Every concept

The graph has 236 nodes and 230 declared edges. Each one has a page — its position on the prism, what must come first, what it unlocks, how it fails, and every connection it has. The list is ordered by centrality: how many edges touch it, which is a rough measure of how much of the rest of the field leans on it.

the atlas →by kind →reading routes →

#conceptlinks

01Agent harnessSystemevery decision you can make without retraining the model — loop, context, tools, memory13

02Reinforcement learningFieldlearning what to do by trying, from delayed reward10

03AgentSystema model plus a harness, running a loop toward a goal9

04AttentionMechanismlet each token weigh every other token, instead of only the last one9

05EntropyQuantitythe average surprise of a source, in bits9

06Tool useMechanismthe model asks for a function to be run, and reads the result back9

07Diffusion modelArchitecturelearn to undo noise, and you can make an image from nothing8

08Model familyPropertyone line of models at several sizes, sharing a recipe and a name8

09PretrainingAlgorithmlearn from raw text at scale by predicting what comes next8

10RLHFAlgorithmtilt a pretrained model toward human preference — how raw prediction becomes behaviour7

11TransformerArchitectureattention and feed-forward blocks stacked — the architecture that swallowed the field7

12GeneralizationPropertydoing well on data you have never seen — the only thing that counts6

13Word embeddingMechanismwords as coordinates, so that similar meanings sit near each other6

14Artificial neuronMechanismadd the inputs, then pass the total through a simple curve5

15Bayes' ruleAlgorithmturn a prior and a likelihood into a posterior — updating belief with evidence5

16Fine-tuningAlgorithmcontinue training a pretrained model on a smaller, targeted set5

17InferenceMechanismrunning a trained model to get an answer out5

18Language modelArchitecturea model that puts a probability on the next piece of text5

19Model Context ProtocolProtocola standard way to plug tools into a model, so every model need not learn every tool5

20Retrieval-augmented generationAlgorithmlook facts up and put them in the context, instead of storing them in the weights5

21Scaling lawResultloss falls as a smooth power law in compute, data and parameters — so you can plan ahead5

22TokenizerMechanismthe rule that cuts text into tokens — it shapes everything downstream5

23BackpropagationAlgorithmassign blame for the error backwards through the layers, then nudge every weight4

24BenchmarkBenchmarka fixed task suite and metric — the yardstick, and the thing people overfit to4

25Convolutional neural networkArchitecturestacked convolutions that see images through learned local filters4

26Cross-entropyObjectivethe cost of coding data with the wrong model — the loss nearly everything trains on4

27DeepSeek-V3Modela 671B MoE (37B active) trained in FP8 — frontier-adjacent at a fraction of the assumed cost4

28Mixture of expertsMechanismroute each token to a few small experts — huge capacity, small active cost4

29Multilayer perceptronArchitecturestacked neuron layers that can bend a boundary no single perceptron could4

30OrchestrationSystemthe layer that decides what runs, in what order, and what happens when it fails4

31PerceptronMechanismthe first machine that learned weights to separate two classes4

32Reasoning modelArchitecturea model trained to spend more thinking before it answers4

33Recurrent neural networkArchitecturea network that carries a hidden state forward, one step at a time4

34Agent loopMechanismthink, act, look at the result, repeat until the job is done3

35ComputabilityPropertywhether some step-by-step method can settle a question at all3

36Compute-optimalPropertythe best split of a fixed compute budget between model size and data3

37Control theoryFieldthe mathematics of steering a system toward a target without it wobbling3

38EvaluationFielddeciding what a number means — validity, contamination, ablations3

39GPUHardwarethousands of simple cores doing the same maths at once — the accident that trains everything3

40Gradient descentAlgorithmwalk downhill on the loss by following its slope3

41GRPOAlgorithmscore a group of answers against each other, so no separate value network is needed3

42InterpretabilityFieldfinding out what the insides are actually doing3

43Next-token predictionTaskthe one training game behind a language model: guess the next token3

44ObservabilityFieldbeing able to see what a running system is doing and why it failed3

45Omni-modelArchitectureone model for text, audio, images and video, instead of a relay of specialists3

46ProbabilityFieldthe mathematics of belief under uncertainty3

47QuantizationMechanismstore weights in fewer bits, so a model fits on hardware you own3

48Symbolic AIFieldintelligence as rules over written symbols, added by hand3

49World modelArchitecturean internal simulator of how a situation evolves — imagining without acting3

50AlignmentFieldmaking a system do what was meant, not merely what was written2

51Any-to-anyPropertyone model that takes in text, image, audio or video and emits any of them2

52Backpropagation through timeAlgorithmunroll a loop over time, then run backprop through the whole tape2

53CapacityQuantityhow many different patterns a model is able to fit2

54ChannelMechanismthe pipe a message travels through, which may corrupt it2

55ChinchillaResultfor a fixed budget, most big models of 2020 were too large and under-trained2

56CompressionHeuristicintelligence is compression — predicting well and compressing well are one thing2

57ConvolutionMechanismslide one small filter across the input and reuse it everywhere — built-in translation symmetry2

58DeepSeek-R1Modelopen reasoning weights under MIT — R1-Zero learned to reason with no supervised examples2

59Diffusion transformerArchitecturea transformer as the image denoiser, usually trained with flow matching2

60Durable executionMechanisma long run that survives crashes and restarts where it left off2

61Embedding spacePhenomenonthe geometry where similar things end up near each other2

62Hidden layerMechanisma layer of neurons between input and output that builds its own features2

63Huawei AscendHardwarea non-Nvidia AI accelerator line, pushed hard by export controls2

64Inference scalingPropertytrading more thinking at answer time for better answers2

65InformationQuantityhow much uncertainty a message removes2

66KL divergenceQuantityextra bits wasted by believing q when the truth is p — never negative, never symmetric2

67Knowledge representationMechanismwriting what the system knows in a form a machine can reason over2

68Latent diffusionArchitecturerun the diffusion steps in a small squeezed space, then decode back to pixels2

69Learning rateQuantityhow big a step to take downhill — too small is slow, too big overshoots2

70LikelihoodQuantityhow probable the data you saw would be under a given hypothesis2

71Loss functionObjectivethe number training tries to make small — the formal statement of what you want2

72LSTMArchitecturea gated loop that remembers over long stretches and eases the vanishing gradient2

73Mechanistic interpretabilityFieldtaking the model apart to name the parts and what each one does2

74Model-based RLMechanismlearn a model of the world, then plan inside it instead of acting blindly2

75Mutual informationQuantityhow much knowing one thing tells you about another2

76Open weightsPropertythe numbers released, letting anyone run and fine-tune — but not rebuild2

77OverfittingFailureModememorising the training set instead of learning the pattern2

78PermissionMechanismthe grant that lets a program read this file or call that service — and nothing else2

79PlanningMechanismchoosing a whole sequence of steps in advance of doing them2

80Prefix codeMechanisma code where no codeword is the start of another, so it is readable as it arrives2

81PriorDistributionwhat you believed before seeing the evidence2

82RateQuantitythe number of bits used per symbol — the score compression tries to lower2

83RegularizationMechanismany penalty that keeps a model simple so it stops memorising2

84RewardObjectivea single number saying how well things just went2

85SandboxMechanisman isolated place for an agent to act, so a wrong step is survivable2

86SearchAlgorithmtry possibilities in a smart order instead of all of them2

87Self-supervised learningMechanismhide part of the data and let the model teach itself to recover it2

88Sequence modelArchitectureany model built to read or write things in order2

89Source coding theoremTheoremyou cannot compress below the entropy — and you can get arbitrarily close2

90TokenQuantitya chunk of text — often a word-piece — that the model reads as one unit2

91Training setDatasetthe examples the model is allowed to learn from2

92Turing testThoughtExperimentcan you tell the machine from a person in conversation? — a test of behaviour, not mind2

93Vector databaseSystema store that finds the nearest vectors fast, even among billions2

94Activation functionMechanismthe bend a neuron applies to its weighted sum — without it, layers collapse into one1

95APIProtocola fixed doorway one program uses to ask another for something1

96Auxiliary-loss-free load balancingMechanismkeep experts evenly used by nudging a bias, instead of adding a penalty to the loss1

97BatchingMechanismserving many requests together so the hardware is never idle1

98Bayesian inferenceAlgorithmreasoning that keeps a full distribution over hypotheses and updates it1

99BiasQuantitya neuron's resting offset — how eager it is to fire with no input1

100Bias–variance tradeoffPropertyerror splits into being too simple and being too jumpy; lowering one raises the other1

101BitQuantityone yes-or-no of uncertainty — the atom of information1

102Byte-pair encodingAlgorithmbuild a vocabulary by repeatedly gluing together the most common pair of pieces1

103CANNSystemHuawei's software stack for Ascend chips — the counterpart to CUDA1

104Chain of thoughtMechanismwriting the working out, which gives the model more room to be right1

105Channel capacityTheoremthe most information a noisy channel can carry — a hard ceiling1

106CheckpointingMechanismsaving progress at safe points so a failure does not undo everything1

107ChunkingMechanismcutting long documents into pieces small enough to retrieve and fit1

108Classifier-free guidanceMechanismsteer generation harder toward the prompt by contrasting with and without it1

109Code executionMechanismletting the model write and run real code instead of guessing the answer1

110Combinatorial explosionFailureModeoptions multiply so fast that brute force stops being possible1

111Computer useMechanismdriving a real screen with mouse and keyboard, like a person would1

112ConstitutionDocumenta written set of principles a model is trained to critique itself against1

113ContaminationFailureModetest questions that leaked into training, making the score a memory test1

114Context managementMechanismdeciding what to keep, summarise or drop so the model sees the right things1

115Context windowPropertythe most text a model can hold in mind at once1

116ControlNetArchitecturea side branch that lets a sketch or pose steer an image model without retraining it1

117CyberneticsFieldthe 1940s study of control and communication in animals and machines alike1

118DeepSeek-V4Modelthe 2026 follow-up — 1.6T/49B and 284B/13B MoE, one-million-token context, hybrid attention replacing MLA1

119DenoisingMechanismlearning to remove a little noise, so many tiny steps turn noise into an image1

120Direct preference optimisationAlgorithmlearn straight from preference pairs, skipping the separate reward model1

121Distributional hypothesisHeuristicwords used in similar company end up meaning similar things1

122DivergencePropertya measure of how far two distributions are apart, without being a distance1

123Dot productQuantitymultiply matching entries and add — a number for how aligned two vectors are1

124DriftPhenomenonthe world shifts, so a model that was right slowly stops being right1

125Emergent abilityPhenomenona skill that barely shows, then appears suddenly as the model grows1

126EnvironmentSystemeverything outside the agent that it can sense and change1

127Error-correcting codeAlgorithmadd redundancy so the receiver can undo the channel's damage1

128Evaluation harnessSystemthe machinery that runs a suite of tests and scores the results the same way every time1

129Expert systemSystema hand-built rulebook that answers like a specialist in a narrow domain1

130ExplorationMechanismdeliberately trying something new so you learn what you are missing1

131Export controlsDocumentrules limiting chip sales across borders — the constraint that bent the field toward efficiency, then partly loosened1

132FeatureVariablea direction inside the model that stands for one human-readable idea1

133Feature mapQuantitythe grid of activations a filter produces as it slides across an image1

134FeedbackMechanismmeasure the gap to the goal and use it to correct the next step1

135Flow matchingMechanismlearn a straight path from noise to data instead of a slow curve1

136FP8Formatan 8-bit number format for training and serving, roughly half the memory of 16-bit1

137Frame problemFailureModehard to say which facts an action leaves unchanged, without listing them all1

138Fully open modelPropertyopen weights plus open data, code and logs — reproducible end to end1

139GGUFFormata single self-contained file holding quantised weights and their metadata1

140GLMModelZhipu's model family, from GLM-130B through the GLM-4 line1

141GuardrailMechanisma check that catches and blocks a bad output before a user sees it1

142Halting problemResultno program can decide, for every program, whether it will ever stop1

143Harness engineeringFielddesigning the loop, tools, memory and permissions around a model — now a discipline of its own1

144Huffman codingAlgorithmbuild the optimal prefix code by repeatedly merging the two least likely symbols1

145Human in the loopMechanisma person approves the risky step before the agent takes it1

146HunyuanModelTencent's open mixture-of-experts line, hundreds of billions of parameters1

147Inductive biasPropertythe assumptions a learner makes so it can guess about unseen cases1

148Instruction tuningMechanismfine-tune on examples written as tasks, so following requests becomes the habit1

149IntelligencePropertyreaching goals in an environment, across many different situations1

150InterconnectMechanismthe links between chips — often the real limit on training a large model1

151InternLMModelShanghai AI Laboratory's open model line, released with full technical reports1

152InteroperabilityPropertytools and agents from different makers working together without custom glue1

153JEPAArchitecturepredict in a learned abstract space, not pixel by pixel1

154KimiModelMoonshot's long-context line, with reasoning trained by large-scale RL1

155Knowledge distillationMechanismtrain a small model to copy a big one's answers, gaining skill cheaply1

156KV cacheMechanismstore the keys and values you already computed, so each new token is cheap1

157Latent spacePropertya squeezed code where each point stands for a whole image, far smaller than pixels1

158Least privilegeHeuristicgive every part the smallest power it needs to do the job1

159LlamaModelthe family that made open weights mainstream — under a community licence with a user cap1

160Local servingSystemrunning a model on your own machine, with no network round-trip1

161Loop policyHeuristicthe rules for when to keep going, ask a human, or stop1

162LoRAMechanismfreeze the model and train a tiny add-on — fine-tuning on one GPU1

163Markov decision processMechanismstates, actions and rewards, where the next state depends only on now1

164MatrixQuantitya grid of numbers acting as one machine that turns vectors into vectors1

165MCP clientSystemthe agent side that connects to servers and offers their tools to the model1

166MCP serverSystema small program that exposes tools and data through the shared protocol1

167Memory bandwidthQuantityhow fast numbers move from memory — the limit that decoding usually hits first1

168Memory wallPhenomenoncompute got fast but memory did not, so moving numbers is now the bottleneck1

169MiniMaxModela Chinese lab building long-context models on lightning attention and mixture of experts1

170MistralModela European lab whose small open models pushed quality per parameter1

171Modality encoderArchitecturethe part that turns one kind of input — sound, image, video — into shared tokens1

172Multi-agentSystemseveral agents splitting a job, talking to each other as they go1

173Multi-head latent attentionMechanismcompress the per-token memory attention must keep, so long contexts cost far less1

174Multi-token predictionMechanismtrain the model to guess several next tokens at once, not just one1

175NoisePhenomenonrandom damage the channel adds to the message1

176Occam's razorHeuristicprefer the simplest explanation that fits — a guess, not a theorem1

177OLMoModela fully open model line that releases data, code and training logs too1

178Open image modelModelan image generator you can download and run yourself1

179Open-weight licenceDocumentthe terms attached to downloadable weights — open weights are not always open source1

180ParallelismPropertydoing many small sums at once instead of one after another1

181PerplexityQuantityhow many equally likely choices the model felt it was facing — cross-entropy, exponentiated1

182PolicyMechanismthe rule that says which action to take in each situation1

183PoolingMechanismshrink a feature map by keeping the strongest value in each patch1

184Positional encodingMechanisma pattern added so the model knows where each token sat in the order1

185PosteriorDistributionwhat you believe after seeing the evidence1

186PredictionMechanismsaying what comes next before you are told1

187Prompt injectionFailureModeinstructions hidden in content the model reads, obeyed as if you had typed them1

188Quantised weightsFormatweights stored at 8, 4 or fewer bits so a big model fits in small memory1

189Query, key, valueMechanismeach token asks a question, every token advertises what it holds, and answers are blended by fit1

190QwenModelAlibaba's open model line, released across many sizes and modalities1

191ReActAlgorithminterleave a line of reasoning with a tool call, so thought follows evidence1

192Red teamingMechanismattacking the system on purpose to find how it breaks before others do1

193Reward hackingFailureModegetting the reward without doing the thing — the optimiser finds the loophole1

194Reward modelModela model trained on human preferences that scores answers so training can use it1

195Self-supervised videoMechanismlearn from raw video by predicting what happens next in it1

196Semantic searchMechanismfinding things by meaning rather than by matching words1

197Sparse autoencoderArchitecturea tool that pulls many tidy, mostly-silent features out of a crowded layer1

198SpecificationDocumentwriting down what we actually want the system to do1

199Speculative decodingAlgorithma small fast model drafts, a big model checks — same output, fewer big steps1

200Speech tokenQuantitya chunk of sound turned into something the same machinery can treat like a word1

201SuperpositionPhenomenona model packs more ideas than it has directions, by sharing them at angles1

202SurpriseQuantityhow unlikely the thing that happened was — the raw material of entropy1

203Tensor coreHardwarea circuit that does a whole small matrix multiply in one instruction1

204Test setDatasetexamples held back, used only to check whether learning generalised1

205Test-time computeQuantitythe effort spent while answering, not while training1

206Tool schemaFormata written description of a function — its name, inputs and outputs — a model can be handed1

207TracingMechanismrecording every step, input and output so a run can be replayed1

208Turing machineMechanisman imagined tape-reading machine that pins down what "computable" means1

209UncertaintyPropertyhow many answers are still possible1

210UnderfittingFailureModethe model is too simple to capture the pattern, so it is wrong everywhere1

211Value functionEstimatora guess of how much reward is still to come from here1

212Vanishing gradientFailureModein a deep stack the learning signal fades to nothing before it reaches the bottom1

213VectorQuantitya list of numbers treated as one point or one arrow1

214Vector arithmeticPhenomenonking − man + woman ≈ queen: directions in embedding space carry meaning1

215vLLMSystema serving engine whose paged memory made long-context batching practical1

216VocabularyDatasetthe fixed set of tokens a model can read and write1

217WeightQuantityhow much a neuron cares about one input — the number training changes1

218word2vecAlgorithmlearn a vector per word by predicting its neighbours, cheaply and at scale1

219WorkflowMechanisma fixed sequence of steps, with the model doing only the fuzzy parts1

220Acceptable use policyDocumenta list of things you may not do with a model, attached to its licence0

221AGIOpenQuestiona system that can do any cognitive job a person can — nobody agrees on the test0

222AI winterPhenomenona period when promises outran results and the money left0

223BarbellPhenomenonthe field splits toward a few huge closed models and many small open ones0

224CommoditisationPhenomenoncapability that was rare becomes cheap, so value moves to what you build on top0

225EU AI ActDocumentthe European Union law that sorts AI uses into risk tiers and sets duties for each0

226Full-duplexPropertylisten and speak at the same time, so talking feels like a conversation0

227GPAIPropertygeneral-purpose AI — the EU category for large models usable for many tasks0

228Model cardDocumenta short factual page describing what a model is, how it was made and its limits0

229No free lunchTheoremaveraged over all problems, no learner beats any other — so assumptions are everything0

230Open-source AIFieldbuilding AI in the open, where anyone can read, run and change it0

231Open-weight baselineBenchmarkthe best open model, used as the yardstick everyone else compares against0

232Open-weight riskOpenQuestiondoes releasing weights let the dangerous uses scale faster than the defensive ones?0

233Openness ladderPropertya scale from weights-only to fully open data, code and logs0

234Permissive licenceDocumenta licence that lets you use, change and sell with almost no conditions0

235SuperintelligenceOpenQuestiona system smarter than us in every way — a hypothesis, not an observation0

236The scaling debateOpenQuestionis more scale enough for general capability, or is something still missing?0

17 nodes carry no edges yet — they are in the graph ahead of their connections. Nothing on this page is hand-listed: it is the node array, sorted by degree.