During this year’s Spring Festival, an “AI rising star” from Hangzhou quietly emerged, named DeepSeek. It struck like a sudden lightning bolt, illuminating the global AI night sky and bringing a mysterious “Eastern power” to the open-source community.
As DeepSeek gained popularity, more and more people began to use this AI tool. So how can we make the best use of it and unleash its powerful capabilities?
It starts with getting to know and understanding it — it’s like a well-read friend who sometimes unconsciously “fills in the blanks”. It may confidently fabricate non-existent data, link unrelated information, or generate seemingly reasonable but actually incorrect content. Therefore,we must remain vigilant while using it,avoiding blindly trusting AI’s “nonsense”.
Additionally, we shouldconstantly learn “how to communicate with AI”. After all, even a powerful AI like DeepSeek can only provide vague answers if we cannot clearly express what we want.
Today, we have compiled a quick guide to DeepSeek for ordinary people, hoping to help everyone utilize this powerful AI tool effectively.
This image was generated under the guidance of DeepSeek
How does DeepSeek actually “think”?
If we were to describe DeepSeek with a metaphor, it would be like a very knowledgeable friend who has not only read countless books but can also instantly establish connections between various pieces of knowledge in their mind, answering your questions without fail (except for illegal matters, of course). This is how modern large language models operate, and the core of this capability is the Transformer architecture pioneered by Google’s team in 2017.
The most powerful aspect of Transformer is its attention mechanism. For example, when you read a book, an ordinary person needs to read it from start to finish, while the Transformer is like a “super reader” that can instantly find the most critical information in the text and quickly understand the relationships between them.
In its growth process, DeepSeek seems like a knowledge-hungry scholar, “reading” the vast knowledge accumulated by humanity—from the dry Wikipedia to beautiful literary works, from cutting-edge academic papers to professional technical documents. By continually predicting the missing words in sentences, it gradually mastered the rules of language. This process is similar to how infants learn their mother tongue through listening and speaking; the large model learns to predict and generate the content we want by hearing questions.
However, DeepSeek’s “thinking” method also has its uniqueness; itis like an impromptu speaking master, where every word spoken is calculated precisely, ensuring that the content is coherent and creative. But because of this spontaneity, its responses can vary like humans, sometimes being eloquent and at other times falling short, and it often makes mistakes—this is the phenomenon of “AI hallucination”.
Before DeepSeek, foreign models like ChatGPT and Claude had already shone brightly, while domestic models like Wenxin Yiyan, Tongyi Qianwen, and Kimi also showcased their brilliance. So why did DeepSeek manage to surprise the world after this Spring Festival?
The answer can be summarized in three keywords: strong capabilities, cost-effectiveness, open-source sharing.
Strong Capabilities: The “Technical Expert” in AI
What sets DeepSeek apart is its “reasoning brain”. Compared to traditional instruction models, DeepSeek has particularly enhanced its reasoning ability in its model design. By leveraging advanced technologies like reinforcement learning, it resembles a thoughtful scholar who not only understands what you’re saying but also knows how to think deeply and innovate.
Currently, DeepSeek R1 has become the “strongest king” among open-source large models, closely chasing OpenAI’s star players and becoming the only powerhouse in the open-source camp that can “wrestle” with closed-source giants.
Swipe up and down to see more. Compared to other mainstream models, DeepSeek R1 is shown in the image: source provided by the author
Affordable Price: The “Cost-Performance King” in AI
DeepSeek not only excels in technology but also impresses with its cost control. By optimizing the training method, they have reduced the training cost of the V3 model to 5.576 million dollars—this number is even lower than the annual salary of many executives at large model companies.
This high efficiency has directly resulted in ultra-low prices; currently, inputting a million tokens only costs 4 yuan, and output costs just 16 yuan, leading some to jokingly call it the “Pinduoduo of the large model world”!
Image source: provided by the author
Open Source Sharing: The “Wulin Alliance Leader” in AI
When other large models set heavy barriers due to commercial interests, DeepSeek chose to fully open its code and allow free commercial use. It’s like a martial arts master openly sharing their lifelong skills, allowing the entire community to enhance their abilities.
This open-sharing philosophy has resonated globally. As a popular saying goes online, “The magical reality of 2025: We bought a model for $200 a month from a nonprofit organization (OpenAI), but received a free open-source AI from a quantitative institution.”
How to Really Make the Most of DeepSeek?
The key to interacting with AI lies in “communication”. AI is like a genius with its eyes covered; it knows a lot but cannot directly perceive the world. Only through our descriptions can it understand what reality looks like.
It’s like describing the sound of waves to someone who has never seen the ocean. If you say, “whoosh,” the other person might interpret it as the sound of rain; if you say, “like countless large silks shaking in the air,” the imagery is entirely different. Similarly, communicating with AI requires this kind of “descriptive ability”.
When communicating with AI, prompts serve as your “expressive ability.” Just like learning a new language, this ability requires constant practice to improve. Even a powerful AI like DeepSeek can only provide vague answers if we cannot clearly express what we want.
Another important point is the “hallucination” issue of AI. It’s like a knowledgeable friend who sometimes unconsciously “fills in the blanks.” It may confidently fabricate non-existent data, link unrelated information, or generate seemingly reasonable but actually incorrect content, so we must be the “sword bearers” against AI hallucinations, remaining vigilant at all times.
In other words, AI is a powerful tool, but the effectiveness of using the tool depends on the user’s ability. Through continuous practice and vigilance, we can truly unleash the value of AI while avoiding falling into the trap of “hallucinations”.
So, how can we make the best use of this “AI tool”? Here are a few tips.
1
Regular Question Template
In the past, when conversing with instruction models like ChatGPT, we often used very complex prompt templates. But DeepSeek is different; it’s more like a thinking partner, where simple and clear expressions work better.I highly recommend this simple yet super practical “four-step questioning method”: Background + Task + Requirements + Supplement. For example:
[Background] I am the owner of a newly opened café;
[Task] I need a promotional plan for the opening;
[Requirements] Within a budget of 10,000 yuan, mainly targeting university students;
[Supplement] Our shop is located in a university town and specializes in premium coffee.
The brilliance of this method lies in:
Background information helps AI understand your situation;
Specific tasks clarify your needs;
Constraints indicate key requirements;
Supplementary information adds important details.
This way of questioning not only helps DeepSeek better understand your needs but also avoids misunderstandings or off-topic answers. This is like talking to a smart friend; once you clarify the situation, they can naturally provide more targeted suggestions.
Remember, when talking to DeepSeek, you don’t need too many fancy techniques; keeping it simple and clear is key. Its reasoning ability is strong; as long as you explain your needs clearly, it can understand your intentions and provide surprising answers.
Practical Demonstration
First, let’s look at an incorrect example:
Question: Write a promotional plan for a store opening.
Output:
Swipe up and down to see more. Image source: provided by the author
Now, let’s look at a correct example:
Question: I am the owner of a newly opened café and need a promotional plan for the opening. The budget is within 10,000 yuan, mainly targeting university students. Our shop is located in a university town and specializes in premium coffee.
Output:
Swipe up and down to see more. Image source: provided by the author
2
Decomposing Complex Tasks
Currently, due to context and model capability issues, AI often tends to “cut corners” when handling complex tasks. At such times, we need to help AI break down tasks.
Next, using the example of writing an article on “the future development of remote work,” I will show you how to guide AI step by step to complete high-quality writing.
Step 1: Have AI write an outline
Example: “I want to write an article about ‘the future development of remote work’; please help me list a detailed outline, including an introduction, 3-4 main points, and a conclusion. Each main point should include 2-3 specific subsections.”
Step 2: Have AI elaborate on the content of the first part based on the outline
Example: “Based on the outline above, please help me detail the introduction part. It needs to include the following points: an introduction to the current state of remote work, why this topic is important now, and the main directions the article will discuss. Please support with vivid examples and data, and the language should be smooth and natural.”
Step 3: Have AI continue to output the content of the following parts based on the outline
Example: “Please continue to complete the first main point section in the outline. Ensure that the content naturally connects with the introduction, and provide specific case support. Each subsection should be around 300-400 words, and pay attention to transitions between paragraphs. If data is involved, please clearly indicate the source.”
Step 4: Have AI check the article for errors and verify information
Example: “Please carefully read the entire article, focusing on:
1) Whether the data and facts mentioned in the article are accurate;
2) Whether the logical relationships between sections are smooth;
3) Whether there are any contradictory viewpoints;
4) Whether the use of technical terms is accurate. If any issues are found, please specify and provide modification suggestions.”
Step 5: Have AI read the generated article and optimize it
Example: Please optimize this article from the following aspects:
1) Increase the readability of the article by using more vivid analogies and examples;
2) Optimize transitions between paragraphs to make the article smoother;
3) Check and improve the expression of the article to avoid repetition and verbosity;
4) Appropriately add some new insights or viewpoints to make the article deeper. Please modify it paragraph by paragraph and explain the reasons for each modification.
3
Continuous Questioning Techniques
Do not expect AI to complete the content you want in one go. When using AI, adjust your mindset and treat it like a knowledgeable but inexperienced intern. It has a wide range of knowledge but needs your guidance to deliver satisfactory work.
For example, the opening plan for the café written above will certainly not meet our needs in the first draft; we need to continue asking questions to optimize the content until it meets our requirements.
Using the café opening plan as an example, the first version provided by DeepSeek did not highlight the target group of “university students” well. At this point, don’t rush to dismiss it; instead, you can continue the dialogue like this:
The overall framework of the plan is good, but I think it can be more aligned with the university student group. Can you optimize it from the following aspects:
– How to design the store space based on students’ learning and social needs?
– Considering students’ spending power, how to design a more attractive pricing strategy?
– Can we design marketing activities around campus events like finals and club activities?
It’s like guiding an intern, telling them the specific areas that need improvement. DeepSeek will adjust the plan based on this new information.If the revised plan is still not satisfactory, you can continue to ask:
The plan has become more targeted, but I think it can be further refined in these areas:
– Can we add a membership points system suitable for students?
– How to use social media to attract young customers?
– Considering the daily routines of students, should the business hours be adjusted?
This process is like a brainstorming session, where each round of dialogue helps the plan become more refined. The key is to:1. Be patient and not expect instant results2. Provide clear directions for improvement3. Gradually improve, one aspect at a time4. Summarize and consolidate in a timely manner to ensure the direction is correct
4
Skillfully Use Different AIs
Each AI is like an assistant with a distinct personality; understanding their strengths allows them to shine in the most suitable roles.
For example, DeepSeek R1 is like a meticulous strategic advisor. Give it a task, and it will think deeply from various angles, not only completing your requested content but also proactively discovering and supplementing details you might have overlooked. Especially in planning, it always provides surprisingly complete ideas.
However, when it comes to literary creation, DeepSeek behaves like an overly enthusiastic writer, often piling on ornate modifiers, which can detract from the overall effect of the article. At this point, Claude, with its relatively restrained personality, can come into play; its writing is more concise and elegant, making it more suitable for tasks requiring literary finesse.
Knowing who is best suited for what allows the team to maximize efficiency. Similarly, when using AI, selecting the most appropriate “assistant” based on the nature of the task often yields better results.
|
Name |
Usage |
Characteristics |
|
ChatGPT |
Copywriting, idea organization, knowledge inquiries, emotional support |
Text is somewhat rigid, has no significant weaknesses, but lacks standout features |
|
Claude |
Long-form writing, code generation, idea organization |
Generates the highest quality Chinese content, virtually free of AI flavor |
|
DeepSeek |
Long-form writing, code generation, mathematical calculations, idea organization |
Very suitable for organizing ideas and plans, has beautiful Chinese language skills, excels at poetry, but has a notable tendency for AI hallucinations |
|
Doubao AI |
Multimodal analysis, daily office assistance, emotional support |
Comprehensive functionality, suitable for daily work use, also very good for emotional support |
|
Recraft |
Image generation |
Professional AI image generation tool, high-quality output |
|
Midjourney |
Image generation |
High-quality visuals, strong playability |
|
Whisk |
Image generation |
Google’s image tool can generate different images based on characters and scenes, with good image quality |
|
Keling AI |
Image and video generation |
One of the most advanced video generation tools currently, with high video generation quality |
|
Hailuo AI |
Voice and video generation |
Generates high-quality voice quickly and is free |
Table: Designed by myself
Planning and Production
Author | Tian Wei, AI tool researcher
Reviewed by | Yu Naigong, Head of Robotics Engineering at Beijing University of Technology, Director of the Robotics Research Center at the Beijing Academy of Artificial Intelligence, PhD supervisor
This article is authorized and reprinted from the WeChat public account of Science Popularization China
Series Links:
Technology Frontier | “Nezha 2” Breaks Global Box Office Records! How Technology Creates the “Nezha Universe”?
Technology Frontier | What is “DeepSeek,” which everyone is discussing? What does it look like in Fengtai?
Technology Frontier | Starting with a flip! Directly hitting the J-35A air show debut
Technology Frontier | New “Space Business Trip Trio” resumes published
Technology Frontier | “Starship” Successfully Recovered for the First Time! What’s Difficult About “Chopsticks Clamping a Rocket”?
This issue was edited by: Beijing Science and Technology News, Jia Shuo Rong
Organized by:Fengtai District Science and Technology Association, Beijing
Undertaken by: Beijing Science and Technology News | Beike Media