The map of AI: where LLMs fit

In one line: "AI" is a set of nested ideas — artificial intelligence contains machine learning, which contains deep learning, which contains generative AI, of which large language models (LLMs) are the text branch — and agents are LLMs wired into a loop with tools. This guide teaches the LLM + agent slice, because that's where most of the useful 2026 work is.

In plain English

"AI" has been a moving target since the 1950s — it has meant chess programs, spam filters, recommendation feeds, and self-driving cars at different times. Today, when most people say "AI," they mean one recent breakthrough: large language models like the one behind ChatGPT. That's a branch of a much older tree, not the whole tree. This page draws the tree once, so that every later lesson — tokens, embeddings, agents — has a place you can point to and say "I am here."

The nested map

Each ring below is contained in the one outside it. Bigger ring = broader, older idea; inner ring = newer, more specific.

Reading it from the outside in:

• Artificial Intelligence (AI) — any technique that makes a machine do something that seems to need intelligence. The umbrella. Includes old rule-based "expert systems," search and planning algorithms, robotics, and everything below.
• Machine Learning (ML) — systems that learn patterns from data rather than being explicitly programmed. The shift that made modern AI work. ML learns in three ways:
  • Supervised — learn from labeled examples ("this email is spam, that one isn't"). Most classical ML, and the instruction-tuning step of an LLM.
  • Unsupervised — find structure in unlabeled data (clustering customers, or the embeddings you'll meet soon).
  • Reinforcement learning (RL) — learn by trial and error against a reward signal (game-playing AIs — and the RLHF that aligns LLMs to human preferences).
• Deep Learning — ML built on many-layered neural networks. What made vision, speech, and language suddenly work in the 2010s.
• Generative AI — deep-learning models that produce new content (text, images, audio, video) instead of only labeling or scoring existing data. The two big families are diffusion models (images/video) and LLMs (text).
• LLMs — generative models that work over tokens of text; the transformer is their architecture. This is the guide's spine.
• Agents — an LLM run in a loop with tools and memory so it can take actions, not just talk. The guide's throughline.

Where classical AI/ML still wins

Generative AI did not delete the rest of the map. For huge classes of problems, an older, smaller, cheaper model is still the right answer:

• Tabular prediction (fraud, churn, credit scoring) — gradient-boosted trees beat LLMs on structured data and are thousands of times cheaper per call.
• Recommendations at scale — collaborative filtering still carries the steady state; LLMs mostly help with cold-start.
• Structured lookups — a SQL SELECT or a regex is faster, deterministic, and auditable.

Knowing the map is what lets you say "this is a gradient-boosting problem, not an LLM problem." The full treatment lives in When not to use AI.

What this guide covers (and what it doesn't)

This is a guide to the LLM + agent slice of AI and the engineering around it — building, evaluating, shipping, and operating LLM-powered and agentic systems. It teaches the deep-learning ideas you need to use these models well (next: neural networks), but it is not a course in training classical ML models, computer vision from scratch, or robotics. Those are their own fields; this page is your pointer to where they sit on the map so the boundary is honest.

Why it matters

• Orientation. Every later lesson is a zoom-in on the inner rings. When something feels abstract, come back here and locate it.
• Vocabulary. "Deep learning," "generative AI," and "LLM" are not synonyms — they're nested. Using them precisely marks you as someone who understands the field, not just the hype.
• Judgment. The single most valuable habit is recognizing when a problem belongs to an outer ring (a rule, a query, a classical model) rather than reaching for the newest, biggest tool by reflex.

Common pitfalls

Where beginners trip

• Thinking "AI = LLMs." LLMs are one branch of one sub-field. The umbrella is far older and wider.
• Believing generative AI replaced classical ML. It didn't — for tabular and structured-data problems, classical ML still wins on cost, speed, and accuracy.
• Confusing the training with the model you call. Building an LLM (deep learning at huge scale) happens once at a lab; your daily work is calling the finished model. See training vs. inference.
• Skipping the map because you "just want to build." Without it, you'll reach for an LLM on problems a regex would solve, and you'll mix up paradigms when reading docs and papers.

🤔 Quick checkQuick check

---

→ Next: Before you start: the programming you need — the small coding bar this guide assumes.