Open curriculum · No sign-up · No paywall

Learn how modern AI
actually works.

A structured path from linear algebra to production agents. Every concept is written to be read once and understood, then drilled with flashcards until it sticks, then extended with a deep dive that shows how it plays out in real systems.

The map

Eight domains, ordered from first principles up.

The order is deliberate. Domain 01 assumes nothing; domain 08 assumes the rest. You do not have to follow it, but if you are starting cold, start at the top.

01

Foundations

The mathematics and neural-network mechanics everything else assumes.

Mathematical FoundationsTensors & Neural PlumbingDeep Learning Building BlocksInformation Theory for Language
35 concepts 170 cards 4.6h reading
02

Transformer Internals

Open the box. How a language model actually turns text into predictions.

TokenisationEmbeddings & RepresentationsAttention InternalsPositional EncodingTransformer AnatomyTraining ObjectivesDecoding & GenerationContext & In-Context Learning
89 concepts 408 cards 11.3h reading
03

Training & Fine-Tuning

From raw web crawl to an aligned model — data, dynamics, scale and adaptation.

Pretraining Data PipelinesSynthetic DataTraining Dynamics & ScalingDistributed TrainingParameter-Efficient Fine-TuningAlignment & Post-Training
102 concepts 687 cards 13.1h reading
04

Reinforcement Learning

Classical RL, then the specific dialect of it that post-trains language models.

RL FoundationsRL for Language Models
41 concepts 282 cards 5.3h reading
05

Inference, Systems & Hardware

Where the model meets the silicon, the memory bus and the latency budget.

Inference OptimisationAccelerator ArchitectureKernels & CompilersServing Systems
55 concepts 379 cards 7.4h reading
06

Applied LLM Engineering

Building things people use: retrieval, prompts, agents and production architecture.

Retrieval & RAGPrompt EngineeringAgents & Tool UseClaude Certified Architect
33 concepts 193 cards 4.5h reading
07

Reasoning, Evaluation & Safety

Models that think longer, the evals that measure them, and the failure modes that matter.

Reasoning ModelsEvaluation & MLOpsSafety & Alignment
31 concepts 157 cards 4.3h reading
08

Multimodal & Applications

Beyond text — vision, speech, robotics and scientific discovery.

Vision & MultimodalSpeech RecognitionSpeech SynthesisRobotics & Embodied AIAI for Science
60 concepts 395 cards 7.9h reading
Start here

If you are new, these come first.

Beginner-level concepts that assume no prior machine learning.

Transformer Anatomy

Anatomy of a Transformer Block

The exact sequence of operations inside one transformer block, from tensor shapes to parameter counts, and why every frontier model is just this same function stacked dozens of times.

8 min · 5 cards
Attention Internals

Attention as Soft Dictionary Lookup

The mental model that makes attention click before the matrix algebra does - a lookup table where the key match is a matter of degree, not an exact hit or miss.

6 min · 5 cards
Speech Recognition

Audio Features and Spectrograms

Raw audio waveforms are rarely fed directly to speech models; this concept explains how and why they are first converted into spectrogram-based representations that compress perceptual information into a learnable 2-D grid.

8 min · 7 cards
Context & In-Context Learning

Autoregressive Generation

How a language model turns next-token prediction into a paragraph, and why generation is a loop that feeds its own output back as input.

6 min · 4 cards
Tensors & Neural Plumbing

Broadcasting and Vectorisation

Broadcasting is the rule that lets a bias vector add to every row of a batch without an explicit loop, and understanding its shape-matching logic prevents the class of bugs that produce wrong answers without an error.

6 min · 5 cards
Prompt Engineering

Chain of Thought Prompting

Why telling the model to think step by step radically improves reasoning, and when it actively hurts.

5 min · 3 cards
Embeddings & Representations

Cosine Similarity vs Dot Product

Two near-identical looking formulas that answer different questions, one measures direction alone, the other measures direction and magnitude together, and picking the wrong one silently breaks a search or ranking system.

6 min · 4 cards
Deep Learning Building Blocks

Dropout and Modern Regularisation

Why dropout was the dominant regulariser for a decade and why modern LLM training mostly skips it in favour of letting data do the work.

6 min · 4 cards