Skills by Role

ML Engineer: build the model muscles

The ML Engineer path is the most math- and theory-heavy of the three, but you do not need all of it before you start. Build a working intuition first, then deepen the theory as specific problems demand it.

1. 1Solid Python + data toolsGet fluent with Python, NumPy, and pandas. You should be able to load, clean, and explore a dataset without friction.
2. 2Core ML with a libraryLearn supervised learning end to end with scikit-learn: train/test splits, a few algorithms, and honest evaluation metrics. Ship a small prediction model.
3. 3The math that explains itLayer in linear algebra, probability, and the calculus behind gradient descent, enough to understand *why* training works, not just how to call it.
4. 4Deep learning fundamentalsMove to PyTorch (or similar): build and train a small neural network, understand backprop, loss, and optimizers.
5. 5Make it realDeploy a model behind an API and learn the basics of monitoring. A model that runs for someone else is worth ten in a notebook.

AI Engineer: build products on top of models

The AI Engineer path leans on software engineering you may already have, then adds the applied skills for working with model APIs. You can build something genuinely useful in week one, which makes this a fast and motivating on-ramp.

1. 1Strong software fundamentalsBe comfortable with a real language, version control (Git), HTTP/JSON, and calling APIs. This is the bedrock the rest sits on.
2. 2Prompting and the model APILearn to call an LLM API and get reliable behavior: system prompts, structured output, and handling failures and cost.
3. 3Retrieval (RAG)Connect a model to your own data: chunking, embeddings, a vector store, and retrieval. Build a document-Q&A app.
4. 4EvaluationLearn to measure whether your app is actually good, test sets, judging outputs, and catching regressions. This is what separates hobby projects from real ones.
5. 5Agents and toolsGive the model tools to act: function calling, multi-step workflows, and guardrails. Add real capability only once you can evaluate it.

Researcher: build depth and rigor

The research path trades breadth for depth. It rewards patience, mathematical maturity, and the discipline to read, reproduce, and write. It is the slowest path to a first result, but the one that pushes the field forward.

1. 1Mathematical maturityGo beyond recipes: linear algebra, probability, and optimization deeply enough to follow derivations in papers.
2. 2Read papers activelyLearn to read a paper in passes, skim, understand, then scrutinize. Keep notes on the problem, idea, and evidence.
3. 3Reproduce resultsRe-implement a paper's core result from scratch. Nothing teaches a method like making it run and match the numbers.
4. 4SpecializePick one narrow area (e.g. interpretability, RL, efficient inference) and go deep. Research rewards depth over breadth.
5. 5Write and shareWrite up your work clearly, a blog post, a reproduction, eventually a paper. Communication is half of research.

Choosing your first trail

Whichever you choose, the meta-rule is the same: learn just enough to build the next concrete thing, build it, then let the gap you hit tell you what to learn next. Skills stick when they are attached to a project you cared about finishing.

Further reading

• fast.ai, Practical Deep Learning for Coders, a free, project-first course that gets you building models quickly.
• Google, Machine Learning Crash Course, a solid, free foundation in core ML concepts.
• Andrej Karpathy, Neural Networks: Zero to Hero, build neural nets and a small language model from scratch.
• DeepLearning.AI, courses and specializations, structured paths across ML, deep learning, and applied AI.
• Andrew Ng, How to Build Your Career in AI, practical guidance on sequencing your learning and projects.