What AI actually is (and isn't)
Before we build anything, let's fix the mental model. Most confusion about AI comes from thinking it's a search engine, a person, or a database. It's none of those. It's something new and it helps to understand what.
What is a language model?
A large language model (LLM) is a very large math function that, given a chunk of text, predicts the most likely next word. That's the whole thing. It doesn't "think," "know," or "want." It predicts. Everything else — chat, agents, coding assistants — is built on top of that one behavior.
The models you'll use in this course — GPT-5, Claude Sonnet, Gemini — are trained on huge amounts of text and then fine-tuned with human feedback. That fine-tuning is what makes them feel conversational.
An LLM is like a brilliant intern who has read the internet, can't Google anything, has no memory between conversations, and will confidently make things up if you don't give them the material. Everything we build on top solves for those four things.
Tokens and context
The model doesn't see words — it sees tokens. A token is roughly ¾ of a word. "Understanding" is two tokens. A comma is one token. This matters because:
- You pay per token in and out.
- Every model has a context window — how many tokens it can hold at once. Modern models range from 128K to 2M tokens. That's your working memory.
- When you exceed the context window, the model forgets the beginning. This is why long chats degrade.
Rough conversions
- 1 page of prose ≈ 500 tokens
- A 20-page RFP ≈ 10,000 tokens
- A 200K-token context ≈ a 400-page book
Temperature and randomness
LLMs sample from probability distributions. The temperature setting controls how adventurous they are.
- Temperature 0 — always pick the most likely next word. Deterministic, predictable, boring. Use for extraction, classification, structured JSON.
- Temperature 0.7 — the default for chat. Balanced.
- Temperature 1.0+ — creative, unpredictable. Use for brainstorming, copy variations.
What they genuinely can't do (yet)
- Access anything they weren't given. Real-time data, your email, your files — unless you plug them in via tools (Module 3).
- Remember previous conversations. Every session starts blank unless the app you're using persists context.
- Do reliable math. They're guessing digits. Use a calculator tool.
- Be sure. They will fabricate confidently. Grounding (Module 6) is the fix.
The model zoo (as of 2026)
| Family | Best for | Notes |
|---|---|---|
| OpenAI GPT-5, GPT-5-mini | General-purpose, agents, coding | Wide tool ecosystem, strong function calling. |
| Anthropic Claude Opus, Sonnet | Long documents, writing, careful reasoning | 200K+ context. Strong at nuance and refusal quality. |
| Google Gemini 2.5 Pro/Flash | Multimodal, huge context, cheap Flash tier | Up to 2M context. Great for whole-codebase or whole-RFP tasks. |
| Open Llama, Mistral, DeepSeek | On-prem, privacy, cost | Runs on your own GPU or in Cloudflare Workers AI. |
Rule of thumb: start with the smart one, then downgrade for cost. Prove the prompt works on Claude Opus or GPT-5. Then move to a cheaper model like Gemini Flash or GPT-5-mini once you know the ceiling.
Feel the temperature dial
Open any AI chat you have. Ask the same question three times:
- "Give me a name for a design agency focused on accessibility."
- "Give me another."
- "Now give me 10 wildly different ones."
Notice how much the answer shape changes with how you phrase the ask. You just controlled temperature-in-spirit without touching a slider. This is why Module 2 exists.