AI Atlas
From first principles to production systems.
An executable map of intelligent systems: from vectors, probability, search, objectives, gradients, and learned representations to inference, retrieval, agents, evaluation, safety, and operations.
AI is observable computation, not a magic box.
Each note asks what is represented, computed, optimized, learned, measured, hidden by frameworks, and changed by production constraints.
Choose a route; keep the computation visible.
Every route names prerequisites and crosses theory, implementation, evidence, failure modes, and production behavior.
First Principles
Math → probability → optimization → autodiff → neural networks → transformers
02AI Engineer
ML → deep learning → data → training → inference → evaluation → MLOps
03LLM Systems
Transformers → language models → context → retrieval → agents → evals → security
04Research Literacy
Math → papers → reproduction → experiments → interpretation → evidence
05Product & Production
Problem framing → data → baseline → evals → product → deployment → monitoring
Glassbox AI Lab
Build the stack progressively: scalars and probability → autodiff → models → runtimes → retrieval → agents → production evidence.
Numbers, vectors & probability
Autodiff from scratch
Neural network from scratch
Tensor framework & image model
Tokenizer & mini-transformer
Training & adaptation
Inference runtime
Retrieval system
Tool-using agent
Multimodal pipeline
Production AI system
Depth and editorial state
Orientation
How to use the atlas, read its notation, validate claims, and choose a route.
Start Here — AI from First Principles to Production
How to navigate AI Atlas, read its mathematics, run Glassbox labs, validate claims, and choose a learning path.
ExploreMust Know — Twelve Rules for Thinking Clearly About AI
Twelve first-principles rules that prevent magical thinking about objectives, data, uncertainty, evaluation, deployment, agents, and oversight.
ExploreAI Index
Root map for AI from mathematical and computational foundations to models, agents, evaluation, safety, product, and production systems.
ExploreFoundations
Mathematics, computation, classical reasoning, statistical learning, data, assumptions, and uncertainty.
Mathematics for AI
Linear algebra, calculus, probability, statistics, information theory, optimization, and numerical stability.
Computation & Autodiff
Tensors as arrays, vectorization, compute graphs, reverse-mode autodiff, floating point, parallelism, and reproducibility.
Classical AI & Reasoning
Search, heuristics, constraints, planning, logic, knowledge representation, probabilistic reasoning, and decision theory.
Learning Foundations
Problem formulation, generalization, inductive bias, distribution shift, leakage, metrics, and when learning is the wrong tool.
Statistical Machine Learning
Regression, classification, trees, ensembles, kernels, clustering, dimensionality reduction, calibration, and error analysis.
Data for AI
Data-centric AI: quality, labeling, dataset design, synthetic data, contamination, documentation, and pipelines.
Learning & Models
Neural learning, sequential decisions, architectures, language, vision, audio, and multimodality.
Deep Learning
Neural networks, representation learning, optimization, embeddings, and scaling behavior.
Reinforcement Learning
Sequential decisions, Bellman equations, value learning, policy gradients, exploration, offline RL, world models, and alignment links.
Model Architectures
CNNs, sequence models, attention, transformers, mixture of experts, state-space models, diffusion, and architectural trade-offs.
Language & Foundation Models
Tokenization, decoder transformers, pretraining, data mixtures, scaling, decoding, instruction tuning, reasoning behavior, and model families.
Vision, Audio & Multimodal AI
Images, vision models, detection, segmentation, speech, multimodal fusion, diffusion, video, evaluation, data, and safety.
Training & Inference
Training systems, model adaptation, inference runtimes, hardware, latency, throughput, and cost.
Training & Adaptation
Training loops, distributed execution, mixed precision, checkpointing, SFT, LoRA/QLoRA, preference optimization, distillation, and continual learning.
Inference Systems
Model loading, kernels, memory bandwidth, quantization, prefill/decode, KV cache, batching, routing, edge inference, latency, throughput, and cost.
Context & Agency
Prompts, assembled context, retrieval, memory, tool use, agents, permissions, and recovery.
Context Engineering
Prompting techniques plus context engineering: what the model knows — system prompts, examples, structured output, memory, and assembled context.
Agents & Tools
Tool calling, planning loops, memory, handoffs, autonomy boundaries, and operator controls.
Retrieval & Knowledge
Sparse and dense retrieval, chunking, embeddings, ranking, query transformation, grounding, citations, memory, and retrieval evaluation.
Measurement & Trust
Evaluation, interpretability, safety, security, governance, uncertainty, and release evidence.
Evaluation & Measurement
Splits, metrics, calibration, uncertainty, human evaluation, statistical evidence, product evals, regressions, and error taxonomies.
Interpretability
Feature attribution, probes, activations, representations, circuits, sparse autoencoders, causal interventions, and faithfulness limits.
AI Safety & Security
Threat modeling, misuse, prompt injection, data leakage, safety cases, and deployment controls.
AI Ethics & Governance
Fairness and bias, transparency, model documentation, privacy, regulation (EU AI Act), governance frameworks, and societal impact.
Product & Operations
Product framing, MLOps, observability, monitoring, feedback, reliability, and human review.
Labs, Research & Playbooks
Glassbox labs, paper reproduction, research logs, and repeatable playbooks.
Research & Experimentation
Read papers, reconstruct claims, reproduce results, test statistical evidence, detect benchmark gaming, and maintain a research log.
AI Playbooks
Repeatable procedures for building, evaluating, debugging, and shipping AI systems.
Transformer Attention Map
A short conceptual note on attention as routing over token representations.
Registries and operating memory
Registries keep external references, decisions, datasets, tools, and evaluation assets normalized behind short atomic notes.