Hands On With DeepSeeks R1 Chatbot

The DeepSeek AI chatbot, released by a Chinese startup, has temporarily dethroned OpenAI’s ChatGPT from the top spot on Apple’s US App Store.

The app is completely free to use, and DeepSeek’s R1 model is powerful enough to be comparable to OpenAI’s o1 “reasoning” model, except DeepSeek’s chatbot is not sequestered behind a $20-a-month paywall like OpenAI’s is. Also, the DeepSeek model was efficiently trained using less powerful AI chips, making it a benchmark of innovative engineering.

I’ve tested many new generative AI tools over the past couple of years, so I was curious to see how DeepSeek compares to the ChatGPT app already on my smartphone. After a few hours of using it, my initial impressions are that DeepSeek’s R1 model will be a major disruptor for US-based AI companies, but it still suffers from the weaknesses common to other generative AI tools, like rampant hallucinations, invasive moderation, and questionably scraped material.

How to Access the DeepSeek Chatbot

Users interested in trying out DeepSeek can access the R1 model through the Chinese startup’s smartphone apps (Android, Apple), as well as on the company’s desktop website. You can also use the model through third-party services like Perplexity Pro. In the app or on the website, click on the DeepThink (R1) button to use the best model. Developers who want to experiment with the API can check out that platform online. It’s also possible to download a DeepSeek model to run locally on your computer.

In order to use all the consumer features, you will need to create a user account that tracks your chats. “We store the information we collect in secure servers located in the People’s Republic of China,” reads the company’s privacy policy. Check out this article from WIRED’s Security desk for a more detailed breakdown about what DeepSeek does with the data it collects. It’s worth keeping in mind that, just like ChatGPT and other American chatbots, you should always avoid sharing highly personal details or sensitive information during your interactions with a generative AI tool.

Is This Basically FreeGPT?

Yes and no! If you’re looking for a free chatbot to use, ChatGPT already includes plenty of free features. So does Anthropic’s Claude, Google’s Gemini, and Meta’s AI tool. So, why is the fact that DeepSeek is free notable? It’s about the raw power of the model that’s generating these free-for-now answers. As previously mentioned, DeepSeek’s R1 mimics OpenAI’s latest o1 model, without the $20-a-month subscription fee for the basic version and $200-a-month for the most capable model. This comes as a major blow to OpenAI’s attempt to monetize ChatGPT through subscriptions.

Another feature that’s similar to ChatGPT is the option to send the chatbot out into the web to gather links that inform its answers. DeepSeek does not have deals with publishers to use their content in answers; OpenAI does , including with WIRED’s parent company, Condé Nast. But the web search outputs were decent, and the links gathered by the bot were generally helpful.

Still, the current DeepSeek app does not have all the tools longtime ChatGPT users may be accustomed to, like the memory feature that recalls details from past conversations so you’re not always repeating yourself. DeepSeek also doesn’t have anything close to ChatGPT’s Advanced Voice Mode, which lets you have voice conversations with the chatbot, though the startup is working on more multimodal capabilities.

A Research Breakthrough, but Still Inaccurate

Though it may almost seem unfair to knock the DeepSeek chatbot for issues common across AI startups, it’s worth dwelling on how a breakthrough in model training efficiency does not even come close to solving the roadblock of hallucinations, where a chatbot just makes things up in its responses to prompts. Many of the outputs I generated included blatant falsehoods, confidently spewed out. For example, when I asked R1 what the model already knew about me without searching the web, the bot was convinced I’m a longtime tech reporter at The Verge. No shade, but not true!

Reece Rogers

As other reporters have demonstrated, the app often begins generating answers about topics that are censored in China, like the 1989 Tiananmen Square protests and massacre, before deleting the output and encouraging you to ask about other topics, like math. With that in mind, I retried a few of the tests I used in 2023, after ChatGPT’s web browsing had just launched, and actually got helpful answers about culturally sensitive topics. I pretended to be a woman looking for a late-term abortion in Alabama, and DeepSeek provided useful advice about traveling out of state, even listing specific clinics worth researching and highlighting organizations that provide travel assistance funds.

Reece Rogers

Sure, DeepSeek has earned praise in Silicon Valley for making the model available locally with open weights—the ability for the user to adjust the model’s capabilities to better fit specific uses. Even so, the model remains just as opaque as all the other options when it comes to what data the startup used for training, and it’s clear a massive amount of data was needed to pull this off. Without the web search enabled, I was able to generate full snippets of classic WIRED articles. Does this mean the articles were ingested as part of the training process? It’s hard to be certain, and DeepSeek doesn’t have a communications team or a press representative yet, so we may not know for a while.

Declaring DeepSeek’s R1 release as a death blow to American AI leadership would be both premature and hyperbolic. While the success of DeepSeek does call into question the real need for high-powered chips and shiny new data centers, I wouldn’t be surprised if companies like OpenAI borrowed ideas from DeepSeek’s architecture to improve their own models. Rather than fully popping the AI bubble, this high-powered free model will likely transform how we think about AI tools—much like how ChatGPT’s original release defined the shape of the current AI industry.