DeepSeek R1 First Impressions: I’ve Seen How the Machine Thinks – CNET
After a week of using DeepSeek’s R1 reasoning model, I’m highly impressed, even if some questions are off-limits.
If you ever wondered what it would look like to see someone’s stream of consciousness in real time, DeepSeek R1 might be the closest you’ll get. When DeepSeek R1 landed on the internet in late January, it caught the AI industry off guard. Suddenly, there was a freely available and open-source “reasoning” AI model that could rival the equivalent $200 per month offering from OpenAI, creator of ChatGPT. Almost immediately, tech stocks plummeted. Investors wondered why they were throwing billions of dollars at OpenAI, Google and Microsoft when a scrappy team out of China was seemingly able to do just as much with significantly fewer resources. Nvidia, the company making the chips powering AI development, saw the single largest market value drop in US history, losing more than $600 billion in a day. Reasoning AI models can handle more-complex questions and do so by breaking problems down step-by-step, whereas the free version of ChatGPT tries to predict the next best word in a sentence.Essentially, DeepSeek R1 commodified reasoning-level AI models overnight — imagine Toyota suddenly producing Ferrari-level engines en masse. I spoke with AI experts, and the most interesting or amusing characteristic of DeepSeek R1 is being able to see its chain of thought. If you ask DeepSeek to tell you a bedtime story about a tenacious kitten, say, you’re able to see all the considerations it’s making while it’s processing your request. Basically, you can see DeepSeek think. And it’s fascinating.DeepSeek will literally sometimes say, “hmm,” which is both quirky and bewildering. OpenAI doesn’t let users see ChatGPT’s train of thought, so that it doesn’t give away its secrets (though now it’s showing small snippets). DeepSeek R1 is also open source, meaning any company can use its AI model, which is a major threat to OpenAI.An example of DeepSeek R1’s train of thought.I’ve been using DeepSeek on desktop for the past week. For security reasons, I opted not to install the app on my phone. Experts advise against installing the app on your personal device, given DeepSeek’s Chinese roots and the extensive data collection mentioned in its terms of service. (It would’ve been nice if word about that had gotten out before the DeepSeek app shot to the top of the App Store.) While using DeepSeek, I’ve been asking more-general and abstract questions, not uploading any sensitive data, or asking questions that could put me, or CNET, in a compromising position. Journalism is largely a job of synthesizing information from multiple sources and distilling it for a lay audience. In the past this required me to go to Google Scholar, skim through research papers and email professors to ultimately collect relevant quotes to include in an article. AI has changed that process significantly. Now I can ask an AI model specific research questions and it can scrub the internet and give me an answer, usually with sourcing. I continue to interview professors, of course, and I do have to go back and double-check the information the AI serves up, because AI bots can sometimes hallucinate and present something false as correct. Despite that, using AI is often still faster than if I had to tackle research papers manually. For a feature story I’m working on, I asked DeepSeek to help me find information about TikTok’s moderation policies and how they’ve been applied both correctly and incorrectly in the past. I also asked it to help me consider some unforeseen circumstances of overly strong social media moderation policies. This type of search isn’t entirely novel — older versions of ChatGPT could also handle these sorts of queries. But I was impressed by DeepSeek’s overall accuracy and synthesis of multiple articles. Where ChatGPT-4o felt like a really advanced version of autocomplete, DeepSeek R1 actually seems to be parsing through information and coming to conclusions. It’s not that ChatGPT-4o’s responses are majorly worse. But reasoning models like DeepSeek R1 are much better in terms of quality and more-extensive text output, sourcing and nuance. It really does feel like DeepSeek is “thinking” more.That’s not to say DeepSeek is perfect. Given its Chinese roots, some questions are off-limits. A commonly cited example in the media are queries regarding the 1989 massacre in Beijing’s Tiananmen Square. You can see DeepSeek trying to begin answering before it cancels out and tells you to ask something else. I tried asking it, very broadly, to give me background information about the Uyghur minority in China over the last few years. I saw it generating paragraphs of thought before totally canceling itself out. But if I asked about issues in America or other parts of the world, DeepSeek had little problem returning results. Example of DeekSeek R1’s train of thought on China’s past human rights abuses.An example of DeepSeek R1 cancelling out a response the Chinese government wouldn’t like.The release of AI chatbots, starting with OpenAI’s ChatGPT in late 2022, is changing how humans gather information online. Instead of you having to read multiple articles and synthesize information in your head, AI can do the connection-making for you. Reasoning AIs are an evolution of that. Using one feels much closer to having an expert inside your house. In the past, if I found myself stuck on a problem, I’d ask my friends or post a thread on Reddit for answers. This is obviously still an option, but it requires me having to bother others and hope they respond. DeepSeek R1 and ChatGPT o1 essentially relieve me of that social anxiety. Plus, it takes seconds to get an answer, versus whatever the other person’s schedule allows for. Knowing if the output is correct, however, is another problem entirely.
