Learn About DeepSeek: China's New Super Powerful AI Model

Artificial intelligence (AI) is evolving rapidly, and the Chinese-developed AI chatbot DeepSeek is one of the latest technologies making waves. There are many advantages to DeepSeek's quick and detailed responses, but there are also some disadvantages that may not be the AI ​​tool you're looking for.

What is DeepSeek?

DeepSeek is a new AI technology developed by a Chinese technology company. Its flagship model, DeepSeek-V3, uses a unique Mixture-of-Experts (MoE) architecture. Think of it as a 'team' of specialized AI systems, where only the most relevant experts are 'activated' to handle specific tasks, where each 'expert' is a specialized neural network.

DeepSeek claims that its LLM V3 was trained on a whopping 14.8 trillion tokens, with 1 million tokens being equivalent to about 750,000 words. DeepSeek V3 is also a massive LLM overall, with 671 billion parameters across its entire MoE architecture, with 37 billion of those parameters being activated for each token. That means each prompt has massive processing power behind it, resulting in faster and more efficient performance.

Furthermore, it is an open-source AI model, meaning anyone can explore, experiment, and build on it for free. This sets it apart from competitors who keep their models behind paywalls.

Check DeepSeek

To see how DeepSeek-V3 performs, we tested it on three real-world tasks: creative writing, code support, and productivity. Its performance demonstrated both versatility and raw power, with a few minor hiccups along the way.

Task 1: Write a description of a fictional character

The author of the article asked DeepSeek to write a detailed description of a fictional character: A rebellious queen leading a resistance against an evil empire. The output was vivid and engaging.

Task 2: Debug a JavaScript function

Test DeepSeek's programming skills by feeding it a faulty JavaScript function that calculates the factorial of a number. Here is the error code it gets:

​​​​function factorial(n) { if (n = 1) { return 1; } return n * factorial(n - 1); } 

DeepSeek immediately spotted the problem: The equal sign (=) in the condition caused a logic error. It provided the corrected code and also explained the problem:

This fix works perfectly and its explanation is clear enough for a JavaScript newbie to understand.

Task 3: Productivity - Create a meeting agenda

To test productivity, ask DeepSeek to draft a concise agenda for a project team meeting about a new product launch.

DeepSeek excels at creative, programming, and productivity tasks, delivering polished output across the board. As with any AI tool, it's important to consider the accuracy of the output, especially when accuracy is a critical factor.

Pros, Cons, and Things to Note in DeepSeek

After putting DeepSeek to the test, some of the AI ​​model's strengths and weaknesses became apparent. It also surprised us in some unexpected ways.

DeepSeek handles a wide range of tasks with ease. From creating vivid compositions to debugging complex code, it has proven to be as versatile as leading AI models. Its Mixture-of-Experts architecture lives up to its promise, making DeepSeek fast without sacrificing output quality. For example, it generates detailed outputs like character descriptions or meeting agendas in seconds. DeepSeek adapts its tone and style with ease. Whether working on formal documents or compositions, it can tailor its response to the task.

Like many AI models, DeepSeek sometimes provided incorrect or incomplete information, especially when asked about deep or very specific topics. For example, in a test related to historical events, it confidently gave an incorrect answer.

While DeepSeek is open source, its support documentation and tutorials are not as robust as those of tools like GPT-4, making it difficult for new users to maximize the model's potential. Because it originates in China, some users may be concerned about how their data is handled or stored. While there is no evidence that their data has been misused, these concerns are worth considering for sensitive tasks (like any AI chatbot!).

Perhaps most surprisingly, for a model developed in China, DeepSeek demonstrated an incredible understanding of global cultures and perspectives. It showed deep understanding of topics ranging from European literature to African history. Another unexpected strength of DeepSeek is its creativity. When asked to describe a character, it not only provided a description, but also created context around that character and how that character related to other characters.

David Pac

Update 14 January 2025