Domestic AI company DeepSeek recently released its latest model, DeepSeek V3, which is said to be one of the strongest open-source AI models available, similar to the famous ChatGPT , Claude The models are pretty much the same. Now the web version of the DeepSeek V3 is available for free and also has deep thinking and internet search features similar to ChatGPT. This post will actually test the three main features of DeepSeek and then compare them with ChatGPT's answers.
Conclusion: following language styles to answer questions, external search, and open knowledge is still slightly stronger with ChatGPT.
What is DeepSeek? How does it work?
DeepSeek was founded in 2023, and most of the team members are current students or recent graduates from Tsinghua University and Peking University, which are the top universities in China. after releasing the DeepSeek V2 model in May 2024, it has received a lot of attention because of its great performance in Chinese language and cheap API price. Moreover, its models are all open source, so anyone can download and modify the code for free.
At the end of 2024, DeepSeek released the latest DeepSeek V3 model, which is claimed to be better than the current open-source models such as Llama 3.1 and Qwen 2.5, and can also compete with the world's top closed-source models such as GPT-4o and Claude-3.5. DeepSeek V3 is now available on DeepSeek's web site, and can be used for free by clicking "Start a Conversation" on DeepSeek's official web site.
In addition to normal text input or file uploading, DeepSeek V3 also has a "Deep Thinking" function similar to OpenAI o1, which allows for multi-step reasoning and thinking before answering a question; and a "Networked Search" function similar to GPT search, which allows you to give answers based on what you've found online. There is also a "network search" function similar to GPT search, which can give answers based on the contents searched on the Internet.
Test upload attachment function: abstract English paper
Let DeepSeek V3 organize the key points of the English paper, and then compare the results with GPT-4o. The results show that the focuses organized by both models are complete and clearly structured, but although DeepSeek can answer in traditional Chinese, its wordings are still based on the usage in China. For example, "movie" is written as "video", "message" is written as "information", and so on. If you want it to answer in Taiwanese terms, you need to specify it in the command.
In contrast, ChatGPT uses terms closer to those used in Taiwan.
Testing the Deep Thinking Feature: Reasoning about Computational Problems
Using DeepSeek's "Deep Thinking" feature, ask it a question that requires reasoning and calculation: "If A stands for 1 and B stands for 2, what is the sum of the numbers represented by the word "rabbit"?" Before giving the answer, DeepSeek will also share its "thinking" process, and you can see that it didn't just solve the question in the simplest way, but carefully interpreted every clue of the question, and even suspected that there was a possibility of "traps". In the end, it took 15 seconds to give the calculated answer.
As for OpenAI o1, it only took 2 seconds to successfully solve the problem without sharing the detailed thought process.
Test the online search function: check out recent tech news
Use DeepSeek's "Internet Search" feature to look up the five most important tech news stories of the week. In addition to organizing the summaries of these five news items, it also provides 43 web pages that it queries. A closer look at the sources of these pages reveals that most of them are from domestic media, and the content is indeed related to recent tech news.
However, there is also some obviously unrelated content mixed in, such as design material sites and the like.
As for the results of GPT Search, it provided 15 web page sources, and labeled the top two web pages as the sources cited in the answers, and the rest as supplementary information. All of these sources are media in Taiwan, and the content is related to recent technology news.
Conclusion: both answers are of similar quality, DeepSeek takes longer to think about it
Overall, DeepSeek and ChatGPT answered about the same quality of these life-like test requirements. The main difference is that DeepSeek mainly uses domestic Chinese sources and terms, while ChatGPT is closer to users in Taiwan. However, when it comes to the deep thinking feature, DeepSeek takes more time to give the final answer compared to ChatGPT, although it can list the detailed thinking process.
In addition, DeepSeek has an unavoidable downside, like theTechCrunch.As mentioned by some media, DeepSeek, as a domestic Chinese company, is restricted by policy from answering "sensitive" questions.