← All roadmaps
Study Roadmap
ML / AI & Nature of Code
creative coding → simulation → learning algorithms → ML engineering
A path from creative coding to working machine learning: make natural systems move on a screen, build the classic learning algorithms by hand, then cross into deep learning and shipping real ML. This roadmap curates the outside world's best resources; the taught, from-first-principles version of the deep material lives in the Zero Series' AI Zero book, which each theory stage links to as its canonical home.
13 Stages
0/56 Resources done
— Day streak
00
STAGE 00 · numbers before symbols
Math You Can See
you finish able to → vector, gradient, and probability intuition you can picture
Web
AI Zero — the algebra, calculus & probability primers
series canonical home: the numbers-before-symbols primers this roadmap is built around
01
STAGE 01 · make something move
Creative Coding Foundations
you finish able to → a habit of turning an idea into a running sketch the same day
Book
1. The Nature of Code (Shiffman)own the print/PDF to work through offline
Book
2. Generative Design (Bohnacker et al.)form from rules — typography, colour, motion systems
Lab
Sketch-a-day journal↗
Lab setup
Software: the p5.js Web Editor (link) — nothing to install; or VS Code + the p5.vscode extension for local work.Hardware: any laptop; a drawing tablet is optional for interactive pieces.Cadence: one small sketch per Nature of Code section, each committed to a git repo so the progression is visible.Guardrail: keep sketches short and self-contained — this stage is about reps, not one giant project.
02
STAGE 02 · systems that come alive
Simulation & Emergence
you finish able to → flocking, cellular automata, and physics you built yourself
Book
1. The Computational Beauty of Nature (Flake)fractals, chaos, cellular automata, L-systems — the deep reference
03
STAGE 03 · evolution as search
Evolutionary Computation
you finish able to → a genetic algorithm evolving a solution in front of you
Book
1. An Introduction to Genetic Algorithms (Mitchell)the classic, readable treatment
Lab
Evolve-a-creature lab↗
Lab setup
Software: p5.js or plain Python (numpy) — no framework needed for a first GA.Build: evolve a phrase, then smart rockets, then a walking creature so you can watch fitness climb.Hardware: any laptop; GAs are CPU-light at this scale.Guardrail: log and plot best/mean fitness per generation — an evolution you can't see converging is a bug you can't find.
04
STAGE 04 · before the deep end
Classical Machine Learning
you finish able to → the ability to frame a problem, train a model, and read its errors
Book
1. Hands-On Machine Learning, 3e (Géron)the practitioner's bible — scikit-learn then Keras
Web
Python Zero — Python, data structures & tooling
series canonical home for the Python this stage assumes
05
STAGE 05 · no framework yet
Neural Nets from Scratch
you finish able to → backprop implemented by hand, so the framework holds no mystery
Book
1. Neural Networks from Scratch in Python (Kinsley & Kukieła)every line of a net, no libraries
Web
AI Zero — the taught neural-network curriculum
series canonical home for the theory, derived slowly from the primers
06
STAGE 06 · now use the tools
Deep Learning with a Framework
you finish able to → real models trained on a GPU, checkpointed and evaluated
Book
1. Deep Learning with PyTorch (Stevens, Antiga, Viehmann)
Lab
GPU training lab↗
Lab setup
Free tier: Google Colab or Kaggle Notebooks give a free GPU with zero local setup — start here.Local stack: Python via uv, PyTorch, and a CUDA-capable NVIDIA GPU (a used RTX 3060 12GB is the budget sweet spot) once you outgrow free tiers.Hardware note: VRAM is the ceiling — 8GB trains small models, 12–24GB for real work.Guardrail: checkpoint every epoch and watch a validation curve; free GPU sessions time out, so save often.
07
STAGE 07 · the two big architectures
Vision & Sequence Models
you finish able to → hands-on CNNs for images and RNNs for sequences
08
STAGE 08 · back to art, with nets
Generative & Creative AI
you finish able to → your own image/audio generators, wired back into sketches
Web
ml5.js↗ free — friendly ML in the browser; pairs directly with your p5.js sketches
Book
1. Generative Deep Learning, 2e (Foster)VAEs, GANs, diffusion, and transformers for making things
Web
Distill.pub↗ free — the clearest visual explanations of how these models work
09
STAGE 09 · the current frontier
Transformers & LLMs
you finish able to → a small language model you built and trained yourself
Book
1. Build a Large Language Model (From Scratch) (Raschka)a working GPT, chapter by chapter
Web
AI Zero — LLM & transformer chapters
series canonical home for the taught path; follow the link, don't duplicate
10
STAGE 10 · learning from reward
Reinforcement Learning
you finish able to → an agent that learns to act, from Q-learning to policy gradients
Lab
Agent-in-a-world lab↗
Lab setup
Software: Gymnasium (the maintained fork of OpenAI Gym) for environments, plus Stable-Baselines3 for reference algorithms.Build: solve CartPole with tabular Q-learning first, then a policy-gradient agent, then something with pixels.Hardware: CPU is fine for the classic control tasks; a GPU only matters for pixel-based envs.Guardrail: log episode reward and seed everything — RL is famously noisy, so compare runs, not single episodes.
11
STAGE 11 · ship it, keep it working
ML Engineering & MLOps
you finish able to → a trained model served, monitored, and reproducible
Book
1. Designing Machine Learning Systems (Huyen)data, features, deployment, monitoring — the practitioner's map
Web
Made With ML↗ free — MLOps, testing, CI/CD for ML, production patterns
Web
AI Ops Zero — model serving, evals, drift & guardrails
series canonical home for running models in production
Lab
Serve-a-model lab↗
Lab setup
Software: FastAPI to wrap the model from Stage 06, containerized with podman (not docker), served locally.Hardware: any laptop for CPU inference.Stretch: deploy the container to a homelab k8s cluster or a free-tier cloud runtime.Guardrail: pin dependencies (a uv lockfile) and record input/output schemas so the served model matches what you trained.
+
STAGE + · adjacent interest
Side track · Safety & Interpretability