🧠 The Complete AI/ML Mastery Roadmap

From Zero to the Person Who Knows It All

This is not a course list. This is a life plan for mastering the deepest field humanity has ever built.

The Core Philosophy

Most people who study AI learn what things are.
World-class researchers understand why they are, when they break, and what comes next.

The gap between "knowing about AI" and "knowing AI" is not more courses.
It is the quality and depth of your thinking.

"You don't understand something until you can derive it yourself, break it yourself, and fix it yourself."

The 4 Levels of AI/ML Mastery

Level 1 — LITERATE       You can read about AI and understand it
Level 2 — PRACTITIONER   You can build AI systems that work
Level 3 — RESEARCHER     You can find and solve problems others don't see
Level 4 — FRONTIER       You define what the field studies next

Most people stop at Level 2. This roadmap takes you to Level 4.

The 24-Month Mastery Curriculum

PHASE 0 — The Foundation Audit (Week 1–2)

Self-Assessment Checklist:
- [ ] Can I prove backpropagation from scratch?
- [ ] Can I derive gradient descent from a Taylor expansion?
- [ ] Can I explain why attention is O(n²) and what flash attention changes?
- [ ] Can I state the PAC learning theorem and what it implies?
- [ ] Can I explain the difference between KL divergence and Wasserstein distance?
- [ ] Can I derive the ELBO for a VAE from Bayes' theorem?
- [ ] Can I explain what an induction head is in a transformer?
- [ ] Can I write a transformer from scratch in NumPy?

Rule: For every "No" — that gap is your first priority.

PHASE 1 — The Mathematical Core (Months 1–4)

Month 1 — Linear Algebra & Calculus

Topic	Resource	Output
SVD, eigendecomposition, matrix calculus	Gilbert Strang (MIT OCW)	Derive PCA from scratch
Multivariable calculus, Jacobians, Hessians	3Blue1Brown (Essence of Calculus) + Stewart	Derive Newton's method
Automatic differentiation	Baydin et al. (2018)	Build autograd in Python

Month 2 — Probability & Statistics

Topic	Resource	Output
Measure theory basics	Durrett "Probability" Ch 1–3	Prove SLLN
Bayesian inference	MacKay Ch 2–4	Implement MCMC sampler
Information Theory	Cover & Thomas Ch 1–5	Prove data processing inequality
Concentration inequalities	Wainwright "High-Dim Stats"	Derive Hoeffding's bound

Month 3 — Optimization Theory

Topic	Resource	Output
Convex analysis	Boyd & Vandenberghe (free PDF)	Solve QP with KKT conditions
Non-convex landscapes	Dauphin et al. 2014	Explain why saddle points ≠ local minima
SGD convergence theory	Bottou et al. 2018	Derive convergence rate
NTK Theory	Jacot et al. 2018	Explain infinite-width limit

Month 4 — Statistical Learning Theory

Topic	Resource	Output
VC Theory	Shalev-Shwartz "Understanding ML"	Prove VC shattering
Rademacher Complexity	Bartlett & Mendelson 2002	Derive generalization bounds
PAC-Bayes	McAllester 1999	Implement a PAC-Bayes bound
Double Descent	Belkin et al. 2019	Reproduce the experiment

Mastery Test: Take any top NeurIPS paper from 2020–2024. If you can follow every derivation — you pass.

PHASE 2 — Neural Architecture Mastery (Months 5–8)

Month 5 — From Perceptron to Deep Networks

[ ] Full neural network in pure NumPy (no PyTorch, no JAX)
[ ] Derive backpropagation through a 5-layer network by hand
[ ] Implement and explain BatchNorm, LayerNorm, Dropout theoretically
[ ] Resource: Andrej Karpathy's "Neural Networks: Zero to Hero" (YouTube)

Month 6 — The Transformer

[ ] Read "Attention Is All You Need" (Vaswani et al.) — annotate every equation
[ ] Implement a full Transformer from scratch (pure NumPy)
[ ] Implement Flash Attention — understand IO complexity argument
[ ] Implement Rotary Position Embeddings (RoPE) from scratch
[ ] Read Anthropic's mechanistic interpretability papers

Month 7 — Convolutional & Graph Architectures

[ ] Implement CNN; derive convolution as group convolution
[ ] Read "Geometric Deep Learning" (Bronstein et al., 2021)
[ ] Implement a basic GNN; prove Weisfeiler-Lehman connection
[ ] Implement S4 (state space model)

Month 8 — Generative Models

[ ] VAE from scratch — derive ELBO yourself first
[ ] DDPM from scratch — derive every step of forward/reverse process
[ ] Read "Score-Based Generative Modeling through SDEs" (Song et al., 2021)
[ ] Implement Real-NVP normalizing flow

Mastery Test: Reproduce any architecture from a landmark paper using only the paper. No tutorial, no existing code. It must converge.

PHASE 3 — Learning Paradigms (Months 9–12)

Month 9 — Self-Supervised & Contrastive Learning

[ ] Derive InfoNCE from mutual information maximization
[ ] Implement SimCLR and BYOL — understand why BYOL works without negatives
[ ] Read DINO, DINOv2, MAE papers

Month 10 — Reinforcement Learning

[ ] Implement Q-Learning and SARSA from scratch
[ ] Derive the Policy Gradient Theorem — full proof
[ ] Implement PPO from scratch
[ ] Read Sutton & Barto — full book

Month 11 — RLHF & Alignment Methods

[ ] Read InstructGPT (Ouyang et al., 2022)
[ ] Derive DPO from RLHF objectives
[ ] Understand Constitutional AI (Anthropic, 2022)
[ ] Implement a toy RLHF pipeline

Month 12 — Advanced Paradigms

[ ] MAML — derive bi-level optimization; implement on Omniglot
[ ] EWC — derive the Fisher Information regularizer
[ ] IRM — derive the invariant risk minimization objective
[ ] LoRA — understand low-rank decomposition theory

Mastery Test: Full RLHF training loop from scratch. It must converge.

PHASE 4 — Large Language Models (Months 13–16)

Month 13 — LLM Internals

[ ] Study GPT-2 architecture line by line (nanoGPT)
[ ] Rederive the Chinchilla scaling laws compute-optimal curve
[ ] Implement BPE tokenization from scratch

Month 14 — Reasoning & Test-Time Compute

[ ] Read DeepSeek-R1 — RL for reasoning
[ ] Study Chain-of-Thought mechanistically
[ ] Read "Scaling LLM Test-Time Compute" (Snell et al., 2024)

Month 15 — Multimodal Models

[ ] Implement CLIP contrastive pretraining
[ ] Study Flamingo cross-attention architecture
[ ] Read GPT-4V, Gemini technical reports critically

Month 16 — Mixture of Experts

[ ] Derive the load balancing loss for sparse MoE
[ ] Switch Transformer paper — read in full
[ ] Study routing collapse and prevention

PHASE 5 — AI Safety & Interpretability (Months 17–20)

Month 17 — Mechanistic Interpretability

[ ] Read "Toy Models of Superposition" (Anthropic)
[ ] Study Sparse Autoencoders — implement one
[ ] Read "In-context Learning and Induction Heads"
[ ] Read "Scaling Monosemanticity"

Month 18 — Alignment Theory

[ ] Read Stuart Russell's "Human Compatible"
[ ] Study mesa-optimization and inner alignment
[ ] Study deceptive alignment detection

Month 19 — Formal Safety Methods

[ ] Adversarial robustness: randomized smoothing
[ ] Formal verification: CROWN, α,β-CROWN
[ ] Study red-teaming techniques

Month 20 — Governance & Risk

[ ] Read Anthropic RSP, OpenAI Preparedness Framework
[ ] Read "Situational Awareness" (Aschenbrenner, 2024)
[ ] Critical analysis: What does the field get wrong about risk?

PHASE 6 — Frontier Specialization (Months 21–24)

Pick 2–3 and go deeper than anyone else:

Frontier Area	Key Papers	What You'll Build
Geometric Deep Learning	Bronstein 2021	Equivariant protein model
Causal AI	Pearl, IRM paper	Causal discovery algorithm
World Models	LeCun JEPA, DreamerV3	Latent dynamics model
Physical AI / Robotics	RT-2, π0	Imitation learning pipeline
Scientific AI	AlphaFold 2, ESM3	Sequence model
AI for Math	AlphaProof, Lean4	Formal proof assistant
AI Agents	ReAct, Voyager	Tool-using agent system

Daily Habits of a World-Class AI Researcher

Morning Protocol (1–2 hours)

07:00 — Read 1 paper from arxiv (cs.LG, cs.AI, stat.ML)
07:45 — Write 3 bullets: main contribution, key assumption, one weakness
08:00 — Implement or derive one thing from yesterday's paper

Evening Protocol (1–2 hours)

21:00 — Deep work: derive, implement, or write about one concept
22:00 — Update your Known Unknown Map
22:30 — Queue tomorrow's paper

Weekly Rituals

Monday: Read one classic paper you've been avoiding
Wednesday: Implement something with only the paper — no tutorials
Friday: Write a 1-page explanation of one concept (Feynman method)
Saturday: Study one topic completely outside AI
Sunday: What do you now know that you didn't on Monday?

The Feynman Loop

1. Study a concept until you think you understand it
2. Close every resource
3. Explain it as if teaching a smart 15-year-old
4. Find every "kind of" or "sort of" — those are your gaps
5. Go back and fill them
6. Repeat until zero vague words remain

The Known Unknown Map

Create known_unknown_map.md and update it weekly:

## I Know Deeply (Can Derive + Implement + Explain)
- Backpropagation
- Attention mechanism

## I Know Superficially (Can Explain but Not Derive)
- NTK theory
- Wasserstein GAN

## I Know Exist But Don't Understand
- Free Energy Principle
- Kolmogorov complexity

## I Don't Know I Don't Know
(Filled by reading broadly — these surprises are the most valuable)

This map is your real measure of progress — not certifications or courses completed.

Master Reading Stack

Tier 1 — Read Every Word

Book	Why
Understanding Machine Learning — Shalev-Shwartz	Best theoretical ML text
Deep Learning — Goodfellow, Bengio, Courville	Architecture theory bible
Reinforcement Learning — Sutton & Barto	Complete RL foundation
Information Theory, Inference & Learning — MacKay	Unified information view
The Book of Why — Pearl	Causal thinking
Human Compatible — Stuart Russell	Alignment from first principles

Tier 2 — Follow Actively

Source	What You Get
arxiv.org/list/cs.LG/recent	Daily frontier papers
Papers With Code	Benchmarks, implementations
Distill.pub	Clearest explanations in ML
Anthropic Research Blog	Interpretability frontier
Alignment Forum	AI safety research discourse

The Implementation Portfolio

ml-from-scratch/
├── foundations/
│   ├── autograd_engine.py        # Your own autograd
│   └── optimization/             # SGD, Adam, natural gradient
├── architectures/
│   ├── transformer_numpy.py      # Full transformer, no PyTorch
│   ├── diffusion_ddpm.py         # DDPM from scratch
│   └── gnn_basic.py              # Message passing GNN
├── learning_paradigms/
│   ├── ppo_from_scratch.py       # PPO implementation
│   └── meta_maml.py              # MAML
└── theory/
    └── generalization_bounds.py   # PAC, VC, Rademacher

This repository is your proof of knowledge — better than any degree.

Mastery Benchmarks

Level	Test
Level 1	Read any NeurIPS abstract and explain the contribution plainly
Level 2	Implement a transformer from scratch in < 200 lines
Level 3	Read 10 papers in one area; identify the 3 most important open problems
Level 4	Propose a novel research direction: formal problem statement + hypothesis + experimental protocol + expected theoretical contribution

How to Enter the Research Community

Month 6+ — Write about one paper per week publicly. Build your network.
Month 10+ — Reproduce a paper. Post a replication report on GitHub.
Month 16+ — Pick ONE open problem. Spend 3 months on it.
Month 18+ — Submit to a NeurIPS/ICML/ICLR workshop.
Always — Email researchers with specific technical questions after reading their work.

Mental Models That Separate Experts from Everyone Else

1. The Abstraction Stack

For every idea: What problem does this solve? What assumptions? What breaks? What's better?

2. The Unification Lens

Diffusion ↔ SDEs ↔ score matching ↔ Langevin dynamics
Attention ↔ kernel regression ↔ memory retrieval
RL ↔ optimal control ↔ Bayesian inference

Always ask: "Where have I seen this structure before?"

3. The Falsification Mindset

What experiment would prove this wrong? Has it been run? Why not?

4. Read What's Not Said

What they didn't compare against
Which ablations they skipped
What the limitations section glosses over

Your First 30 Days — Start Here

Day	Task
1–3	Take the self-assessment. List every gap.
4–7	Implement backpropagation from scratch in NumPy. No tutorials.
8–14	Derive and implement the attention mechanism from scratch.
15–20	Read "Attention Is All You Need" — annotate every equation.
21–25	Build your Known Unknown Map. Be brutally honest.
26–30	Implement DDPM from the paper alone.

After 30 days, you will know more about AI than 95% of ML engineers.
After 24 months, you will be at the frontier.

The Final Truth

There is no certification for knowing everything about AI.
There is no moment where you "arrive."

The person who knows the most about AI is the person who:

Has the most precise understanding — can derive, not just describe
Has the widest map — connects the most domains
Has the deepest curiosity — never stops asking why
Contributes back — because knowledge unexpressed is knowledge that dies

The goal is not to finish this roadmap.
The goal is to become the kind of person who never stops.

"The master has failed more times than the beginner has tried."

This is a living document. Update it as you learn. The version of you at Month 24 will find things in Month 1 that need correcting — that's how you know it's working.