How to Learn AI From Scratch in 2026: A Realistic Roadmap

The field moves fast and most "AI roadmaps" online are out of date the day they are published. This one is current for 2026 and assumes you are starting from zero.

Stage 1 — Python (2-4 weeks)

You do not need to be a Python expert, but you do need to be comfortable with:

• Variables, lists, dicts, loops, functions, classes
• Reading and writing files
• Installing packages with pip or uv
• Using Jupyter notebooks

The Python for AI tutorial on ModelRefs covers exactly what you need without padding.

Stage 2 — Numerical Python (1-2 weeks)

Learn numpy and pandas. Almost every model you will ever train expects data as a tensor or a dataframe, and these are the libraries that get you there.

Stage 3 — Machine learning fundamentals (4-6 weeks)

Focus on the intuition, not the proofs. Learn:

• Train/validation/test splits and why they matter
• Overfitting and how to fight it
• Linear and logistic regression
• Decision trees and gradient boosting
• The bias-variance tradeoff

scikit-learn is your friend here.

Stage 4 — Deep learning (6-8 weeks)

Pick one framework — PyTorch is the modern default — and learn:

• Tensors and autograd
• Building and training a simple neural network
• Convolutional networks for images
• Transformers for text (see our Transformers article)

Stage 5 — LLMs and applied AI (ongoing)

This is where the field is hottest in 2026. Learn:

• Prompt engineering and prompt structure
• Retrieval-augmented generation (RAG)
• Fine-tuning with LoRA / QLoRA
• Agent frameworks and tool use

Realistic timeline

If you study 10 hours per week, you can be productive with applied AI in 4-6 months. You will never be "done" learning — the field reinvents itself yearly — but you will be useful long before that.

The honest shortcut

The single fastest way to learn is to build things. Pick a tiny project (a chatbot for your notes, a classifier for your photos) and build it badly. Then build it less badly. Then again. That is how everyone in this field actually got good.