✳ GENESIS THE AI ATLAS
An interactive learning experience
FROM FIRST PRINCIPLES TO GENERATIVE SYSTEMS

Inside AI.

Understand the basics. Follow the data. Explore what comes next.

Training turns examples into learned weights. Inference turns a prompt into new tokens.
Data & representationsLearningGeneration
DRAG TO ORBIT · SCROLL TO ZOOM · CLICK TO EXPLORE
GO DEEPER

The reading room

Start with AI, machine learning, neural networks, deep learning and matrix multiplication. World models connect prediction to action and observation. The generative journey explains a typical decoder-only language model. Additional chapters cover post-training, application systems and other generative model families. Every scene is a conceptual 3D map, not a physical layout or a replica of a specific frontier model.

What is machine learning?Google introductory course: examples and learning signals ↗Neural networksGoogle: layers, weights and nonlinear activations ↗Linear algebra for deep learningGoodfellow, Bengio and Courville: vectors and matrix products ↗World ModelsHa and Schmidhuber: learned environment representations and dynamics ↗Attention Is All You Need Original Transformer paper ↗Training a causal language model Hugging Face LLM Course ↗Parallelism strategies NVIDIA Megatron Core ↗Learning from human feedbackInstructGPT research · RLHF ↗Direct Preference OptimizationPreference optimization without an online RL loop ↗DeepSeek-R1An example of reinforcement learning for reasoning ↗Context engineeringManaging information across agent turns ↗Building effective agentsWorkflows, tools, feedback and stopping conditions ↗ReActInterleaving reasoning and environment actions ↗Workflow graphs and agentsExecution paths and state ↗Retrieval-augmented generationRetrieval as evidence for generation ↗Microsoft GraphRAGKnowledge graphs and retrieval ↗Mixtral of ExpertsSparse expert routing ↗GPT-4 technical reportA documented example of image-and-text inputs ↗Denoising diffusion probabilistic modelsTraining with noise and iterative sampling ↗Context windowsLimits, input and output tokens ↗Prompt engineeringTasks, success criteria and examples ↗Application memoryState and external persistence ↗Agent evaluationsOutcome and trajectory checks ↗

Terminology: “loop engineering” is treated here as agent-loop design. “Graph engineering” is split into workflow graphs and knowledge graphs. These are not presented as universal names for new model architectures.

All token IDs, vectors, probabilities, hardware layouts and learning curves in the experiments are illustrative. The gradient experiment computes real updates for a tiny mathematical model. The foundations add exact one-parameter fitting, a hand-built feature circuit, matrix arithmetic and a deterministic cart simulation. No frontier model is running here.