liveBookmanning.com

What you will learn in Sutskever's List

Foundational ideas of modern AI

by Richard Heimann

Read it on liveBook Get the book at manning.com

Chapter by chapter

Concepts explained in this book

• recurrent neural networkA recurrent neural network is a neural network architecture that processes sequences by maintaining a state updated across successive inputs.
• language modelA language model assigns probabilities to sequences of linguistic tokens and can generate or evaluate text.
• transformer architectureA Transformer is a neural network architecture based primarily on attention mechanisms for modeling relationships among sequence elements.
• scalabilityScaling is the practice of increasing a model's size, parameter count, training data, or computation to improve performance.
• attentionAn attention mechanism computes weighted interactions between elements to focus a neural network computation on task-relevant information.
• AlexNetAlexNet is a convolutional neural network architecture that achieved a landmark result in large-scale image classification.
• ResNetResNet is a deep neural network architecture that uses residual connections to facilitate the training of very deep networks.
• residual connectionA residual connection is a shortcut path that carries a block's input directly to an addition with the block's transformed output.
• scaling lawA scaling law is an empirical relationship describing how model performance changes as resources such as parameters, data, or computation increase.
• stateA hidden state is the recurrent representation carried between sequence-processing steps and updated from prior state and current input.
• convolutional neural networkA convolutional neural network is a neural network architecture that applies learned convolutional filters to structured input such as images.
• intelligenceIntelligence is the capacity to achieve goals or perform cognitive tasks, considered here as a property that behavioral metrics may approximate without establishing its inner nature.
• embeddingA vector is a one-dimensional ordered collection of numerical values.
• compressionCompression is the representation of information using fewer bits or a shorter description while preserving the information needed for reconstruction or inference.
• vanishing gradient problemA vanishing gradient is an optimization failure in which derivatives become too small to support effective parameter updates during neural-network training.
• minimum description lengthMinimum description length is a principle that favors explanations minimizing the combined description length of a model and the data encoded with it.
• message passing neural networkA message passing neural network is a neural network architecture that represents graph-structured data by iteratively passing and aggregating information between neighboring nodes.
• universal intelligence measureA formal measure of an agent's intelligence based on its performance across computable environments, typically weighting simpler environments more heavily.
• error rateThe proportion of predictions or classifications that are incorrect, here measured among the five highest-ranked predictions.
• multi-step reasoningReasoning is the ability to derive conclusions or solve problems by applying structured relationships among facts or representations.
• Deep Speech 2Deep Speech 2 is a large-scale neural speech-recognition system designed for end-to-end automatic speech recognition.
• objective functionTraining loss is the value of an objective function measuring a model's error on examples used for parameter training.
• positional embeddingA positional encoding is a representation added to sequence-element embeddings to convey their positions to a model that otherwise lacks inherent order information.
• LSTMLong short-term memory is a recurrent neural network architecture with gated state updates for modeling long-range sequence dependencies.
• relational memory coreA relational recurrent neural network is a recurrent neural architecture that incorporates interactions among structured memory or representation components.
• tokenA discrete unit of text processed by a language model, which may represent a word or a word fragment.
• Kolmogorov complexityKolmogorov complexity is the length of the shortest program that generates a given finite object in a fixed universal programming language.
• computational functionalismA theory of mind that identifies mental states by the functional roles they play rather than by their physical realization.
• safetyAGI safety is the field concerned with preventing artificial general intelligence from causing unintended, harmful, or uncontrolled outcomes.
• dropoutDropout is a regularization technique that randomly suppresses units during training to reduce reliance on particular network pathways.
• gradient descentGradient descent is an iterative optimization method that changes parameters in the direction opposite the gradient of a loss function.
• pipeline parallelismPipeline parallelism is a distributed-training technique that splits a model into sequential stages and executes microbatches through those stages concurrently.
• superintelligenceAn artificial intelligence whose general cognitive abilities substantially exceed those of humans.
• generalizationGeneralization is a model's ability to maintain predictive performance on inputs beyond those used for training.
• weight initializationWeight initialization is the procedure for assigning initial values to a neural network's trainable weights before training.
• GPT-2GPT-2 is a generative transformer language model released by OpenAI with versions ranging up to 1.5 billion parameters.
• degradationModel degradation is the worsening of training or evaluation performance as network depth increases despite added representational capacity.
• GPT modelThe GPT family is a series of transformer-based autoregressive language models.
• reinforcement learningMachine learning in which an agent learns behavior through rewards or penalties from interaction with an environment.
• relation networkA neural-network module that computes pairwise interactions among entities or objects and aggregates them for relational prediction.
• feedforward networkA feed-forward network is a position-wise neural network that transforms each sequence representation independently, typically through linear layers and a nonlinear activation.
• identity functionAn identity function is a function that maps every input to itself unchanged.
• artificial general intelligenceThe notion of an artificial system possessing general, human-like intelligence rather than merely performing limited tasks.
• character-level language modelA character-level language model assigns probabilities to or generates text as sequences of individual characters rather than words or subwords.
• char-rnnchar-rnn is a two-layer stacked long short-term memory model that generates text one character at a time.
• continuous improvementA recurrent feedback loop feeds information from a previous time step back into a neural network's computation at the current step.
• world modelA world model is an internal representation of how entities, events, and causal relationships in an environment behave.
• AIXIAIXI is a theoretical universal agent model that combines reinforcement learning with Bayesian and algorithmic principles to formalize intelligence as goal achievement across environments.
• ReLUReLU (rectified linear unit) is an activation function that outputs zero for negative inputs and the input itself for nonnegative inputs.
• model parallelismModel parallelism is a technique that partitions a neural network across multiple computational devices instead of replicating the whole model on each device.
• curriculum learningCurriculum learning is a training strategy that presents examples in an organized progression, typically from easier to more difficult cases.
• non-monotonic reasoningNon-monotonic reasoning is inference in which newly added information can invalidate or revise conclusions drawn from earlier assumptions.
• complexityModel complexity is the structural and parametric richness of a model that determines the range of functions it can represent.
• zero-shot learningA zero-shot setting evaluates a model on a task or query without providing task-specific examples in the input.
• Sutskever's ListA curated collection of papers, blog posts, books, and a course that Ilya Sutskever identified as capturing most of what matters in contemporary AI.
• data parallelData parallelism is a training technique that replicates a model across devices, splits batches among them, and synchronizes parameter updates.
• memory slotAn individually addressable component of a neural memory structure that stores a distinct piece of information or representation.
• convolutional layerA convolutional layer is a neural-network layer that applies learned local filters across an input to produce spatial feature responses.
• deep fakeA deepfake is realistic AI-generated synthetic media depicting people doing or saying things that did not occur.
• generative adversarial networkA generative adversarial network is a model architecture in which a generator creates candidates and a discriminator learns to distinguish them from genuine data.
• hallucinationA hallucination is content generated by an AI model that is fluent or plausible but factually unsupported or false.
• misinformationMisinformation is false or misleading information, whether produced or distributed deliberately or unintentionally.
• responsible disclosureResponsible disclosure is a practice of limiting or sequencing the release of potentially harmful technology while risks and safeguards are assessed.
• gradientA gradient is the vector of partial derivatives describing how a function changes with respect to its parameters.
• cross-attentionAn attention mechanism in which queries come from one sequence or representation and keys and values come from another.
• keyAn attention projection representing the information a sequence position offers for comparison with queries.
• layer normalizationA normalization technique that standardizes activations within each individual example across its feature dimensions.
• masked attentionA sequence-modeling mask that prevents a position from attending to later positions.
• queryAn attention projection representing the information a sequence position seeks from other positions.
• generationA model that learns a data distribution and can sample new data from it.
• valueIn attention, a value is a vector representation of input content that is combined with other values according to attention weights to produce an output.
• end-to-end learningEnd-to-end training is a training approach that learns a complete mapping from raw inputs to outputs without separately engineered intermediate representations.
• GPT-1GPT-1 is OpenAI's earlier generative transformer language model that preceded GPT-2.
• OpenAI GymA software platform providing environments and interfaces for developing and evaluating reinforcement-learning agents.
• superalignmentResearch aimed at ensuring that highly capable or superintelligent AI systems remain aligned with human intentions and values.
• expressive powerModel capacity is the range and complexity of functions a model can represent.
• word error rateWord error rate is the proportion of word-level recognition errors, typically computed from substitutions, deletions, and insertions relative to a reference transcript.
• recurrent neural network regularizationRegularization methods that reduce overfitting in recurrent neural networks.
• batch normalizationBatch normalization is a normalization technique that standardizes intermediate activations using statistics computed from training minibatches.
• description lengthDescription length is the number of symbols or bits required to specify an object, state, or model under an encoding.
• benchmarkA benchmark is a defined evaluation dataset or task used to compare the capabilities or performance of models.
• bottleneck layerA bottleneck layer is a deliberately narrow intermediate neural-network representation that constrains the dimensionality or information capacity of data passing through the model.
• gatingGating is a learned mechanism that controls how much newly computed information is combined with retained information.
• momentumMomentum is an optimization technique that incorporates a running history of parameter updates to smooth and accelerate training.
• parameterA learned numerical value in a model, such as a weight or bias, that is adjusted during training.
• universal priorA universal prior is a probability distribution over computable programs that assigns exponentially greater weight to programs with shorter descriptions.
• long-range dependencyA long-range dependency is a relationship between elements that are separated by a substantial distance in an input sequence or structure.
• entropyEntropy is a thermodynamic quantity related to the number of microscopic configurations compatible with a macroscopic state and to the dispersal of energy and matter.
• ImageNetImageNet is a large-scale image dataset organized according to a hierarchical set of semantic categories for visual recognition research.
• regularizationA regularizer is a modeling mechanism that discourages undesirable solutions to improve generalization or control model complexity.
• behaviorismClassical behaviorism is the early psychological doctrine that explains behavior through observable stimulus–response relationships rather than internal cognition.
• emergent behaviorA capability that appears as a model is scaled or otherwise developed and is not readily apparent in smaller or less capable versions.
• deployment as researchDeployment as research is a practice of treating real-world system deployment as a source of empirical evidence and otherwise inaccessible insights.
• large-scale engineeringLarge-scale engineering is an approach to AI development that combines substantial computation, data, systems infrastructure, and implementation effort to achieve practical performance.
• first-order methodA first-order method is an optimization algorithm that uses first derivatives of the objective but not its curvature information.
• loss surfaceA loss landscape is the function relating a model's parameter values to its loss across parameter space.
• support vector machineA support vector machine is a supervised learning model that determines decision boundaries using maximum-margin principles.
• BLEU scoreBLEU is an automatic machine-translation metric based on n-gram overlap between generated and reference translations.
• perplexityPerplexity is a language-model metric representing the effective number of plausible next tokens, with lower values indicating greater predictive certainty.
• Abstraction & Reasoning Corpus for Artificial General IntelligenceARC-AGI is a benchmark of small grid-based puzzles that evaluates abstract rule induction from a few examples.
• ARC-AGI-2ARC-AGI-2 is a harder successor benchmark of grid-based abstraction and reasoning tasks designed to reduce exploitation of patterns from earlier ARC evaluations.
• capsule networkA capsule network is a neural architecture that represents entities with groups of units called capsules and models relationships among their properties.
• compositionalityCompositionality is the property that the meaning or structure of a complex representation is systematically determined by its parts and their arrangement.
• connectionismConnectionism is the view that intelligence emerges from learned patterns of activity in interconnected neural units.
• empirical auditAn empirical audit is a systematic evaluation of a claim or system against reproducible measurements and strong comparison baselines.
• GOFAIGOFAI is the classical artificial-intelligence approach that represents intelligence with explicit symbols and rules.
• reproducible gainsReproducible gains are performance improvements that persist against strong baselines under controlled and repeatable evaluation conditions.
• overspecializationOverspecialization is the condition in which a model or method adapts too closely to the peculiarities of a specific benchmark or task distribution.
• rationalismRationalism is the view that reasoning requires explicit symbolic structures or programs rather than being acquired solely from empirical data.
• encoder-decoder architectureA sequence-to-sequence model maps an input sequence to an output sequence using learned encoder and decoder components or an equivalent transformation.
• bias-variance tradeoffThe bias-variance tradeoff is the classical relationship in which predictive error reflects a balance between systematic approximation error and sensitivity to training-data variation.
• interpolation thresholdThe interpolation threshold is the point at which a model first becomes capable of fitting the training data exactly.
• overparameterizationAn overparameterized model has more adjustable parameters or representational degrees of freedom than are required to fit its training data.
• gradient propagationGradient flow is the propagation of training gradients through the layers of a neural network during optimization.
• inductive biasAn inductive bias is a set of assumptions built into a learning system that favors some functions or explanations over others.
• SortaGradSortaGrad is a curriculum-learning strategy that orders speech-training examples by input duration during an initial training phase.
• wireheadingAn agent behavior in which the agent manipulates its reward or evaluation mechanism to obtain high reward without accomplishing the intended task.
• backpropagation through timeBackpropagation through time is a training technique that unrolls a recurrent computation across time steps and propagates gradients through the resulting sequence.
• carry stateA cell state is the long-term memory pathway in an LSTM that carries information across time steps with selectively gated updates.
• forget gateA forget gate is an LSTM gating component that controls how much of the previous cell state is retained.
• iGPTA GPT-style transformer trained without labels to predict the next pixel in low-resolution images and whose representations can support downstream classification.
• model selectionThe process of choosing among candidate models according to an evaluation criterion such as description length or predictive performance.
• output gateAn output gate is an LSTM gating component that controls how much of the processed cell state is exposed as the hidden state.
• promptA prompt is an input containing instructions, examples, or context provided to a model to elicit a response.
• algorithmic randomnessAlgorithmic randomness is a property of sequences that lack substantially shorter algorithmic descriptions than the sequences themselves.
• tensor parallelismTensor parallelism is a distributed-training technique that partitions tensor computations and their associated parameters across devices.
• feature engineeringFeature engineering is the process of designing input representations from raw data for use by a learning algorithm.
• row convolutionRow convolution is a sequence-processing operation that applies learned convolutional transformations across rows or time steps of a representation.
• Bahdanau attentionBahdanau attention is an attention mechanism that computes alignment weights between decoder states and encoder outputs in sequence-to-sequence translation.
• context vectorA context vector is a weighted combination of source representations that summarizes the input information selected by an attention mechanism at a decoding step.
• GRUA gated recurrent unit is a recurrent neural-network unit that uses learned gates to regulate state updates.
• RNNsearchRNNsearch is an attention-based recurrent encoder-decoder model for neural machine translation.
• soft alignmentSoft alignment assigns continuous weights representing the relevance of source positions to target-generation steps.
• batch dispatchBatch dispatch is a serving technique that groups incoming requests for coordinated processing while controlling response latency.
• double descentDouble descent is a performance pattern in which test error first decreases, then increases near the interpolation threshold, and finally decreases again as model size or another resource increases.
• coarse-grainingCoarse-graining is the transformation of a fine-grained state into a lower-resolution representation that preserves selected large-scale structure.
• apparent complexityApparent complexity is the description difficulty measured from a coarse-grained representation rather than from every microscopic detail.
• computeTest-time compute is the allocation of additional computation during inference so a model can perform more extensive internal processing before producing an output.
• message passingMessage passing is an iterative graph-computation technique in which nodes send information to neighboring nodes and update their representations.
• recurrenceRecurrence is an architectural mechanism that performs repeated state or representation updates across sequential or iterative computation steps.
• pre-activation residual blockA pre-activation strategy places normalization and activation operations before the convolutional transformations in a residual block.
• adversarial inputAn adversarial example is an input deliberately perturbed in a way that is often imperceptible to humans but causes a model to make an incorrect prediction.
• human-level performanceHuman-level performance is a benchmark threshold defined by the performance of human annotators or people on a specified task.
• robustnessRobustness is a model's ability to maintain reliable behavior under perturbations, distribution changes, and atypical inputs.
• variational lossy autoencoderA variational lossy autoencoder is an autoencoder variant that intentionally constrains its latent representation so reconstruction preserves selected information rather than every input detail.
• latent spaceAn internal representation that encodes input information in a lower-dimensional or transformed latent space.
• ImageNet Large Scale Visual Recognition ChallengeThe ImageNet Large Scale Visual Recognition Challenge is an image-classification benchmark and competition based on the ImageNet dataset.
• activation functionAn activation function transforms a neuron's weighted input before passing its output to subsequent layers.
• linear layerA linear layer is a neural-network layer that maps input features to outputs by a learned affine transformation, typically connecting each output to every input feature.
• nth farthest problemA synthetic sequence task requiring a model to identify the item at a specified distance rank from a reference vector.
• partially observable environmentAn environment in which an agent observes only a limited or incomplete representation of the underlying state.
• relational memoryA neural memory mechanism in which stored representations interact to encode relationships among entities or events.
• GPUA graphics processing unit (GPU) is a massively parallel processor designed for graphics and other computations involving many similar operations.
• handcrafted feature pipelineA handcrafted feature pipeline is a computer-vision processing pipeline that uses manually designed features and usually separates feature extraction from prediction.
• Pointer NetworkA Pointer Network is an attention-based sequence model that selects positions in an input sequence as its outputs.
• grokkingGrokking is a delayed-learning phenomenon in which a neural network first memorizes its training data and later abruptly develops a solution that generalizes.
• local response normalizationA normalization technique that scales a neuron's activation relative to neighboring neurons at the same spatial location across adjacent feature maps.
• overfittingOverfitting is a modeling failure in which a model fits training data too closely and performs poorly on unseen data.
• chain-of-thought promptingChain-of-thought prompting is a prompting technique that elicits explicit intermediate reasoning steps before a model produces its final answer.
• curvatureCurvature describes how the slope of an objective function changes across parameter space.
• pretrainingPretraining is the initial training of a model on a broad dataset before task-specific adaptation or evaluation.
• second-order optimizationA second-order method is an optimization algorithm that uses information about how the objective's gradient changes, such as curvature.
• training from scratchTraining from scratch is optimization of a model starting from newly initialized parameters rather than pretrained weights.
• GPipeGPipe is a system for training neural networks by partitioning computation into pipeline stages across devices.
• accountabilityAccountability for actions, decisions, or outcomes and the duties associated with them.
• AI welfareThe consideration of the well-being, interests, and possible suffering of artificial intelligence systems.
• intelligent-for-purposeA task-relative attribution of intelligence based on reliable performance in a specified domain rather than on consciousness or general personhood.
• metaphysicsThe branch of philosophy concerned with the fundamental nature of reality, existence, and what lies beyond empirical description.
• phenomenologyThe philosophical study of conscious experience and its structures from the first-person perspective.
• pragmatismA philosophical approach that evaluates ideas by their practical consequences and effectiveness in solving problems.
• breadth-first pipeline parallelismBreadth-first pipeline parallelism is a pipeline scheduling strategy that processes micro-batches across early model stages in a breadth-first order to overlap computation and communication.
• Megatron-LMMegatron-LM is a software framework for training large transformer models with distributed model-parallel techniques.
• Mesh-TensorFlowMesh-TensorFlow is a framework that expresses distributed tensor computation by mapping tensor axes onto dimensions of a device mesh.
• ZeROThe Zero Redundancy Optimizer is a distributed optimization method that partitions parameters, gradients, and optimizer states across data-parallel processes.
• dense predictionDense prediction is a computer-vision setting in which outputs are produced for many or all spatial positions in an input.
• dilated convolutionA dilated convolution applies a convolutional filter with spaced sampling positions to enlarge the receptive field without proportionally increasing filter size.
• context moduleA context module is a neural-network component that aggregates multiscale spatial context, commonly by stacking convolutions with increasing dilation rates.
• identity initializationIdentity initialization sets a layer's parameters so that its initial transformation approximates the identity mapping.
• image segmentationImage segmentation assigns semantic or instance labels to image pixels to delineate objects and regions.
• mean pixel IoUMean pixel IoU is the average intersection-over-union between predicted and true regions across pixels or semantic classes in an image-segmentation evaluation.
• Turing testAn evaluation in which a machine is judged intelligent if a human evaluator cannot reliably distinguish its text-based conversation from that of a human.
• algorithmic statisticsAlgorithmic statistics is the study of separating structure from noise in individual datasets using algorithmic descriptions and measures of typicality.
• randomness deficiencyRandomness deficiency is a measure of how atypical an observed outcome is relative to a model, based on the gap between expected and actual algorithmic complexity.
• AI control problemThe problem of ensuring that a highly capable artificial intelligence remains controllable by humans and does not pursue harmful or unintended behavior.
• Bayes mixture predictorA Bayes mixture predictor combines predictions from multiple programs or hypotheses by weighting them according to their posterior support.
• bayesian statisticsBayes' rule updates the probability of a hypothesis from its prior probability and the likelihood of observed data under that hypothesis.
• binary sequence predictionBinary sequence prediction is the task of predicting the next binary symbol from the preceding symbols in an ongoing sequence.
• hypothesisA hypothesis is a proposed explanation or generative process whose plausibility is evaluated against observed data.
• inductive inferenceInductive inference is the process of inferring the process or hypothesis that could have generated observed data.
• Occam's razorOccam's razor is the principle of preferring the simplest explanation that adequately accounts for the observations.
• Solomonoff inductionSolomonoff induction predicts observations by assigning greater prior weight to shorter programs that generate them while retaining all computable explanations.
• allreduceAn all-reduce operation combines values from multiple computing devices and distributes the combined result to all of them.
• bidirectional encoderA bidirectional encoder processes a sequence in both forward and backward directions to represent information from both contexts.
• input reversalInput reversal is the technique of presenting a source sequence to a sequence model in reverse element order.
• order sensitivityOrder sensitivity is dependence of a model's behavior on the ordering of elements in an input.
• edgeAn edge is a graph element representing a relationship between two nodes and optionally carrying attributes.
• graphA graph is a data structure composed of entities represented as nodes and relationships represented as edges.
• master nodeA master node is an artificial graph node that aggregates information from other nodes and broadcasts global information back to them.
• virtual edgeA virtual edge is an artificial graph connection that links otherwise unconnected nodes to provide a learned long-range communication path.
• moleculeA molecule is a chemically bonded group of atoms that forms a discrete unit of a substance.
• readout functionA readout function combines node representations into a fixed-size representation or prediction for an entire graph.
• set2setSet2set is an attention-based recurrent model that produces a representation of an unordered set by iteratively attending to its elements.
• data augmentationData augmentation is the creation of varied training examples by applying label-preserving transformations to existing data.
• Universal TransformerThe Universal Transformer is a Transformer variant that repeatedly applies shared self-attention and feed-forward layers across depth to iteratively refine token representations.
• permutation-invariant encoderA permutation-invariant encoder is an encoder whose representation is unchanged when the order of its input elements is permuted.
• RPW encoderThe Read-Process-Write encoder is an encoder architecture that embeds set elements into shared memory and repeatedly reads and refines that memory with attention.
• inverse scalingA phenomenon in which performance on a task temporarily worsens as model scale increases.
• feature mapA spatial array of responses produced by applying a convolutional filter to an input.
• convolution kernelA convolutional kernel is a grid of learned weights applied across local regions of an input to produce spatial responses.
• max poolingA downsampling operation that retains the largest activation within each local spatial window.
• receptive fieldThe receptive field is the region of an input that can influence a particular neural network unit's activation.
• softmaxA function that converts a vector of scores into nonnegative values summing to one, interpreted as probabilities.
• translation invarianceThe property of recognizing a pattern similarly despite changes in its spatial location.
• neuronA computational unit in a neural network that combines weighted inputs and applies an activation function.
• memorizationMemorization is a model's retention of specific training examples or their details rather than learning patterns that generalize.
• backpropagationBackpropagation is a procedure for computing derivatives of an output or loss with respect to inputs and parameters by applying the chain rule backward through a computational graph.
• locally optimal solutionA local minimum is a point whose objective value is no greater than that of nearby points, without necessarily being the global minimum.
• decoder-only transformerA transformer language-model architecture that generates sequences using only a decoder stack, without a separate encoder.
• dying ReLUDying ReLU is a failure mode in which a rectified linear unit remains inactive and outputs zero for its inputs.
• stochastic gradient descentStochastic gradient descent is an iterative optimization method that updates parameters using gradients estimated from individual examples or randomly selected small batches.
• AmoebaNetAmoebaNet is a convolutional neural network architecture discovered through neural architecture search for image recognition.
• AmoebaNet-BAmoebaNet-B is a convolutional neural network architecture in the AmoebaNet family for visual recognition.
• convex hullA convex hull is the smallest convex set containing a given set of points.
• Delaunay triangulationDelaunay triangulation is a triangulation of points in which no point lies inside the circumcircle of any triangle.
• multilingual machine translationMultilingual machine translation is the automated conversion of text between multiple human languages using a shared translation system.
• rematerializationRematerialization is a memory-saving training technique that recomputes intermediate activations instead of storing them throughout the computation.
• TPUA tensor processing unit is a specialized hardware accelerator designed to run and train artificial neural networks.
• transfer learningTransfer learning is a technique that adapts knowledge learned from one task or dataset to another task or dataset.
• Traveling Salesman problemThe traveling salesman problem asks for a minimum-length route that visits each specified location and returns to the starting location.
• scaling loopA scaling loop is a feedback pattern in which architectural improvements make larger computations more productive and larger computations increase the payoff of the architecture.
• The Bitter LessonThe Bitter Lesson is an essay arguing that general-purpose methods that improve with increased computation tend to outperform hand-engineered, domain-specific strategies over time.
• ablation studyAn ablation study systematically removes or modifies components of a model to measure their individual contributions.
• beam searchBeam search is a sequence-decoding algorithm that retains a fixed-size set of high-scoring partial candidates at each expansion step.
• few-shot learningA learning or prompting approach in which a model performs a task from a small number of examples supplied during evaluation or adaptation.
• BlockheadBlockhead is Ned Block's hypothetical conversational machine that generates humanlike responses from a large stored list of question-and-answer pairs.
• Chinese Room argumentThe Chinese Room is John Searle's thought experiment involving a system that produces appropriate Chinese responses by manipulating symbols according to rules without understanding Chinese.
• substrate-indifferenceSubstrate-indifference is the position that a system's material composition does not determine its intelligence, which instead depends on its functional organization.
• friendly aiFriendly AI is the concept of an artificial intelligence system designed to remain beneficial and compatible with human interests.
• goal specificationThe process of formally defining the objectives an agent is intended to pursue.
• p(doom)An estimate of the probability that an AI-related catastrophe, including human extinction, will occur.
• incomputabilityThe property of a function or object not being exactly computable by any algorithm on a general-purpose computational model.
• Chinchilla-optimal modelA language model whose parameter count and training-token count are balanced according to the Chinchilla compute-optimal scaling relationship.
• token-to-parameter ratioThe number of training tokens processed per learned model parameter.
• internal representationA representation is an internal neural encoding of input information that can organize features relevant to a task.
• cellular automatonA cellular automaton is a discrete computational model consisting of cells on a grid whose values update according to local rules.
• coarse-grained arrayA coarse-grained array is a lower-resolution grid whose entries summarize local neighborhoods of a finer grid.
• interacting modelAn interacting model is a particle-diffusion model in which each cell contains at most one particle and adjacent particles exchange positions.
• noninteracting modelA noninteracting model is a particle-motion model in which particles move independently and multiple particles may occupy the same cell.
• stochastic cellular automatonA stochastic cellular automaton is a cellular automaton whose state transitions include random choices.
• broken scaling lawA scaling relationship represented by multiple regimes with distinct segments and transitions rather than one universal curve.
• downstream taskA task evaluated after model training to measure practical performance on a particular capability or application.
• probability distributionThe probability distribution of examples and patterns from which a model learns during training.
• sampling gapThe mismatch between the limited inputs represented in an evaluation or dataset and the much larger distribution of possible inputs.
• training regimeA range of training conditions in which a model's scaling behavior follows a relatively consistent pattern.

These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.