# 20 AI concepts worth understanding, from basic to advanced

A progressive field guide from simple model vocabulary to how modern AI systems actually work.

**Audience:** Students, professionals, creators, and curious builders

**Outcome:** Learn the concepts that make AI products, agents, and trade-offs easier to reason about.

## The foundation

- [ ] **Model** — A trained system that maps an input to a prediction or generated output.
- [ ] **Training vs inference** — Training adjusts the model; inference uses it to answer requests.
- [ ] **Tokens** — Pieces of text the model processes, not necessarily whole words.
- [ ] **Context window** — The amount of material available during a single interaction.
- [ ] **Hallucination** — A fluent but unsupported or incorrect model output.

## Getting useful answers

- [ ] **Prompting** — Instructions and context that shape a model's response.
- [ ] **Temperature** — A parameter influencing the randomness of sampling.
- [ ] **Embeddings** — Vector representations used to compare meaning or similarity.
- [ ] **Retrieval-augmented generation** — Supplying retrieved external information at answer time.
- [ ] **Tool calling** — Letting a model request operations performed by external software.

## Building systems

- [ ] **Fine-tuning** — Further training a model on task-specific examples.
- [ ] **Quantization** — Using reduced numerical precision to shrink model memory and compute.
- [ ] **Latency vs throughput** — Time per response versus total work processed over time.
- [ ] **Evaluation sets** — Repeatable examples used to assess performance and regressions.
- [ ] **Agent orchestration** — Coordinating planning, tools, and multi-step execution.

## Advanced ideas

- [ ] **Attention** — A mechanism for weighing relationships between token representations.
- [ ] **Mixture of experts** — Routing computation through selected model sub-networks.
- [ ] **Distillation** — Training a smaller model to approximate a larger system.
- [ ] **Inference-time compute** — Spending more computation to improve a response at runtime.
- [ ] **Alignment and guardrails** — Methods to guide behavior and constrain risky actions.

---

Agentic Agency · https://agenticagency.io