Under the Hood
Peek inside how AI actually works (optional technical module)
What You'll Learn
- ✓The high-level flow: Data → Training → Model → Application
- ✓What tokens are and why they matter for cost and context
- ✓The concept of embeddings (turning words into numbers)
- ✓What parameters mean and why GPT-4 is bigger than GPT-3
Key Ideas
AI follows a flow: (1) Collect massive amounts of text/image data, (2) Train a model by learning patterns, (3) Fine-tune for specific tasks, (4) Deploy as an application (like ChatGPT). Understanding this helps you grasp AI's strengths and limitations.
Examples:
- • Data: Books, websites, conversations
- • Training: Learning word relationships and patterns
- • Fine-tuning: Making it helpful, harmless, and honest
- • Application: The chatbot you interact with
AI doesn't see words; it sees tokens (word chunks). 'Hello' might be one token, 'ChatGPT' might be two. This matters for costs (priced per token) and limits (max tokens per conversation).
Examples:
- • Roughly 1 token ≈ 4 characters or 0.75 words
- • A typical conversation uses thousands of tokens
- • Longer prompts = higher costs and slower responses
Parameters are like the 'brain cells' of an AI model. More parameters generally mean better capability but also more compute power needed. GPT-4 has hundreds of billions of parameters.
Suggested Resources
Coursera
Try This Now
Put your learning into practice with these hands-on exercises. Copy the prompts and try them in your favorite AI tool.
Visit OpenAI's GPT-4 technical report or Anthropic's Claude documentation. Read the introduction and summarize 3 key capabilities or limitations in plain English.
Use OpenAI's tokenizer tool to see how different texts are broken into tokens. Compare: 'Hello' vs 'Supercalifragilisticexpialidocious' vs 'ChatGPT'.
Find a model card for GPT-4, Claude, or Gemini. What does it say about the model's training, intended use, and limitations?
Reflection Questions
- 1.How does understanding the technical side change how you use AI tools?
- 2.What technical concepts do you want to learn more about?
- 3.How might this knowledge help you explain AI to others?