#Rao's Thought Notes · No. 454

01 

Both DeepSeek and OpenAI's o1 large model have super strong reasoning capabilities. Although they use different technical routes, their questioning methods can be used as reference for each other.

Last year, OpenAI and Stanford University professor Andrew Ng, a famous artificial intelligence scholar, launched a course called "Reasoning with o1" on the use of o1. The other lecturer was Colin Jarvis, director of strategic solution architecture at OpenAI. This should be the official explanation of the use of o1 by OpenAI.

In the course, the tips for o1 are as follows:

-The picture content is summarized by Kimi-k1.5 long thinking model

According to Andrew Ng's course content and my experience using o1 in the past month, as well as my experience using DeepSeek in recent days, although the two have different technical routes, they can learn from each other and refer to each other in terms of the way of asking questions. In addition, I think it can be simpler when asking questions to this kind of thinking model.

02 

Since DeepSeek has super reasoning ability, we treat it as a super-capable employee. We don't need to teach this employee how to do things. We only need to focus on the two parts of "input" and "output" to greatly call on the capabilities of DeepSeek.

"Input" is enough background information to give DeepSeek to solve a problem.

"Output" is the rules that DeepSeek must follow when it is asked to give an answer.

"Input" refers to the detailed background information you provide in response to a question, which allows DeepSeek to understand more and give more targeted answers.

For example, if you regard DeepSeek as a personal fitness coach, and you don’t provide DeepSeek with your own body data and exercise preferences, how can it give more reasonable suggestions?

But we don't know what information DeepSeek needs to solve the problem. At this time, you need to let DeepSeek tell you what information it needs to solve the problem. Then let DeepSeek compile a questionnaire, fill it out by yourself, and send it to DeepSeek.

At this time, a universal formula can be used:

You want to solve a certain problem, so what background information do you need?

I usually use this question as a starting point for asking DeepSeek, and the answers I give later are not too bad.

Refer to the following figure:

-The above content is screenshot from DeepSeek

Whenever you ask any questions in the future, DeepSeek can give you fitness suggestions tailored to your personal information.

Similarly, this way of asking questions is also suitable for making plans at work, raising children in life, and other aspects.

"Output" is the rules that DeepSeek must follow when giving an answer.

Some rules are related to language style, and you can use these prompts to make requirements for DeepSeek.

For example, "speak human language", "use language that even elementary school students can understand", "all non-professionals can understand" , answer by imitating someone's writing style, or "this is a plan for the boss", etc.

If you want to write a speech for your leader, you can copy and paste all the content of the leader's previous speeches into a Word file, then upload it to DeepSeek as an attachment, and then ask him to imitate the style of the article in your attachment. The effect is generally better at this time.

There are also writing rules that make specific requirements on the answer format. You can ask DeepSeek to output in a certain order and format.

For example, answering in the format of "what, why, and how" or "summarizing in three key points" works well. If you are writing an article with a fixed format, such as a paper, you can directly let DeepSeek output according to the specified article structure.

There are also some requirements for DeepSeek's way of thinking.

The prompts I personally like to use include "Please answer a question with a counterintuitive point of view" and "Please answer with a critical thinking style." Sometimes I will ask DeepSeek to first give a process for solving a problem, and then ask it to answer according to this process.

For example, you can ask "What is the general process that professional fitness coaches follow when making plans for novices?" and then ask DeepSeek to answer using this process. I have tried it myself and the effect is quite good.

03

In our previous experience, we often use the prompt word model of "personality + background + task + requirements" to write prompt words, which is actually also applicable to DeepSeek.

Character: Who do you want to help you solve this problem?

DeepSeek has too much knowledge. If we don’t limit the application of certain knowledge points, the answers will be rather vague.

Background: What background information do you need to know in order to solve this problem?

DeepSeek is an extremely capable employee, but if you don’t tell it background information, it can only answer questions based on the knowledge it was trained on, which will also be untargeted.

Task: Try to ask only one question in each conversation to avoid multiple questions interfering with each other's answers.

DeepSeek has a maximum word limit for each answer. If there are too many questions and the answers to each question are not detailed, it will be like scratching the surface.

Requirements: Rules or precautions that you must follow when answering.

DeepSeek's answer itself has a set of rules, but some companies have their own writing framework style, such as weekly and monthly report format requirements. At this time, if you tell DeepSeek in advance, you can get a better answer.

The structure of this prompt is also the simplest content. In the "Background" section, you can directly ask DeepSeek what background information is needed to solve this task. If there is no "Requirement", you can leave it blank.

The question model above also optimizes DeepSeek in terms of "input" and "output".

04

To sum up, for a large model like DeepSeek with super strong reasoning capabilities, there are only two things we can do:

"Input" should include enough background information, and "output" requirements such as language format should be clearly defined.

For the intermediate reasoning process, we try to intervene as little as possible and let DeepSeek think on its own.

If the answer given is not ideal, then change the dialog box, ask a few more times, and you will always get a suitable answer.

"Seeing" is the problem, "doing" is the answer.

Use DeepSeek more often and summarize constantly, and you will be able to find a questioning method that is more suitable for you.

The truth is simple, just use it more.

-End