AI AI-00.03
The map of modern AI
How the AI track is organized: classical AI in the first half, modern LLM engineering in the second, and which ML lessons to take first.
The big idea
Modern AI looks like one giant topic, but it splits cleanly into two halves. The classical half is about how a machine can decide what to do: search, logic, planning and reasoning with probabilities. These ideas are older, exact and still everywhere, in map routing, game engines, schedulers and the planning loops inside AI agents. The modern half is about large language models and the systems built around them: how text becomes numbers, how transformers work, how models are trained and aligned, and how engineers build reliable products with prompting, retrieval, tools, evaluation and serving.
You do not have to read in order. If you want to build with LLMs this month, you can start at the second half. If you want to understand why an agent's plan fails, the first half explains it. This lesson shows you where everything is.
How it works
Step 1: The map
Each station is a module. Lit stations are open now; dots show lessons you have finished on this device.
Step 2: The classical half (modules 1 to 3)
- Agents and search: framing problems as states and actions, then finding paths with breadth-first search, A*, game-tree search and constraint solving.
- Knowledge, logic and planning: representing facts and rules, drawing conclusions, and planning sequences of actions.
- Reasoning under uncertainty: probability, Bayesian networks and decisions when the world is noisy.
Step 3: The modern half (modules 4 to 15)
- Representations and transformers: embeddingsLists of numbers that represent words, images or other items so that similar things end up close together., tokenization, attention and how a large language model produces text.
- Training and alignment: pre-training, fine-tuning, LoRA, and making models follow instructions and preferences.
- Building with LLMs: prompt and context engineering, retrieval augmented generation (RAG), tools and agents.
- Running it for real: evaluation, inference and serving, multimodal models, safety and security, product design, and capstone projects that combine everything.
Step 4: Where the ML track fits
The AI track explains how transformers work, but it assumes you know what a neural network is and how training by gradient descent works. If you do not, take these ML lessons first:
The matching lessons are Vectors and vector spaces , Matrices and linear transformations , The perceptron and the artificial neuron , Multi-layer networks and Backpropagation .
Going deeper
Three suggested routes
- "I want to build an LLM app now." Read module 7 (prompting), module 8 (RAG), module 9 (agents) and module 10 (evaluation), then come back for the internals.
- "I want to understand how LLMs work." Take the ML lessons above, then modules 4 to 6 in order.
- "I am preparing for AI engineering interviews." Modules 5, 8, 10 and 11 cover the questions that come up most: attention, RAG design, evaluation and serving cost.
Every lesson has the same rhythm
Big idea, mechanics with diagrams you can step through, a hands-on widget, trade-offs, common mistakes, a real story, a quiz and a recap. The quizzes explain every answer, so getting one wrong is part of the lesson.
Common mistakes
- Skipping evaluation. Building an AI feature without a way to measure it is the most common reason it disappoints in production.
- Treating classical AI as obsolete. Agents still plan, search and satisfy constraints, often with an LLM proposing steps.
- Jumping into transformer internals with no neural network background. Take the short ML path first; it saves time.
Check yourself
-
Which topics belong to the classical half of the track?
Show answer
B. Search, logic, planning and probability Modules 1 to 3 cover agents and search, knowledge and planning, and reasoning under uncertainty.
-
You want to ship a document question-answering app quickly. Which modules matter most?
Show answer
B. Prompting, RAG, agents and evaluation Modules 7 to 10 cover building and measuring LLM applications.
-
What should you study before transformer internals if you have never trained a neural network?
Show answer
B. The ML track's lessons on vectors, matrices, the perceptron, multi-layer networks and backpropagation Transformers are neural networks trained by backpropagation, so those ML lessons make module 5 much easier.
Recap
- The AI track has a classical half (modules 1 to 3) and a modern half (modules 4 to 15).
- Classical AI: search, logic, planning and probability, still used inside real systems and agents.
- Modern AI: embeddings, transformers, training, prompting, RAG, agents, evaluation, serving and safety.
- Take the short ML path on neural networks before transformer internals.
- Pick a route by goal: build now, understand deeply, or prepare for interviews.