From Foundation to Frontier
A curriculum for understanding modern AI.
Start with the pieces that do not change quickly: Python, arrays, vectors, probability, gradients, optimization, and neural networks. Use them to understand language models, transformers, reinforcement learning, generative models, and frontier systems.
Where should I start?
Choose the first useful step.
The path is cumulative, but not everyone enters at the same point. Pick the first stage that exposes a real gap, then move forward.
Foundations
These tracks are for beginners and intermediate readers. They form the path from basic computation to modern LLMs, generative models, and reinforcement learning.
Python
Programming basics, functions, files, NumPy, plotting, and small scientific experiments.
plannedMath
Mathematical language, vectors, matrices, geometry, linear systems, probability, gradients, optimization, and information theory for machine learning.
Open MathavailableDeep Learning
Neural networks as tensor maps: shapes, gradients, losses, optimization, and debugging.
Open Deep LearningdraftLanguage Modeling
Text as data: context, next-token prediction, tokens, n-grams, sequence probability, embeddings, and neural language models.
Open Language ModelingdraftTransformers and LLMs
Tokenization, embeddings, attention, transformer blocks, and training objectives.
plannedGenerative Models
Autoencoders, VAEs, GANs, diffusion, flow matching, and generation as learned structure.
plannedRL Foundations
States, actions, rewards, policies, value functions, Bellman equations, Q-learning, and basic policy gradients.
plannedFrontier
These tracks are for intermediate and advanced readers. They will develop carefully because research language changes faster than the underlying ideas.
RL for LLMs
Reward modeling, preference learning, RLHF, RLAIF, DPO-style methods, and policy optimization around LLM behavior.
advancedWorld Models
Learning models of environments, planning with learned dynamics, and links to agents and reasoning.
advancedModern LLM Systems
Frameworks, APIs, retrieval, agents, evaluation, deployment, and newer ideas once they are clear enough to teach.
futureWhat you build
Projects make ideas inspectable. They are not proof that a reader can call a library function. They show the computation behind the call.
Visualizations and small tools
Plots, array inspections, and tiny utilities that make quantities visible.
Numerical building blocks
Small implementations of linear algebra, gradients, losses, and optimization ideas.
Training loops
Readable experiments that show models learning, failing, and improving.
Language model pieces
Tokenizers, embeddings, attention blocks, and sequence-modeling experiments.