How to Become an AI Engineer in 2027: A Free Map of 229 Concepts That Starts Where You Are

You search "AI roadmap" and every video hands you the same list: Python, math, machine learning, transformers, RAG, agents. The list starts everyone at step one, and it never answers the only question that matters: what should I learn next? Me. Today.

So we stopped making lists, and we built a map. The AI Engineer Map has 229 concepts. Every one is connected to what you need to learn first, and every one has a free resource and a free video. You pick a goal, you tick what you already know, and the map lights up your exact next step. It's free, with no signup.

Why roadmaps stop working

Every roadmap feels like it's working, and then suddenly it isn't. That's because learning isn't a line. It's a web.

Take attention, the idea inside ChatGPT, Claude and Gemini. To really get attention, you need dot products, softmax and word embeddings. Embeddings need vectors and a basic neural network. The neural network needs logistic regression, which needs gradient descent, which needs derivatives. Miss one of those links, and the video you're watching just stops making sense.

That's not you being slow. That's a missing prerequisite (something you have to know first). A list can't show you that. A map can. So we drew every one of those links: 355 of them.

The whole map: 12 tracks

The 229 concepts sit in 12 tracks, in three layers.

• The foundations: Programming Foundations (24 concepts, from your first line of Python to Git, APIs and Docker), Math for AI (22: vectors, matrices, derivatives, probability) and Working With Data (14: NumPy, pandas, cleaning, and parsing PDFs).
• The engine room: Classical Machine Learning (25), Deep Learning (23) and Transformers (24), the architecture behind every chatbot you use.
• The part companies are hiring for right now: Building With LLMs (24), Retrieval and RAG (17), AI Agents (17), Fine-tuning (11), Evals and Safety (14) and Shipping to Production (14).

The longest chain on the whole map is 19 steps. It starts at "what is a vector" and ends at AI agents talking to other AI agents. Nobody walks all 19 in one go. With this map, you don't have to.

How the map works (3 steps)

1. Pick your goal. Say you want to build an AI agent. The 229 concepts collapse into the 28 that this goal actually needs. Everything else is hidden, because right now, it's noise.
2. Tick what you already know. Python basics, tick. The command line, tick. Each tick fills your progress bar, and the concepts that just became learnable turn orange. Orange means ready: you have everything you need for it.
3. Open anything that's ready. You get one line on why it matters, a free video that plays right there on the page, a free resource to read, what you need first, and what it unlocks. When you've got it, hit "mark as known," and the map moves forward with you.

Your progress saves in your browser, so you can close it and come back tomorrow. It also works on your phone.

How much math do you need?

This is the question everybody asks. The honest answer, straight from the map: it depends on where you're going.

To build an AI agent, you need 3 math concepts: vectors, matrices and tensors. That's about 40 minutes of video. An MCP server needs the same 3. A RAG chatbot needs 9. Fine-tuning needs 12. And if you want to understand how GPT works from the inside, 13.

So math isn't a wall at the start. It's a toll you pay depending on the road you pick, and you only pay it when you get there.

All 6 goal paths, with hours

1. Build an AI agent

An agent is an AI that does tasks for you, step by step. This path covers 6 tracks, and at about 13 and a half hours of video, it's two weekends. The spine goes: calling an LLM, then prompts, then structured outputs, then tool use, then the agent loop. You finish with designing tools and keeping a human in the loop.

2. Build a RAG chatbot

RAG is the kind of chatbot that answers questions from your own documents. The path covers 8 tracks. This is where embeddings, chunking and semantic search come in, and it finishes with reranking and citations, so your bot shows its sources.

3. Build an MCP server

An MCP server lets Claude, ChatGPT or Cursor use your own data and tools. It's the most programming-heavy path: 11 of the 27 concepts are pure programming.

4. Fine-tune a model

This path covers 9 tracks. It's a big one, because you need deep learning before LoRA (a common fine-tuning method) makes sense. The good news: the last step is fine-tuning on a free Colab GPU, so it costs you nothing.

5. Understand how GPT actually works

This path has the most math of any. It ends with Andrej Karpathy building a GPT from scratch, line by line. If you do one deep project in your life, make it this one.

6. Get hired as an AI engineer

This is the widest path, covering 11 of the 12 tracks, because interviews test breadth. It finishes with a deployed portfolio project and the AI system design interview.

Same map, 3 different people

Here's why a map beats a list.

• Person one works in IT support. They know some Python, SQL, Git, Linux and APIs. They pick "Get hired" and tick what they know. They're 17% done before they've watched a single video. The map says: learn vectors, type hints, functions and graphs, and testing. Not transformers. Not yet.
• Person two is a developer who already calls the ChatGPT API at work. They pick "Build an AI agent." They're 75% done. Their next step is literally what an agent is, versus a workflow. A list would have sent them back to Python lesson one.
• Person three is a student starting from zero who wants to understand how GPT works. They're at 0%, and that's fine. The map gives them 4 places to start: Python, vectors, functions and probability. Not 229. Four.

Same map. Three completely different next steps. That's the whole point.

The one concept under 78% of the map

It's not Python. Python unlocks 153 concepts. This one unlocks 179 of the 229.

It's the vector. Just an arrow: a list of numbers with a direction. Every word a model reads becomes a vector. Embeddings are vectors. Attention is vectors talking to each other. Semantic search, RAG, recommendations: all vectors. 3Blue1Brown explains it in 10 minutes, and it sits right at the very start of the map. If you do one thing after reading this, watch that.

226 free videos, checked by hand

The map links 226 videos from 78 channels. That's 86 hours of free teaching, from IBM Technology, StatQuest, 3Blue1Brown, Andrej Karpathy, Corey Schafer, freeCodeCamp and more.

To be fair about how we got there: the first automatic search got about 50 matches wrong. Searching for "training loop" returned a video about Microsoft Loop. "Distillation" returned a chemistry lesson on separating liquids. Every one of those got replaced by hand. Every reading link was tested too, and 224 out of 224 load.

How to use it this week

1. Pick one goal. Just one.
2. Tick honestly. Only what you could explain to a friend.
3. Learn only what's orange. Watch the video, skim the reading, then mark it known.
4. Every five ticks, build something tiny with what you just learned. Then come back, and the map will have moved.

Open source, plus the printable map

The whole map is open source on GitHub. Every concept is one small file. If you know a better video, or a concept we missed, open a pull request and it goes on the map for everyone.

If you want it offline, the printable version has the full map on one poster, plus all 6 goal paths with their first 3 steps. Comment NEXTSTEP on the video, or grab it below. Then open the map, pick your goal, and go learn the thing that's glowing.