What is DeepSeek?

DeepSeek, fully known as Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., was established on July 17, 2023. It is an innovative technology company focused on developing advanced large language models (LLM) and related technologies. In Silicon Valley, DeepSeek has long been referred to as “The Mysterious Power from the East”, and is also one of the highly discussed entities known as “Hangzhou’s Six Little Dragons” online.

DeepSeek’s weekly download volume reaches about 2.4 million times, possibly surpassing the downloads achieved by Doubao through advertising during the previous year’s Spring Festival;
The significant drop in the US stock market was triggered by two models, DeepSeek-V3 and DeepSeek-R1, with the former being a class 4o model and the latter a class o1 model;
DeepSeek-V3’s training cost is only 1% of Llama 3’s, and DeepSeek-R1’s inference cost is only 3% of OpenAI o1’s;
Many innovations in DeepSeek-V3 are related to overcoming memory and bandwidth limitations caused by using H800 (instead of H100);
The R1 series models abandoned the HF (human feedback) part of RLHF, leaving only pure RL (reinforcement learning), during which the model exhibited the ability to “reflect”;
DeepSeek also fine-tuned Alibaba’s Qwen model with its 800,000 thought chain data, resulting in improved inference capabilities for the latter;
DeepSeek has not yet launched a financial investment large model, but this is just a matter of time.

The Correct Way to Use DeepSeek

1. The Most Important Secret: Throw Away Your Prompt Templates

If you are still using various “professional prompts” and “templates,” you are heading in the wrong direction.

DeepSeek does not operate this way at all.

Why?

Because its core is a reasoning large model, not an instruction large model.

It’s like having two interns:

A little bookworm who needs you to arrange every task step in detail. (Instruction-based)
A clever little one who can think for themselves as long as you explain the purpose. (Reasoning-based)

To illustrate with a real case, one of our community’s operation members tested it for analyzing the new energy industry, preparing for negotiations with BYD suppliers.

Traditional Approach:

Please act as a new energy industry analyst and analyze according to the following steps:
1. Market size
2. Competitive landscape
3. Technical route
4. Future trends
Requirements: 800 words for each part, citing authoritative data...

Result: A dry report that looks AI-generated at first glance.

The Correct Way to Use DeepSeek! — AI-generated, with no real use.

Correct Approach:

I need to negotiate with BYD's suppliers next week, but I don't understand power batteries.

Help me explain in the simplest language:
1. Where are their technical advantages
2. How much might they charge
3. What professional terms can I use during negotiations

The key is to make it understandable for me, so I can sound like an insider.

Result: DeepSeek provides a grounded analysis along with negotiation phrases.

This is the biggest difference:

DeepSeek does not need you to write “professional prompts,”

It needs real scenarios and specific needs.

Here’s a universal formula for you:

I want xx, to be used for xx, hoping to achieve xx effect, but worried about xx issues…

2. The Most Underestimated Feature: Make It “Speak Human”

Many people complain that DeepSeek’s responses are too abstract, like reading a celestial book.

But what you may not know is that with just a simple prompt, this issue can be completely changed.

This magical prompt is:

Speak human.

That’s right, just these three words.

Original Response:

Response after adding “speak human”:

It instantly becomes relatable, right?

Because DeepSeek is very sensitive to the phrase “speak human.”

3. The Most Powerful Skill: Deep Thinking

A free domestic AI is making the $200-a-month GPT-o1 uneasy.

Why?

Because DeepSeek’s way of thinking is more intelligent than GPT-o1.

To compare with a real case:

GPT-o1’s response:

DeepSeek’s response:

This is the biggest difference:

GPT-o1 lists linearly, like a high-level document tool
DeepSeek thinks deeply, like a thinking partner

The free DeepSeek has directly made the entire Silicon Valley AI companies work overtime,

Fortunately, they don’t have to celebrate the Spring Festival.

However, recently, a phenomenon has been observed:

Due to a surge in users, DeepSeek has clearly adjusted its response strategy:

Thinking time has dropped from 20 seconds to 5 seconds
Response depth has noticeably declined
Reflective ability is limited

This is an understandable temporary measure, after all, computing power costs money.

But for us users, how can we continue to stimulate its deep thinking ability?

I have compiled three core prompts, which I call the deep thinking trio:

Please conduct critical thinking at least 10 rounds during your analysis process, be thorough

Please consider your answer from the opposite perspective at least 10 rounds during your analysis process, be thorough

Please review your answer at least 10 rounds during your analysis process, be thorough

In this way, deep thinking will recover from 5 seconds to about 20 seconds.

The italicized parts can be freely replaced with forms you are good at, or can be combined,

but the core is reflection.

4. The Most Powerful Style Converter

Using DeepSeek, I wrote a Hanfu to praise Wang Xing’s girlfriend for her loyalty, wisdom, and courage. Wang Xing is the actor who was recently sold to northern Myanmar.

This use of classical references and parallelism is truly unmatched.

After posting it in the group, it directly shocked three Chinese teachers…

Some friends even wrote texts that professional screenwriters said were mature.

In fact, this is DeepSeek’s fourth secret weapon: the style converter.

The prompt is very simple, you can even dictate it:

Imitate xxx’s writing style to write a xx genre article about xxxxx.

However, note two limitations:

On one hand, it is more suitable for imitating classic authors, as training data is abundant, and style characteristics are distinct.

On the other hand, do not expect a 100% restoration; an 80% similarity is already impressive, and the key is to capture the essence.

I believe DeepSeek has definitely trained on a large amount of literary data, especially Chinese classical literature.

For example, Lu Xun and Mo Yan are both vividly imitated.

Remember that universal formula:

I want xx, to be used for xx, hoping to achieve xx effect, but worried about xx issues…

For example:

I want to write a poem praising Wang Xing’s clever and brave girlfriend, to showcase on Xiaohongshu, hoping to focus on imitating Wang Bo’s style, with the goal of impressing my literary talent and getting a billion likes on Xiaohongshu, but worried that others might find it too obscure to understand…

Similar style transfers can also be applied to contemporary authors, such as imitating Liu Run’s New Year’s speech:

Note to first provide the original text (detailed, generally no less than 8000 words), and then directly request imitation.

Of course, for better results, it’s best to use the universal formula.

5. Usage Restrictions: When Not to Use It

Having discussed so many advantages, we must also talk about its limitations.

The following scenarios are not recommended for using DeepSeek:

1. Long Text Writing

Articles exceeding 4000 words are likely to exhibit logical breaks; it is advisable to use Claude200k.

Because DeepSeek defaults to 64k, it is insufficient for long texts.

2. Sensitive Content

After all, it is a domestic AI, and the built-in review standards are excessively strict.

Many times you don’t know which sentence triggered the review.

How to solve this situation?

Because DeepSeek has a post-review system, there are the following three solutions:

1. Click to edit your question, and submit several times; one of them will surely not trigger a review.

2. During the response generation, keep clicking the copy response button. This ensures that when a review is triggered, your clipboard contains the previous response.

3. For a permanent solution, switch to the big three. (GPT, Claude, Gemini)

3. Personal Style Writing

This does not need elaboration; this is a reasoning model, suitable for problem-solving and imitation.

However, it is challenging to ensure the desired style writing effect through precise control.

This is not really a disadvantage of DeepSeek, but rather a characteristic.

-end

__________________________

Article sources: CCTV, New Cortex NewNewThing, He Zhuzhu

If there is any infringement, please contact the editor to delete it