05. Taxonomy

This document defines the topic tree that holds the nodes, the entry indexes (tasks, models, benchmarks, risks, laws), the facets, the levels, the core list, the prerequisites and paths, the crosswalks to outside classifications, and the border with algorithm, math2 and programming.

A reader comes with one of five questions:

1. "What is X?": they know the name (an acronym, a model, a paper). Search answers it (17).
2. "How do I do this?": they know the task, not the method. The task index answers it (5.3).
3. "What is in this field?": they want a map. The area tree answers it (5.2).
4. "What should I learn next?": they want an order. Prerequisites, paths and courses answer it (5.7).
5. "What is known, by whom, and since when?": they want the record. The catalogues of models, benchmarks, results, works, incidents and laws answer it (5.4), with the evidence labels of 07.

5.1 The principles

• One primary area for each node. The primary area gives the directory (04.4), the breadcrumb and the hue of the page. Other areas are secondary (areas) and list the node too. This is algorithm's rule.
• The tree is three levels. Group (A to I), top area, sub-area; then the node. The group is a heading on the map, not a directory level: directories are nodes/<top>/<sub>/<slug>/, as in algorithm. A finer split is a facet or a section of the sub-area page.
• Tasks are first-class. Each method, architecture and model that is used for something solves at least one task (04.7), and each task page lists them with the results that compare them (10.7). This is the AI form of algorithm's "problems are first-class".
• Ideas and artifacts are separate. The tree holds ideas (concepts, methods, architectures, theorems, findings, arguments, risks). Artifacts (models, datasets, benchmarks, software, hardware, systems, policies) also have a home area, but readers mostly reach them through the catalogues of 5.4, which cut across the tree.
• Methods cut across areas. Attention appears in NLP, vision, speech and graphs. A method has one home (deep-learning-architectures/attention-mechanisms), and the other areas link it. They do not repeat it.
• Crosswalks, not adoption. The tree is our own. Each area maps to ACM CCS 2012, the CS2023 AI knowledge area, arXiv categories, the OpenAlex topics, Wikidata and the Wikipedia outlines (5.8), so that every outside list can be compared with it. No outside scheme fits teaching, research and reference use together; OpenAlex has 4,516 topics with no hierarchy below its 252 subfields, and CCS 2012 has no node for transformers (01.3).

5.2 The area tree

topics/areas.yaml defines the tree. Each top area and sub-area is also an area node with a page. Each sub-area slug is unique over the global namespace (04.3).

The first draft has 9 groups, 52 top areas and 615 sub-areas. The sizes are Estimates at M5 (18.8) from the inventory of 01.4.

A. Foundations

Top area	Sub-areas (first draft)	Crosswalk
ai-foundations	definitions-of-ai, agents-and-environments, rationality-and-bounded-rationality, decision-theory-for-ai, problem-formulation, paradigms-of-ai, measuring-intelligence, general-intelligence-concepts, computability-and-limits-of-ai	CCS I.2.0; CS2023 AI-Introduction; AIMA ch. 1 to 2
learning-theory	statistical-learning-framework, pac-learning, vc-dimension-and-growth, rademacher-and-covering, generalization-bounds, pac-bayes, stability-and-generalization, online-learning-theory, learning-in-the-limit, computational-hardness-of-learning, no-free-lunch-and-inductive-bias, deep-learning-theory, kernel-regime-and-ntk, implicit-bias-and-overparameterization	CCS "Theory of computation › Machine learning theory"; arXiv stat.ML, cs.LG
math-for-ai	linear-algebra-for-ml, matrix-calculus-and-autodiff, probability-for-ml, statistics-for-ml, information-theory-for-ml, high-dimensional-geometry, concentration-and-random-matrices, numerical-methods-for-ml, graphs-for-ml, functional-analysis-for-ml	mostly bridge cards to math2 (5.9); MML book table of contents
optimization-for-learning	empirical-risk-minimization, gradient-descent-and-variants, stochastic-gradient-methods, momentum-and-acceleration, adaptive-optimizers, second-order-and-quasi-newton, constrained-and-proximal-methods, nonconvex-optimization-theory, learning-rate-schedules, large-batch-and-distributed-optimization, zeroth-order-and-black-box, hyperparameter-optimization, bilevel-and-meta-optimization	arXiv math.OC ∩ cs.LG; bridge to algorithm optimization/gradient-methods

B. Symbolic and classical AI

Top area	Sub-areas (first draft)	Crosswalk
search-and-evolutionary-computation	state-space-search, uninformed-search, informed-search-and-heuristics, local-search, metaheuristics, genetic-algorithms, genetic-programming, evolution-strategies-and-cma, swarm-intelligence, neuroevolution, quality-diversity-and-open-endedness, artificial-life	CCS I.2.8; arXiv cs.NE; AIMA ch. 3 to 4; algorithm games-and-search bridges
constraints-and-satisfiability	constraint-satisfaction-problems, constraint-propagation, constraint-programming, sat-in-ai, smt-and-optimization-modulo-theories, distributed-constraints, constraint-learning	AIMA ch. 6; algorithm logic-and-constraints bridges
planning-and-scheduling	classical-planning, planning-languages, heuristic-planning, hierarchical-task-networks, temporal-and-resource-planning, planning-under-uncertainty, epistemic-and-multi-agent-planning, scheduling-in-ai, plan-recognition, generalized-planning, planning-with-language-models	CCS I.2.8; ICAPS topics; Ghallab-Nau-Traverso
knowledge-representation	propositional-and-first-order-logic-for-ai, description-logics-and-ontologies, semantic-web, knowledge-graphs, frames-and-semantic-networks, commonsense-knowledge, nonmonotonic-reasoning, belief-revision, reasoning-about-action-and-change, spatial-and-temporal-reasoning, answer-set-programming, argumentation-frameworks, neuro-symbolic-ai	CCS I.2.4; KR conference topics
automated-reasoning	automated-theorem-proving, interactive-theorem-proving-and-ai, logic-programming, inductive-logic-programming, program-synthesis, neural-theorem-proving, ai-for-formal-verification	CCS F.4.1, I.2.3; arXiv cs.LO ∩ cs.AI
probabilistic-reasoning	uncertainty-and-probability-in-ai, bayesian-networks, markov-random-fields, exact-inference, sampling-based-inference, variational-inference, temporal-probabilistic-models, filtering-and-tracking, probabilistic-programming, decision-networks-and-value-of-information, dempster-shafer-and-other-calculi	AIMA ch. 12 to 16; Koller-Friedman; algorithm machine-learning/probabilistic-inference bridges
classical-ai-systems	expert-systems, rule-based-production-systems, case-based-reasoning, fuzzy-logic-and-soft-computing, cognitive-architectures, blackboard-and-agent-architectures, natural-language-systems-before-statistics	CCS I.2.1, I.2.5; history sources (13.2)
games-and-multi-agent-systems	game-theory-for-ai, adversarial-search-in-ai, monte-carlo-tree-search, imperfect-information-games, general-game-playing, game-ai-milestones, agent-architectures-and-bdi, coordination-and-cooperation, negotiation-and-bargaining, mechanism-design-and-auctions, computational-social-choice, agent-communication, emergent-communication	CCS I.2.11; arXiv cs.MA, cs.GT; AAMAS topics

C. Machine learning

Top area	Sub-areas (first draft)	Crosswalk
supervised-learning	learning-problem-setup, linear-regression, logistic-regression-and-glms, nearest-neighbours, decision-trees, naive-bayes-and-generative-classifiers, support-vector-machines, kernel-methods, ensembles-and-bagging, boosting, gradient-boosted-trees, model-selection-and-validation, classical-regularization, feature-engineering, imbalanced-and-cost-sensitive, structured-prediction, learning-to-rank-classical, tabular-learning	CCS I.2.6, I.5; ESL 2e; algorithm machine-learning/supervised-learning bridges
unsupervised-learning	clustering-in-practice, dimensionality-reduction, manifold-learning, density-estimation, matrix-factorization, topic-models, anomaly-detection, independent-component-and-source-separation, classical-representation-learning	algorithm clustering, dimensionality-reduction bridges
probabilistic-machine-learning	bayesian-inference-for-ml, latent-variable-models, mixture-models-and-em, gaussian-processes, bayesian-nonparametrics, bayesian-neural-networks, uncertainty-quantification, conformal-prediction, approximate-bayesian-computation	Murphy PML 1 and 2; Bishop PRML
learning-settings	semi-supervised-learning, weak-supervision, positive-unlabeled-learning, active-learning, transfer-learning, domain-adaptation, multi-task-learning, meta-learning, few-shot-and-zero-shot, continual-learning, federated-learning, online-and-streaming-learning, curriculum-learning, learning-with-noisy-labels	arXiv cs.LG; NeurIPS subject areas
time-series-and-sequential-data	forecasting-classical, deep-forecasting, time-series-classification, sequence-anomaly-detection, spatio-temporal-learning, time-series-foundation-models, event-sequences-and-point-processes	arXiv stat.ML; Monash and other archives (13.2)
reinforcement-learning	markov-decision-processes, dynamic-programming-in-rl, monte-carlo-and-td-learning, function-approximation-in-rl, deep-q-learning, policy-gradient-methods, actor-critic-methods, model-based-rl, exploration, multi-armed-bandits, contextual-bandits, offline-rl, imitation-learning, inverse-rl-and-reward-learning, hierarchical-rl, multi-agent-rl, partially-observable-rl, rl-theory, reward-design-and-shaping, rl-environments-and-benchmarks	Sutton-Barto 2e chapters; CCS I.2.6; algorithm machine-learning/reinforcement-learning bridges
causal-inference-and-discovery	causal-graphical-models, structural-causal-models, identification-and-do-calculus, potential-outcomes, treatment-effect-estimation, causal-discovery, counterfactual-reasoning, causal-representation-learning, causality-and-invariance-in-ml	Pearl; Peters-Janzing-Schölkopf

D. Deep learning and foundation models

Top area	Sub-areas (first draft)	Crosswalk
deep-learning-foundations	neural-network-basics, activation-functions, loss-functions, backpropagation-and-autodiff, initialization, expressivity-and-approximation, depth-and-width, inductive-biases-of-networks, biological-inspiration	Goodfellow ch. 6; Prince UDL ch. 1 to 7; D2L
deep-learning-training	regularization, normalization, data-augmentation, training-dynamics, batch-size-and-learning-rate, mixed-precision-and-numerics, training-stability, loss-landscapes, grokking-and-double-descent, hyperparameter-transfer, training-recipes	Goodfellow ch. 7 to 8; Prince UDL ch. 8 to 9
deep-learning-architectures	feedforward-networks, convolutional-networks, recurrent-networks, residual-and-skip-connections, attention-mechanisms, transformers, position-encodings, state-space-and-linear-recurrent-models, mixture-of-experts, graph-neural-networks, equivariant-and-geometric-networks, memory-augmented-networks, neural-odes-and-implicit-models, hypernetworks-and-dynamic-networks, hybrid-architectures	Goodfellow ch. 9 to 10; Bronstein et al. GDL; arXiv cs.LG, cs.NE
generative-models	autoencoders, variational-autoencoders, generative-adversarial-networks, normalizing-flows, diffusion-models, flow-matching, autoregressive-models, energy-based-models, discrete-and-masked-generative-models, latent-generative-models, guidance-and-conditioning, consistency-and-distillation-of-samplers	Prince UDL ch. 14 to 18; Bishop-Bishop 2024
self-supervised-learning	pretext-tasks, contrastive-learning, masked-modelling, self-distillation, joint-embedding-predictive-architectures, embeddings-and-metric-learning, word-and-sentence-embeddings	arXiv cs.CV, cs.LG
language-models	statistical-language-models, neural-language-models, tokenization, pretraining-objectives, encoder-models, decoder-only-models, encoder-decoder-models, scaling-laws, in-context-learning, long-context, decoding-and-sampling, emergent-and-surprising-behaviour, knowledge-and-memorization-in-lms, hallucination, language-model-evaluation	SLP3 ch. 3, 7 to 12; arXiv cs.CL
post-training	supervised-fine-tuning, instruction-tuning, reward-modelling, rlhf, preference-optimization, ai-feedback-and-constitutional-methods, parameter-efficient-fine-tuning, model-merging, safety-training-and-refusals, system-prompts-and-personas, continued-pretraining-and-domain-adaptation	arXiv cs.CL, cs.LG
reasoning-and-test-time-compute	prompting, chain-of-thought, self-consistency-and-voting, search-at-inference, reinforcement-learning-with-verifiable-rewards, process-and-outcome-reward-models, test-time-scaling, reasoning-faithfulness, mathematical-reasoning-in-models, planning-in-language-models	arXiv cs.CL, cs.AI
multimodal-models	vision-language-pretraining, vision-language-models, text-to-image, text-to-video, text-to-3d, audio-language-models, any-to-any-models, multimodal-grounding, multimodal-evaluation	arXiv cs.CV ∩ cs.CL
efficient-ai	quantization, pruning-and-sparsity, knowledge-distillation, efficient-attention, kv-cache-and-memory-reduction, speculative-and-parallel-decoding, efficient-architectures-for-devices, neural-architecture-search, early-exit-and-adaptive-compute, energy-and-carbon-of-ai	arXiv cs.LG, cs.AR
foundation-models	model-families-and-lineage, frontier-closed-models, open-weight-llms, small-language-models, code-models, vision-foundation-models, speech-foundation-models, scientific-foundation-models, multimodal-model-families, model-releases-and-cards, open-and-closed-access	Epoch AI, the AI Index, the Hub (13.2); the home of model nodes

E. Perception, language and action

Top area	Sub-areas (first draft)	Crosswalk
natural-language-processing	text-preprocessing, text-classification, sequence-labelling, syntactic-parsing, semantic-parsing, lexical-and-compositional-semantics, information-extraction, entity-linking, question-answering, machine-translation, summarization, dialogue-systems, natural-language-generation, sentiment-and-affect, discourse-and-pragmatics, multilingual-and-low-resource-nlp, computational-linguistics-foundations	CCS I.2.7; arXiv cs.CL; ACL Anthology areas; SLP3
speech-and-audio	speech-recognition, speech-synthesis, speaker-recognition-and-diarization, spoken-language-understanding, speech-translation, audio-classification-and-tagging, audio-generation, music-information-retrieval, music-generation, source-separation-and-enhancement	arXiv eess.AS, cs.SD; Interspeech areas
computer-vision	image-formation-and-features, image-classification, object-detection, semantic-and-instance-segmentation, open-vocabulary-and-promptable-vision, 3d-reconstruction-and-geometry, neural-rendering, video-understanding, tracking, human-pose-and-motion, faces-and-biometrics, ocr-and-document-understanding, low-level-vision, image-generation-and-editing, visual-reasoning, remote-sensing	CCS I.4, I.5; arXiv cs.CV; Szeliski 2e; CVF topics
robotics-and-embodied-ai	robot-perception, localization-and-slam, motion-planning-in-robotics, manipulation, locomotion, robot-learning, sim-to-real, vision-language-action-models, human-robot-interaction, autonomous-driving, aerial-and-field-robotics, embodied-ai-simulators	CCS I.2.9; arXiv cs.RO; Thrun et al.
ai-agents	agent-loops-and-architectures, tool-use-and-function-calling, agent-memory, agent-planning-and-reflection, multi-agent-llm-systems, web-and-computer-use-agents, coding-agents, research-agents, agent-protocols, agent-sandboxing-and-permissions, agent-evaluation	arXiv cs.AI, cs.CL, cs.MA
retrieval-and-recommendation	information-retrieval-models, sparse-retrieval, dense-retrieval, vector-indexes-in-practice, reranking, retrieval-augmented-generation, recommender-systems, collaborative-filtering, sequential-recommendation, search-and-ranking-systems	CCS H.3; arXiv cs.IR; algorithm searching/similarity-search bridges

F. Applications

Top area	Sub-areas (first draft)	Crosswalk
ai-for-science	protein-structure-and-design, drug-discovery, chemistry-and-materials, genomics-and-biology, weather-and-climate, physics-and-simulation, astronomy-and-cosmology, mathematics-and-ai, earth-and-environment, autonomous-laboratories, science-agents	arXiv q-bio, physics.comp-ph, cond-mat ∩ cs.LG; the AI Index science chapter
ai-applications	healthcare-and-medicine, medical-imaging, software-engineering-and-code, law-and-legal-practice, finance-and-trading, education, creative-arts-and-design, writing-and-media, games-and-entertainment, transportation-and-logistics, manufacturing-and-industry, energy-and-utilities, agriculture-and-food, public-sector-and-government, military-and-defence, marketing-and-commerce, customer-service, accessibility-and-assistive, cybersecurity-applications	CCS "Applied computing"; the AI Index applications chapters

G. Systems and engineering

Top area	Sub-areas (first draft)	Crosswalk
data-for-ai	web-crawls-and-corpora, data-curation-and-filtering, deduplication, data-labelling-and-annotation, synthetic-data, data-mixtures-and-curricula, data-documentation, dataset-licensing-and-consent, multimodal-datasets, evaluation-data-construction, data-attribution-and-valuation	the datasheets and data cards literature; Hub datasets (13.2)
ml-systems	training-systems, data-parallelism, tensor-and-sequence-parallelism, pipeline-parallelism, expert-parallelism, memory-optimization, compilers-and-graph-optimization, gpu-kernels, inference-serving, batching-and-scheduling, fault-tolerance-and-checkpointing, mlops-and-deployment, experiment-tracking-and-reproducibility, cluster-and-datacenter-scale	arXiv cs.DC ∩ cs.LG; MLSys topics
ai-hardware	gpus, tpus-and-custom-asics, other-accelerators, wafer-scale-and-dataflow, interconnects, memory-and-hbm, ai-datacenters-and-power, edge-and-mobile-npus, neuromorphic-and-analog, historical-ai-hardware, compute-trends-and-supply-chain	arXiv cs.AR; Epoch AI hardware data
ai-software	deep-learning-frameworks, classical-ml-libraries, model-hubs-and-weight-formats, training-libraries, serving-and-inference-engines, local-inference-runtimes, llm-application-frameworks, agent-frameworks-and-sdks, evaluation-harnesses, data-and-labelling-tools, rl-and-simulation-environments, notebooks-and-experiment-tools	the AI software that programming leaves to us (5.9); GitHub, PyPI (13.2)
evaluation	evaluation-methodology, classification-and-regression-metrics, ranking-and-retrieval-metrics, text-generation-metrics, image-and-video-generation-metrics, speech-metrics, statistical-testing-for-ml, benchmarks-and-their-design, leaderboards-and-comparability, human-evaluation-and-arenas, model-graded-evaluation, contamination-and-leakage, calibration-evaluation, capability-and-dangerous-capability-evaluations, efficiency-and-cost-evaluation	the home of metric and benchmark nodes (10)

H. Safety, ethics and governance

Top area	Sub-areas (first draft)	Crosswalk
ai-safety-and-alignment	the-alignment-problem, specification-gaming-and-reward-hacking, goal-misgeneralization, scalable-oversight, deceptive-alignment-and-scheming, sycophancy-and-honesty, power-seeking-and-instrumental-convergence, ai-control, model-organisms-of-misalignment, catastrophic-and-existential-risk, agent-foundations, cooperative-ai, corrigibility, situational-awareness, safety-cases	the Alignment Forum and safety lab research (13.2); Hendrycks
interpretability	interpretable-models, feature-attribution, explainable-ai-methods, probing, mechanistic-interpretability, circuits, superposition-and-features, sparse-autoencoders-and-dictionaries, representation-engineering-and-steering, concept-based-explanations, interpretability-evaluation	Molnar (interpretable-ml); Distill and Transformer Circuits (13.2)
robustness-security-and-privacy	adversarial-examples, certified-and-adversarial-robustness, distribution-shift-and-ood, jailbreaks, prompt-injection, data-poisoning-and-backdoors, model-extraction, membership-and-attribute-inference, differential-privacy-in-ml, machine-unlearning, watermarking-and-content-provenance, synthetic-media-and-deepfakes, red-teaming, ai-supply-chain-security, ai-threat-taxonomies	arXiv cs.CR ∩ cs.LG; MITRE ATLAS; OWASP LLM Top 10
fairness-and-ethics	ethics-of-ai, ai-principles-and-guidelines, fairness-definitions, bias-measurement, bias-mitigation, transparency-and-documentation, accountability-and-auditing, human-oversight-and-contestability, human-ai-interaction-and-trust, ai-and-human-rights, ai-and-labour-ethics, ai-welfare-and-moral-status	FAccT topics; arXiv cs.CY
ai-governance-and-law	eu-ai-law, united-states-ai-policy, china-ai-regulation, uk-ai-policy, other-national-ai-policies, international-ai-governance, ai-standards, frontier-ai-policies-and-commitments, compute-and-export-controls, copyright-and-ai, liability-and-product-safety, data-protection-and-ai, sector-regulation, ai-incidents-and-reporting, ai-safety-institutes	the policy sources of 13.2 (EUR-Lex, OECD.AI, NIST, legislation registers, AIID)

I. Context

Top area	Sub-areas (first draft)	Crosswalk
history-of-ai	precursors-and-automata, cybernetics-and-early-computing, the-birth-of-ai, early-symbolic-ai, first-ai-winter, knowledge-based-systems-era, second-ai-winter, statistical-and-probabilistic-turn, connectionism-revival, deep-learning-revolution, transformer-and-foundation-model-era, histories-by-country	Wikipedia "History of AI" and timelines; Nilsson; McCorduck; Russell-Norvig ch. 1
philosophy-and-cognitive-science	philosophy-of-ai, can-machines-think, consciousness-and-machine-minds, meaning-and-understanding, symbol-grounding, embodiment-and-situated-cognition, cognitive-science-and-ai, neuroscience-and-ai, computational-models-of-mind, theories-of-intelligence	PhilPapers "Philosophy of Artificial Intelligence" categories; SEP entries (names only)
ai-and-society	economics-of-ai, labour-and-automation, ai-industry-and-investment, ai-geopolitics, open-source-and-open-weights-movement, public-opinion-and-discourse, ai-in-culture-and-fiction, ai-forecasting-and-timelines, ai-and-education-systems, ai-and-information-ecosystem, ai-and-the-environment	the AI Index economy and policy chapters; arXiv cs.CY, econ
people-and-institutions	researchers, engineers-and-builders, labs-and-companies, universities-and-research-groups, government-bodies-and-regulators, standards-bodies, funders, conferences-and-venues, awards-and-prizes, competitions-and-challenges, landmark-works, notable-events	the home of person, organization, event and canon work nodes

Rules:

• A top area or sub-area is a change to topics/areas.yaml in its own pull request, with a note on what it holds and what it does not, as in algorithm.
• A sub-area with more than 200 nodes splits, and one with fewer than 5 nodes after M3 merges into a neighbour (rule T09 reports both). The limit is 200, not algorithm's 150, because many AI sub-areas hold model and benchmark nodes that the reader reaches through the catalogues.
• Each sub-area slug is reviewed against the crosswalks at each release: every arXiv category of 5.8, every OpenAlex topic inside subfield 1702 with more than 2,000 works, and every Wikipedia "Outline of artificial intelligence" heading must map to some area (18.4).
• The tree is a draft. M0 and M1 test it against the inventory (13.3): the curate step of 14.3 places 2,000 random inventory items, and any item with no good home is a tree gap.

5.3 The task index

The task index is the "how do I do this" entrance, as algorithm's problem index is. It lists the task nodes by area, each with a one-line description in plain words, its input and output, its main metrics, its main benchmarks, and the methods and models used for it, with dated results:

Computer vision › Object detection
  2D object detection            image → boxes + labels      mAP@[.5:.95]   COCO 2017, LVIS v1        two-stage detectors, one-stage detectors, DETR family
  Open-vocabulary detection      image + text → boxes        AP (novel)     LVIS v1 (rare), ODinW     grounded vision-language detectors
  3D object detection (LiDAR)    point cloud → 3D boxes      mAP, NDS       nuScenes, Waymo Open      voxel, pillar and point methods
Natural language processing › Question answering
  Extractive QA                  question + passage → span   EM, F1         SQuAD 1.1, SQuAD 2.0       encoder models, span heads
  Open-domain QA                 question → answer           EM             Natural Questions, TriviaQA retrievers + readers, RAG, LLMs

Each task node also has facet keys that the "I need to…" search (17.6) filters on: modality (text, image, video, audio, speech, tabular, graph, time series, 3D, multimodal, action), output_type (label, labels, span, sequence, box, mask, scalar, ranking, policy, program, structure), supervision (labelled, unlabelled, weak, reward, preference, none), setting (offline, online, interactive, embodied), and compute_hint (a rough class of the compute that common methods need, from the results register).

The task list starts from the inventory: the Papers with Code task list in its final data dump (about 4,000 tasks, many duplicates; 01.2), the Hugging Face Hub task taxonomy, the ACL and CVF topic lists, the Kaggle competition categories, and the canon. The curate step merges duplicates and gives each surviving task a place in the tree (14.3).

5.4 The catalogues

Readers reach artifacts mostly through catalogues, not the tree. Each catalogue is a collection page over a kind, with facets (5.5), computed tables, a timeline and a dataset download (16.6). The site routes of 04.3 belong to them.

Catalogue	Kind	Route	Default sort	Key facets
Models	model	/models/	release date, newest first	developer, country, architecture family, parameters (log buckets), training compute (log buckets), modality, access, licence, notable (18.5)
Benchmarks	benchmark	/benchmarks/	task area, then introduction date	task, modality, metric, saturated (10.8), contamination reports, licence
Datasets	dataset	/datasets/	area, then size	modality, size, source type, licence, documentation present
Results	result	/results/ and each benchmark version	per benchmark version, score	model, setting, evidence (reported, measured here, replicated), date
Works	work	/works/	year	venue, area, canon, citations (computed, dated)
Systems	system	/systems/	year	domain, developer, status
Software	software	/software/	area	category, language, licence
Hardware	hardware	/hardware/	release date	vendor, type, peak FLOP/s, memory
Risks	risk	/risks/	area	source taxonomy, affected part, incidents (count)
Incidents	incident	/incidents/	date, newest first	risk, sector, harm type, country, system
Laws and policies	policy	/policy/	jurisdiction, date	jurisdiction, type, status (computed), topic
Standards	standard	/standards/	body	body, status, topic
People	person	/people/	name	area, era, roles, awards
Organizations	organization	/orgs/	name	kind, country, area
Events	event	/timeline/	date	type, area
Glossary	definition, and the lead sentence of every concept	/glossary/	alphabetical	area, level

5.5 Facets

Every list page (area, task, catalogue, search results) can filter and sort on these facets. They come from the front matter, the cards and the build, never from a separate index.

Facet	Values	From
kind	the 28 kinds of 04.2	kind
group, area	the tree	area, areas
task	task slugs	solves, task
modality	the values of 5.3	modality, inputs, outputs
paradigm	symbolic, statistical, connectionist, probabilistic, evolutionary, hybrid	topics/paradigms.yaml, set per area and overridden per node
learning signal	supervised, unsupervised, self-supervised, reinforcement, preference, none	front matter
level	intro, core, advanced, research	level (5.6)
era	decade of history.year, and the named eras of history-of-ai	history
evidence	proved (F2+ or derivation), measured here (R2+), replicated, reported, contested	the claim labels of 07, at the node's main claim
implementation	frameworks with X2 or better	build
scale	parameters and training compute of models, in log buckets	facts
access and licence	open weights, open source, API, closed; SPDX	facts
people, organizations	slugs	people, orgs, developer
jurisdiction	ISO 3166 codes, eu, intl	policy nodes and cards

5.6 Levels

Level	Who it is for	Example
intro	a reader with school mathematics and some programming, in a first AI or ML course	agents, search, linear regression, gradient descent, overfitting, a neural network, a confusion matrix, the Turing test
core	a student who has finished an undergraduate AI and ML sequence, or a working ML engineer	backpropagation, SVMs, EM, Q-learning, CNNs, the Transformer, BERT, LoRA, BLEU, Bayesian networks, PPO, RLHF
advanced	a graduate student or specialist	PAC-Bayes, normalizing flows, the neural tangent kernel, ZeRO, FlashAttention, mechanistic circuits, do-calculus identification
research	a researcher at the edge of a field, or an item less than two years old whose standing is not settled	a method from a recent conference with one replication; an open problem; a new benchmark

level is an editorial field, set by the curate step and checked by the audit (rule T07 samples it). An item younger than two years starts at research unless the canon (5.10) already teaches it.

5.7 Prerequisites, paths and courses

• Prerequisites. prereqs edges form a DAG (rule E03). Each page shows a "Read first" strip of its direct prerequisites, and a full reading path on request, as in algorithm and math2. A core node must have a prereqs path back to ai-foundations, math-for-ai or deep-learning-foundations (rule E09). Prerequisites may point to bridge cards (5.9), so a path can cross into math2 for the linear algebra and come back.
• Order. As in algorithm 05.4: sub-areas and items are listed in a reading order, from the prereqs DAG first, then by level, then by a hand-set order field.
• Paths. A path is the topological order of the prerequisites of one target node. The site computes it (16.12). Readers can ask for "from zero to the Transformer" or "from zero to PPO".
• Courses. A course is a curated, ordered list of nodes with section headings, in topics/courses/<slug>.yaml. The first courses, each cross-checked against public syllabi (names and orders only, 19.3):

Course	Basis	Size (Estimate)
Introduction to AI	the chapter order of AIMA 4e, cross-checked with CS2023 AI and with public "Introduction to AI" syllabi	250 nodes
Machine learning	the order of a standard ML course, cross-checked with ESL 2e, Bishop PRML and Murphy PML 1	250 nodes
Deep learning	the order of Prince UDL and D2L, cross-checked with public deep learning syllabi	220 nodes
Language models from scratch	tokenization to pretraining to post-training to evaluation, with the from-scratch code of 08 at every step	150 nodes
Reinforcement learning	the chapter order of Sutton-Barto 2e, then deep RL	160 nodes
Computer vision	Szeliski 2e and deep vision	180 nodes
NLP	SLP3 chapter order	180 nodes
Probabilistic graphical models	Koller-Friedman order	120 nodes
Learning theory	Shalev-Shwartz and Ben-David, then Mohri et al.	100 nodes
AI safety and alignment	the order of Hendrycks's textbook, then the alignment sub-areas	140 nodes
AI policy and law	the governance sub-areas, with the provisions that a reader needs	100 nodes
History of AI	the eras of history-of-ai, in time order	120 nodes
The proved core	the nodes with F2, in an order for reading the Lean proofs (11)	80 nodes

A course page shows progress (stored in the reader's browser only, as in programming Q-15), and each node in a course shows a step strip with the previous and the next node.

5.8 Crosswalks

The build keeps crosswalks to the classifications that readers, libraries and other sites use. Each crosswalk is a register; each row maps one outside id to one or more of our tags, with the kind of match (exact, broader, narrower, related), as in programming 5.7.

Crosswalk	Outside system	Licence of the outside ids	Used for
crosswalk/acm-ccs-2012.yaml	ACM CCS 2012, "Computing methodologies › Artificial intelligence" and "› Machine learning", and the AI parts of "Theory of computation" and "Applied computing"	free to use with credit	linking papers to areas; the coverage check of 18.4
crosswalk/cs2023-ai.yaml	the CS2023 Artificial Intelligence knowledge area (AI) and its knowledge units, and the AI parts of the SEP and SEC areas	names of a public curriculum, cited	the coverage report and the learning paths
crosswalk/arxiv.yaml	arXiv categories cs.AI, cs.LG, cs.CL, cs.CV, cs.NE, cs.RO, cs.MA, cs.IR, cs.SD, cs.CY, cs.CR (AI part), stat.ML, eess.AS, eess.IV	arXiv metadata CC0	the daily job (14.9) and the harvest priority
crosswalk/openalex.yaml	OpenAlex domain 3, field 17, subfield 1702 (Artificial Intelligence) and its topics, plus the AI-related topics of subfields 1707 (computer vision and pattern recognition), 1710 and others	CC0	the work universe, the "missing" signal (18.6), citation counts
crosswalk/wikidata.yaml	Wikidata QIDs for every item that has one	CC0	same_as, people, organizations, models, software
crosswalk/wikipedia.yaml	English Wikipedia articles in the "Outline of artificial intelligence", "Outline of machine learning", "Glossary of artificial intelligence", "Timeline of artificial intelligence" and the AI and ML category trees	titles only (CC BY-SA for text, which we never copy)	the concept universe (18.3)
crosswalk/hf-tasks.yaml	the Hugging Face Hub task taxonomy and its pipeline tags	names cited	task cards and the model scope rule
crosswalk/pwc.yaml	the tasks, methods, datasets and areas of the final Papers with Code data dump	CC BY-SA 4.0 for the dump; we keep ids and names	the task and method universes; the historical results (10.6)
crosswalk/epoch.yaml	Epoch AI model, hardware and benchmark ids	CC BY 4.0	model, hardware and result cards
crosswalk/aiid.yaml	AI Incident Database incident ids	the AIID terms, ids as facts	incident cards
crosswalk/mit-ai-risk.yaml	the domains and subdomains of the MIT AI Risk Repository	CC BY 4.0	risk nodes
crosswalk/iso-22989.yaml	the terms of ISO/IEC 22989 (AI concepts and terminology)	names only; the standard's text is not free (19.3)	the glossary check
crosswalk/nist-ai-100.yaml	the NIST AI terms and the AI RMF functions and categories	public domain	policy provisions and the glossary
crosswalk/philpapers.yaml	the PhilPapers categories under "Philosophy of Artificial Intelligence"	names cited	the philosophy sub-areas
crosswalk/algorithm.yaml, crosswalk/math2.yaml, crosswalk/programming.yaml	the tags of our neighbours	CC0, our own	the border (5.9) and bridge cards

Rules:

• The crosswalks drive the "no missing" checks of 18.4. A row of an outside list that maps to no item is a gap, and the universe register (18.3) takes it in with an outcome.
• An outside system's text is never copied into a node. Only ids and names are kept, with the source (19).

5.9 The border with algorithm, math2 and programming

The four sites are one library with four wings. The border rules:

Question	Goes to	Example
Is it a procedure with an input and output contract whose correctness and cost can be stated and proved without a learning experiment?	algorithm	Lloyd's algorithm, the junction tree algorithm, alpha-beta pruning, A*, DPLL, CDCL, Viterbi, the simplex method, reverse-mode automatic differentiation, HNSW construction and search
Is it judged mainly by its empirical results when it learns from data, or is it a model, a trained artifact, a dataset, a benchmark, an AI system, a risk, a law, or an idea about intelligence?	ai	dropout, Adam as a training method, LoRA, the Transformer, BERT, ImageNet, MMLU, AlphaGo, reward hacking, the EU AI Act, the Chinese room
Is it a mathematical object or theorem that AI uses but that does not belong to AI?	math2	eigendecomposition, the KL divergence as mathematics, Jensen's inequality, Markov chains, convex duality
Is it a general language, library or tool, or how programmers work?	programming	Python, NumPy as a general library, Docker, Git, CUDA C++ as a language

• Items on two sides get two pages, each about its own side, and they link each other with same_as.al, same_as.m2 or same_as.pg. Examples: "k-means" on algorithm (Lloyd's algorithm, its convergence, its cost, k-means++ seeding and its O(log k) guarantee) and "k-means clustering" here (when it is used, how to choose k in practice, its empirical record, its role in vector quantization and in tokenizers); "backpropagation" on algorithm (reverse-mode differentiation of a computation graph, its cost) and here (its history from Linnainmaa to Rumelhart, Hinton and Williams, vanishing gradients, its role in training, its biological plausibility debate); "attention" here only.
• The test for a second page is that the AI view adds at least three facts with sources that the algorithm page does not have and should not have (rule T10). Otherwise the item is a bridge card (04.1), and the universe register records the outcome elsewhere:algorithm with the target tag.
• Algorithm's existing ML, search and logic nodes. Algorithm has 419 nodes in machine-learning, 209 in games-and-search and 325 in logic-and-constraints (01.6). At M0, aik bridge sync al makes a bridge card for each of them from algorithm's .json twins, with its tag, slug, name, summary and URL. The curate step then decides for each one whether this site needs its own node (by the T10 test). The Estimate is that about 250 of the 419 ML nodes get a second page here, and that most of the search and logic nodes stay bridge cards.
• Optimizers. SGD, momentum, Adam and their variants are on both sides. Algorithm owns them as optimization algorithms with convergence analysis. This site owns them as training methods: defaults, interactions with normalization and batch size, the empirical record, the variants that the literature compares. The two pages link each other.
• AI software. PyTorch, JAX, TensorFlow, scikit-learn, Hugging Face Transformers, vLLM, llama.cpp, DeepSpeed, LangChain and other AI software belong here (ai-software). Programming keeps a card for each with same_as.ai. NumPy and pandas belong to programming, and this site keeps bridge cards.
• Mathematics. math-for-ai holds bridge cards to math2 for the mathematics, and nodes only for the AI-specific views that math2 does not have (the reparameterization trick, the log-sum-exp trick in practice, matrix calculus conventions of ML).
• Bridge cards are refreshed at each release of the neighbour (20.6), and a neighbour's rename becomes an alias here.

5.10 The core list

The core is the set of items that must be complete before the site is called "complete for the core". It answers "what does every student and practitioner of AI need?" It is the measure for M2 and M3 (21). The rule is the one of algorithm: an item is core when at least two canon works name it (the canon match step, 14.3).

The canon for the core rule (13.3):

Work	Covers
Russell and Norvig, Artificial Intelligence: A Modern Approach, 4th ed. (AIMA 4e)	the whole field, symbolic and statistical
Goodfellow, Bengio and Courville, Deep Learning	deep learning foundations
Bishop, Pattern Recognition and Machine Learning	probabilistic ML
Bishop and Bishop, Deep Learning: Foundations and Concepts (2024)	modern deep learning
Murphy, Probabilistic Machine Learning: An Introduction and Advanced Topics	ML, both volumes
Sutton and Barto, Reinforcement Learning: An Introduction, 2nd ed.	RL
Hastie, Tibshirani and Friedman, The Elements of Statistical Learning, 2nd ed.	statistical learning
Shalev-Shwartz and Ben-David, Understanding Machine Learning	learning theory
Mohri, Rostamizadeh and Talwalkar, Foundations of Machine Learning, 2nd ed.	learning theory
Prince, Understanding Deep Learning	deep learning
Zhang, Lipton, Li and Smola, Dive into Deep Learning	deep learning with code
Jurafsky and Martin, Speech and Language Processing, 3rd ed. draft	NLP and speech
Szeliski, Computer Vision: Algorithms and Applications, 2nd ed.	vision
Koller and Friedman, Probabilistic Graphical Models	graphical models
MacKay, Information Theory, Inference, and Learning Algorithms	inference and information
Thrun, Burgard and Fox, Probabilistic Robotics	robotics
Pearl, Causality, 2nd ed.	causal inference
Ghallab, Nau and Traverso, Automated Planning and Acting	planning
Hendrycks, Introduction to AI Safety, Ethics, and Society	safety, ethics, governance

Rules:

• The canon is a register (registers/canon.yaml) with each work's table of contents and index terms, entered by hand or extracted from public tables of contents (names only, 19.3). It changes by pull request with a reason.
• The canon covers the stable core. It does not cover the last two years. So the core list has a second part, the frontier core: items that the AI Index of the current year, at least two of the survey works of registers/surveys.yaml, or at least two of the courses of 5.7 name. The frontier core is recomputed each year, and the page of an item shows why it is core.
• Core sizes (Estimate): about 2,400 concepts, methods, architectures, theorems and tasks from the canon; about 600 from the frontier core; about 300 notable models, 200 benchmarks and 150 datasets by the scope rules of 18.5; the core policies (the EU AI Act and the Digital Omnibus on AI with their main provisions, the NIST AI RMF, the main national acts and executive acts, the Council of Europe convention, the main frontier commitments); the 80 people and 60 organizations that the canon names most often. The total is about 4,000 items.
• The core is computed and stored in registers/core.yaml with the reason for each item, as in algorithm. A person approves it at the end of M1.

5.11 Names and disambiguation

• One English name. Shared names get qualified slugs and a disambiguation entry (04.3).
• Eponyms use a hyphen in this project's style (no en dash, 06.2): "Bellman equation", "Kullback-Leibler divergence", "Nadaraya-Watson estimator". Slugs drop accents and use hyphens.
• Acronyms are aliases, and a search for one goes straight to its page when it is unique (rlhf, dpo, vae, gan, ppo, mcts, rag). When an acronym is ambiguous (ppo, sae, moe are unique; gp, rl, ner, lm are mostly unique; ica, pca, svm are unique; ssm, ar are not), search shows the candidates.
• Model names keep the developer's spelling and capitals (GPT-4o, Llama 3.1, Gemini 2.5 Pro, DeepSeek-R1, Qwen2.5), and the slug is the lower-case form with hyphens (gpt-4o, llama-3-1, gemini-2-5-pro, deepseek-r1, qwen2-5).
• Paper titles are the titles of the works, and slugs are short forms with the first author and year when the title is generic (vaswani-2017-attention, with the alias "Attention Is All You Need").

Source of this page