What is DeepSeek?

• AI + Domestic + Free + Open Source + Powerful
  DeepSeek is a Chinese tech company specializing in General Artificial Intelligence (AGI), focusing on the development and application of large models.
  DeepSeek-R1 is its open-source reasoning model, excelling in handling complex tasks and available for free commercial use.

DeepSeek from entry to mastery (Tsinghua University) PDF Downlod

What Can DeepSeek Do?

Text Generation

• Structured Generation: Tables, lists (e.g., schedules, recipes)

• Document Writing: Code comments, documentation

• Creative Writing: Articles, stories, poetry, marketing copy, social media content, scripts, etc.

• Summarization and Rewriting: Long text summaries (papers, reports), text simplification, multilingual translation and localization

Natural Language Understanding and Analysis

• Knowledge Reasoning: Logical problem-solving (math, common sense reasoning), causal analysis (event correlation)

• Semantic Analysis: Sentiment analysis (reviews, feedback), intent recognition (customer service, user queries), entity extraction (names, locations, events)

• Text Classification: Topic labeling (e.g., news categorization), spam content detection

Programming and Code-Related Tasks

• Code Generation and Completion: Code snippets (Python, JavaScript), auto-completion with comments

• Code Debugging: Error analysis and repair suggestions, performance optimization tips

• Technical Documentation: API documentation, codebase explanation and example generation

Conventional Drawing

• (Not explicitly mentioned but implied through general capabilities)

How to Use DeepSeek?

Access:
https://chat.deepseek.com

From Beginner to Master:
When everyone can use AI, how can you use it better and more effectively?

Reasoning Models

• Example: DeepSeek-R1, GPT-3 excel in logical reasoning, mathematical reasoning, and real-time problem-solving.
  Reasoning models are models that enhance reasoning, logical analysis, and decision-making capabilities on top of traditional large language models. They often incorporate additional technologies such as reinforcement learning, neuro-symbolic reasoning, and meta-learning to strengthen their reasoning and problem-solving abilities.
  Non-reasoning models are suitable for most tasks. General models typically focus on language generation, context understanding, and natural language processing, without emphasizing deep reasoning capabilities. These models usually grasp language patterns through extensive text data training and can generate appropriate content, but they lack the complex reasoning and decision-making abilities of reasoning models.

Dimension Comparison

Example Models

• GPT-3, GPT-4 (OpenAI), BERT (Google): Mainly used for language generation, language understanding, text classification, translation, etc.

Fast Thinking vs. Slow Thinking

• Fast Reaction Models (e.g., ChatGPT-4): Quick response, low computational cost, based on probability prediction through extensive data training

• Slow Thinking Models (e.g., OpenAI-1): Slow response, high computational cost, based on chain-of-thought reasoning to solve problems step-by-step

• Decision-Making: Fast reaction models rely on pre-set algorithms and rules, while slow thinking models can make autonomous decisions based on real-time analysis

• Creativity: Fast reaction models are limited to pattern recognition and optimization, while slow thinking models can generate new ideas and solutions

• Human Interaction: Fast reaction models follow pre-set scripts and struggle with human emotions and intentions, while slow thinking models can interact more naturally and understand complex emotions and intentions

• Problem-Solving: Fast reaction models excel in structured and well-defined problems, while slow thinking models can handle multi-dimensional and unstructured problems, providing creative solutions

• Ethical Issues: Fast reaction models as controlled tools have minimal ethical concerns, while slow thinking models raise discussions on autonomy and control

CoT Chain-of-Thought

The emergence of CoT chain-of-thought divides large models into two categories: "probability prediction (fast reaction)" models and "chain-of-thought (slow thinking)" models. The former is suitable for quick feedback and immediate tasks, while the latter solves complex problems through reasoning. Understanding their differences helps in choosing the appropriate model for the task to achieve the best results.

Prompt Strategy Differences

• Reasoning Models: Prompts should be concise, focusing directly on the task goal and requirements (as reasoning logic is internalized). Avoid step-by-step guidance, as it may limit the model's capabilities.

• General Models: Prompts need to explicitly guide reasoning steps (e.g., through CoT prompts), otherwise, the model may skip key logic. Rely on prompt engineering to compensate for capability shortcomings.

Key Principles

1. Model Selection: Choose based on task type, not model popularity (e.g., reasoning models for math tasks, general models for creative tasks).

2. Prompt Design:

   • Reasoning Models: Use concise instructions, focus on the goal, and trust the model's internalized reasoning capabilities. ("Just say what you want.")

   • General Models: Use structured and compensatory guidance. ("Fill in what's missing.")

    

3. Avoid Pitfalls:

   • Do not use heuristic prompts (e.g., role-playing) with reasoning models, as it may interfere with their logical mainline.

   • Do not over-trust general models (e.g., directly asking complex reasoning questions); instead, validate results step-by-step.

    

From "Giving Instructions" to "Expressing Needs"

Strategy Types

Task Demand and Prompt Strategy

How to Express Needs to AI

Prompt Examples

Decision-Making Demand:
"Two options are available to reduce logistics costs:
① Build a regional warehouse (high initial investment, low long-term costs)
② Partner with a third party (pay-as-you-go, high flexibility)
Please use the ROI calculation model to compare the total costs over 5 years and recommend the optimal solution."

Verification Demand:
"Here is a conclusion from a paper: 'Neural network model A is superior to traditional method B.'
Please verify:
① Whether the experimental data supports this conclusion;
② Check if there is any bias in the control group setup;
③ Recalculate the p-value and determine significance."

Analytical Demand:
"Analyze the sales data of new energy vehicles over the past three years (attached CSV), and explain:
① The correlation between growth trends and policy;
② Predict the market share in 2025 using the ARIMA model and explain the basis for parameter selection."

Execution Demand:
"Convert the following C code to Python, with the following requirements:
① Maintain the same time complexity;
② Use numpy to optimize array operations;
③ Output the complete code with time test cases."

Creative Demand:
"Design a smart home product to address the safety issues of elderly people living alone, combining sensor networks and AI early warning. Provide three different technical route prototype sketches with explanations."

Verification Demand:
"Below is a conclusion from a paper: 'Neural network model A is superior to traditional method B.'
Please verify:
① Whether the experimental data supports this conclusion;
② Check if there is any bias in the control group setup;
③ Recalculate the p-value and determine significance."

Do We Still Need to Learn Prompts?

Prompts are the instructions or information that users input into an AI system to guide it to generate specific outputs or perform specific tasks. Simply put, prompts are the language we use to "converse" with AI. They can be a simple question, a detailed instruction, or a complex task description.
A prompt consists of three basic elements:

• Instruction (Instruction): The core of the prompt, explicitly telling the AI what task to perform.

• Context (Context): Providing background information to help the AI better understand and execute the task.

• Expectation (Expectation): Clearly or implicitly expressing the requirements and expectations for the AI's output.

Types of Prompts

• Instructional Prompts: Directly tell the AI what task to perform.

• Question-Answer Prompts: Pose questions to the AI, expecting corresponding answers.

• Role-Playing Prompts: Require the AI to assume a specific role and simulate a particular scenario.

• Creative Prompts: Guide the AI to perform creative writing or content generation.

• Analytical Prompts: Require the AI to analyze and reason about given information.

• Multimodal Prompts: Combine text, images, and other forms of input.

The Essence of Prompts

Article from:
Team: Yu Menglong, Postdoctoral Fellow

Tsinghua University School of Journalism and Communication

New Media Research Center, Metaverse Culture Lab