The taxonomy

The same 236 concepts, sorted the other way: by what kind of thingthey are, then by where they live and how they are looked at. This is the facet view of the graph — nothing here is typed by hand.

← the atlas (by job and layer)

Kinds — 29, in 6 families

A concept has exactly one kind. Kinds are closed, so the validator can reject anything that does not fit the contract.

Evidence & claims · 16

Theorem3

Channel capacity Source coding theorem No free lunch

Result3

Scaling law Halting problem Chinchilla

Heuristic5

Compression Occam's razor Distributional hypothesis Least privilege Loop policy

ThoughtExperiment1

Turing test

OpenQuestion4

The scaling debate AGI Superintelligence Open-weight risk

Formal objects · 48

Distribution2

Prior Posterior

Variable1

Feature

Objective3

Cross-entropy Loss function Reward

Estimator1

Value function

Methods · 92

Algorithm18

Bayes' ruleBayesian inferenceGradient descent Backpropagation Pretraining Fine-tuning RLHF GRPO Retrieval-augmented generation Huffman coding Error-correcting code Search Backpropagation through time word2vec Byte-pair encoding Direct preference optimisation ReAct Speculative decoding

Architecture17

Convolutional neural network Recurrent neural network LSTM Transformer Diffusion model Diffusion transformer World model Omni-model Multilayer perceptron Sequence model Language model Sparse autoencoder Reasoning model JEPA Latent diffusion ControlNet Modality encoder

Artifacts · 50

Model14

DeepSeek-V3 DeepSeek-R1 Llama Reward model DeepSeek-V4 Qwen GLM Kimi MiniMax InternLM Hunyuan Mistral OLMo Open image model

Dataset3

Training set Test set Vocabulary

Benchmark2

Benchmark Open-weight baseline

Task1

Next-token prediction

System13

Agent Agent harness Environment Expert system Vector database MCP server MCP client Orchestration Multi-agent vLLM Evaluation harness CANN Local serving

Protocol2

Model Context Protocol API

Kernel0

empty

Format4

Tool schema FP8 GGUF Quantised weights

Hardware3

GPU Tensor core Huawei Ascend

Document8

Export controls Specification Constitution Permissive licence Open-weight licence Acceptable use policy Model card EU AI Act

Actors & collectives · 12

Person0

empty

Institution0

empty

Field12

Probability Reinforcement learning Alignment Interpretability Evaluation Symbolic AI Control theory Cybernetics Mechanistic interpretability Observability Open-source AI Harness engineering

Diagnostics · 18

FailureMode8

Overfitting Reward hacking Prompt injection Underfitting Combinatorial explosion Frame problem Vanishing gradient Contamination

Phenomenon10

AI winter Noise Embedding space Vector arithmetic Emergent ability Superposition Drift Memory wall Commoditisation Barbell

Facets

Every concept also carries a volatility (how fast it goes stale) and sits at one or more layers. Bars are counts.

volatility

timeless149

decade41

frontier46

Frontier concepts carry an asOf date and must be re-checked; timeless ones do not.

mode

Represent46

Infer19

Learn43

Search & planning7

Act22

Measure39

69 concepts are artifacts with no mode of their own — a protocol, a format, a hardware part.

layer

L0 Foundations53

L1 Model63

L2 Data9

L3 Compute9

L4 Training32

L5 Serving15

L6 Systems,39

L7 Ecosystem25

overlay

Efficiency, cost & energy25

Governance, policy & ethics20

Evaluation & methodology12

Safety & alignment11

Security, privacy & robustness6

Interpretability & transparency3

Overlays cut across layers: safety, security, interpretability, evaluation, efficiency, governance.