a $20 monthly subscription fee, DeepSeek’s chatbot remains accessible without such a barrier. Moreover, the DeepSeek model was adeptly trained using less powerful AI hardware, showcasing a remarkable feat of engineering.
Having explored numerous new generative AI tools in recent years, I was intrigued to assess how DeepSeek stacks up against the ChatGPT app already installed on my smartphone. After a few hours of interaction, my initial evaluation suggests that DeepSeek’s R1 model has the potential to significantly disrupt US-based AI firms. However, it still grapples with familiar shortcomings that plague other generative AI tools, including rampant hallucinations, overly cautious moderation, and issues with questionable data sourcing.
How to Use the DeepSeek Chatbot
For those keen on experiencing DeepSeek, the R1 model can be accessed via the startup’s smartphone applications (both Android and Apple) and on the company’s desktop website. Users can also engage with the model through third-party platforms like Perplexity Pro. Within the app or website, users should select the DeepThink (R1) button to access the most advanced model. Developers interested in testing the API can find it on the company’s online platform, and there is also an option to download a DeepSeek model for local use on personal computers.
To access all consumer features, users must create an account to keep track of their chats. According to the company’s privacy policy, “We store the information we collect in secure servers located in the People’s Republic of China.” For a more in-depth analysis of how DeepSeek manages data, refer to this article from WIRED’s Security desk. It’s important to remember, just as with ChatGPT and other American chatbots, users should steer clear of sharing highly personal or sensitive information during their interactions with generative AI tools.
Is This Essentially FreeGPT?
In some ways, yes and in others, no! If you’re looking for a no-cost chatbot, ChatGPT already offers a variety of free features. Similarly, Anthropic’s Claude, Google’s Gemini, and Meta’s AI tool also provide free options. So what makes DeepSeek’s free status noteworthy? It’s the robust capabilities of the model generating these currently free responses. As mentioned earlier, DeepSeek’s R1 closely resembles OpenAI’s latest o1 model but without the standard $20 monthly subscription fee for its basic version and the $200 monthly charge for the most advanced model. This presents a significant challenge to OpenAI’s strategy of monetizing ChatGPT through subscriptions.
Another aspect reminiscent of ChatGPT is the feature that allows the chatbot to search the web for links that inform its answers. However, DeepSeek has not established agreements with content publishers to utilize their material in its responses, whereas OpenAI has, including with WIRED’s parent company, Condé Nast. Nonetheless, the web search results were satisfactory, and the links compiled by the bot were generally useful.
That said, the current DeepSeek app lacks some tools that long-time ChatGPT users might expect, such as a memory feature that recalls details from past conversations to avoid repetition. DeepSeek also does not offer a voice conversation capability comparable to ChatGPT’s Advanced Voice Mode, although the startup is working on enhancing its multimodal functions.
A Breakthrough in Research, but Still Not Perfect
While it may seem unfair to criticize the DeepSeek chatbot for issues widespread among AI startups, it’s worth noting that a significant advancement in training efficiency does not remedy the common problem of hallucinations, where a chatbot fabricates information in its responses. Many outputs I generated contained outright inaccuracies presented with unwarranted confidence. For instance, when I inquired what R1 knew about me without searching the web, the bot mistakenly asserted that I was a longtime tech reporter for The Verge. No offense, but that’s simply not the case!
Reece Rogers
