LLMs & Feedback
What is an LLM?
LLM (Large Language Model) — a super-intelligent "text predictor."
How it works:
- Training phase: Reads vast amounts of books, articles, and conversations from the internet
- Learning ability: Masters patterns and rules of language
- Generation ability: Predicts the most appropriate response based on input
Key Insight
LLMs don't truly "understand" — they become extremely good at "imitating" human language expression through massive learning.
What Happens When We Talk to an LLM?
Step 1: Tokenization
A Token is the smallest unit an LLM uses to understand text. English tokens can be words, word roots, or characters.
Step 2: Inference
Step 3: Text Generation
LLMs generate responses word by word, each step based on:
- Previously generated content
- The original Prompt requirements
- Language patterns learned during training
Prompt: The Art of Talking to AI
Prompt = the instruction or question you give to an LLM.
Types of prompts:
- Simple conversation:
"What's the weather like today?" - Task instruction:
"Summarize the main points of this article" - Role-playing:
"You are a poet, write a poem in classical style" - Chain-of-thought:
"Analyze this math problem step by step"
From LLM to Agent
Agent = an AI system built on LLMs capable of executing complex tasks.
| Feature | Simple LLM | AI Agent |
|---|---|---|
| Interaction | Single Q&A | Multi-step dialogue |
| Capabilities | Text generation | External tool use |
| Task complexity | Simple tasks | Complex task decomposition |
| Error handling | No self-correction | Can retry and optimize |
The Chinese Room Thought Experiment
Philosopher John Searle's famous thought experiment (1980):
The key question: People outside would think someone inside understands Chinese, but the person inside has no understanding of Chinese — they're just mechanically following rules.
Executing the correct program ≠ True understanding
The connection to LLMs: ChatGPT processes text like the person in the room processes Chinese characters — both follow preset rules, both produce convincing output, but neither may truly "understand."
Cybernetic Analysis of LLMs
Inference as a Cybernetic Loop
- Input Processing (Sensor): Tokenization, positional encoding, context awareness
- Attention Mechanism (Controller): Weight allocation, feature extraction, pattern recognition
- Output Generation (Actuator): Probability calculation, token selection, text generation
- Context Update (Feedback): Historical information integration, state maintenance
AI Agent as a Cybernetic Model
From a cybernetic perspective, both LLMs and AI Agents exhibit classic cybernetic characteristics: the attention mechanism serves as the controller, regulating the generation process through feedback.
Key Concepts Summary
- Token: AI's "dictionary" — breaking text into small pieces
- Prompt: The "art of conversation" with AI — how you ask matters
- LLM: AI's "brain" — the core of understanding and generation
- Agent: AI's "assistant" — an upgraded version capable of complex tasks