AI can feel overwhelming. But once you understand the core building blocks, everything clicks. Here are the 20 most important concepts, short, clear, and straight to the point. With the added bonus, you look like the smart AI cookie in the room.
The basics
1. Neural networks — Layers of connected “neurons” that transform input into predictions. Each connection has a weight, and training adjusts billions of these until the model gets things right.
2. Transfer learning — Instead of training from scratch, you take a model already trained on broad data and adapt it for a specific task. Like learning to ride a motorcycle when you already ride a bike.
How transformers work
3. Tokenisation — AI doesn’t read words; it reads tokens. “Playing” might split into “play” + “ing.” This keeps the vocabulary manageable and lets the model handle words it’s never seen before.
4. Embeddings — Tokens get converted into lists of numbers. Words with similar meanings land close together in this space. That’s how the model understands meaning — through distance and direction.
5. Attention — A word’s meaning depends on context. “Apple” means something different near “shares” vs “juice.” Attention, let’s look at every word and decide what’s relevant.
6. Transformer — The architecture combining all of the above. Introduced in 2017. Powers GPT, Claude, Gemini; basically all modern AI.
Large language models
7. LLM — A transformer trained on trillions of tokens with one goal: predict the next word. Do this enough times, and the model learns to write, reason, and code, without being explicitly taught any of it.
8. Context window — The maximum text a model can hold at once, its short-term memory. Bigger is better, but costs more. Models also tend to “lose” things buried in the middle.
9. Temperature — Controls how creative or predictable the output is. Low = focused and safe. High = more surprising, but less reliable.
10. Hallucination — When a model generates convincing text that’s simply wrong. It doesn’t know truth from fiction; it just predicts what sounds right. Always verify important outputs.
Training and optimization
11. Fine-tuning — Continue training a pretrained model on a smaller, specific dataset to specialise it, like legal documents or medical Q&A. Powerful, but expensive.
12. RLHF — Reinforcement Learning from Human Feedback. Humans compare responses and pick better ones. The model learns what “helpful” and “safe” actually look like. This is why chatbots feel conversational.
13. LoRA — Low-Rank Adaptation. Fine-tunes a model by adding tiny trainable layers on top, rather than changing the whole thing. Cuts costs dramatically; you can run it on a single GPU.
14. Quantisation — Compresses a model by storing weights with less precision. A model that normally needs a server can suddenly run on a laptop. Slight quality drop, massive efficiency gain.
Prompting and reasoning
15. Prompt engineering — The way you ask matters. “Explain APIs” gives a generic answer. “Explain REST API authentication with an example for a beginner” gives a useful one. Clarity beats length.
16. Chain of thought — Instead of jumping to an answer, ask the model to reason step by step. This dramatically improves accuracy on logic, math, and complex tasks. Give it room to think.
Building AI systems
17. RAG — Retrieval-Augmented Generation. Before answering, the system fetches relevant documents and passes them as context. Reduces hallucinations without retraining. Update docs, not the model.
18. Vector database — Stores text as numerical embeddings and searches by meaning, not keywords. “Heart attack” still finds documents about “cardiac arrest.” The engine behind RAG systems.
19. AI agents — Models that don’t just respond, they act. They can use tools, run code, call APIs, and loop through decisions until a task is done. Powerful, but reliability is still a real challenge.
20. Diffusion models — Image generators that learn by corrupting images with noise, then reversing it. At runtime, start with pure noise and clean it up step by step until an image appears. Behind Midjourney, DALL-E, and Stable Diffusion.
That’s it. Twenty concepts that cover 95% of what you’ll encounter in AI today.
