The whole graph
Every concept
482 concepts across 47 domains. 479 reviewed and merged, 3 awaiting that review — pick anywhere and follow the thread.
Adversarial8
Attacks
Foundations
Agents8
Control
Interface
Alignment14
Evaluation
- Dangerous Capability EvaluationBefore deploying a model, test whether it can do the specific things that would be harmful if it could — not whether…verified
- Model-Written EvaluationWriting evaluations by hand is slow, and the behaviours worth measuring keep multiplying.verified
- Safety Benchmark ValidityA benchmark labelled "safety" is only measuring safety if a model can score well on it without simply being more…verified
Preference
- DPORLHF trains a reward model and then optimises against it with reinforcement learning — two models, two stages, and a…verified
- PPOPolicy gradient methods improve a policy by pushing up the probability of actions that scored well.verified
- Reward ModelNobody can write down a loss function for 'a helpful answer'.verified
- RLHFInstruction tuning teaches a model to imitate good answers.verified
Research
- Deceptive AlignmentA model that behaves well while it is being tested and differently when it is not.verified
- Scalable OversightEvery method that trains a model on human judgement assumes a human who can judge.verified
- Weak-to-Strong GeneralizationFine-tune a strong model on labels produced by a weak one.verified
Architecture9
Blocks
- Decoder-Only TransformerThe original transformer had an encoder that read and a decoder that wrote.verified
- Encoder-DecoderWhen the input and the output are different things — a French sentence and its English translation — one stack can…verified
- Feed-Forward NetworkAttention decides which tokens a position should look at.verified
- Residual ConnectionInstead of asking a layer to produce its output from scratch, a residual connection asks it only for the change, and…verified
- Transformer BlockThe transformer is one block repeated.verified
Normalization
Attention13
Core
- AttentionA model reading a sentence needs to decide, for each word, which other words matter.verified
- Causal MaskingA model trained to predict the next word must not be allowed to see it.verified
- Cross-AttentionCross-attention is attention between two different sequences.verified
- Multi-Head AttentionOne attention pass has to settle on a single set of weights, so it averages away anything it cannot decide between.verified
- Query, Key, ValueAttention borrows its vocabulary from databases.verified
- Self-AttentionSelf-attention is attention pointed at its own input: the queries, keys and values all come from the same sequence.verified
Efficiency
- FlashAttentionStandard attention writes the whole score matrix to memory, then reads it back to apply the softmax, then reads it…verified
- Grouped-Query AttentionMulti-head attention gives every head its own keys and values; multi-query gives them all one set.verified
- Multi-Query AttentionAttention heads each keep their own keys and values, and at generation time all of them sit in memory.verified
Position
- ALiBiALiBi drops positional encoding entirely and instead penalises attention for reaching far.verified
- Positional EncodingAttention has no idea what order anything is in.verified
- Rotary Position EmbeddingRather than adding position to a token's representation, RoPE rotates it — by an angle proportional to where it sits.verified
Behaviour8
Access
Deployment
Causality8
Foundations
- Causal InferencePrediction asks what usually happens together.verified
- ConfoundingIce cream sales and drownings rise together, and neither causes the other — summer causes both.verified
- CounterfactualWould this patient have recovered without the drug they in fact took? The question is about a world that did not…verified
- InterventionThere is a difference between seeing a value and setting it.verified
Methods
Code8
Generation
Retrieval
Compilers8
Methods
- AutotuningNobody can predict which tile size, unroll factor and thread arrangement will be fastest, so the compiler tries many…verified
- Custom KernelSometimes the compiler will not produce what you need and you write the kernel yourself.verified
- Operator Fusion Pass`kernel-fusion` is the thing being done — merging operations so intermediate values never reach memory.verified
- TilingA matrix multiply that streams straight through memory is slow whatever the arithmetic peak says.verified
Compression8
Methods
- Knowledge DistillationTrain a small model to copy a large one's outputs rather than the original labels.verified
- Low-Rank FactorizationReplace a big weight matrix with the product of two thin ones.verified
- PruningMost weights in a trained network can be set to zero without hurting it.verified
- Structured SparsityZeros only help if the hardware can skip them.verified
Context8
Behaviour
Evaluation
Methods
- Context CompressionIf the window is expensive and much of what goes into it is redundant, spend some compute shortening the context…verified
- Long-Context TrainingA model does not get a long context by having its positions rescaled; it gets one by then being trained on long…verified
- Position InterpolationRather than asking a model to handle positions it has never seen, squeeze the new positions into the range it already…verified
Continual8
Foundations
- Catastrophic ForgettingTrain a network on a second task and it does not merely get worse at the first — it can lose it almost entirely…verified
- Continual LearningLearn a sequence of tasks without keeping all the data and without losing the earlier ones.verified
- Stability-PlasticityA system that never changes cannot learn; a system that changes freely cannot remember.verified
Methods
Data8
Evaluation
Methods
- Data FilteringMost of a web crawl is not worth training on — boilerplate, spam, machine translation, pages that are mostly…verified
- Data MixingA training corpus is several corpora — web text, code, books, papers, forums — and how much of each you use is a…verified
- DeduplicationWeb corpora contain the same text many times — mirrored pages, quoted passages, licence boilerplate.verified
- Synthetic DataGenerate training data with a model instead of collecting it.verified
Deployment8
Monitoring
Practice
Evaluation13
Capabilities
Methods
- AblationRemove one piece and measure what happens.verified
- Elo RatingRank things by who beats whom rather than by scoring each alone.verified
- Human EvaluationAsk people whether the output is good.verified
- LLM as JudgeGrading open-ended answers needs a human, and humans are slow and expensive.verified
- Statistical SignificanceWhether a difference between two numbers is bigger than the noise that produced them.verified
Metrics
- CalibrationA well-calibrated model is right about 80% of the time when it says it is 80% confident.verified
- pass@kFor code, whether a single answer is right is the wrong question — a developer can generate several attempts and run…verified
- PerplexityPerplexity asks how surprised a model is by text it has not seen.verified
- Precision and RecallTwo questions about the same detector that pull against each other.verified
Fairness8
Criteria
Data
Foundations
Foundations37
Architecture
- Activation FunctionThe small nonlinear step between one layer and the next.verified
- Long Short-Term MemoryA recurrent cell with a deliberate path for information to travel a long way unchanged.verified
- Recurrent Neural NetworkProcess a sequence one element at a time, carrying a hidden state forward.verified
- Sequence to SequenceRead a whole input sequence, then produce a whole output sequence of a different length.verified
Functions
Generalization
- Bias-Variance TradeoffA model can be wrong because it is too simple to capture the pattern, or because it is so flexible it fits the noise…verified
- Cross-ValidationOne held-out test set gives you one number, and that number has error bars nobody prints.verified
- Inductive BiasThe assumptions a model makes before it sees any data.verified
Instance-Based
Optimization
- Batch NormalizationNormalise each feature across the examples in the batch, so every layer sees inputs with a stable scale however the…verified
- ConvexityA bowl has one bottom.verified
- Vanishing GradientTrain a deep network and the early layers barely move.verified
- Weight InitializationThe numbers a network starts with, before it has seen anything.verified
Representation
- Byte-Pair EncodingByte-pair encoding builds a vocabulary by repetition.verified
- Token EmbeddingA token id is just a number with no meaning — id 400 is not twice id 200.verified
- TokenizationA model does not see text, it sees numbers.verified
- Word EmbeddingA word becomes a vector, and words used in similar contexts end up near each other — not because anyone described…verified
Statistics
- Bayesian InferenceStart with what you believed before, weigh it by how well each possibility explains what you saw, and end with what…verified
- Expectation-MaximizationFitting a model when some of what you need is missing.verified
- Hidden Markov ModelA sequence of states you cannot see, each producing an observation you can.verified
- Maximum LikelihoodPick the parameters that make the data you actually observed as probable as possible.verified
- N-Gram ModelPredict the next word by counting how often it followed the previous few in a large pile of text.verified
Supervised
- Active LearningIf labels are expensive, do not label at random — let the model choose what to ask about.verified
- Data AugmentationMake more training data out of the data you have, by changing it in ways that do not change the answer.verified
- Decision TreeAsk a series of yes-or-no questions about a data point until you reach an answer.verified
- Early StoppingWatch performance on data the model is not training on, and stop when it stops improving.verified
- Ensemble MethodsCombine several models instead of trusting one.verified
- Linear RegressionFit a straight line through points, choosing the one whose distances to them are smallest overall.verified
- Logistic RegressionThe same idea as fitting a line, except the answer has to be a probability.verified
- Semi-Supervised LearningA few labelled examples and an enormous pile of unlabelled ones — which is the situation almost everybody is actually…verified
- Support Vector MachineOf all the lines separating two classes, take the one with the widest empty corridor around it.verified
Training
- Exposure BiasThe consequence of `teacher-forcing`: a model has only ever seen correct prefixes, so when its own output drifts it…verified
- Label SmoothingInstead of training toward a target of exactly one for the correct class and zero for everything else, aim slightly…verified
- Teacher ForcingDuring training, feed the model the correct previous tokens rather than the ones it just produced.verified
- Transfer LearningLearn something on one problem and reuse it on another.verified
Generative8
Architecture
Objectives
Graphs8
Architecture
Attention
Foundations
- Graph ExpressivityThere are pairs of graphs a message-passing network simply cannot tell apart — not because it was trained badly, but…verified
- Graph Neural NetworkSome data is not a sequence and not a grid — molecules, road networks, citations, social graphs.verified
- Message PassingThe operation nearly every graph network is built from: each node sends a message to its neighbours, each node adds…verified
Hardware9
Compute
- Accelerator UtilizationA cluster's peak throughput is a number on a specification sheet, and a training run reaches somewhere between a…verified
- FLOPsA count of arithmetic operations — the currency compute is budgeted in.verified
- Memory Bandwidth WallEvery accelerator generation adds far more arithmetic throughput than memory bandwidth.verified
- Sparsity HardwareZeros in a weight matrix are only free if the silicon can skip them, and general unstructured sparsity is very hard…verified
- Tensor CoreMost of the arithmetic in a neural network is one operation — multiply a small matrix by another and add the result —…verified
Fabric
Numerics
Inference20
Adaptive
- Adaptive ComputationSpend more on hard inputs than on easy ones.verified
- Early ExitStop partway up the stack when the answer is already clear.verified
- Mixture of DepthsLet each token choose how many layers it passes through.verified
- Model CascadeAsk a cheap model first.verified
- Model RoutingDecide which model should answer before any of them does.verified
- Semantic CacheReturn a stored answer when a new question means the same as an old one.verified
Decoding
- Autoregressive DecodingA language model does not write a sentence, it writes a token and then reads what it wrote.verified
- Beam SearchGreedy decoding takes the most likely next token every step and can walk into a corner, because a mediocre token now…verified
- Nucleus SamplingTop-k always keeps the same number of tokens, whatever the model thinks.verified
- Sampling TemperatureOne dial between always taking the most likely next token and picking almost at random.verified
- Speculative DecodingGenerating one token at a time wastes most of an accelerator: the model is enormous, the work per token is tiny, and…verified
- Top-k SamplingSampling from the full distribution occasionally picks something absurd, because the tail holds thousands of tokens…verified
Memory
- Context WindowA model can only attend to what fits in its context window — the maximum number of tokens it can hold at once.verified
- KV CacheGenerating a token requires attending to everything before it.verified
- Paged AttentionA KV cache has to be contiguous, and nobody knows how long a sequence will get, so servers reserve the maximum up…verified
- Quantized KV CacheStore the attention cache in fewer bits.verified
Serving
Information8
Foundations
Interpretability8
Methods
- Activation PatchingTo find out whether a component causes a behaviour, replace its activations with the ones it had on a different input…verified
- Feature AttributionWhich parts of the input were responsible for this output? Attribution methods answer with one number per input…verified
- Probing ClassifierTrain a small classifier on a model's internal activations to predict some property — part of speech, sentiment…verified
- Sparse AutoencoderIf features are packed into a space too small for them, widen the space.verified
Knowledge8
Foundations
- Knowledge ConflictYou retrieve a document saying the population is nine million; the model learned eight.verified
- Knowledge GraphFacts written as triples — subject, relation, object — so that "Ada Lovelace" and "born in" and "London" become an…verified
- Parametric vs Retrieved KnowledgeA model knows things two ways: baked into its weights during training, and handed to it in the prompt.verified
Reasoning
- Logical ConstraintSome things must never be output: a schedule with two events in one room, a molecule that violates valence, an answer…verified
- Multi-Hop Question Answering"Who directed the film that won Best Picture the year Ada was born?" needs three facts joined in order, and no single…verified
- Neuro-SymbolicNeural networks are good at perception and bad at guarantees; symbolic systems are the reverse.verified
Landscape8
Foundations
- Loss LandscapeThe loss is a function of every weight at once, so its surface lives in a space with millions of dimensions.verified
- Mode ConnectivityTwo networks trained from different random seeds end up in what look like separate valleys — but you can often find a…verified
- Saddle PointThe old worry about non-convex optimisation was getting stuck in a bad local minimum.verified
- SharpnessTwo solutions with the same training loss can sit in very different neighbourhoods: one in a narrow valley where…verified
Generalization
Multilingual8
Foundations
Interface
Methods
Multimodal16
Foundations
- Audio-Visual LearningVideo comes with its own soundtrack, and the two are aligned for free — nobody had to label anything for the audio at…verified
- Modality FusionTwo streams of information have to become one prediction, and where they meet decides most of what the system can do.verified
- Modality ImbalanceTrain a model on two modalities and one of them usually wins.verified
Generation
- Classifier-Free GuidanceA conditioned diffusion model follows its prompt loosely.verified
- Diffusion ModelTake an image and add noise until nothing is left.verified
- Latent DiffusionRunning diffusion on raw pixels is enormously expensive, and most pixels carry no information anyone would miss.verified
- Video GenerationGenerating a video is not generating many images.verified
Objectives
Optimization26
Generalization
- Compute-Optimal TrainingGiven a fixed compute budget, you can spend it on a bigger model or on more data.verified
- OverfittingA model that keeps improving on its training data while getting worse on everything else has stopped learning and…verified
- Scaling LawsLoss falls predictably as models, data and compute grow — smoothly, along a straight line on a log-log plot, over…verified
Objectives
Regimes
- Fine-TuningA pretrained model knows a great deal but will happily continue text rather than answer a question.verified
- LoRAFine-tuning every parameter of a large model means storing a whole new model per task.verified
- PretrainingPretraining is the long, expensive first phase where a model learns language itself by predicting the next token over…verified
- Prompt TuningInstead of changing a model to fit a task, prompt tuning changes what you put in front of it — but learns those…verified
- QLoRALoRA makes the trainable parameters small, but the frozen base model still has to sit in memory at full precision…verified
Regularization
- DropoutDropout switches off a random subset of units on every training step, so no unit can rely on any particular other one…verified
- RegularizationAnything you do to a model to stop it memorising the training set.frontier
- Weight DecayA model with enormous weights has usually memorised something rather than learned it.verified
Schedules
Training
- AdamDifferent parameters need different step sizes, and nobody can tune them individually.verified
- AdamWAdamW is Adam with the weight decay applied separately from the gradient rather than folded into it.verified
- BackpropagationTo improve, a model needs to know how much each of its millions of parameters contributed to being wrong.verified
- Batch SizeBatch size is how many examples the model looks at before each update.verified
- Gradient ClippingOccasionally a batch produces an enormous gradient — a strange example, a numerical accident — and one full-size step…verified
- Gradient DescentThe gradient points in the direction the loss increases fastest, so stepping the other way makes it smaller.verified
- MomentumPlain gradient descent forgets everything between steps, so it zig-zags across narrow valleys instead of running…verified
- MuonMomentum-based optimizers take a step in whatever direction the gradient points, and for a weight matrix that…verified
- Stochastic Gradient DescentComputing the gradient over the whole dataset for every step is unaffordable.verified
- Straight-Through EstimatorA discrete step — rounding, picking an index, thresholding — has a derivative of zero almost everywhere, so gradients…verified
Planning8
Methods
- Hierarchical RLDeciding every muscle twitch and deciding where to go are different problems at different timescales.verified
- Imitation LearningRather than specify a reward, show the behaviour you want and have the model copy it.verified
- Model-Based RLModel-free methods learn what to do; model-based methods learn how the world works and then work out what to do.verified
Regimes
Platform8
Data
Practice
Serving
Privacy8
Attacks
- Gradient LeakageFederated learning sends gradients instead of data on the assumption that a gradient reveals little.verified
- Membership InferenceGiven a trained model and a record, decide whether that record was in the training set.verified
- Training Data ExtractionA language model can be made to emit sequences it memorised verbatim — addresses, keys, phone numbers — by prompting…verified
Provenance8
Cost
- Carbon AccountingEnergy is not carbon.verified
- Energy Cost of InferenceTraining happens once; serving happens every time anyone asks a question.verified
- Energy Cost of TrainingA training run is a physical event: some number of accelerators drawing some number of watts for some number of weeks.verified
Foundations
Rights
- Membership in Training DataGiven a document and a model, was that document in the model's training set? Authors want to know it about their…verified
- Output WatermarkingBias a model's sampling so its output carries a statistical signature only someone with the key can detect, without…verified
- Training Data LicensingWhether a model may be trained on a piece of text is a legal question, it is answered differently in different…verified
Reasoning8
Foundations
- Inference Scaling LawAccuracy improves predictably with compute spent at inference, in the same regular way it improves with compute spent…verified
- Reasoning TraceThe tokens a model generates on the way to an answer, treated as a first-class object rather than as output.verified
- Test-Time ComputeThere are two ways to make a model answer harder questions: build a bigger one, or let the one you have think for…verified
Selection
Recommenders8
Foundations
Regimes
Reinforcement13
Foundations
- Bellman EquationThe value of where you are is the reward you get now plus the value of where you end up.verified
- Credit AssignmentSomething good happened.verified
- Markov Decision ProcessAn agent in a state picks an action, the world moves it to a new state and pays it a reward, and the loop repeats.verified
- Value FunctionHow good is it to be here? A value function answers that with one number per state — the total reward you expect to…verified
Methods
- Actor-CriticTwo learners working together: an actor that chooses what to do and a critic that judges how good the situation is.verified
- Advantage EstimationWas this action better than what I would normally have done here? That is the advantage — the value of an action…verified
- Policy GradientRather than learning what each state is worth and acting greedily, adjust the behaviour itself: make the actions that…verified
- Q-LearningLearn a number for every state-action pair saying how good that action is there, and always update as though you will…verified
- Reward ShapingAdd extra rewards along the way so the agent gets feedback before the end.verified
- Self-PlayA system improves by playing against itself.verified
- Temporal Difference LearningWait until the end of the game to learn what your moves were worth and you learn slowly and rarely.verified
Retrieval8
Indexing
Ranking
Representation
Robotics8
Foundations
- Contact DynamicsEverything a robot does that matters involves touching something, and touching is where the physics stops being…verified
- ManipulationPicking things up is the problem that has stayed hard.verified
- ProprioceptionA robot knows where its own joints are, precisely and at high frequency, without looking.verified
Sequence8
Architecture
- Dilated ConvolutionA convolution sees a few neighbouring steps.verified
- Selective State SpaceA linear state space model treats every input the same way, which is efficient and also why it cannot decide that one…verified
- State Space ModelInstead of attending over everything you have seen, carry a fixed-size state forward and update it at each step.verified
Foundations
- SeasonalityElectricity demand has a daily shape, a weekly shape and a yearly shape, all at once.verified
- Time Series ForecastingPredict what a measured quantity does next — demand, load, price, temperature.verified
- Trend DecompositionSplit a series into the slow part, the repeating part, and what is left.verified
Objectives
Speech8
Compression
Representation
Structure8
Decoding
- Beam Search TradeoffsWidening the beam finds sequences the model likes better.verified
- Grammar-Constrained DecodingAt every step, work out which tokens could still lead to a valid output and set the probability of all the others to…verified
- Schema ConformanceThe output has to be JSON matching a declared schema — the right keys, the right types, the enum values you listed.verified
Inference
Synthesis8
Representation
Systems23
Distributed
- Collective CommunicationParallel training is mostly agreement: every worker computes something, and then they all have to end up holding the…verified
- Data ParallelismThe simplest way to use many accelerators is to put a full copy of the model on each, give each a different slice of…verified
- Expert ParallelismA mixture-of-experts model holds more parameters than one accelerator can, so the experts are spread across devices.verified
- Pipeline ParallelismGive each device a consecutive group of layers and pass activations along the chain.verified
- Sequence ParallelismTensor parallelism splits the large matrix multiplies but leaves the normalization and dropout between them…verified
- Tensor ParallelismWhen a single layer is too large for one device, split the layer.verified
- ZeROData parallelism keeps a complete copy of the optimizer state on every worker, which is the same numbers stored N…verified
Hardware
- Arithmetic IntensityEvery kernel does some arithmetic and moves some bytes.verified
- GPU Memory HierarchyThe memory closest to the arithmetic is tiny and enormously fast; the memory large enough to hold a model is far away…verified
- Kernel FusionA chain of simple operations — add a bias, apply an activation, scale the result — each reads its input from main…verified
Memory
Operations
- Determinism and ReproducibilityRun the same training script twice with the same seed and you will usually get two different models.verified
- Job SchedulingA cluster is shared, and the decision about which job runs next is where a lot of a research team's effective…verified
- ThroughputHow much work finishes per second, as opposed to how long any one piece takes.frontier
- Training ObservabilityThe failure that costs the most is the one nobody notices.verified
Reliability
- Checkpointing StrategyA run across thousands of accelerators for six weeks will be interrupted many times, so the only question is how much…verified
- Elastic TrainingIf a job must have exactly 1,024 accelerators, losing one stops it.verified
- Failure RecoveryAt the scale large models are trained, hardware failure is not an incident — it is the weather.verified
- Loss SpikeA long run is going well and then the loss jumps — sometimes recovering on its own over a few thousand steps…verified
- StragglerIn a synchronous run every worker must finish its step before any of them can start the next, so the whole cluster…verified
Tools8
Interface
Methods
Regimes
Uncertainty8
Foundations
Methods
Unsupervised8
Objectives
Representation
- AutoencoderTrain a network to output its own input, but force everything through a narrow layer in the middle.verified
- Variational AutoencoderAn ordinary autoencoder learns a latent it can decode, with no promise that the space between two latents decodes to…verified
- Vector QuantizationInstead of a continuous latent, keep a fixed dictionary of vectors and replace each encoding with whichever entry is…verified
Vision9
Dense Prediction
- Instance SegmentationSemantic segmentation says these pixels are all 'person'.verified
- Promptable SegmentationRather than training a model to segment a fixed list of categories, train one to segment whatever a person points at.verified
- Semantic SegmentationLabel every pixel with what it belongs to — road, building, sky — without distinguishing one car from the car parked…verified
Foundations
Weights8
Foundations
- Merge InterferenceMerging two models degrades both a little.verified
- Merge vs EnsembleBoth take several models and combine them; the difference is when.verified
- Model MergingTake two models fine-tuned from the same base for different things, average their weights, and get one model that…verified
- Permutation SymmetrySwap two neurons in a hidden layer, and swap the corresponding weights that read from them, and the network computes…verified
Methods
- Fisher-Weighted MergingA plain average treats every weight as equally important to every model, which is obviously false — some weights a…verified
- Model SoupFine-tune the same base many times with different hyperparameters, then average the weights of the ones that came out…verified
- Task ArithmeticSubtract a base model's weights from its fine-tune and you get a task vector — a direction in weight space that means…verified
- Weight AveragingAverage the weights a single run passes through, rather than the weights of separate runs.verified