Chapter 3: Overview of AI Language Models

🎯 Learning Objectives

After completing this chapter, you will be able to:

• Explain how large language models (LLMs) work at a high level
• Compare major AI models and their unique strengths
• Identify which model is best suited for different tasks
• Understand key limitations of current AI models

📚 Introduction

To become an effective prompt engineer, you need to understand the tools you're working with. In this chapter, we'll explore how AI language models work and compare the major players in the field.

How LLMs Work (Simplified)

Large Language Models are AI systems trained on vast amounts of text to understand and generate human-like language.

The Basic Concept

┌─────────────────────────────────────────────────────────────┐
│                    HOW LLMs WORK                            │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│   INPUT (Your Prompt)                                       │
│         │                                                   │
│         ▼                                                   │
│   ┌─────────────┐                                          │
│   │ TOKENIZATION │ → Break text into tokens                │
│   └──────┬──────┘                                          │
│          │                                                  │
│          ▼                                                  │
│   ┌─────────────┐                                          │
│   │  EMBEDDING  │ → Convert tokens to numbers              │
│   └──────┬──────┘                                          │
│          │                                                  │
│          ▼                                                  │
│   ┌─────────────┐                                          │
│   │ TRANSFORMER │ → Process through neural network         │
│   └──────┬──────┘                                          │
│          │                                                  │
│          ▼                                                  │
│   ┌─────────────┐                                          │
│   │ PREDICTION  │ → Predict most likely next token         │
│   └──────┬──────┘                                          │
│          │                                                  │
│          ▼                                                  │
│   OUTPUT (AI Response)                                      │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Key Concepts

Concept	Explanation	Why It Matters
Token	Basic unit of text (roughly 4 characters)	Affects pricing and context limits
Context Window	Maximum input + output length	Determines how much info AI can consider
Parameters	Model's learned weights (billions)	More parameters often = more capable
Temperature	Randomness in output	Higher = more creative, Lower = more consistent
Training Data	Text the model learned from	Affects knowledge and biases

Understanding Tokens

"Hello, how are you?" 

Tokenized: ["Hello", ",", " how", " are", " you", "?"]
           = 6 tokens

Token count varies by model:
• GPT: ~4 characters per token
• Claude: Similar to GPT
• Some models use different tokenization

Why This Matters

Understanding tokens helps you:

• Estimate costs
• Stay within context limits
• Write more efficient prompts

Major AI Models Compared

Let's explore the major players in the LLM space:

OpenAI Models (GPT Family)

GPT-4 / GPT-4 Turbo

Feature	Details
Company	OpenAI
Context Window	8K-128K tokens
Strengths	Reasoning, coding, general knowledge
Best For	Complex tasks, analysis, creative writing
Access	ChatGPT Plus, API

Example Prompt Style:

You are an expert Python developer. Review this code 
and provide specific improvements with explanations:

[code here]

GPT-3.5 Turbo

Feature	Details
Context Window	4K-16K tokens
Strengths	Speed, cost-efficiency
Best For	Simple tasks, high-volume applications
Cost	~20x cheaper than GPT-4

Anthropic Models (Claude Family)

Claude 3 (Opus, Sonnet, Haiku)

Variant	Strength	Use Case
Opus	Most capable	Complex analysis, research
Sonnet	Balanced	General purpose, coding
Haiku	Fastest, cheapest	Quick tasks, high volume

Claude's Unique Features:

• Constitutional AI (built-in safety)
• Excellent at following complex instructions
• Strong performance with XML-structured prompts
• Very long context window (200K tokens)

Example Prompt Style:

<context>
You are a helpful assistant analyzing customer feedback.
</context>

<task>
Categorize the following review and extract key themes:
</task>

<review>
[customer review text]
</review>

<output_format>
- Category: [positive/negative/neutral]
- Key themes: [bulleted list]
- Summary: [one sentence]
</output_format>

Meta Models (Llama Family)

Llama 2 & Llama 3

Feature	Details
Company	Meta
License	Open source (mostly)
Sizes	7B, 13B, 70B parameters
Strengths	Customizable, self-hostable, free
Limitations	Less capable than GPT-4/Claude 3

Best For:

• Organizations needing data privacy
• Custom fine-tuning
• Cost-sensitive applications
• Research and experimentation

Other Notable Models

Model	Company	Notable Feature
Gemini	Google	Multimodal, integrated with Google services
Mistral	Mistral AI	High quality open-source alternative
Command	Cohere	Enterprise-focused, RAG optimized
Grok	xAI	Real-time information access

Model Comparison Matrix

┌──────────────────────────────────────────────────────────────┐
│                    MODEL COMPARISON                          │
├─────────────┬───────────┬──────────┬─────────┬──────────────┤
│             │ GPT-4     │ Claude 3 │ Llama 3 │ Gemini Pro   │
├─────────────┼───────────┼──────────┼─────────┼──────────────┤
│ Reasoning   │ ★★★★★    │ ★★★★★   │ ★★★★   │ ★★★★★       │
│ Coding      │ ★★★★★    │ ★★★★★   │ ★★★★   │ ★★★★        │
│ Creative    │ ★★★★★    │ ★★★★    │ ★★★    │ ★★★★        │
│ Context     │ ★★★★     │ ★★★★★   │ ★★★    │ ★★★★        │
│ Speed       │ ★★★      │ ★★★★    │ ★★★★★  │ ★★★★        │
│ Cost        │ ★★       │ ★★★     │ ★★★★★  │ ★★★         │
│ Privacy     │ ★★★      │ ★★★     │ ★★★★★  │ ★★★         │
└─────────────┴───────────┴──────────┴─────────┴──────────────┘

Understanding Limitations

All AI models have limitations that affect how you should prompt them:

1. Knowledge Cutoff

Models are trained on data up to a certain date:

Model	Approximate Cutoff
GPT-4	April 2023 (varies)
Claude 3	Early 2024
Llama 3	December 2023

Implication

Always verify time-sensitive information. The model doesn't know recent events.

2. Hallucinations

AI models can generate plausible-sounding but incorrect information:

❌ Prompt: "What is the population of Mars?"

❌ Bad Output: "The population of Mars is approximately 
150,000 as of the latest census data."

This is completely fabricated but sounds confident!

Prevention Strategies:

• Ask for sources
• Verify critical facts
• Use prompts that encourage uncertainty acknowledgment

3. Context Window Limits

Each model has a maximum amount of text it can process:

Model	Context Window	Practical Meaning
GPT-3.5	4K-16K tokens	~6-24 pages
GPT-4	8K-128K tokens	~12-192 pages
Claude 3	200K tokens	~300 pages

4. Reasoning Limitations

While impressive, AI reasoning has gaps:

• Struggles with complex multi-step math
• Can be fooled by tricky logic puzzles
• May not understand implicit context
• Can be inconsistent across similar questions

5. Bias and Safety

Models may reflect biases in their training data:

• Cultural biases
• Stereotypes
• Potential for harmful content

Choosing the Right Model

Decision Framework

┌─────────────────────────────────────────────────────────────┐
│              WHICH MODEL SHOULD I USE?                       │
└─────────────────────────────────────────────────────────────┘

Is the task complex (analysis, coding, reasoning)?
├─ YES → Is cost a major concern?
│        ├─ YES → Claude 3 Sonnet or GPT-4 Turbo
│        └─ NO → GPT-4 or Claude 3 Opus
│
└─ NO → Is speed critical?
         ├─ YES → GPT-3.5 Turbo or Claude 3 Haiku
         └─ NO → Do you need privacy/local hosting?
                  ├─ YES → Llama 3 or Mistral
                  └─ NO → GPT-3.5 or Claude 3 Sonnet

Task-Based Recommendations

Task	Recommended Model	Why
Complex coding	GPT-4, Claude 3 Opus	Best reasoning
Quick answers	GPT-3.5, Haiku	Speed + cost
Long documents	Claude 3	200K context
Customer chatbots	GPT-3.5, Sonnet	Balance of quality/cost
Private data	Llama 3	Self-hosted option
Creative writing	GPT-4, Claude 3	Nuanced outputs

Practical Exercise

Exercise: Model Testing

Try the same prompt across different models and compare outputs:

Test Prompt:

Explain quantum computing to a 10-year-old in 3 sentences, 
then explain it to a physics PhD student in 3 sentences.

Compare:

1. Quality of simplification
2. Technical accuracy
3. Tone adjustment
4. Overall helpfulness

Key Takeaways

Summary

1. LLMs predict text token by token based on training data
2. Major models include GPT-4, Claude 3, Llama 3, and Gemini
3. Each model has strengths: GPT-4 (reasoning), Claude (context), Llama (privacy)
4. Limitations are real: hallucinations, knowledge cutoffs, reasoning gaps
5. Choose based on task: Consider complexity, cost, speed, and privacy needs

What's Next?

Congratulations! You've completed Module 1. Now it's time to test your knowledge with the Module 1 Quiz.

---

← Previous: Why Prompt Engineering Matters | Take the Quiz →