AI Starter #

This guide is designed for anyone starting their journey into AI. We will explore key questions to build a solid foundation, including:

The “Why”: Why is AI development often compared to alchemy, and where does its unpredictability come from?

The “How”: What is the process of training a large language model, and what does an “AI Alchemist” (ML Engineer) actually do?

The “When”: Why has AI seen explosive growth specifically in recent years?

The “Who”: What is a Prompt Engineer, and why is this role critical today?

By answering these questions, we bridge the gap between complex theory and practical understanding, lowering the technical barrier to empower everyone with AI knowledge.

AI Development and Training #

The core paradigm of artificial intelligence shifted from rule-based Symbolism to data-driven Connectionism, constructing the basic building block of neural networks (“linear transformation + activation function”), using gradient descent and backpropagation algorithms to find optimal parameters, and in the constant battle against challenges like overfitting and gradient problems, evolved specialized architectures like CNNs and RNNs, until ultimately the Transformer’s attention mechanism broke through the bottleneck of processing sequential data, laying the foundation for modern large models.

1. Core Idea: From Simple Rules to Complex Functions #

The basic building blocks of a Neural Network (NN).

• Philosophical Shift: From Symbolism (based on rules and logic) to Connectionism (based on learning through connections of simple units).
• Core Component: The Neuron. Its basic operation is Output = Activation Function(Linear Transform(Input)), i.e., f(x) = g(wx + b).
• Deepening the Network: A single neuron can’t solve complex problems. Connecting multiple neurons in layers forms an Input Layer -> Hidden Layer -> Output Layer structure. Multiple hidden layers can be used, each layer’s output becoming the next layer’s input.
• Universal Approximation: By increasing the number of hidden layers and neurons, the network can fit extremely complex non-linear functions.
• Learning Objective: The network’s “knowledge” is stored in all the connection weights (w) and biases (b). The goal of learning (training) is to find a suitable set of w and b so that the network function f(x) can produce the desired output y for a given input x.
• Forward Propagation: The process where data starts from the input layer and is calculated layer by layer until the output layer, akin to signal propagation in a biological neural network.

2. Learning Mechanism: How to Find the Optimal Parameters #

After defining the network structure, an algorithm is needed to automatically find the optimal parameters.

• Evaluating Performance: Loss Function
  • A criterion is needed to measure the difference between the network’s current predictions y_pred and the true values y_true. Mean Squared Error (MSE) is a commonly used loss function.
• Optimization Strategy: Gradient Descent
  • The goal is to minimize the loss function. By calculating the partial derivative (the gradient) of the loss function with respect to each parameter (w, b), we can determine the direction (increase or decrease) and magnitude (controlled by the learning rate) each parameter should be adjusted to reduce the loss.
• Efficient Calculation: Backpropagation
  • Neural networks are layered and nested, making calculating gradients for all parameters very complex. The Backpropagation algorithm cleverly uses the chain rule to calculate the gradient for each parameter layer by layer backwards from the output layer, doing so efficiently and elegantly.
• Training Loop: One complete training iteration consists of:
  1. Forward Propagation: Calculate the current prediction and loss.
  2. Backpropagation: Calculate the gradients for all parameters.
  3. Gradient Descent: Update the parameters using the gradients. This cycle repeats until the loss decreases to an acceptable level.

3. Core Challenges & Solutions (Overfitting & Training Difficulties) #

Greater network power brought new problems.

• Overfitting: The model performs perfectly on training data but poorly on unseen data (test data). This means the model has poor generalization ability; it has merely “memorized” the training data.
• Strategies to Combat Overfitting:
  • Reduce Model Complexity: Use a smaller, simpler network.
  • Data Augmentation: Artificially expand the dataset by applying operations like rotation, cropping, and adding noise to training data (e.g., images).
  • Regularization: Add a penalty term (e.g., L1, L2) to the loss function to prevent parameters from becoming too large or complex, encouraging the model to learn more general patterns.
  • Dropout: Randomly “drop” (temporarily disable) a portion of neurons during training, forcing the network not to rely too heavily on any single neuron or combination, thereby enhancing robustness.
• Other Training Challenges:
  • Vanishing/Exploding Gradients: In deep networks, gradients can become extremely small (preventing updates) or large (causing instability) during backpropagation. Solutions include: improved activation functions, Residual Connections (ResNet), gradient clipping, etc.
  • Convergence Speed & Computation: Use better optimizers (e.g., Adam), proper weight initialization, Batch Normalization, and Mini-batch training to speed up training and stabilize the process.

4. Specialized Architecture I: Convolutional Neural Networks (CNN) #

A specialized network architecture designed for processing grid-like data such as images.

• Core Idea: Use “convolution” and “pooling” operations to process images efficiently and automatically learn hierarchical spatial features.
• Convolutional Layer: Uses a kernel (a small weight matrix) that slides over the image performing matrix operations to extract local features (e.g., edges, corners). Key Advantages: Parameter sharing (one kernel detects the same feature across the entire image) and sparse connectivity (each output connects only to a small local region of the input).
• Pooling Layer (Subsampling): Reduces the dimensionality of feature maps, compressing data and parameters while retaining the most salient features (e.g., Max Pooling).
• Typical Structure: Input Layer -> [Convolutional Layer -> Activation Function -> Pooling Layer] * N -> Fully Connected Layer -> Output Layer.
• Fatal Flaw: Primarily used for spatial data but is weak at modeling sequential data and long-range dependencies.

5. Specialized Architecture II: Recurrent Neural Networks (RNN) & Their Evolution #

Designed for processing sequential data (e.g., text, speech, time series).

• Word Embedding: First, words need to be converted into meaningful numerical vectors (word vectors). Word embedding methods (e.g., Word2Vec) learn vectors that capture semantic relationships, far superior to simple One-hot encoding. The matrix of all word vectors is called the embedding matrix, and the space these vectors inhabit is called the latent space.
• RNN Core: Has a “memory” function, passing the hidden state from previous steps to the next, enabling it to handle contextual information.
• RNN Shortcomings:
  • Difficulty capturing long-range dependencies: As the time steps increase, the influence of early information tends to vanish (vanishing gradient problem).
  • Inability to parallelize computation: Calculations must be done sequentially step-by-step.
• Improvements: LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit) introduced “gating” mechanisms (forget gate, input gate, output gate) to selectively remember and forget information, effectively mitigating the long-range dependency issue.

6. Revolutionary Architecture: Transformer #

It fundamentally changed how sequence data is processed and became the foundation for modern Large Language Models (LLMs).

• Core Innovation: Self-Attention Mechanism
  • Allows any element in a sequence to directly interact with and assign weights to all other elements, efficiently capturing global dependencies regardless of distance.
  • Achieved by calculating Query (Q), Key (K), and Value (V) vectors.
  • Multi-Head Attention: Uses multiple sets of Q/K/V matrices, allowing the model to attend to information from different representation subspaces simultaneously.
• Advantages:
  • Solves long-range dependency problem completely: Global relationships are captured directly in a single step.
  • Highly parallelizable: No longer requires sequential computation, greatly enhancing training speed.
• Impact: Transformer-based models (e.g., GPT, BERT) achieved revolutionary success in Natural Language Processing (NLP), excelling not only in understanding and judgment but also demonstrating powerful generation and decision-making capabilities.

Large Model Comparison #

🔑 Fine-Tuning a Large Language Model (LLM) #

Core idea: You don’t train a model from scratch. Instead, you take a pre-trained model (like a chef who already knows basic cooking) and fine-tune it on specific tasks (like mastering one dish).

🧩 Six Steps of Fine-Tuning #

1. Choose Fine-Tuning Method ⚙️

   • LoRA / QLoRA → update only part of the parameters (low hardware cost)
   • Full Parameter Tuning → update all parameters (needs strong GPUs / cloud servers)

2. Select a Pre-Trained Model 📚

   • Source: Hugging Face (the “GitHub of AI”)
   • Choose a reliable model (preferably from big tech companies)

3. Define Training Goal 🎯

   • What should the model learn?
   • Examples: logical reasoning, imitating writing style, domain-specific Q&A

4. Prepare Training Samples 📝

   • Create datasets matching your goal
   • Often generated using another LLM
   • Use clear prompts + examples, then save as dataset file

5. Start Training 🖥️

   • Environment: Local GPU (WSL2/Linux) or Cloud GPU server
   • Upload dataset + run training script
   • Output model saved in output/ folder

6. Validate the Model ✅

   • Load fine-tuned model
   • Ask test questions to check if it meets the goal

🧑‍🍳 Analogy: Chef Training #

• Pre-trained Model = Fresh graduate chef (knows many ingredients & techniques, but average at cooking)
• Fine-tuning = Chef focuses on one dish (e.g., Mapo Tofu) until mastery
• Result = Expert chef in that dish → Fine-tuned model for a specific task

MCP Protocol #

1. What is the MCP Protocol? #

MCP (Model Context Protocol) is an open protocol introduced by Anthropic in 2024. It standardizes how AI models interact with external tools, data sources, and systems.

In simple terms:

• Previously, each platform had its own way for LLMs to call tools (databases, APIs, file systems) → no interoperability.
• MCP defines a unified communication standard, so different models, applications, and tools can talk in the same way.
• Just like HTTP is for the web, MCP is for AI tool invocation and contextual interaction.

2. What is MCP Used For? #

• Unified interface: Consistent way for LLMs to call APIs, databases, knowledge bases, etc.
• Security isolation: Defines access permissions to prevent LLMs from having unrestricted system access.
• Cross-model interoperability: Works across Claude, GPT, Gemini, or any LLM that supports MCP.
• Developer simplicity: No need to write different adapters for each model or platform.

3. Use Cases #

1. Enterprise Knowledge Retrieval

   • MCP connects LLMs to enterprise document stores, CRM, ERP systems.
   • Query: “Show me the latest financial report” → MCP fetches securely from the database.

2. Software Development Assistant

   • IDEs (like VSCode) expose tools via MCP: file read/write, run tests.
   • LLM uses MCP to interact with the development environment.

3. Multimodal Interaction

   • MCP lets LLMs call image-processing or speech-synthesis tools.
   • An AI agent could process images and generate spoken responses via MCP.

4. Personal Assistant

   • MCP gives AI access to calendar, email, to-do APIs.
   • Query: “Schedule a meeting with John” → LLM uses MCP to update the calendar.

4. MCP and AI Agents #

MCP = the “action protocol” for AI agents.

• AI agents need: Perception (inputs) → Reasoning (thinking) → Action (tool use).
• MCP solves the “action” layer by standardizing tool invocation.
• Previously, each agent framework had its own plugin or API wrapper system. MCP unifies this layer, enabling interoperability.

Analogy:

• AI agent = Web browser
• Tools / Data sources = Websites
• MCP = HTTP protocol

Prompt Engineer #

A Prompt Engineer is someone who “uses language instructions to control AI, making it smarter and more reliable at completing tasks.”

1. Core Responsibilities #

• Design Prompts: Craft input instructions or context for AI models (like GPT, Claude, Gemini, etc.) so the model understands the task.
• Optimize Outputs: Refine prompts by adding constraints, examples, or guidance to improve the quality and accuracy of AI-generated content.
• Test and Iterate: Evaluate AI performance with different prompts, analyze errors or biases, and optimize prompting strategies.
• Cross-Domain Application: Apply AI capabilities to practical business scenarios such as content creation, data analysis, customer support, and code generation.

2. Required Skills #

• Understanding AI Models: Knowledge of how large language models work, including context windows and generation logic.
• Business Insight: Ability to translate business needs into prompts that AI can understand.
• Language Precision: Skill in crafting clear, precise, and guiding instructions.
• Data Analysis: Analyze AI outputs and iteratively improve prompt strategies.

3. Practical Examples #

• Text Generation: Guiding the model to produce marketing copy consistent with brand style.
• Coding Assistance: Designing prompts to generate correct, runnable code snippets.
• Q&A Systems: Optimizing prompts so AI answers FAQ questions accurately and coherently.

AI-Agent #

An AI Agent is like an “AI employee” that can think for itself and get things done.

You don’t need to micro-manage it step-by-step. You just give it an end goal, and it can plan the steps, use tools, and complete the task on its own.

An Analogy to Help You Understand #

Imagine you want to plan a family trip.

• A standard AI assistant (like ChatGPT): It’s like a smart consultant. You ask it: “What’s a good travel guide for Sanya?” It will give you a detailed text list of attractions, food. But booking flights, hotels, and tickets—you have to do all that yourself on various websites.

• An AI Agent: It’s like a full-capability personal travel concierge. You just tell it: “Help me plan and book a 5-day, 4-night family trip to Sanya for next week, with a budget of 5,000 RMB per person.”

  • This “concierge” will then spring into action on its own:
    1. Planning: It first thinks (uses its brain/model), searches for information, and drafts an itinerary.
    2. Using Tools: It operates the computer itself (calls tool APIs), for example:
       • Opens airline websites to check flights and book them.
       • Logs into hotel booking platforms to pick a suitable room and reserve it.
       • Even goes to the attraction’s official website to buy your tickets.
    3. Confirmation: Finally, it compiles the booked flight, hotel, and ticket information into a spreadsheet and sends it to you, saying: “Boss, all done. Here is your itinerary for review.”

This “AI concierge” that can autonomously understand goals, plan tasks, and use various tools (software, websites, calculators, etc.) to execute them is an AI Agent.

The Core Features of an AI Agent #

1. Has a “Brain”: It uses a Large Language Model (like GPT-4) as its core to understand your intent, think logically, and make plans.
2. Has “Hands and Feet”: It can call upon tools (Tools). This is the key difference from a standard chat AI. It can use API interfaces to operate software, search the web, run code, etc., thereby actually “doing things,” not just “answering.”
3. Has a “Sense of Purpose”: It works towards a final goal, autonomously breaking it down into steps and executing them one by one until the task is complete or cannot proceed further.

Real-Life Examples #

• Research Assistant: You tell the Agent, “Help me write a paper about black holes.” It won’t just copy and paste; it will first search online for the latest literature, download and read it, summarize key points, draft the paper itself, and finally polish and edit it for you.
• Shopping Assistant: You tell it, “Buy me a birthday gift suitable for my mom under 300 RMB.” It will automatically go to e-commerce sites to search, compare prices, read reviews, and finally place the order.
• Office Assistant: You tell it, “Analyze last week’s sales data and make a PPT report.” It can automatically pull data from the database, analyze it using code, and then use presentation software to generate a slideshow with charts and conclusions.

Case-TradingAgents #

TradingAgents is an open-source multi-agent framework for financial trading with large language models (LLMs), developed by TauricResearch. It simulates how a real trading firm organizes teamwork: different agents analyze markets, evaluate risks, and make joint trading decisions. (github.com)

🧱 System Architecture & Roles #

The system defines several specialized agents, each with its own role:

✅ Strengths

1. Division of labor — mirrors real trading firms, combining fundamentals, sentiment, technical, and macro analysis.
2. Modularity & flexibility — easily swap models, adjust settings, or configure debate mechanics.
3. High research & educational value — demonstrates how AI agents collaborate in trading contexts.
4. Open-source (Apache-2.0) — well-documented with examples and default configs.

⚠️ Limitations

1. Resource intensive — multiple agents calling LLMs + live data + debates → high compute/API costs.
2. Latency — multi-step analysis and debate may be too slow for high-frequency trading.
3. Data/model dependency — output quality depends on real-time data reliability and LLM accuracy.
4. Overfitting risk — backtest success does not guarantee robustness in live markets.
5. Risk management complexity — extreme events may fall outside the assumptions of the system.

Application Tools: The AI Tool Ecosystem #

AI technology is driving the emergence of diverse tools and applications across industries. By application scenario, they can be classified as follows:

• Content Creation:

  • Text: ChatGPT, Claude, Gemini and other conversational AI writing assistants.
  • Image Generation: Midjourney, DALL·E, Stable Diffusion.
  • Video Generation & Editing: Runway, Pika, and Sora (by OpenAI).
  • Audio Generation: ElevenLabs and other AI-based voice synthesis tools.

• Work Productivity:

  • Programming: GitHub Copilot (powered by OpenAI Codex) for code autocompletion, windsurf, Qodo.
  • Document Processing & Knowledge Management: Notion AI, Microsoft Copilot (integrated into Office).
  • Presentations: Gamma, Beautiful.ai for automated slide and presentation creation.

• Professional Applications:

  • Research Assistants: Semantic Scholar, Elicit for academic literature search and summarization.
  • Legal Field: Casetext, Ross Intelligence and other AI tools for legal search and analysis.
  • Healthcare: Emerging AI-assisted diagnostic systems that analyze medical images or records to support diagnosis.

• Development Tools:

  • API and platform services: OpenAI/Anthropic APIs.
  • Model deployment & fine-tuning libraries: Hugging Face.
  • Cloud services: Amazon Bedrock, Azure AI, etc., enabling developers to easily integrate LLM capabilities.

Future Potential: Directions of AI Development #

• Technological Breakthroughs: Advancing toward more general intelligence involves:

  • Multimodal integration — enabling models to efficiently understand images, text, speech, and more.
  • Enhanced reasoning ability — supporting complex causal reasoning and long-form logical inference.
  • Efficiency optimization — reducing deployment costs through model compression, pruning, quantization, and mixture-of-experts (MoE); lowering latency with edge computing; and enabling customized fine-tuning or personalized models for individual users.

  A recent example is Anthropic’s Claude 3.5 Sonnet, which balances performance (intelligence) with reasoning cost. Compared with its predecessor, it achieves significant benchmark improvements while keeping cost per million tokens at a relatively affordable level. These advances stem from architectural optimizations and training improvements, showing how new-generation LLMs continue to optimize the trade-off between intelligence, speed, and cost.

• Application Expansion: AI will penetrate deeper into various domains:

  • Autonomous Agents: Smarter AI assistants capable of planning and executing multi-step tasks, orchestrating workflows (e.g., handling customer queries, automating document analysis and drafting).
  • Scientific Research: Accelerating drug discovery, materials design, and solving advanced mathematical problems.
  • Education: Delivering personalized tutoring, auto-generating and updating knowledge graphs, and giving teachers and students more precise learning resources.
  • Creative Industries: AI will support game storyline design, film post-production, music and art creation, providing both inspiration and productivity for creators.

• Social Impact:

  • Job Market: Shifts toward high-skill and AI-collaborative roles, with repetitive jobs increasingly automated.
  • Education Systems: Adjustments to meet new skill demands, fostering innovation and digital literacy.
  • Ethics & Regulation: Policies are required to ensure privacy, fairness, and transparency. Examples include the EU AI Act and China’s draft Regulation on Generative AI.

Challenges and Limitations #

• Technical Challenges:

  • Hallucination: Models generating plausible but incorrect outputs.
  • Black-box complexity: Poor interpretability, making decision processes hard to understand.
  • Data dependency: Reliance on large-scale, high-quality datasets; bias or scarcity reduces reliability and safety.

• Social Challenges:

  • Privacy & Security Risks: Risks of personal data being used in training or inference.
  • Algorithmic Bias: Potential to amplify existing inequalities.
  • Market Concentration: Big tech dominance in foundational models may hinder innovation and competition.

• Ethical Considerations:

  • Safety & Reliability: Ensuring AI works dependably in critical domains.
  • Human–AI Roles: Defining responsibility and authority of AI decisions, especially in medicine or law.
  • Responsible Use: Preventing misuse (e.g., generating disinformation) and guiding AI development toward beneficial directions.

AI Terminology Guide #