Free AI Roadmap 2026

Don't jump into code yet. Spend Week 1 just understanding the landscape. Watch videos, read articles, and form your own mental model of what AI is and isn't.

Focus on: lists, dictionaries, functions, loops, and classes. Practice on LeetCode Easy or HackerRank. Google Colab is your best friend, free, browser-based, no setup needed. And if you are weighing Python against R: Python won for AI and production systems, R remains lovely for pure statistics.

Don't get stuck here for months. Spend 1-2 weeks getting the intuition from 3Blue1Brown videos, then move on. You'll reinforce the math naturally as you code ML models. Intuition > memorization.

StatQuest on YouTube is the single best resource for statistics. Josh Starmer explains everything with humor and clarity. Watch his playlists before touching any textbook.

Download a real dataset from Kaggle and play with it. Load it in Pandas, explore with .head(), .describe(), .value_counts(). Plot charts. This hands-on experience is worth more than 10 tutorials.

Data scientists spend 60-80% of their time on data preparation. Master Pandas deeply: .merge(), .groupby(), .pivot_table(), .apply(), handling missing values with .fillna() and .interpolate(). Learn basic SQL, most ML data comes from databases.

Start with scikit-learn, NOT TensorFlow/PyTorch. Build simple models first: Linear Regression → Logistic Regression → Decision Trees → Random Forest. Do Kaggle's Titanic competition, it's the Hello World of ML.

NEVER just use accuracy, it's misleading for imbalanced datasets. Learn when to use precision vs recall (hint: false positives vs false negatives have different costs). Start tracking experiments from day 1 with Weights & Biases, it's free for individuals.

Do at least 3 Kaggle competitions: Titanic (classification), House Prices (regression), and one tabular playground. Read winning solutions, they teach you more than any course.

Start with 3Blue1Brown's 4-part neural network series, it gives you the visual intuition that textbooks can't. Then move to fast.ai which teaches top-down (build working models first, understand theory after). Choose PyTorch over TensorFlow for learning.

Andrej Karpathy's 'Neural Networks: Zero to Hero' playlist is the single best resource for learning deep learning from scratch. He builds everything from raw Python, no library shortcuts. Watch it all.

Don't build from scratch, use transfer learning. Take a pretrained ResNet or EfficientNet from torchvision, freeze early layers, fine-tune on your data. You'll get 95%+ accuracy on most image tasks with very little data.

RNNs/LSTMs are largely being replaced by Transformers, but understanding them is crucial for the historical context and for time-series problems where they still shine. Read Chris Olah's LSTM blog post, it's the best explanation ever written.

This is the most important topic in modern AI. Karpathy's 'Let's build GPT from scratch' video (2hrs) is THE best way to learn it, he codes a working GPT character by character. Pair it with Jay Alammar's Illustrated Transformer blog post for the visual intuition.

Start with Hugging Face's free NLP course, it's practical and modern. spaCy is the go-to library for production NLP. For learning, the Stanford CS224n lectures are unmatched in depth.

The most underrated AI skill. Start with Anthropic's prompt engineering guide, it's the best practical resource. Key techniques: (1) Be specific and structured (2) Use XML tags (3) Chain-of-thought for reasoning (4) Few-shot examples (5) System prompts. Practice daily.

Karpathy's LLM intro talk is the best 1-hour overview that exists. Then try running a model locally with Ollama (literally one command). Understanding how to USE LLMs well (prompt engineering) is just as important as understanding how they work internally.

RAG is the most practical skill for 2025-2026. It lets you connect LLMs to your own data without expensive fine-tuning. Start by building a simple RAG app: embed your documents in ChromaDB, retrieve relevant chunks, pass to an LLM. LangChain or LlamaIndex make this easy.

Diffusion models have largely replaced GANs for image generation. Start by running Stable Diffusion locally with ComfyUI, it's the best way to understand the generation pipeline. For audio, OpenAI's Whisper is incredibly good and open-source.

AI Agents are where the industry is moving in 2025-2026. Start by building a simple ReAct agent with LangGraph, give an LLM access to a calculator and web search, watch it reason and act. Then build something real: a research agent, a coding assistant, or a data analysis agent.

Build a portfolio project end-to-end: train a model, wrap it in FastAPI, containerize with Docker, deploy to Hugging Face Spaces or Railway. Gradio and Streamlit are the fastest ways to create a shareable demo. Chip Huyen's MLOps book is essential reading.

David Silver's course is the gold standard for RL theory. For hands-on, use Hugging Face's Deep RL course, you train agents in game environments. Understanding PPO is especially important because it's the algorithm behind RLHF.

Your GitHub IS your resume. Build 3 standout projects: (1) an end-to-end ML app with a live demo, (2) a Kaggle competition with a top 10% finish, (3) a novel project that shows creativity. Write blog posts explaining what you built, this shows communication skills.

This isn't optional anymore. Companies are hiring for Responsible AI roles, and every ML engineer is expected to understand bias, fairness, and safety. The EU AI Act is now law, understanding AI regulation is a career advantage. Read 'Weapons of Math Destruction'.

Don't read papers cover-to-cover, read Abstract → Conclusion → Figures → Methods. Yannic Kilcher's channel is the best for paper walkthroughs. Subscribe to The Batch (Andrew Ng) and Import AI (Jack Clark).

Start with CLIP zero-shot classification in Colab, it is ten lines of code and the aha moment of the whole field: images and text living in one embedding space. Everything else in multimodal is elaboration on that trick.

Transcribe a podcast episode with Whisper on your own machine in week one, it works out of the box and makes the whole field concrete. Then learn what a mel spectrogram is: audio models are image models wearing headphones.

Simulation first: MuJoCo plus CS285 teaches you control without owning a robot. The field's big bet is that imitation learning plus VLA models does for robots what GPT did for text, watch pi.website and LeRobot to follow it live.

Run a 7B model quantized to 4-bit on your own machine, then serve it with vLLM and measure tokens per second. Inference cost is the number that decides which AI products survive, engineers who can cut it are never unemployed.

Build a MovieLens recommender from scratch before touching a library: user-item matrix, cosine similarity, done in an afternoon. Then read Google's course on why production recommenders are retrieval plus ranking, not one model.

The embarrassing secret of the field: a seasonal naive forecast beats fancy models on many real series. Always compute it first, and read Hyndman's fpp3 cover to cover, it is the rare free book that is also the field's standard reference.

Start with the Distill article, it is the best visual explanation in the field. Then do Cora node classification in PyTorch Geometric, the MNIST of graphs. This site's own Graph Types for AI Agents teardown pairs well as the applied view.