Tsinghua University's Third Bullet: How Ordinary People Capture the DeepSeek Dividend.pdf

In the field of artificial intelligence technology. DeepSeek is a high-profile Chinese company specializing in the development and application of general artificial intelligence (AGI). DeepSeek The core business of the company revolves around big models and strives to make breakthroughs at the forefront of innovation in AI technology.

DeepSeek's core product: DeepSeek-R1

DeepSeek Star Products DeepSeek-R1 is an open source inference model. This model performs well when dealing with complex tasks, and more commendably, it allows free commercial use. DeepSeek-R1 The performance is very impressive, comparable to OpenAI's o1 release, which makes it stand out from the crowd of open source models.

DeepSeek Reinforcement learning techniques were employed on a large scale in the post-training phase of model training. Surprisingly, even with very little labeled data, the DeepSeek-R1 The reasoning ability has also been significantly improved. On critical tasks such as math, code, and natural language reasoning, the DeepSeek-R1 Demonstrates excellent performance similar to the official OpenAI-o1 version.

DeepSeek-R1 model architecture

DeepSeek's capability mapping: a wide range of application scenarios

DeepSeek Not only directly for end users, but also provide strong support for developers, its application scenarios are very wide, mainly including:

• intelligent dialog
• Text Generation
• semantic understanding
• computational reasoning
• Code Generation and Completion
• Networked Search and Deep Thinking Models
• File upload: able to scan and read the text content in all kinds of documents and pictures.

DeepSeek capability mapping

Text generation: an enabler of creativity and efficiency

DeepSeek It excels in text generation and can be applied to a variety of scenarios:

Text creation: unleashing creative inspiration

• Essay, story, and poetry writing: whether it's a long tome or a short, sharp DeepSeek All are competent.
• Marketing copy, tagline generation: quickly produce eye-catching marketing content.
• Social media content (e.g. tweets, posts): easily create engaging social media content.
• Script or dialog design: Provide script writing and dialog design support for film, television, game and other industries.

Abstracting and rewriting: improving the efficiency of information processing

• Long text summaries (papers, reports): quickly distill the core content of long documents.
• Text Simplification (Reduced Complexity): Converts specialized text into an easy-to-understand version.
• Multi-language translation and localization: breaking down language barriers to achieve multi-language distribution of content.

Structured generation: efficient organization of information

• Form and list generation (e.g., scheduling, recipes): quickly generate structured data for easy information management and use.
• Code comments, documentation writing: improve code readability and development efficiency.

Natural language understanding and analysis: insight into the meaning behind the text

DeepSeek It is also very powerful in natural language understanding and analysis, able to dig deeper into the semantic information of the text:

Semantic analysis: understanding the essence of language

• Semantic parsing: accurately understand the intent of the user's query.
• Sentiment analysis (comments, feedback): analyzing the text for sentiment tendencies, e.g. positive, negative or neutral.
• Intent recognition (customer service dialog, user query): accurately determine the user's needs and purpose.
• Entity Extraction (names, places, events): Extract key information from text to build a knowledge graph.

Text classification: efficient information categorization

• Text categorization: divides text into predefined categories.
• Topic tag generation (e.g. news categories): automatically add topic tags to text for easy information retrieval and management.
• Spam Content Detection: Filtering undesirable information and maintaining a healthy online environment.

Intellectual reasoning: exploring the mysteries of logic

• Knowledge Reasoning: logical deduction based on known knowledge to discover hidden associations.
• Logical Problem Solving (Math, General Knowledge Reasoning): Solve problems that require logical thinking, such as math problems and general knowledge quizzes.
• Causal analysis (event correlation): analyzes the causal relationship between events, e.g., whether the occurrence of event A causes event B to occur.

Routine plotting: a data visualization tool

DeepSeek A variety of plotting functions are also supported, allowing data and information to be presented in a more visual way:

• SVG Vectors
• Mermaid Chart
• React Charts

Basic Graphics: Quickly Create Visual Elements

• Simple illustrations: used to decorate documents or presentations.
• Flowchart: Used to show the process steps.
• Organizational Chart: Used to show the organizational structure.

Types of charts: to meet diversified display needs

• flow chart
• chronology
• class diagram
• state chart
• Entity Relationship Diagram
• mind map

Data charts: professional data analysis presentation

• Line graphs: show the trend of data over time.
• Bar charts: compare the size of data in different categories.
• Pie charts: show the proportion of each component's data to the total.
• Scatterplot: shows the relationship between two variables.
• Radar charts: comparative analysis of multidimensional data.
• Combined Charts: Combine multiple chart types together for a more comprehensive presentation of data.

Programming and code-related: the programmer's right-hand man

DeepSeek Code Capability Example 1

DeepSeek Code Capability Example 2

Code Generation: Improving Development Efficiency

• Generate code snippets on demand (Python, JavaScript, etc.): quickly generate commonly used code to save development time.
• Auto-completion and comment generation: Improve code writing efficiency and readability.

Code Debugging: Reducing the Bug Rate

• Error analysis and repair suggestions: help developers quickly locate and fix code errors.
• Code Performance Optimization Tip: Improve the efficiency of code operation.

Technical Documentation Processing: Streamlining Documentation

• API document generation: automatically generate API documents for developers to use.
• Code base explanation and example generation: help developers quickly understand and use the code base.

DeepSeek's way of thinking: unique and deep

Case Study: "Spice Trio" Eggplant and Pepper Sausage Rolls with Yogurt Mint Sauce

To better understand DeepSeek of the deep thinking process, let's look at an interesting case study - creative conceptualization of a dish. DeepSeek An example of a unique way of thinking about a dish called Spice Trio.

Inspiration Map: Balkan spices × Mediterranean freshness × Japanese tempura technique

structural deconstruction::

1. thinly sliced green pepper: Deep fried using the Japanese tempura technique until translucent, forming an amber-colored crispy net.
2. eggplant mousse: Incorporate Bulgarian paprika with Greek yogurt for an airy mousse.
3. sausage crisp: Iberian chorizo sausage baked at low temperature and crushed into golden crispy grains.
4. Peppermint Cold Essence Oil: Emerald-colored essential oil extracted from liquid nitrogen-frozen peppermint leaves.

taste storm::

• 1st floor: Green pepper crisps release the characteristic sweetness of bell peppers.
• 2nd floor: Smoked paprika blasting through the mousse for warmth.
• 3rd floor: Crumbled chorizo brings a savory, slightly spicy graininess.
• final chapter: Peppermint Cold Essence Oil explodes on the tongue with a polarizing coolness.

Molecular Cuisine Techniques::

1. backspin styling: The traditional rolls are turned inside out and out, with the crispy mesh exposed to encase the tender interior.
2. The temperature difference game: 65°C warm mousse and -196°C cold extracted oil on the same plate.
3. qualitative paradox: The seemingly hard crunchy pieces crumble in your mouth, and the soft mousse conceals blasting crunchy particles.

DeepSeek Not only can you understand complex dish concepts, but you can also create innovative ideas based on user needs.

Innovative Ideas for Dishes: New Possibilities for Eggplant, Peppers and Sausage

User questions: Please help me conceptualize a new dish that will use eggplant, peppers, and sausage. It has to be different from the dishes that are available now.

DeepSeek's Deep Thoughts (takes 18 seconds)::

users want DeepSeek Help conceptualize a new dish that involves eggplant, peppers, and sausage and is different from an existing dish.

Analyze the characteristics of the ingredients:

• eggplant (Solanum melongena L.): Absorb flavor and have a soft texture.
• Capsicum annuum: Crisp, with a hint of sweetness or spice.
• ham: Savory, possibly smoky.

Think about the direction:

• Avoidance of existing dishes: You can't go Chinese with fish-flavored eggplant or ground pork, or Western with baked goods or anything like that.
• A blend of seasonings from different regions: Consider Southeast Asian spices, South American flavors, or Middle Eastern seasonings.
• Innovative cooking methods: For example, eggplant is pureed, peppers are used as containers, and sausage is used as filling.

Suggested Recipes:

Cuisine: Thai Coconut Eggplant and Green Pepper Boat

ingredients::

• eggplant (Solanum melongena L.)
• Capsicum annuum
• ham
• coconut milk
• curry powder
• rosemary
• fish sauce
• mint

move::

1. Handling Eggplant: Bake and hollow out the eggplant, mix in the diced sausage, coconut milk and curry powder and fill back into the eggplant.
2. Green Pepper Processing: Cut peppers in half, remove seeds, and blanch to maintain crispness.
3. padding: Fill peppers with eggplant mixture.
4. flavored: Drizzle with a sauce made with lemongrass, fish sauce and mint.
5. plating: Garnish with mint leaves.

Dish Name Suggestion:

• Thai Coconut Eggplant and Green Pepper Boat
• Mediterranean Spice Aubergine Box

How to use DeepSeek at work?

Scenario 1: 10,000 word project book in one hour challenge

take: At 3pm, you suddenly receive a notice from your leader: "A 10,000-word proposal for an intelligent logistics park project must be delivered by 4pm tonight, with a temporary early meeting with the client!"

reflect in depth::

Users need to complete a large amount of text work in a very short period of time, with tight deadlines and a high demand for content quality.

prescription::

Core Strategy: Framework Replication + Module Population + Data Grafting

Phase 1: 5 Minutes - Quickly Build a Framework with AI (Goal: 3,000 words)

1. input::
   Example of a DeepSeek command

2. manipulate::
   • Directly copy the table of contents as a document skeleton for fast word count (about 2000-3000 words).

Stage 2: 20 Minutes - Batch Fill Module with AI (Goal: 6000 words)

• Ask separate questions for each subsection, for example:
  "Write a paragraph on '2.1 Functional Partitioning', which is required to contain the technical parameters of the Automated Stereo Warehouse, AGV Dispatch Center, and Cold Chain Zone, presented as a list of data."

Key Tips::

• data grafting: In the absence of concrete data, simply let the AI generate reasonable dummies (labeled "examples" to avoid risk).
• Template Copying: Use the same command template for similar sections (e.g. 3.1/3.2/3.3), replacing only keywords.
• mandatory format: Require AI output with numbered subheadings, sub-points, and tables, and then paste them directly into the text for a "professional look and feel".

Phase 3: 20 Minutes - Completing Soft Content with AI (Goal: 1000 words)

• Fill in the "false but necessary" part:
  • policy endorsement: "Generate 5 original 2023 national-level policies supporting smart logistics parks (with issue numbers) and interpret the guidance for this case."
  • risk assessment:: "List the top 3 technology risks common to smart logistics parks (e.g., AGV system downtime), with a 100-word response for each."
  • Benefit Measurement: "Using the formula: cost reduction and efficiency gains within 3 years of the park's completion, assuming a reduction in labor costs of 301 TP3T and a reduction in sorting error rates of 251 TP3T."

Phase 4: 10 Minutes - Optimization and Formatting with AI

• standardize one's speech::
  "Rewrite the following paragraph in the style of a government report, adding keywords such as 'digital transformation' 'dual carbon strategy': {paste original}"
• Generate Charts::
  Directive: "Convert the 'Device Configuration Table' above to a three-line table in LaTeX format."
• final touches::
  "Examine the following logic gaps in the proposal, list 3 points that could be questioned by the client, and provide answers to address them."

Key reminders::

• life-saving priority: Pile on the words before optimizing, and focus on "getting the document to 10,000 words" for the first 30 minutes.
• Virtual Data Labeling: Suffix "(example)" to all AI-generated data to avoid liability.
• trick of the eye: More tables, charts, numbered lists, visually masking the relatively thin text content of the problem. Title font enlargement, paragraph spacing widening, rapid "expansion" of the number of pages.
• If I have time, I'll refine it.The

Scenario 2: New employees quickly familiarize themselves with the company and the electronic components industry

take: Li has just joined a company that manufactures electronic components. As a new employee, he needs to familiarize himself with the company's product lines, organizational structure, internal systems, and the industry in general in a short period of time.

prescription::

Quickly learn about the industry and market trends

• procedure::
  1. Upload industry reports, market analysis documents, competitor information, etc. DeepSeekThe
  2. Use the search function to query keywords such as "current situation" and "future development trend" of the electronic components industry and generate concise analysis reports.
• dominance::
  • Higher efficiency:: DeepSeek By integrating information sources and intelligent searching, the time spent manually finding and organizing information is reduced, and new employees can get up to speed quickly in a short period of time.
  • more effective: Information is more accurate and comprehensive, reducing misunderstandings and errors caused by incomplete information.
  • Lower cost: Reduces reliance on training resources, new employees can be DeepSeek Self-directed learning to reduce training costs.

How DeepSeek works in your life

Scenario 1: Early Morning Productivity Boost for Working Moms

scene restoration7:15 a.m.: I was awakened by a message from the kindergarten parent group, realizing that it was my turn to take care of the craft materials for my class today. At the same time, I remembered that my husband asked me to pick up my clothes from the dry cleaner before he went on a business trip, the fridge was empty and I needed to buy milk, I needed to prepare my PPT for the departmental presentation at 3pm, and the water on the stove was about to boil.

prescription::

1. prioritize: Kindergarten Affairs > Meeting Preparation > Life Purchase.
2. Generate optimal routes: Map showing the location of the kindergarten/dry cleaner/supermarket in relation to the company.
3. Instant Service Docking::
   • Calling the Community Runner API to place an order for manual material delivery
   • Access to the dry cleaner's smart cabinet system appointment pickup code
   • Automatic replenishment of milk after price comparison on fresh food platforms
4. Preparation for the Conference::
   • Automatically extract last week's sales data to generate a visual charting framework
   • Pulling historical report templates for semantic reorganization
5. risk warning::
   • Cooktop timer synchronized with phone vibration alert
   • Real-time monitoring of commute conditions (triggers backup plan if traffic jam lasts more than 15 minutes)

technology dividend::

• Increased time utilization 40%
• Decreased morning pressure values 65%
• Completion rate of critical services 100%

Scenario 3: Critical Incident Response - Pregnant Women in Distress, Data Center Emergencies

scene restoration: A typhoon strikes, leaving a 34-week pregnant wife stranded in the suburbs, a data center with a 4-hour backup power supply, an urgent need to relocate an elderly man who lives alone and has lost his mind, and a shortage of supplies due to a community shopping spree.

prescription::

DeepSeek Emergency Response Protocol:

1. Resource heat map::
   • Real-time integration of weather bureau data/road collapse reports/hospital reception status
   • Material prediction algorithms target unsaturated convenience stores within 3 kilometers
2. lifeline project::
   • Maternity Rescue Channel::
     • Automatic generation of QR codes for medical records
     • Drone survey of possible routes
     • Coordinate with civilian rescue teams. GPS location.
   • Elderly Transfer Program::
     • Retrieve the historical activity track of the smart bracelet
     • Community Volunteer Network Instant Radio
3. Enterprise level emergency response::
   • Initiate edge computing nodes to transfer critical data
   • Generate government disaster report templates (auto-populated with damage assessments)
4. Social collaboration::
   • Creation of a temporary material exchange blockchain ledger
   • Automatic generation of multi-language requests for assistance (interfacing with consulate systems)

technology dividend::

• 3.2 times faster rescue response
• Decrease in loss of assets 78%
• Crisis duration compression 56%

DeepSeek Helps Maintain Social Relationships

Scenario 1: AI helps you skillfully cope with the New Year's call for marriage.

Scene Description: Chinese New Year family gatherings, facing the urging of relatives and elders to get married.

goal: Handle the issue of marriage reminders appropriately without hurting the feelings of elders and maintain family harmony while protecting personal space and choices.

Conversation Skills:

• Use of "I" statements:: Avoid accusations, as in "I feel...".
• Use humorFor example, "For now, I'll focus on my work, and then I'll come back to trouble you when I'm with someone!"
• Avoiding conflicts: Finding common ground such as "You're right, marriage is important and I'll give it serious consideration."
• advance communication: Discuss with parents in advance to minimize surprises.

Proper handling strategy:

1. Express gratitude and build bridges::
   • typical example: Smiling, I responded, "Thank you for your concern, I appreciate your love."
2. Open communication and expression of personal positions::
   • typical example: A gentle clarification, "I have some goals I want to achieve at work right now, and the marriage thing will take time and serious choices."
3. Setting boundaries and clarifying personal bottom lines::
   • typical example: Politely but firmly, "I'll give it serious consideration, but would like to have the space to make my own choices."
4. Change the subject and lighten the mood skillfully::
   • typical example: Take the initiative and ask, "What's new with you lately? I hear you've been studying photography lately?"
5. Seek family support::
   • typical example: privately to my parents, "I know everyone cares about me, but I want time to find the right partner, and I hope you will understand and support me."

Scenario 2: Intergenerational communication in mother-in-law/daughter-in-law relationships

contexts: You and your husband live with your in-laws after marriage. Due to generational differences, you and your mother-in-law have friction from time to time in the education of your children, the distribution of household chores, living habits, and so on.

goal: Ease conflicts between mothers-in-law and daughters-in-law through effective communication and mutual understanding, find solutions acceptable to both parties, and improve the family atmosphere.

Conversation Skills:

1. Use of "I" statements:: Avoid blaming the other person, as in "I am concerned that...".
2. Avoiding arguments: Focus on solving problems, not arguing about right and wrong.
3. Finding common goals: Emphasize that both parties want happy children and family harmony.
4. advance communication: Discuss with your mother-in-law ahead of time, when you are emotionally stable, about education styles and distribution of household chores.

Proper handling strategy:

1. Calm emotions and avoid impulsive reactions::
   • typical example: Take a deep breath and leave the scene of the argument for a moment to give yourself time to calm down.
2. Express feelings and avoid direct accusations::
   • typical example: Expressed in "I" statements, "I am concerned because I want my child to have more freedom to explore."
3. Listen to each other's point of view and find common ground::
   • typical example: Ask your mother-in-law, "What do you think? Why do you think it's better?"
4. Searching for consensus and proposing compromises::
   • typical example: suggests that "we can try to combine our approaches and find a balance."
5. Setting communication boundaries and emphasizing mutual respect::
   • typical example: gently but firmly, "I understand your intentions, but I want us to respect each other."

Scenario 3: Maintaining Good Working Relationships with Leaders

contexts: You are a newcomer to the workplace and have been on the job for a year. You are a hard worker, but during a recent performance evaluation, you disagreed with your leader.

goal: Resolve differences with leaders, maintain good working relationships, and find mutually acceptable solutions through effective communication and mutual understanding.

Example of a specific dialog:

1. Calm your emotions and avoid emotional responses::
   • typical example: After the meeting, go back to your office to calm down and do not respond immediately.
2. Express personal feelings and avoid direct accusations::
   • typical example: The next day, ask to speak with your leader and express your feelings, "I feel a little disappointed that some of my contributions were not fully recognized in my performance review."
3. Listening to the leader's point of view and finding common goals::
   • typical example: Ask your leader, "What specifically do you think I can improve in the areas of teamwork and communication? I'd love to hear your suggestions."
4. Seek consensus and propose improvements::
   • typical example: suggested, "Maybe we can work together to set some specific goals for improvement and communicate my progress on a regular basis."
   • further onwards: "If I excel in some areas but fall short in others, we can work together on how to balance those performances."
5. Setting expectations and clarifying communication boundaries::
   • typical example: gently but firmly, "I would like to get more feedback and support in the future so I can better understand your expectations."

Scenario 4: Inheritance Difficulties in Cross-Cultural Families

contexts: You are a Chinese man who grew up overseas with a Chinese father and an American mother, and your family members are of different religions (Buddhist father, Christian mother). Your grandfather (father) is elderly and has an inheritance and family business in China.

goal: To properly deal with the issue of family inheritance rights while respecting traditional culture and modern laws, to maintain family harmony and to avoid conflicts arising from cultural differences and family secrets.

Proper handling strategy:

1. Stay calm and avoid getting emotional::
   • typical example: When you learn of your grandfather's decision, calm down and don't react immediately.
2. Express feelings and avoid blame::
   • typical example: Communicate with grandfather and parents to express feelings, "I understand my grandfather's wishes, but there are complexities involved in this matter and I am a little confused."
3. Listening to perspectives and seeking common ground::
   • typical example: Ask your grandfather, "Why do you want me to inherit the family business? Was this always the idea?"
   • further onwards: "Mom, what do you think should happen to the estate? What are your thoughts?"
4. Searching for consensus and proposing compromises::
   • typical example:: Suggests that "it might be possible to combine traditional and modern ways of dealing with the estate. One part to be distributed according to family tradition and the other part to be distributed equitably according to modern law."
   • further onwards: "We might also consider disclosing family secrets to avoid future disputes."
5. Setting boundaries and emphasizing mutual respect::
   • typical example: gently but firmly, "I understand the sentiment, but I hope we can respect each other's cultural backgrounds and legal concepts."

Example of a specific dialog:

• you: Grandfather, Mom, Dad, we need to have a good conversation about the estate. I understand Grandfather's wishes, but it's a complicated matter, and I hope we can find a solution that is acceptable to everyone.
• grandfathers: I want you to inherit the family estate. You're the only grandchild, and I want you to remember our roots.
• motherhoodBut it's not fair. According to modern law, inheritance should be distributed equally and should not be given to just one person. And the illegitimate son, he's entitled to it too.
• paternity: I understand everyone's concerns, but it's also important to respect my grandfather's wishes. He fought for his family all his life and wants the family to continue.
• you: I understand how people feel and how complicated things are. Perhaps a balance can be found. Part of it should be distributed according to tradition, and part according to modern law and fairness. As for family secrets, I think we should be honest about them to avoid bigger problems later.
• grandfathers: I can consider this proposal, but you must take on the family responsibility of passing on culture and tradition.
• motherhood: I agree with the suggestion, but want to be fair and have something for everyone.
• paternity: I think it's a good way to work together and find a solution that everyone is happy with.

Conversation Skills:

1. Use of "I" statements:: Avoid accusations, as in "I feel...".
2. Avoiding arguments: Focus on solving problems, not arguing about what's right or wrong.
3. Finding common goals: Emphasize family harmony and equitable distribution of inheritance.
4. advance preparation: Organize your opinions and feelings before communicating and avoid getting emotional.
5. Bringing in the professionals: Seek the help of a lawyer or family counselor, if necessary, to ensure that the distribution of the estate is legal and fair.

Cueing Strategies: Reasoning Models vs Generic Models

From "giving instructions" to "expressing needs": the evolution of cueing strategies

Boundaries and Breakthroughs in AI Content Generation

AI Content Generation Boundaries

AI content generation has boundary effects

Exploration of Knowledge Generation Boundaries for Intelligent Bodies

• research objective: Understanding whether intelligences can break out of established knowledge frameworks for creative knowledge generation.
• methodologies: Constructing measurement systems and analyzing innovation conditions to explore how intelligentsia can drive knowledge generation from routinization to innovation.

Intelligent Body Knowledge Cycle Boundary Study

• impunity: Intelligentsia often experience "knowledge loop boundaries" in long conversations, i.e., the generation of content that is repetitive or limited to a specific pattern.
• rationale: Stems from limitations in training data, algorithmic models, and predefined rules, and is related to the self-referential problem in logic (e.g., Russell's Paradox, Gödel's Theorem).
• Research methodology: Experimentally analyze the effect of the type of problem (fully convergent, semi-convergent, non-convergent) and the number of conversations (50, 100, 150) on the similarity and innovativeness of the generated content.

How do you determine the boundaries of AI content generation?

• Research methodology: Operationalize the knowledge loop boundary of the intelligentsia as the difference value of the generated content, measured as the weighted value of the average similarity and repetition rate of the generated text.

Similarity calculation

• methodologiesThe cosine similarity algorithm is used to transform the text into word frequency vectors, and the ratio of the dot product of the vectors to the product of the mode lengths is calculated to evaluate the similarity between the texts, which takes the value in the range of [-1, 1], and the closer the value is to 1, the higher the similarity is.
• formulas: Similarity = A⋅B||A||×||B||

Repetition rate calculation

• methodologies: Using the n-gram method (n=2), the generated text is divided into consecutive 2-gram fragments and the percentage of duplicated fragments is counted.
• vantage: The ability to identify textual redundancies and assess content diversity is particularly suitable for long text generation.
• final formula::
  Difference value = w1- Similarity + w2× Repetition rate
  where the weights w1 = 0.6 and w2 = 0.4, refer to Kleinberg's (1999) study on social network analysis, which emphasizes the priority of relevance over redundancy.

Combining Adaptive Feedback and Progressive Cue Chaining to Improve AI Content Quality

• impunity: Intelligentsia are prone to hitting cognitive boundaries in multiple rounds of interaction, manifesting themselves in the solidification of generated content and the stagnation of incremental information.
• Results: Highly convergent cues lead to content convergence, while non-convergent cues and diverse designs can break out of the logic loop.
• prescription: The combination of adaptive feedback and progressive cue chaining can push intelligentsia to generate new content, avoiding knowledge loops and broadening the generative space of intelligentsia.

Adaptive feedback and progressive cue chains

How to test the innovativeness and value of AI-generated knowledge?

Testing knowledge after it has been generated: studying the closed loop

• goal: Return to the definition of knowledge: validity and social guidance.
• methodologies: Experimentally explore the performance of intelligentsia-generated content in terms of knowledge accuracy, logical consistency, and content innovativeness. Analyze the adaptability of their generated knowledge in different contexts and task types. Propose criteria and evaluation framework to measure the quality of generated knowledge.

Tests after knowledge generation

Are the prompts still worth learning?

• Prompt: Instructions or information entered by the user into the AI system to direct the AI to generate a specific output or perform a specific task.
• framework: Includes instructions, context, and expectations.
  • Instruction: Tell the AI exactly what tasks you want it to perform.
  • Context: Provide background information to the AI to help it understand and perform tasks more accurately.
  • Expectation: Explicitly or implicitly express your requirements and expectations for AI output.

cue structure

Task Requirements and Prompt Language Strategies

Examples of prompts

Decision-making needs

• practical skill::
  • Analyze demand::
    • typical example::
      "In order to reduce logistics costs, there are two existing solutions: 1 ⊕ Build your own regional warehouse (high initial investment, low long-term cost) 2 ⊕ Cooperate with a third party (pay-as-you-go, high flexibility) Please compare the total cost over 5 years and recommend the optimal solution based on the ROl calculation model."
  • Analyzing data::
    • typical example::
      "Analyze the last three years of new energy vehicle sales data (with CSV), explaining: 1 ⊕ the correlation between growth trend and policy; 2 ⊕ the forecast of market share in 2025, using an ARIMA model and explaining the basis of parameter selection."
  • Design Products::
    • typical example::
      "Design a smart home product that requires: 1 ⊕ solving the problem of security for elderly people living alone; 2 ⊕ combining sensor networks and Al alerts; 3 ⊕ providing sketch illustrations of prototypes for three different technology routes."

verification requirement

• practical skill::
  • Validation of the conclusions::
    • typical example::
      "The following is the conclusion of a paper: 'Neural network model A is superior to traditional method B'. Please verify: 1 ⊕ whether the experimental data support the conclusion; 2 ⊕ check the control group settings for bias; 3 ⊕ recalculate the p-value and determine significance."
  • code switching::
    • typical example::
      "Convert the following C code to Python, requiring: 1 ⊕ Keep the time complexity constant; 2 ⊕ Optimize array operations using numpy; 3 ⊕ Output the complete code with time test cases."

Types of prompts

• Instructional Cues: Directly informs the AI of the tasks it needs to perform.
• question-answer prompt: Ask the AI questions and expect answers.
• Role-playing prompts: AI is asked to play specific roles and simulate specific scenarios.
• Creative Cues: Guide AI for creative writing or content generation.
• Analytical Cues: Requires AI to analyze and reason with given information.
• multimodal cue: Combine multiple forms of input such as text and images.

Mastering Cue Design: Essential Skills for the AIGC Era

Core Skill System for Prompt Design

Advanced skills in prompt design

The secret to tuning up AI: cue multiplication strategies

• Strategy 1: Define tasks precisely and reduce ambiguity.
• Strategy 2: Decompose complex tasks appropriately to reduce AI cognitive load.
• Strategy 3: Introduce guiding questions to enhance the depth of generated content.
• Strategy 4: Control the length of prompts to ensure the accuracy of the generated content.
• Strategy 5: Flexible use of open-ended vs. closed-ended prompts.

How to achieve precise definitions:

• Clarification of core issues
• Concrete Generation Instructions
• Removing redundant information

Tips for breaking down tasks:

• Segment generation
• go deeper and deeper
• Setting up the logical structure

Design Essentials of Introductory Questions:

• Setting up multi-level questions
• Enable AI comparison or argumentation
• Leads to diversity of thought

Tips for controlling the length of prompts:

• Avoid nested complex instructions
• Keep it simple.
• Using Step-by-Step Tips

Open-ended prompts:

• Ask open-ended questions that allow AI to generate from multiple perspectives

Closed Cue:

• Ask specific questions or set clear limits that require precise answers from the AI.

Common Pitfalls and Responses: A Guide to Avoiding Pitfalls in Cue Design

The lack of iteration trap: expecting a one-step solution

• Trap Symptoms::
  • Overly Complicated Initial Prompts
  • Give up if you're not happy with the initial output
  • Lack of analysis and feedback on AI output
• response strategy::
  • Adoption of an incremental approach: Start with the basic prompts and gradually add details and requirements.
  • Proactively seeking feedback: Ask the AI to self-assess the output and provide suggestions for improvement.
  • Preparing for multiple rounds of dialog: Design follow-up questions for clarifying and improving the initial output.

The excessive or vague instruction trap: details overwhelm the point, intentions are not clearly expressed

• Trap Symptoms::
  • Cues are too long or too short
  • AI output deviates significantly from expectations
  • Frequent need to clarify or reinterpret requirements
• response strategy::
  • Balance Detail: Provide sufficient context, but avoid being overly restrictive.
  • Clarify key points: Highlight the most important 2-3 requirements.
  • Using a structured format: Adopt a clear structure for organizing requirements.
  • provide examples: If possible, give a short example of the desired output.

The Assumption Bias Trap: AI Only Says What You Want to Hear

• Trap Symptoms::
  • The prompt contains a clear position or tendency
  • Access to information always supports a particular point of view
  • Lack of opposing or differing viewpoints presented
• response strategy::
  • self-examination: Reflect on your own potential biases when designing prompts.
  • Use of neutral language: Avoid including bias or preconceptions in the prompt.
  • Requires multi-perspective analysis: AI is explicitly asked to provide a different point of view or argument.
  • critical thinking: Be vigilant about AI output and cross-validate important information.

Illusion Generation Trap: AI Confidence "Nonsense"

• Trap Symptoms::
  • AI provides data or facts that cannot be verified
  • Output contains terms or concepts that seem specialized but do not exist
  • Making overly specific predictions about future or uncertain events
• response strategy::
  • Clarifying uncertainty: Encourage AI to be explicit when they are unsure.
  • Fact-checking tips: Requires AI to distinguish between known facts and speculation.
  • Multi-source authentication: Require AI to validate information from multiple perspectives or sources.
  • Request for citations: AI is explicitly asked to provide sources of information for easy verification.

Competency development system in the era of man-machine symbiosis

Competency development system in the era of man-machine symbiosis

Four core competencies

AI Thinking

• algorithmic thinking: Understanding AI Decision Logic.
• Data Insight: Data-driven analytical capabilities.
• Boundary awareness: Grasp the boundaries of AI capabilities.
• cooperative consciousness: Modeling human-computer collaboration.

Core views: Mastering the AI mindset and building a cognitive framework for human-computer collaboration.

integrating force

• cross-domain translation: Translating domain knowledge.
• Creative reorganization: Reconfiguring working methods.
• Organization of resources: Optimizing human-machine collaboration.
• knowledge fusion: Integration of old and new knowledge.

Core views: Combine the advantages of man and machine to create the value of $1+1>2$.

channelling power

• Tip Engineering: Designing Efficient Instructions.
• Dialogue management: Controls the direction of the interaction.
• Breakdown of tasks: Optimize the structure of the problem.
• quality control: Control the quality of output.

Core views: Lead the AI interaction process and ensure that the output meets expectations.

judgment

• recognize the genuine from the fake: Assess content reliability.
• valuation: Judge the value of the application.
• risk prediction: Anticipate potential risks.
• contextual adaptation: Assess scenario applicability.

Core views: Keep thinking independently and be the gatekeeper of AI output.

AI Usage Levels and Breakthrough Paths

Breakthrough Path:

1. Creating a cue word system
2. Design Collaboration Process
3. Developing innovative approaches
4. Creating a personal touch

Unique workflow methodology innovation ⊕ Domain integration

AI Usage Levels and Breakthrough Paths