AI Foundationspredict · compress · act
GENERATED FROM THE GRAPH

The atlas

Six jobs across, eight layers down. Every cell below is filled fromsrc/data/concepts.ts — the single source of truth — not typed by hand. Dashed cells are open ground; hatched cells cannot exist.

236 concepts · 230 edges · 28 cells filled · 46 on the frontier

every concept →by kind →reading routes →

Represent
Infer
Learn
Search & planning
Act
Measure
L0 Foundations & theory
CompressionPrefix codeHuffman codingError-correcting codeVectorMatrixEmbedding spaceFrame problem
ProbabilityBayes' ruleLikelihoodBayesian inferencePredictionChannelPriorPosterior
Gradient descentLoss functionGeneralizationOverfittingNo free lunchBias–variance tradeoffInductive biasOccam's razorFeedback
Turing machineComputabilityHalting problemSearchCombinatorial explosion
Reinforcement learningMarkov decision process
BitEntropyCross-entropyKL divergenceDivergenceChannel capacitySource coding theoremCompressionPerplexityIntelligenceUncertaintyInformationSurpriseRateNoiseMutual informationDot productTuring test
L1 Model
Artificial neuronConvolutionConvolutional neural networkRecurrent neural networkLSTMWord embeddingAttentionTransformerTokenizerMixture of expertsDiffusion modelDiffusion transformerWorld modelOmni-modelActivation functionKnowledge representationHidden layerFeature mapPoolingDistributional hypothesisVector arithmeticPositional encodingLatent spaceFeatureSuperpositionContext windowMulti-head latent attentionLatent diffusionControlNetAny-to-anyModality encoder
World modelQuery, key, valueLanguage modelReasoning model
PerceptronBackpropagationWeightBiasMultilayer perceptronSequence modelword2vecValue functionSparse autoencoderJEPA
open
Policy
CapacityEmergent ability
L2 Data
TokenizerTokenByte-pair encodingChunkingSpeech token
open
open
open
open
Contamination
L3 Compute
n/a
n/a
n/a
n/a
n/a
Memory bandwidthInterconnectMemory wall
L4 Training
Denoising
open
Cross-entropyGradient descentScaling lawBackpropagationMixture of expertsPretrainingFine-tuningLoRARLHFGRPOReward hackingLearning rateUnderfittingRegularizationVanishing gradientBackpropagation through timeSelf-supervised learningChinchillaInstruction tuningDirect preference optimisationSelf-supervised videoAuxiliary-loss-free load balancingMulti-token predictionKnowledge distillationFlow matching
Exploration
RLHFRewardModel-based RL
Scaling lawCompute-optimal
L5 Serving & deployment
AttentionKV cache
Classifier-free guidanceInferenceBatchingSpeculative decodingChain of thought
open
open
Full-duplex
Test-time computeInference scaling
L6 Systems, agents & applications
open
Retrieval-augmented generationSemantic search
open
Planning
Reinforcement learningAgentAgent harnessTool useAgent loopReActCode executionWorkflowDurable executionCheckpointingPermissionLeast privilegeComputer useContext managementLoop policy
BenchmarkRed teamingHuman in the loopTracingDriftGuardrail
L7 Ecosystem & institutions
n/a
n/a
n/a
n/a
n/a
Openness ladderModel familyFully open modelCommoditisationBarbell

swipe to scroll →

A chip links to the chapter that explains the concept. Bold cells hold a concept's primary cell; a concept may also carry a noted second cell. 10 cells are structurally impossible.

Every concept, by kind

The 29 kinds in 6 families (FRAMEWORK §8.2). This is the taxonomy view of the same graph.

Evidence & claims
Channel capacityTheorem · timeless
the most information a noisy channel can carry — a hard ceilingread →zoom →shannon-1948
Source coding theoremTheorem · timeless
you cannot compress below the entropy — and you can get arbitrarily closeread →zoom →shannon-1948
CompressionHeuristic · timeless
intelligence is compression — predicting well and compressing well are one thingread →zoom →shannon-1948
No free lunchTheorem · timeless
averaged over all problems, no learner beats any other — so assumptions are everythingread →zoom →kaplan-2020
Scaling lawResult · decade
loss falls as a smooth power law in compute, data and parameters — so you can plan aheadread →zoom →kaplan-2020hoffmann-2022
The scaling debateOpenQuestion · frontier
is more scale enough for general capability, or is something still missing?read →zoom →kaplan-2020
Occam's razorHeuristic · timeless
prefer the simplest explanation that fits — a guess, not a theoremread →zoom →russell-norvig-2020
Turing testThoughtExperiment · timeless
can you tell the machine from a person in conversation? — a test of behaviour, not mindread →zoom →turing-1950
Halting problemResult · timeless
no program can decide, for every program, whether it will ever stopread →zoom →turing-1936
Distributional hypothesisHeuristic · timeless
words used in similar company end up meaning similar thingsread →zoom →jurafsky-martin-2024
ChinchillaResult · timeless
for a fixed budget, most big models of 2020 were too large and under-trainedread →zoom →hoffmann-2022
Least privilegeHeuristic · timeless
give every part the smallest power it needs to do the jobread →zoom →saltzer-schroeder-1975
AGIOpenQuestion · decade
a system that can do any cognitive job a person can — nobody agrees on the testread →zoom →russell-norvig-2020
SuperintelligenceOpenQuestion · decade
a system smarter than us in every way — a hypothesis, not an observationread →zoom →russell-norvig-2020
Open-weight riskOpenQuestion · frontier
does releasing weights let the dangerous uses scale faster than the defensive ones?read →zoom →olmo-2024
Loop policyHeuristic · frontier
the rules for when to keep going, ask a human, or stopread →zoom →yao-2022
Formal objects
BitQuantity · timeless
one yes-or-no of uncertainty — the atom of informationread →zoom →shannon-1948
EntropyQuantity · timeless
the average surprise of a source, in bitsread →zoom →shannon-1948
Cross-entropyObjective · timeless
the cost of coding data with the wrong model — the loss nearly everything trains onread →zoom →shannon-1948
KL divergenceQuantity · timeless
extra bits wasted by believing q when the truth is p — never negative, never symmetricread →zoom →shannon-1948
DivergenceProperty · timeless
a measure of how far two distributions are apart, without being a distanceread →zoom →shannon-1948
PerplexityQuantity · decade
how many equally likely choices the model felt it was facing — cross-entropy, exponentiatedread →zoom →shannon-1948
LikelihoodQuantity · timeless
how probable the data you saw would be under a given hypothesisread →zoom →shannon-1948
Loss functionObjective · timeless
the number training tries to make small — the formal statement of what you wantread →zoom →rumelhart-1986
GeneralizationProperty · timeless
doing well on data you have never seen — the only thing that countsread →zoom →kaplan-2020
Memory bandwidthQuantity · decade
how fast numbers move from memory — the limit that decoding usually hits firstread →zoom →cuda-2007
Open weightsProperty · decade
the numbers released, letting anyone run and fine-tune — but not rebuildread →zoom →meta-llama-2-2023olmo-2024
IntelligenceProperty · timeless
reaching goals in an environment, across many different situationsread →zoom →russell-norvig-2020turing-1950
RewardObjective · timeless
a single number saying how well things just wentread →zoom →sutton-barto-2018
UncertaintyProperty · timeless
how many answers are still possibleread →zoom →cover-thomas-2006
InformationQuantity · timeless
how much uncertainty a message removesread →zoom →shannon-1948
SurpriseQuantity · timeless
how unlikely the thing that happened was — the raw material of entropyread →zoom →mackay-2003
RateQuantity · timeless
the number of bits used per symbol — the score compression tries to lowerread →zoom →shannon-1948
Mutual informationQuantity · timeless
how much knowing one thing tells you about anotherread →zoom →cover-thomas-2006shannon-1948
PriorDistribution · timeless
what you believed before seeing the evidenceread →zoom →bishop-2006
PosteriorDistribution · timeless
what you believe after seeing the evidenceread →zoom →bishop-2006
VectorQuantity · timeless
a list of numbers treated as one point or one arrowread →zoom →goodfellow-2016
Dot productQuantity · timeless
multiply matching entries and add — a number for how aligned two vectors areread →zoom →goodfellow-2016
MatrixQuantity · timeless
a grid of numbers acting as one machine that turns vectors into vectorsread →zoom →goodfellow-2016
Learning rateQuantity · timeless
how big a step to take downhill — too small is slow, too big overshootsread →zoom →goodfellow-2016
Bias–variance tradeoffProperty · timeless
error splits into being too simple and being too jumpy; lowering one raises the otherread →zoom →bishop-2006goodfellow-2016
Inductive biasProperty · timeless
the assumptions a learner makes so it can guess about unseen casesread →zoom →goodfellow-2016russell-norvig-2020
CapacityQuantity · timeless
how many different patterns a model is able to fitread →zoom →goodfellow-2016
ComputabilityProperty · timeless
whether some step-by-step method can settle a question at allread →zoom →turing-1936
WeightQuantity · timeless
how much a neuron cares about one input — the number training changesread →zoom →goodfellow-2016
BiasQuantity · timeless
a neuron's resting offset — how eager it is to fire with no inputread →zoom →goodfellow-2016
ParallelismProperty · timeless
doing many small sums at once instead of one after anotherread →zoom →cuda-2007
Feature mapQuantity · timeless
the grid of activations a filter produces as it slides across an imageread →zoom →lecun-1998
Value functionEstimator · timeless
a guess of how much reward is still to come from hereread →zoom →sutton-barto-2018
TokenQuantity · timeless
a chunk of text — often a word-piece — that the model reads as one unitread →zoom →kudo-2018jurafsky-martin-2024
Compute-optimalProperty · timeless
the best split of a fixed compute budget between model size and dataread →zoom →hoffmann-2022
Latent spaceProperty · timeless
a squeezed code where each point stands for a whole image, far smaller than pixelsread →zoom →rombach-2022
FeatureVariable · timeless
a direction inside the model that stands for one human-readable idearead →zoom →templeton-2024
Context windowProperty · timeless
the most text a model can hold in mind at onceread →zoom →brown-2020
InteroperabilityProperty · timeless
tools and agents from different makers working together without custom glueread →zoom →mcp-2024
Test-time computeQuantity · timeless
the effort spent while answering, not while trainingread →zoom →snell-2024
Inference scalingProperty · timeless
trading more thinking at answer time for better answersread →zoom →snell-2024
Openness ladderProperty · decade
a scale from weights-only to fully open data, code and logsread →zoom →olmo-2024
Model familyProperty · decade
one line of models at several sizes, sharing a recipe and a nameread →zoom →qwen2.5-2024
Fully open modelProperty · decade
open weights plus open data, code and logs — reproducible end to endread →zoom →olmo-2024
GPAIProperty · frontier
general-purpose AI — the EU category for large models usable for many tasksread →zoom →eu-ai-act-2024
Any-to-anyProperty · frontier
one model that takes in text, image, audio or video and emits any of themread →zoom →qwen3-omni-2025
Full-duplexProperty · frontier
listen and speak at the same time, so talking feels like a conversationread →zoom →qwen3-omni-2025
Speech tokenQuantity · frontier
a chunk of sound turned into something the same machinery can treat like a wordread →zoom →qwen3-omni-2025
Methods
Bayes' ruleAlgorithm · timeless
turn a prior and a likelihood into a posterior — updating belief with evidenceread →zoom →shannon-1948
Bayesian inferenceAlgorithm · timeless
reasoning that keeps a full distribution over hypotheses and updates itzoom →shannon-1948
Gradient descentAlgorithm · timeless
walk downhill on the loss by following its sloperead →zoom →rumelhart-1986
Artificial neuronMechanism · timeless
add the inputs, then pass the total through a simple curveread →zoom →mcculloch-pitts-1943
PerceptronMechanism · timeless
the first machine that learned weights to separate two classesread →zoom →rosenblatt-1958minsky-papert-1969
BackpropagationAlgorithm · timeless
assign blame for the error backwards through the layers, then nudge every weightread →zoom →rumelhart-1986
ConvolutionMechanism · timeless
slide one small filter across the input and reuse it everywhere — built-in translation symmetryread →zoom →krizhevsky-2012
Convolutional neural networkArchitecture · decade
stacked convolutions that see images through learned local filtersread →zoom →krizhevsky-2012
Recurrent neural networkArchitecture · decade
a network that carries a hidden state forward, one step at a timeread →zoom →hochreiter-1997
LSTMArchitecture · decade
a gated loop that remembers over long stretches and eases the vanishing gradientread →zoom →hochreiter-1997
Word embeddingMechanism · decade
words as coordinates, so that similar meanings sit near each otherread →zoom →mikolov-2013
AttentionMechanism · decade
let each token weigh every other token, instead of only the last oneread →zoom →vaswani-2017
TransformerArchitecture · decade
attention and feed-forward blocks stacked — the architecture that swallowed the fieldread →zoom →vaswani-2017
TokenizerMechanism · decade
the rule that cuts text into tokens — it shapes everything downstreamread →zoom →kudo-2018
Mixture of expertsMechanism · decade
route each token to a few small experts — huge capacity, small active costread →zoom →deepseek-v2-2024deepseek-v3-2024
Diffusion modelArchitecture · decade
learn to undo noise, and you can make an image from nothingread →zoom →ho-2020rombach-2022
Diffusion transformerArchitecture · frontier
a transformer as the image denoiser, usually trained with flow matchingread →zoom →peebles-xie-2023esser-2024
World modelArchitecture · frontier
an internal simulator of how a situation evolves — imagining without actingread →zoom →kaplan-2020
Omni-modelArchitecture · frontier
one model for text, audio, images and video, instead of a relay of specialistsread →zoom →qwen3-omni-2025
PretrainingAlgorithm · decade
learn from raw text at scale by predicting what comes nextread →zoom →brown-2020
Fine-tuningAlgorithm · decade
continue training a pretrained model on a smaller, targeted setread →zoom →ouyang-2022
LoRAMechanism · decade
freeze the model and train a tiny add-on — fine-tuning on one GPUread →zoom →ouyang-2022
RLHFAlgorithm · decade
tilt a pretrained model toward human preference — how raw prediction becomes behaviourread →zoom →ouyang-2022bai-2022
GRPOAlgorithm · frontier
score a group of answers against each other, so no separate value network is neededread →zoom →shao-2024deepseek-r1-2025
KV cacheMechanism · decade
store the keys and values you already computed, so each new token is cheapread →zoom →kwon-2023
QuantizationMechanism · decade
store weights in fewer bits, so a model fits on hardware you ownread →zoom →kwon-2023
Tool useMechanism · decade
the model asks for a function to be run, and reads the result backread →zoom →yao-2022
Retrieval-augmented generationAlgorithm · decade
look facts up and put them in the context, instead of storing them in the weightsread →zoom →lewis-2020
SandboxMechanism · decade
an isolated place for an agent to act, so a wrong step is survivableread →zoom →yao-2022
PredictionMechanism · timeless
saying what comes next before you are toldread →zoom →mackay-2003
Prefix codeMechanism · timeless
a code where no codeword is the start of another, so it is readable as it arrivesread →zoom →cover-thomas-2006
Huffman codingAlgorithm · timeless
build the optimal prefix code by repeatedly merging the two least likely symbolsread →zoom →mackay-2003cover-thomas-2006
ChannelMechanism · timeless
the pipe a message travels through, which may corrupt itread →zoom →shannon-1948
Error-correcting codeAlgorithm · timeless
add redundancy so the receiver can undo the channel's damageread →zoom →cover-thomas-2006
RegularizationMechanism · timeless
any penalty that keeps a model simple so it stops memorisingread →zoom →goodfellow-2016bishop-2006
Turing machineMechanism · timeless
an imagined tape-reading machine that pins down what "computable" meansread →zoom →turing-1936
Activation functionMechanism · timeless
the bend a neuron applies to its weighted sum — without it, layers collapse into oneread →zoom →goodfellow-2016
Knowledge representationMechanism · timeless
writing what the system knows in a form a machine can reason overread →zoom →mccarthy-1958
SearchAlgorithm · timeless
try possibilities in a smart order instead of all of themread →zoom →russell-norvig-2020
FeedbackMechanism · timeless
measure the gap to the goal and use it to correct the next stepread →zoom →wiener-1948
Markov decision processMechanism · timeless
states, actions and rewards, where the next state depends only on nowread →zoom →bellman-1957sutton-barto-2018
Hidden layerMechanism · timeless
a layer of neurons between input and output that builds its own featuresread →zoom →rumelhart-1986
Multilayer perceptronArchitecture · timeless
stacked neuron layers that can bend a boundary no single perceptron couldread →zoom →rumelhart-1986goodfellow-2016
PoolingMechanism · timeless
shrink a feature map by keeping the strongest value in each patchread →zoom →lecun-1998
Sequence modelArchitecture · timeless
any model built to read or write things in orderread →zoom →goodfellow-2016
Backpropagation through timeAlgorithm · timeless
unroll a loop over time, then run backprop through the whole taperead →zoom →goodfellow-2016
word2vecAlgorithm · timeless
learn a vector per word by predicting its neighbours, cheaply and at scaleread →zoom →mikolov-2013
PolicyMechanism · timeless
the rule that says which action to take in each situationread →zoom →sutton-barto-2018
ExplorationMechanism · timeless
deliberately trying something new so you learn what you are missingread →zoom →sutton-barto-2018
Query, key, valueMechanism · timeless
each token asks a question, every token advertises what it holds, and answers are blended by fitread →zoom →vaswani-2017
Positional encodingMechanism · timeless
a pattern added so the model knows where each token sat in the orderread →zoom →vaswani-2017
Byte-pair encodingAlgorithm · timeless
build a vocabulary by repeatedly gluing together the most common pair of piecesread →zoom →kudo-2018
Self-supervised learningMechanism · timeless
hide part of the data and let the model teach itself to recover itread →zoom →devlin-2018radford-2019
Language modelArchitecture · timeless
a model that puts a probability on the next piece of textread →zoom →radford-2019brown-2020
Instruction tuningMechanism · timeless
fine-tune on examples written as tasks, so following requests becomes the habitread →zoom →ouyang-2022
Direct preference optimisationAlgorithm · timeless
learn straight from preference pairs, skipping the separate reward modelread →zoom →rafailov-2023
DenoisingMechanism · timeless
learning to remove a little noise, so many tiny steps turn noise into an imageread →zoom →ho-2020
Classifier-free guidanceMechanism · timeless
steer generation harder toward the prompt by contrasting with and without itread →zoom →ho-2020
Sparse autoencoderArchitecture · timeless
a tool that pulls many tidy, mostly-silent features out of a crowded layerread →zoom →templeton-2024
Red teamingMechanism · timeless
attacking the system on purpose to find how it breaks before others doread →zoom →perez-2022
Agent loopMechanism · timeless
think, act, look at the result, repeat until the job is doneread →zoom →yao-2022
ReActAlgorithm · timeless
interleave a line of reasoning with a tool call, so thought follows evidenceread →zoom →yao-2022
PlanningMechanism · timeless
choosing a whole sequence of steps in advance of doing themread →zoom →russell-norvig-2020yao-2022
Code executionMechanism · timeless
letting the model write and run real code instead of guessing the answerread →zoom →chen-2021
ChunkingMechanism · timeless
cutting long documents into pieces small enough to retrieve and fitread →zoom →lewis-2020
Semantic searchMechanism · timeless
finding things by meaning rather than by matching wordsread →zoom →lewis-2020
WorkflowMechanism · timeless
a fixed sequence of steps, with the model doing only the fuzzy partsread →zoom →kleppmann-2017
Durable executionMechanism · timeless
a long run that survives crashes and restarts where it left offread →zoom →kleppmann-2017
CheckpointingMechanism · timeless
saving progress at safe points so a failure does not undo everythingread →zoom →kleppmann-2017
InferenceMechanism · timeless
running a trained model to get an answer outread →zoom →kwon-2023
BatchingMechanism · timeless
serving many requests together so the hardware is never idleread →zoom →kwon-2023
Speculative decodingAlgorithm · timeless
a small fast model drafts, a big model checks — same output, fewer big stepsread →zoom →leviathan-2023
PermissionMechanism · timeless
the grant that lets a program read this file or call that service — and nothing elseread →zoom →saltzer-schroeder-1975
Computer useMechanism · timeless
driving a real screen with mouse and keyboard, like a person wouldread →zoom →anthropic-computer-use-2024
Human in the loopMechanism · timeless
a person approves the risky step before the agent takes itread →zoom →russell-norvig-2020
TracingMechanism · timeless
recording every step, input and output so a run can be replayedread →zoom →sculley-2015
GuardrailMechanism · timeless
a check that catches and blocks a bad output before a user sees itread →zoom →bai-2022
Chain of thoughtMechanism · timeless
writing the working out, which gives the model more room to be rightread →zoom →wei-2022
Reasoning modelArchitecture · frontier
a model trained to spend more thinking before it answersread →zoom →snell-2024
Model-based RLMechanism · timeless
learn a model of the world, then plan inside it instead of acting blindlyread →zoom →hafner-2019sutton-barto-2018
Self-supervised videoMechanism · timeless
learn from raw video by predicting what happens next in itread →zoom →lecun-2022
JEPAArchitecture · frontier
predict in a learned abstract space, not pixel by pixelread →zoom →lecun-2022
Multi-head latent attentionMechanism · frontier
compress the per-token memory attention must keep, so long contexts cost far lessread →zoom →deepseek-v2-2024
Auxiliary-loss-free load balancingMechanism · frontier
keep experts evenly used by nudging a bias, instead of adding a penalty to the lossread →zoom →deepseek-v3-2024
Multi-token predictionMechanism · frontier
train the model to guess several next tokens at once, not just oneread →zoom →deepseek-v3-2024
Knowledge distillationMechanism · timeless
train a small model to copy a big one's answers, gaining skill cheaplyread →zoom →hinton-2015deepseek-r1-2025
InterconnectMechanism · timeless
the links between chips — often the real limit on training a large modelread →zoom →cann-2024
Context managementMechanism · frontier
deciding what to keep, summarise or drop so the model sees the right thingsread →zoom →yao-2022
Latent diffusionArchitecture · timeless
run the diffusion steps in a small squeezed space, then decode back to pixelsread →zoom →rombach-2022
Flow matchingMechanism · frontier
learn a straight path from noise to data instead of a slow curveread →zoom →esser-2024
ControlNetArchitecture · frontier
a side branch that lets a sketch or pose steer an image model without retraining itread →zoom →zhang-2023
Modality encoderArchitecture · frontier
the part that turns one kind of input — sound, image, video — into shared tokensread →zoom →qwen3-omni-2025
Artifacts
GPUHardware · decade
thousands of simple cores doing the same maths at once — the accident that trains everythingread →zoom →cuda-2007krizhevsky-2012
AgentSystem · decade
a model plus a harness, running a loop toward a goalread →zoom →yao-2022
Agent harnessSystem · decade
every decision you can make without retraining the model — loop, context, tools, memoryread →zoom →yao-2022mcp-2024
Model Context ProtocolProtocol · frontier
a standard way to plug tools into a model, so every model need not learn every toolread →zoom →mcp-2024
BenchmarkBenchmark · frontier
a fixed task suite and metric — the yardstick, and the thing people overfit toread →zoom →jimenez-2023
DeepSeek-V3Model · frontier
a 671B MoE (37B active) trained in FP8 — frontier-adjacent at a fraction of the assumed costread →zoom →deepseek-v3-2024
DeepSeek-R1Model · frontier
open reasoning weights under MIT — R1-Zero learned to reason with no supervised examplesread →zoom →deepseek-r1-2025
LlamaModel · frontier
the family that made open weights mainstream — under a community licence with a user capread →zoom →meta-llama-2-2023llama4-2025
Export controlsDocument · frontier
rules limiting chip sales across borders — the constraint that bent the field toward efficiency, then partly loosenedread →zoom →bis-2022bis-2026
EnvironmentSystem · timeless
everything outside the agent that it can sense and changeread →zoom →russell-norvig-2020
Training setDataset · timeless
the examples the model is allowed to learn fromread →zoom →bishop-2006
Test setDataset · timeless
examples held back, used only to check whether learning generalisedread →zoom →bishop-2006
Tensor coreHardware · timeless
a circuit that does a whole small matrix multiply in one instructionread →zoom →cuda-2007
Expert systemSystem · timeless
a hand-built rulebook that answers like a specialist in a narrow domainread →zoom →russell-norvig-2020
VocabularyDataset · timeless
the fixed set of tokens a model can read and writeread →zoom →kudo-2018
Next-token predictionTask · timeless
the one training game behind a language model: guess the next tokenread →zoom →radford-2019
Reward modelModel · timeless
a model trained on human preferences that scores answers so training can use itread →zoom →ouyang-2022
SpecificationDocument · timeless
writing down what we actually want the system to doread →zoom →russell-norvig-2020
ConstitutionDocument · timeless
a written set of principles a model is trained to critique itself againstread →zoom →bai-2022
Tool schemaFormat · timeless
a written description of a function — its name, inputs and outputs — a model can be handedread →zoom →schick-2023
APIProtocol · timeless
a fixed doorway one program uses to ask another for somethingread →zoom →mcp-2024
Vector databaseSystem · timeless
a store that finds the nearest vectors fast, even among billionsread →zoom →malkov-2018
MCP serverSystem · timeless
a small program that exposes tools and data through the shared protocolread →zoom →mcp-2024
MCP clientSystem · timeless
the agent side that connects to servers and offers their tools to the modelread →zoom →mcp-2024
OrchestrationSystem · timeless
the layer that decides what runs, in what order, and what happens when it failsread →zoom →kleppmann-2017
Multi-agentSystem · timeless
several agents splitting a job, talking to each other as they goread →zoom →yao-2022
vLLMSystem · timeless
a serving engine whose paged memory made long-context batching practicalread →zoom →kwon-2023
Evaluation harnessSystem · timeless
the machinery that runs a suite of tests and scores the results the same way every timeread →zoom →liang-2022
DeepSeek-V4Model · frontier
the 2026 follow-up — 1.6T/49B and 284B/13B MoE, one-million-token context, hybrid attention replacing MLAread →zoom →deepseek-v4-2026
Permissive licenceDocument · decade
a licence that lets you use, change and sell with almost no conditionsread →zoom →olmo-2024
FP8Format · frontier
an 8-bit number format for training and serving, roughly half the memory of 16-bitread →zoom →deepseek-v3-2024
QwenModel · frontier
Alibaba's open model line, released across many sizes and modalitiesread →zoom →qwen2.5-2024qwen3-omni-2025qwen38-2026
GLMModel · frontier
Zhipu's model family, from GLM-130B through the GLM-4 lineread →zoom →glm-2024
KimiModel · frontier
Moonshot's long-context line, with reasoning trained by large-scale RLread →zoom →kimi-2025kimi-k2-2025
MiniMaxModel · frontier
a Chinese lab building long-context models on lightning attention and mixture of expertsread →zoom →minimax-2025
InternLMModel · frontier
Shanghai AI Laboratory's open model line, released with full technical reportsread →zoom →internlm2-2024
HunyuanModel · frontier
Tencent's open mixture-of-experts line, hundreds of billions of parametersread →zoom →hunyuan-2024
MistralModel · frontier
a European lab whose small open models pushed quality per parameterread →zoom →mistral-2023
OLMoModel · frontier
a fully open model line that releases data, code and training logs tooread →zoom →olmo-2024
Open-weight baselineBenchmark · frontier
the best open model, used as the yardstick everyone else compares againstread →zoom →olmo-2024
Huawei AscendHardware · frontier
a non-Nvidia AI accelerator line, pushed hard by export controlsread →zoom →cann-2024
CANNSystem · frontier
Huawei's software stack for Ascend chips — the counterpart to CUDAread →zoom →cann-2024
Open-weight licenceDocument · decade
the terms attached to downloadable weights — open weights are not always open sourceread →zoom →meta-llama-2-2023
Acceptable use policyDocument · decade
a list of things you may not do with a model, attached to its licenceread →zoom →meta-llama-2-2023
Model cardDocument · decade
a short factual page describing what a model is, how it was made and its limitsread →zoom →olmo-2024
EU AI ActDocument · frontier
the European Union law that sorts AI uses into risk tiers and sets duties for eachread →zoom →eu-ai-act-2024eu-gpai-2025
GGUFFormat · frontier
a single self-contained file holding quantised weights and their metadataread →zoom →gguf-2024
Local servingSystem · decade
running a model on your own machine, with no network round-tripread →zoom →gguf-2024
Quantised weightsFormat · decade
weights stored at 8, 4 or fewer bits so a big model fits in small memoryread →zoom →gguf-2024
Open image modelModel · frontier
an image generator you can download and run yourselfread →zoom →esser-2024
Actors & collectives
ProbabilityField · timeless
the mathematics of belief under uncertaintyread →zoom →shannon-1948
Reinforcement learningField · timeless
learning what to do by trying, from delayed rewardread →zoom →shao-2024
AlignmentField · decade
making a system do what was meant, not merely what was writtenread →zoom →bai-2022ouyang-2022
InterpretabilityField · decade
finding out what the insides are actually doingread →zoom →ami-2021
EvaluationField · decade
deciding what a number means — validity, contamination, ablationsread →zoom →jimenez-2023
Symbolic AIField · timeless
intelligence as rules over written symbols, added by handread →zoom →mccarthy-1958russell-norvig-2020
Control theoryField · timeless
the mathematics of steering a system toward a target without it wobblingread →zoom →wiener-1948
CyberneticsField · timeless
the 1940s study of control and communication in animals and machines alikeread →zoom →wiener-1948
Mechanistic interpretabilityField · timeless
taking the model apart to name the parts and what each one doesread →zoom →elhage-2022
ObservabilityField · timeless
being able to see what a running system is doing and why it failedread →zoom →sculley-2015
Open-source AIField · decade
building AI in the open, where anyone can read, run and change itread →zoom →olmo-2024
Harness engineeringField · frontier
designing the loop, tools, memory and permissions around a model — now a discipline of its ownread →zoom →yao-2022
Diagnostics
OverfittingFailureMode · timeless
memorising the training set instead of learning the patternread →zoom →kaplan-2020
Reward hackingFailureMode · decade
getting the reward without doing the thing — the optimiser finds the loopholeread →zoom →ouyang-2022
Prompt injectionFailureMode · frontier
instructions hidden in content the model reads, obeyed as if you had typed themread →zoom →mcp-2024
AI winterPhenomenon · timeless
a period when promises outran results and the money leftread →zoom →minsky-papert-1969
NoisePhenomenon · timeless
random damage the channel adds to the messageread →zoom →shannon-1948
Embedding spacePhenomenon · timeless
the geometry where similar things end up near each otherread →zoom →jurafsky-martin-2024
UnderfittingFailureMode · timeless
the model is too simple to capture the pattern, so it is wrong everywhereread →zoom →bishop-2006
Combinatorial explosionFailureMode · timeless
options multiply so fast that brute force stops being possibleread →zoom →russell-norvig-2020
Frame problemFailureMode · timeless
hard to say which facts an action leaves unchanged, without listing them allread →zoom →mccarthy-1958russell-norvig-2020
Vanishing gradientFailureMode · timeless
in a deep stack the learning signal fades to nothing before it reaches the bottomread →zoom →goodfellow-2016
Vector arithmeticPhenomenon · timeless
king − man + woman ≈ queen: directions in embedding space carry meaningread →zoom →mikolov-2013
Emergent abilityPhenomenon · timeless
a skill that barely shows, then appears suddenly as the model growsread →zoom →wei-2022-emergent
SuperpositionPhenomenon · timeless
a model packs more ideas than it has directions, by sharing them at anglesread →zoom →elhage-2022
ContaminationFailureMode · timeless
test questions that leaked into training, making the score a memory testread →zoom →sainz-2023
DriftPhenomenon · timeless
the world shifts, so a model that was right slowly stops being rightread →zoom →sculley-2015
Memory wallPhenomenon · timeless
compute got fast but memory did not, so moving numbers is now the bottleneckread →zoom →wulf-mckee-1995
CommoditisationPhenomenon · frontier
capability that was rare becomes cheap, so value moves to what you build on topread →zoom →olmo-2024
BarbellPhenomenon · frontier
the field splits toward a few huge closed models and many small open onesread →zoom →olmo-2024openrouter-2026

This page is generated on every build byscripts/validate.ts and the graph in src/data/concepts.ts. If a rule in the framework breaks, the build fails and this page never ships stale.