beginner

Glossary & milestone quiz

One-screen recap, then ten questions. Pass with ≥7/10 to call yourself fluent in the foundations.

The whole track in one picture

Every lesson in the track maps to one stage of this end-to-end loop.

Acronym cheat sheet

The tech world loves three-letter labels. Here are the ones you've met (and a few you'll meet next), each tied to its meaning so the letters stop being random.

Acronym	Full form	How to remember
AI	Artificial Intelligence	Machines doing things that needed human intelligence.
ML	Machine Learning	Programs that learn from examples, not from rules.
LLM	Large Language Model	Large (huge) + Language (text) + Model (a learned program).
BPE	Byte-Pair Encoding	Common pairs of bytes (characters) get bundled into single tokens.
API	Application Programming Interface	The "front desk" one Application uses to ask another for something.
JSON	JavaScript Object Notation	Text with curly braces. Universal data format on the web.
TTFT	Time To First Token	The pause before the very First Token appears.
TPS	Tokens Per Second	How fast text streams once it's started.
RAG	Retrieval-Augmented Generation	Retrieve the right passage → Augment the prompt with it → Generate the answer.
CoT	Chain-of-Thought	Intermediate thinking steps before the final answer. "Think step by step."
RLHF	Reinforcement Learning from Human Feedback	RL (rewarding good behaviour) + HF (rewards come from humans rating responses).
MCP	Model Context Protocol	A protocol for giving any Model the right Context from any source. "USB-C for AI tools."
PII	Personally Identifiable Information	Any info that identifies a person — name, email, ID.
OCR	Optical Character Recognition	Optically reading characters (text) inside images.
GPU	Graphics Processing Unit	The chip type that powers AI training and inference.

(Hover any acronym anywhere in the lessons to see its full form and memory hook.)

Plain-English glossary

Term	If you only remember one thing…
LLM	A very, very good autocomplete.
Token	The chunks of text the model bills by. ~0.75 words per token in English.
Context window	The chef's counter. Everything in this conversation has to fit on it.
System prompt	The manager's brief — your rules and persona, set by the developer.
User / Assistant	The customer's question / what the assistant said earlier.
Temperature	Dice heat. 0 = always favourite. Higher = more variety.
Top-p	Roll the dice only among the top X% of probability mass.
Hallucination	Confident-sounding text that's plain wrong.
TTFT	The pause before the model starts speaking. The thing users feel.
Streaming	Speaking while still thinking — same total time, much better UX.
Prompt caching	The translator memorising the briefing instead of re-reading it.
Model routing	Cheap translator for easy questions, senior interpreter for hard ones.
RAG (medium track)	Giving the model the actual document instead of asking from memory.
Embedding (medium track)	A vector that captures meaning — used for semantic search.
Tool calling (medium track)	The model asks your code to run a function; reads the result; replies.
Agent (medium track)	An LLM in a `while` loop that can call tools.
Fine-tuning (advanced)	Continued training to specialise a model.

Cheat sheet: what to reach for

A user asks a question →
├── Generic knowledge → just call the LLM (low temperature, good system prompt)
├── Needs your data → RAG (medium track)
├── Needs an action → tool calling (medium track)
├── Needs many steps → agent loop (medium track)
└── Custom voice or rare task → fine-tuning (advanced track)

Milestone quiz (10 questions)

Check your understanding

1. Roughly how many tokens are in 1,500 words of English prose?
2. Which roles can the developer set in a chat request?
3. You set temperature=0 and run the same prompt twice. Will the outputs match?
4. Stuffing 199k tokens into a 200k context window — what's the most likely problem?
5. Output tokens are typically priced about how much higher than input tokens?
6. Your model returns a citation to 'Smith v. Jones, 412 F.3d 891 (4th Cir. 2009)'. Best next step?
7. Which of these is NOT a real way to reduce hallucinations?
8. Which is most likely a single token in a modern BPE tokenizer?
9. What's the strongest combination for a chatbot that should never invent facts?
10. You move from a frontier model to a mid-tier model — CSAT drops 2 points but cost drops 80%. The right framing?

What's next

You're ready for the Medium track, where you'll learn how production LLM apps are actually built: embeddings, RAG, prompt engineering, tool calling, agents, evals, and caching. Things start to feel less like magic and more like systems.

Related lessons in this track

Ethics & responsible use

Bias, fairness, environmental cost, who decides what AI refuses.

Capstone: use AI for one real task

Apply the whole track to a real task this week. No coding.