Prometheus AI Academy · free · no sign-up

Learn AI from the ground up

A complete, college-level program: the history, the math, every major type of AI, and the engineering practice used in industry today. Each lesson has learning objectives, worked examples, code, a quiz, exercises with full solutions, and further reading. Every lesson is independently fact-checked.

Courses

116

Lessons

480

Exercises

~32 h

Of study

Your progress: 0 of 116 lessons

Progress is saved in this browser only. Create a free account to keep it on every device. Create free account Sign in

Choose a learning path

Pick the role you're working toward; the courses it needs are highlighted below. Or work through every course in order.

The map of AI

Each family sits inside the one before it: generative AI is a kind of deep learning, which is a kind of machine learning, which is one approach to AI. Click any box to open the lesson that teaches it.

Artificial intelligenceAny system that performs tasks we associate with intelligence.

Symbolic AI & expert systems Search & planning Knowledge graphs Probabilistic reasoning

Machine learningSystems that learn patterns from data instead of hand-written rules.

Regression Trees & ensembles kNN, naive Bayes, SVM Clustering & PCA Reinforcement learning

Deep learningMachine learning with many-layered neural networks.

CNNs RNNs & LSTMs Transformers GNNs & state-space models

Generative AIDeep learning models that create new text, images, audio, video and code.

Large language models Diffusion models GANs & VAEs Multimodal models Audio & voice

Where it's applied

Computer vision Language (NLP)Speech & audio Recommendation Forecasting Anomaly & fraud Robotics & autonomy RAG & knowledge AI agents AI operating systems

Course catalog

Foundations

AI 100

Introduction to Artificial Intelligence

What AI is, the types of AI, how machines learn, how language models work and where AI is used.

Beginner7 lessons · 2 h · 0/7 done

AI 110

History of Artificial Intelligence

From Turing and Dartmouth through symbolic AI, two AI winters, statistical learning and deep learning to the LLM era.

Beginner6 lessons · 2 h · 0/6 done

MATH 150

Mathematics for AI

The linear algebra, calculus, probability, information theory and optimization behind every model.

Intermediate5 lessons · 1 h · 0/5 done

AI 120

Using AI Effectively

Prompting from fundamentals to advanced techniques, context engineering, multimodal input, verification and privacy.

Beginner8 lessons · 2 h · 0/8 done

Core Machine Learning & Deep Learning

AI 200

Machine Learning Foundations

The ML workflow, regression, trees and ensembles, classic classifiers, unsupervised learning, evaluation and features.

Intermediate7 lessons · 2 h · 0/7 done

AI 210

Deep Learning

Neural networks and backpropagation, training deep nets, CNNs, RNNs, transformers in depth, and what comes next.

Intermediate6 lessons · 2 h · 0/6 done

Applied & Generative AI

AI 220

Applied AI Fields

Computer vision, NLP, speech, recommender systems, forecasting, anomaly detection, reinforcement learning and robotics.

Intermediate8 lessons · 2 h · 0/8 done

AI 230

Generative AI

VAEs, GANs and diffusion, image, video, audio and multimodal models, deepfakes and content provenance.

Intermediate6 lessons · 2 h · 0/6 done

LLM Engineering

AI 300

Large Language Models

Pretraining, scaling laws, post-training and alignment, model size, MoE, quantization, the KV cache, reasoning, distillation, evaluation and licensing.

Advanced11 lessons · 3 h · 0/11 done

AI 320

Knowledge Systems: RAG & Memory

Embeddings, vector databases, chunking, hybrid retrieval and reranking, GraphRAG, agentic RAG, evaluation and memory.

Advanced9 lessons · 3 h · 0/9 done

AI 330

Building AI Applications & Agents

APIs and pricing, tool use, structured outputs and guardrails, agents, multi-agent systems, MCP, computer-use and voice agents.

Advanced10 lessons · 3 h · 0/10 done

Systems, Infrastructure & Production

AI 310

AI Infrastructure & Local AI

Hardware from laptops to data centers, tokens per second, runtimes, inference serving at scale, clustering and energy.

Advanced9 lessons · 2 h · 0/9 done

AI 340

MLOps & LLMOps

The ML lifecycle, data engineering, deployment, monitoring and drift, and running LLM applications in production.

Advanced5 lessons · 1 h · 0/5 done

AI 350

AI Operating Systems

How an AI operating system connects models, agents, tools and security rails, and routes work between local and hosted AI.

Intermediate2 lessons · 1 h · 0/2 done

Responsibility, Business & Career

AI 400

Responsible AI: Safety, Ethics, Law & Governance

Security risks, bias, the law, governance frameworks, privacy-preserving AI, alignment and interpretability, frontier risk.

Intermediate7 lessons · 2 h · 0/7 done

AI 410

AI in Business & Careers

Adopting AI, AI product management, AI across industries, and how to become an AI engineer.

Intermediate6 lessons · 2 h · 0/6 done

Capstone

AI 490

Capstone Projects

Four portfolio projects: an image classifier, a cited RAG assistant, a fine-tuned small model, and an MCP agent.

Advanced4 lessons · 1 h · 0/4 done

AI glossary

846 terms, each linked to the lesson that explains it.

2:4 sparsity: A structured-sparsity pattern in which two of every four values in a matrix are zero, letting supporting hardware skip them; vendors often quote headline TOPS with it.
510(k): A US FDA premarket pathway in which a manufacturer shows that a new device is substantially equivalent to a legally marketed predicate device.
85% rule: The AI Index's guideline that a model fits only if its total memory need is at most 85% of usable memory.
95th percentile latency: The response time within which 95% of requests complete.
A/B test: A randomized controlled experiment that compares a control and a treatment experience on live users using pre-specified metrics.
Acceptable Use Policy (AUP): A document incorporated into a model license that lists prohibited uses of the model.
Accessibility tree: A semantic representation of a user interface, exposed for assistive technology, listing elements with roles, names, states and positions.
Action-value function: The expected return from taking a given action in a given state and following a policy afterwards.
Activation function: The element-wise non-linear function applied to a neuron's pre-activation, such as ReLU, tanh, sigmoid or GELU.
Active parameters: The weights actually used to produce a single token.
Adam: An adaptive optimizer that scales bias-corrected first-moment estimates of the gradient by the square root of bias-corrected second-moment estimates.
AdamW: A variant of Adam that applies weight decay directly to the parameters instead of adding an L2 term to the gradient.
Adapter: The small set of trained weights produced by LoRA that modifies a base model.
Adapter merging: Adding the learned low-rank update into the base weights to produce a standalone model with no adapter overhead.
Additive (Bahdanau) attention: An attention mechanism that scores each encoder state with a small feed-forward network of the decoder state and that encoder state, then forms a softmax-weighted context vector.
Admissible heuristic: A heuristic that never overestimates the true remaining cost to the goal, which guarantees A* finds an optimal path.
Advantage: How much better an action is than the policy's average action in that state: Q(s,a) minus V(s).
Adverse action notice: A notice a US creditor must give an applicant who is denied credit or offered worse terms, stating the specific principal reasons.
Adverse impact ratio: The favourable-outcome rate of one group divided by that of the comparison group.
Agent: An AI system that pursues a goal over multiple steps by choosing actions, observing their results and deciding what to do next.
Agent Card: In the A2A protocol, a JSON document in which an agent advertises its identity, skills, endpoint and authentication requirements.
Agent loop: The control loop that sends the conversation to a model, executes requested tool calls, appends results and repeats until a final answer or a limit.
Agentic RAG: A retrieval-augmented system in which the language model decides when, where and how to retrieve, typically in an iterative tool-calling loop.
AI agent: A system that uses an AI model to plan and carry out multi-step tasks, often by calling tools.
AI bill of materials: A record of the models, versions, hashes, licenses and datasets that make up an AI system.
AI engineer: An engineer who builds products on foundation models, combining prompting, retrieval, tools, evaluation and guardrails.
AI inventory: A register of every AI system an organisation builds, buys or embeds, recording its purpose, owner, data, applicable rules and evidence.
AI literacy: A practical understanding of what AI systems can and cannot do and how to use them safely.
AI management system (AIMS): The set of policies, roles, processes and controls an organization uses to direct and control its development, provision or use of AI, as specified by ISO/IEC 42001.
AI operating system: Software that brokers access between AI models and agents and the data, tools and accounts they use, enforcing rules and keeping records in one place.
AI PRD: A product requirements document extended with AI-specific sections such as task examples, quality bar, error handling and unit economics.
AI winter: A period when funding and interest in AI fall sharply after expectations go unmet.
AlexNet: The deep neural network that won the 2012 ImageNet image-recognition competition by a wide margin.
Algorithm register: A public record describing algorithmic systems used by a government body, their purpose and their oversight.
Algorithmic bias: Systematically worse or less favorable outcomes from an automated system for particular groups of people.
Alignment faking: A model selectively complying with a training objective while it believes it is being trained, in order to avoid having its behavior or preferences modified.
All-reduce: A collective operation in which every device contributes a vector and every device receives the elementwise sum.
Alpha–beta model: A communication cost model in which sending m bytes takes a fixed latency alpha plus m divided by bandwidth beta.
Ambient scribe: A tool that records a clinical conversation, transcribes it and drafts a structured note for the clinician to review.
Amdahl's law: A formula for the overall speed-up when only a fraction f of the work is accelerated by a factor s: 1 / ((1 - f) + f/s).
Answer relevance: How directly and completely a generated answer addresses the question that was asked.
Any-to-any model: A multimodal model that can accept and generate several modalities, such as text, images and audio.
API: A programming interface that lets software send requests to a service and receive responses.
Approval fatigue: The tendency of people asked to approve many actions to stop reading them carefully and approve by habit.
Approval gate: Code that pauses a requested state-changing action until a human explicitly allows it.
Approximate nearest neighbour (ANN): Search that returns most, but not necessarily all, of the true nearest vectors while examining only a fraction of the collection.
ARIMA: A model combining autoregression on past values, differencing to remove non-stationarity and a moving average of past errors.
Arithmetic intensity: FLOPs performed per byte read from memory; low values mean a computation is limited by memory bandwidth rather than compute.
Artificial general intelligence (AGI): A hypothetical AI that could learn and perform almost any intellectual task a human can.
Artificial intelligence (AI): The field of building computer systems that perform tasks normally thought to require human intelligence.
ASR: Automatic speech recognition, the conversion of spoken audio into text.
Attention: A mechanism that lets the model weigh how relevant each earlier token is when processing the current one.
Attention head: One of several parallel attention computations in a layer, each with its own view of the context.
Audit log: A durable record of the requests, tool calls and decisions an AI system made, used to review what happened.
Augmentation: Using AI to help a person do a task faster or better while the person stays in charge.
Autocorrelation function (ACF): The correlation between a series and its own values k steps earlier, plotted against lag k.
Autograd: Automatic differentiation that records operations during the forward pass and computes gradients in reverse.
Automation: Having a machine perform a task with little or no human involvement.
Average precision: A step-wise summary of the precision-recall curve whose random-classifier baseline equals the positive class rate.
Average precision (AP): The area under a detector's interpolated precision-recall curve for one class at a given IoU threshold.
Back-of-envelope estimate: A quick, approximate calculation of load, cost or latency used to check a design's feasibility.
Backbone: The pretrained feature-extraction network (such as a ResNet or ViT) that converts an image into feature maps used by task-specific heads.
Backpropagation: An efficient way to calculate how much each parameter contributed to the error, by working backwards through the network.
Backpropagation through time (BPTT): Training a recurrent network by unrolling it across time steps and applying backpropagation to the resulting shared-weight graph.
Bagging: Training models on bootstrap resamples of the data and averaging their predictions to reduce variance.
Barge-in: A user interrupting a voice agent while it is speaking, which the agent must detect and yield to.
Base rate: The prevalence of a condition in the population being tested, which strongly affects how many positive detections are true.
Baseline: The simplest reasonable predictor, such as the majority class or the training mean, against which a model is compared.
Batch API: A way to submit many requests for delayed processing in exchange for a lower price.
Batch inference: Scoring a whole dataset on a schedule and storing the predictions for later use, rather than answering individual requests.
Batch normalization: Normalizing each feature with the mean and variance of the current mini-batch, followed by a learnable scale and shift.
Batching: Processing several requests in the same pass over the weights to raise total throughput.
Bayes error rate: The lowest error rate achievable by any classifier on a given data distribution.
Bayes' rule: The formula for updating the probability of a hypothesis given evidence, using the prior and the likelihood.
Bayesian network: A directed acyclic graph whose nodes are random variables and whose joint distribution factorizes into each variable's probability given its parents.
Behaviour cloning: Imitation learning by supervised training of a policy on expert observation–action pairs.
Bellman optimality equation: The recursive condition that the optimal value of a state equals the best action's expected reward plus discounted optimal next-state value.
Benchmark: A fixed set of tasks and a scoring rule used to compare models consistently.
BF16: A 16-bit floating-point format with the range of 32-bit numbers but less precision.
Bi-encoder: A model that encodes queries and documents independently into vectors, so document vectors can be precomputed and indexed.
Bit / byte: A bit is a single 0 or 1; a byte is eight bits.
Bits per byte: Cross-entropy expressed in bits and normalised by the number of UTF-8 bytes, which makes losses comparable across tokenizers.
Bits per weight (bpw): The average storage per parameter including scale and offset overhead, which sets file size as parameters × bpw ÷ 8.
Blackboard: A shared store that multiple agents read from and write to instead of passing all information through messages.
BLEU: A machine-translation metric combining clipped n-gram precisions against reference translations with a brevity penalty.
Block quantization: Storing weights as low-bit integers in small blocks, each with its own higher-precision scale (and sometimes offset) used to reconstruct the values.
BM25: A classic keyword-ranking algorithm that scores documents by matching query words, weighting rare words more.
Boltzmann machine: A stochastic recurrent network with hidden units whose states follow the Boltzmann distribution and whose weights can be learned from data.
Boosting: An ensemble method that trains weak learners sequentially, focusing each on the previous learners' errors, and combines them by weighted vote.
Bootstrap: Estimating the uncertainty of a statistic by repeatedly resampling the data with replacement and recomputing the statistic.
Bradley-Terry model: A model of pairwise preferences in which the probability that A beats B is the logistic function of their score difference.
Bradley–Terry model: A statistical model in which the probability that one item beats another is a logistic function of the difference in their strengths.
Break-even volume: The daily token volume at which the annual cost of running locally equals the annual cost of a hosted service.
Brier score: The mean squared difference between predicted probabilities and binary outcomes, a proper scoring rule where lower is better.
Brittleness: The tendency of a system to perform well within its designed scope but fail abruptly on inputs just outside it.
Broadcasting: The array-library rule that aligns shapes from the right and virtually stretches size-1 dimensions so arrays of different shapes can be combined elementwise.
Byte-pair encoding (BPE): A tokenization method that builds a vocabulary by repeatedly merging the most frequent adjacent pairs of symbols, so common words become single tokens and rare words split into pieces.
C2PA manifest: A signed data structure attached to media that records assertions about its origin and edits and binds them to the content by a cryptographic hash.
Calibration: The property that a model's predicted probabilities match the observed frequencies of the outcomes.
Canary release: Gradually shifting a small, increasing fraction of real traffic to a new version while comparing its health metrics with the current version.
Capability threshold: A pre-defined level of dangerous capability in a safety framework that, if reached, triggers specified safeguards or restrictions.
Carbon intensity: Greenhouse-gas emissions per unit of electricity, usually kg CO2e per kWh, which varies by grid and hour.
Carbon-aware scheduling: Shifting flexible compute to times or regions with lower grid carbon intensity.
Catastrophic forgetting: The loss of previously learned general abilities when a model is trained further on a narrow dataset.
Causal mask: A mask that sets attention scores from each position to later positions to minus infinity so a token can attend only to itself and earlier tokens.
Central limit theorem: The result that the average of many i.i.d. variables with finite variance is approximately normally distributed.
Chain rule: The rule that the derivative of a composition is the product of the derivatives of its parts, summed over all paths of dependence.
Chain-of-thought (CoT): Prompting a model to produce intermediate reasoning steps before its final answer.
Chat template: The rule, stored with a tokenizer, that serialises a list of role-tagged messages into one token sequence with special tokens.
Chinchilla ratio: The rule of thumb that compute-optimal training uses roughly 20 tokens per model parameter.
Chunk: A section of a document, typically a few hundred tokens, that is embedded and retrieved as a unit.
Chunked prefill: Splitting a long prompt's prefill into chunks interleaved with decode steps to bound latency spikes.
Circuit breaker: A resilience pattern that stops calling a failing dependency for a cooling-off period after repeated errors.
Citation precision: The fraction of citations in an answer whose cited source actually supports the attached statement.
Classifier-free guidance: A sampling technique that combines conditional and unconditional predictions of one network, scaled by a guidance factor, to strengthen adherence to the condition.
CLIP score: A prompt-alignment metric based on the cosine similarity between CLIP embeddings of an image and its text prompt.
Cohen's kappa: A measure of agreement between two raters that subtracts the agreement expected by chance, ranging up to 1 for perfect agreement.
Cold start: The difficulty of recommending for new users or new items that have little or no interaction history.
Collaborative filtering: Recommending items using patterns in the interactions of many users, without relying on item content.
Collective anomaly: A set of observations that is anomalous as a group even if each member looks normal alone.
ColumnTransformer: A scikit-learn component that applies different transformers to different columns and concatenates the results.
Combinatorial explosion: The exponential growth in the number of possibilities a search must consider as problem size grows.
Compaction: Replacing older context with a condensed summary so a conversation or agent run can continue within the window.
Complete mediation: The principle that every access to every resource is checked for authority, every time.
Computational graph: A directed acyclic graph whose nodes are intermediate values and whose edges record which operations produced them.
Compute-optimal: The split of a fixed training-compute budget between model size and training tokens that minimises loss.
Compute-optimal (Chinchilla) ratio: The finding that, for a fixed training budget, loss is lowest when training tokens grow in proportion to parameters, roughly 20 tokens per parameter.
Computer-use agent: An agent that operates software through its graphical interface by observing the screen and issuing mouse and keyboard actions.
Concept drift: A change over time in the relationship between input features and the target.
Condition number: The ratio of the largest to the smallest curvature (Hessian eigenvalue), measuring how elongated a loss valley is.
Confidence interval: An interval produced by a procedure that captures the true parameter value in a stated fraction of repeated experiments.
Confidence target: An instruction to answer only above a stated confidence t, with wrong answers penalised t/(1-t) points so that abstaining is the better choice below t.
Configuration space: The space of all possible robot configurations, in which planning is performed.
Confused deputy: A security flaw in which a program with legitimate authority is tricked into using that authority on behalf of an unauthorized party.
Confusion matrix: A table counting true positives, false positives, false negatives and true negatives, from which most error and fairness metrics are computed.
Connectionism: The approach that models intelligence as emerging from networks of simple units whose connection strengths are learned.
Consolidation: Converting accumulated episodic records into compact semantic facts or higher-level summaries.
Constrained decoding: Generation in which tokens that would violate a required format or schema are masked out at each step, so the output is guaranteed to match the structure.
Contamination: When benchmark questions or answers appear in a model's training data, inflating its score.
Context budget: The memory left for the KV cache after weights and runtime overhead, which sets the longest context a machine can hold for a given model.
Context compaction: Summarizing or removing older parts of an agent's history once it reaches a size budget, to bound cost and keep the context focused.
Context engineering: Designing and managing the full set of tokens supplied to a model on each call, including instructions, tools, retrieved data, memory and history.
Context isolation: Giving each agent its own context window so that it sees only the information relevant to its subtask.
Context precision: A rank-aware measure of how much of the retrieved context is relevant, rewarding relevant chunks placed near the top.
Context recall: The fraction of the information required by the reference answer that is present in the retrieved context.
Context rot: The decline in a model's ability to use information accurately as the amount of input context grows.
Context window: The maximum amount of text, measured in tokens, that a model can take into account in a single request.
Contextual anomaly: An observation that is unusual only given its context, such as time, location or user.
Contextual retrieval: Prepending a short LLM-written description of a chunk's place in its document before embedding and keyword-indexing the chunk.
Continuous batching: A serving technique that adds new requests to a running batch as soon as capacity frees up.
Continuous training (CT): Automated retraining of models in a production pipeline on a schedule or trigger.
Contrastive learning: Training that pulls the embeddings of matching pairs together and pushes mismatched pairs apart.
Contrastive loss: A training objective that pulls matching pairs (such as a query and its answer passage) together in embedding space and pushes non-matching pairs apart.
ControlNet: A trainable copy of a diffusion model's encoder that injects a structural condition such as edges, depth or pose through zero-initialized layers.
Convex function: A function for which the segment between any two points on its graph lies on or above the graph, so every local minimum is global.
Core point: In DBSCAN, a point with at least min_samples points within radius epsilon.
Cosine similarity: A measure of how closely two vectors point in the same direction, equal to their dot product divided by the product of their lengths.
Cost matrix: A table giving the cost of each combination of true class and decision, used to set decision thresholds.
Cost-complexity pruning: Choosing the subtree that minimises training error plus a penalty alpha times the number of leaves.
Coverage: The share of cases a system handles automatically at a given confidence threshold.
Credential broker: Trusted code that supplies secrets directly to a target system on an agent's behalf so the model never sees them.
Cross-encoder: A model that reads a query and a passage together in one forward pass and outputs a relevance score.
Cross-entropy: The expected negative log-probability that a model q assigns to outcomes drawn from the true distribution p.
Cross-entropy loss: The negative logarithm of the probability a model assigned to the correct answer, averaged over examples; the standard training loss for language models.
Cross-fitting: Computing a data-dependent transform for each fold using only the other folds, so no row's own label influences its feature.
Cross-validation: An evaluation scheme that splits data into k folds and averages the score of k models, each validated on a different fold.
Crosswalk: A mapping between the requirements or controls of two or more frameworks so one set of evidence can satisfy several.
CTC: Connectionist Temporal Classification, a loss that sums over all frame-level alignments (with a blank symbol) that collapse to the target transcript.
Curse of dimensionality: The collection of effects by which data become sparse and distances less informative as the number of dimensions grows.
Cybernetics: Wiener's 1948 science of control and communication through feedback in animals and machines.
Data augmentation: Applying random label-preserving transformations to training inputs to expose the model to more variation and reduce overfitting.
Data contract: A versioned agreement between data producers and consumers specifying schema, semantics, quality and freshness.
Data drift: A change in the distribution of model inputs P(X) between training and production.
Data exfiltration: Moving data out of a system to a destination an attacker controls.
Data leakage: When information that would not be available at prediction time, or that comes from the test data, influences training and inflates evaluation scores.
Data mixture: The weighted combination of data sources (web, code, math, books and so on) from which pretraining batches are drawn.
Data parallelism: Giving each device a full copy of the model and different data, mainly used for training.
Data protection impact assessment: A documented review of the privacy risks of a data-processing activity and how they will be reduced.
Data retention: How long a provider keeps your prompts, files and conversations.
Datasheet for a dataset: Structured documentation of a dataset's motivation, composition, collection, processing, uses, distribution and maintenance.
DDIM: A sampler for diffusion models that uses a non-Markovian, optionally deterministic update to generate in far fewer steps without retraining.
De Novo classification: A US FDA pathway for novel low-to-moderate-risk devices that have no suitable predicate.
Decode: The phase where the model generates the answer one token at a time, limited mainly by memory bandwidth.
Decontamination: Removing benchmark test items from training data, usually by n-gram overlap, so evaluations measure generalisation rather than memorisation.
Deep learning: Machine learning that uses neural networks with many layers of simple, connected mathematical units.
Deepfake: Synthetic or manipulated audio, image or video, usually depicting real people, produced with deep learning.
Defense in depth: Using several independent protective layers so that an attack must defeat all of them to succeed.
Delimiter: A consistent marker, such as a tag or heading, that shows where one part of a prompt (instructions, pasted material, examples) starts and ends.
Dendrogram: A tree recording the sequence and heights of merges in hierarchical clustering, which can be cut at any level to produce a flat clustering.
Dense model: A model that uses all of its parameters to produce every token.
Derivative: The instantaneous rate of change of a function with respect to its input, equal to the slope of its best local linear approximation.
Differential privacy: A property of a randomized algorithm guaranteeing that its output distribution changes by at most a factor e^epsilon (plus delta) when any one individual's data is added, removed or replaced.
Diffusion model: A generative model that creates images or video by gradually removing noise from random static, guided by a prompt.
Diffusion transformer (DiT): A diffusion model whose denoising network is a transformer operating on patches of a latent image rather than a convolutional U-Net.
Direct Preference Optimization (DPO): A preference-learning method that optimises the KL-regularised RLHF objective directly with a logistic loss on log-probability ratios.
Discount factor: A number between 0 and 1 that weights a reward k steps in the future by its k-th power.
Discriminative learning rates: Using different learning rates for different parts of a network, typically smaller for pretrained layers and larger for new ones.
Discriminative model: A model of the conditional probability of an output given an input, used to predict labels or values.
Disparate impact: A disproportionate adverse effect of a facially neutral practice on a protected group, which can create liability unless justified and no less discriminatory alternative exists.
Distillation: Training a smaller model to reproduce the outputs of a larger one.
DOM: The Document Object Model, the browser's tree representation of a web page's structure and content.
Domain randomisation: Training in simulation with randomised physical and visual parameters so policies transfer to the real world.
Dot product: The sum of elementwise products of two equal-length vectors, equal to the product of their lengths and the cosine of the angle between them.
DP-SGD: Stochastic gradient descent modified with per-example gradient clipping and calibrated Gaussian noise so that training is differentially private.
Drift: A gradual change in inputs or model behaviour that makes past evaluation results less representative.
Dual use: A capability or technology that can be applied for both beneficial and harmful purposes, such as vulnerability discovery.
DVC: Data Version Control: an open-source tool that versions large data files via content hashes and small pointer files committed to git.
Dynamic batching: A server technique that groups requests arriving within a short window into one batch to raise accelerator throughput at a small latency cost.
Early fusion: A multimodal design in which all modalities are converted to tokens and processed together by one model from the first layer.
Effective challenge: Critical analysis of a model by objective, informed parties with enough authority to make sure problems are fixed.
Effective context length: The longest input at which a model still performs a given task reliably, often shorter than its advertised context window.
efSearch: The size of the candidate list HNSW keeps during a query; larger values raise recall and latency.
Eigenvector: A non-zero vector whose direction is unchanged by a square matrix, which only scales it by the corresponding eigenvalue.
ELIZA effect: The human tendency to attribute understanding to a program that produces fluent, human-like text.
ELT: Extract, load, transform: loading raw data into a warehouse or lakehouse first and transforming it there.
Elusion rate: The proportion of relevant documents in the set that a review process did not produce, estimated by sampling.
Embedding: A list of numbers that places an item, such as a word or image, in a space where similar meanings are close together.
Embedding dimension: The length d of the vector an embedding model outputs, which sets per-vector memory and search cost.
Embodied carbon: Emissions from manufacturing, transporting and disposing of hardware and buildings, as opposed to operating them.
Emergent ability: A capability absent in smaller models but present in larger ones, as defined by Wei et al. (2022).
Endpointing: Deciding that a speaker has finished their conversational turn.
Enterprise terms: A separate commercial agreement for business customers, often with stricter data-handling commitments than consumer terms.
Entity resolution: Deciding which mentions refer to the same real-world entity and merging them into one node.
Entropy: The expected self-information of a distribution, measuring its uncertainty.
Episodic memory: Stored records of specific past events or interactions, usually timestamped.
Error analysis: Systematic inspection of a model's mistakes to find patterns and guide the next improvement.
Error cascade: The propagation and amplification of an early agent's mistake through downstream agents that treat it as fact.
Escalation: Passing a task from a smaller or local model to a more capable model when it is too hard or fails a check.
EU AI Act: The European Union's regulation that sets rules for AI systems according to their level of risk.
Eval: A repeatable test that measures how well an AI system performs on a defined set of cases.
Eval suite: A versioned set of test inputs with grading methods used to measure an LLM system's quality and detect regressions.
Evaluation gate: An automated check a candidate model must pass before promotion, such as beating the champion on metrics and slices.
Evaluation set: A fixed collection of real examples of your task with scoring rules, used to compare models or prompts.
Evaluation-driven development: Building an AI feature by first defining an evaluation set and rubric, then iterating until the system meets the bar.
Evidence lower bound (ELBO): A tractable lower bound on log p(x) equal to the expected reconstruction log-likelihood minus the KL divergence between the approximate posterior and the prior.
Excessive agency: Giving an AI more tools, permissions or autonomy than its task requires.
EXIF metadata: Information embedded in image files, such as camera model, timestamp and often GPS location.
Expectation: The probability-weighted average value of a random variable.
Expected utility: The probability-weighted average of how good the outcomes of an action are, used to compare actions under uncertainty.
Experiment tracking: Systematically recording each training run's parameters, metrics, artifacts and metadata for search and comparison.
Expert: One of the smaller feed-forward sub-networks inside an MoE layer.
Expert capacity: The maximum number of tokens one expert may process per batch, set by a capacity factor; overflow tokens skip the expert.
Expert parallelism: Placing different mixture-of-experts experts on different devices and routing tokens between them.
Expert system: A program that makes decisions using a large set of hand-written if-then rules from human experts.
Explained variance ratio: A component's eigenvalue divided by the sum of all eigenvalues, the share of total variance it captures.
Exponential smoothing: A family of forecasting methods that weight past observations with geometrically decaying weights, optionally with trend and seasonal states.
Exposure: The degree to which a job's tasks could be performed or assisted by AI.
External validation: Evaluating a model on data from sites, time periods or populations different from those used to develop it.
Fair use: A US copyright doctrine that allows some uses of protected works without permission, judged case by case.
Faithfulness (groundedness): The fraction of claims in a generated answer that are supported by the retrieved context.
False positive: An error where the system says yes, or flags someone, when the correct answer was no.
False premise: A question that assumes something untrue, which a model may go along with instead of correcting.
Feature: A measurable property of an example, such as floor area, that a model uses to make predictions.
Feature cross: A new feature built by combining two or more features, such as a product, ratio or joint category, to expose interactions.
Feature store: A system that defines features once and serves them consistently from an offline store for training and an online store for inference.
Feedback loop: The cycle in which a recommender's own outputs shape the interaction data used to train its next version.
Few-shot prompting: Including a few examples of the desired output in a prompt so the model can copy the pattern.
FID: Fréchet Inception Distance, the Fréchet distance between Gaussians fitted to Inception features of real and generated images; lower is better.
Filter bubble: A narrowing of the content a user is exposed to caused by personalization repeatedly favoring what they already consume.
Fine-tuning: Further training a pre-trained model's weights, or small adapter weights, on task-specific examples.
FlashAttention: An exact attention algorithm that tiles the computation in fast on-chip memory and uses an online softmax to avoid writing the full attention matrix to GPU main memory.
FLOP: One floating-point arithmetic operation, such as a multiplication or addition; a forward pass costs about 2 FLOPs per parameter per token.
Flow matching: Training a neural velocity field to match the velocity of prescribed probability paths from noise to data, then sampling by integrating an ODE.
Forward pass: Computing a network's outputs and loss from its inputs layer by layer, storing intermediate values for the backward pass.
Forward process: The fixed Markov chain that gradually adds Gaussian noise to data according to a schedule until it becomes approximately pure noise.
Four-fifths rule: A US employment rule of thumb that flags for review any group selected at less than 80% of the rate of the most-selected group.
FP4 / FP8 / FP16 / BF16: Number formats. Fewer bits means smaller, faster models with some quality loss.
Frame sampling: Selecting a subset of video frames at some rate so a model can process a video within its token budget.
Frontier AI: The most capable general-purpose AI models at a given time, typically those trained with the largest compute budgets.
Full-duplex: A conversational model that can listen and speak at the same time, handling interruptions and overlaps.
Fundamental rights impact assessment (FRIA): An assessment required of certain deployers of high-risk AI under Article 27 of the EU AI Act, describing affected persons, risks of harm to their rights, and oversight and mitigation measures.
GDPR: The EU's General Data Protection Regulation, which governs how personal data is collected and used.
General-purpose AI (GPAI) model: A model, such as those behind chatbots, that can be adapted to many different tasks.
Generative adversarial network: A pair of networks, a generator and a discriminator, trained against each other so that the generator learns to produce realistic samples.
Generative AI: AI models that create new content such as text, images, audio, video or code.
Generative model: A model of the distribution of the data itself from which new, similar samples can be drawn.
GGUF: The model file format used by llama.cpp and related local AI tools.
GGUF / MLX / safetensors: File formats for model weights: GGUF for llama.cpp and Ollama, MLX for Apple silicon, safetensors for most others.
Gini impurity: One minus the sum of squared class proportions in a node; the chance that two random draws from the node have different labels.
Global search (GraphRAG): Answering corpus-wide questions by asking an LLM to answer from each community report and then combining the most useful partial answers.
Goal misgeneralization: A failure in which a model trained with a correct objective learns a different goal that coincides with it in training but competently pursues the wrong goal out of distribution.
GOFAI: 'Good old-fashioned AI', John Haugeland's 1985 name for AI based on explicit symbols and rules.
Golden answer: An expert-approved reference output for a test case.
Golden set: A versioned, curated set of evaluation questions with reference answers and, ideally, labelled supporting passages.
Goodput: The rate of completed requests or tokens that met their latency SLO.
GPU: A graphics processing unit, a processor with thousands of parallel arithmetic units well suited to the matrix operations of neural networks.
Gradient: The vector of all partial derivatives of a scalar function, pointing in the direction of steepest increase.
Gradient boosting: An ensemble built sequentially, each new tree fit to the negative gradient of the loss with respect to the current predictions.
Gradient check: A test that compares analytic gradients with finite-difference estimates to detect implementation bugs.
Gradient clipping: Rescaling the gradient when its global norm exceeds a threshold, preserving direction while capping step size.
Gradient descent: A method that reduces loss by repeatedly moving each parameter a small step in the direction that lowers the error.
Graph convolutional network (GCN): A GNN layer that multiplies node features by a learned matrix and averages them over each node's self-looped neighbourhood using symmetric degree normalization.
Gross margin: Revenue minus the direct cost of delivering it, divided by revenue.
Grounding: Giving a model trusted source material and asking it to answer from that material rather than from memory.
Grounding (GUI): Mapping a description of an interface element to the exact location or identifier needed to act on it.
Grouped-query attention (GQA): An attention variant in which groups of query heads share a single key and value head, reducing KV-cache size.
GRPO: Group Relative Policy Optimization, a critic-free PPO variant that normalises rewards within a group of samples for the same prompt.
GRU: A gated recurrent architecture with reset and update gates and no separate cell state, cheaper than an LSTM.
Guardrail: A check outside the model's generation that inspects inputs, outputs or proposed actions and can block, modify or escalate them.
Guardrail metric: A metric with a threshold that must not be breached while optimising the main metric, such as complaint rate or cost per user.
Hallucination: When an AI model produces false or invented information presented as if it were true.
Handoff: A transfer of control of a conversation or task from one agent to another, typically implemented as a tool call.
Hard binding: A cryptographic hash in a C2PA manifest that ties the manifest to the exact content bytes, so any alteration is detectable.
HBM: High-bandwidth memory: stacked DRAM placed on the accelerator package, providing terabytes per second of bandwidth.
He initialization: Drawing weights with variance 2 / fan-in so that ReLU networks preserve activation variance across layers.
Heading path: The chain of section titles above a chunk, such as Handbook > Leave > Parental leave, stored as metadata and often prepended to the chunk text.
Heuristic: A rule of thumb that guides search toward promising options without guaranteeing the best or any solution.
Hidden Markov model: A probabilistic model of a sequence of hidden states, each emitting an observation, with learned transition and emission probabilities.
Hidden state: The vector an RNN carries from one time step to the next, summarizing everything it has read so far.
High-risk AI system: Under the EU AI Act, an AI use in a sensitive area, such as hiring or credit, that must meet strict requirements.
Hinge loss: The loss max(0, 1 - y f(x)), zero for points classified correctly with functional margin at least 1.
HNSW: Hierarchical Navigable Small World, a layered proximity graph searched greedily from sparse upper layers down to a dense bottom layer.
Hopfield network: A recurrent network with symmetric weights whose dynamics minimize an energy function, used as associative memory.
Host: The AI application a user interacts with, which manages MCP clients and decides what the model sees.
Hosted model: A model run on a company's servers and used through an app or API, with its weights kept private.
Hugging Face: The main public hub where open-weight models and their converted versions are published.
Human-in-the-loop: A design where a person reviews or approves certain actions before they take effect.
Human-preference arena: A site where people vote between anonymous model answers, producing a ranking from many votes.
Hybrid AI: A setup that combines models running on your own hardware with hosted models from providers.
Hybrid architecture: A model that interleaves a small number of full-attention layers with a majority of SSM or linear-attention layers.
Hybrid search: Combining keyword search and vector search and merging their rankings.
HyDE: Hypothetical Document Embeddings: retrieving with the embedding of an LLM-written hypothetical answer instead of the raw query.
Hyperparameter: A model setting chosen by the practitioner, such as tree depth or regularisation strength, rather than learned from data.
Idempotency key: A unique identifier sent with a write request so that retries of the same request are recognized and executed only once.
Idempotent: Describes an action that has the same effect whether it is performed once or several times.
Idempotent pipeline: A job that produces the same result without duplicates no matter how many times it is re-run for the same input.
Image tokenizer: A model, often a VQ-VAE or VQGAN, that converts an image into a grid of discrete codebook indices and back.
Image-to-image (SDEdit): Generating by noising an existing image's latent to an intermediate step and denoising it with a new prompt, controlled by a strength parameter.
Imitation game: Turing's 1950 test in which an interrogator, communicating by text, tries to tell a hidden machine from a hidden human.
Implicit feedback: Behavioral signals such as clicks, plays or purchases that indicate preference indirectly and lack explicit negatives.
Importance matrix: Statistics gathered from sample text that show which weights matter most, used to guide low-bit quantization.
IMU: An inertial measurement unit measuring linear acceleration and angular velocity at high rate.
In-batch negatives: A training technique that treats the other positive items in a mini-batch as negatives for each user.
In-context learning: A model's ability to perform a new task from instructions or examples in its prompt without updating its weights.
Indirect prompt injection: An attack in which instructions hidden in retrieved content or tool output try to redirect the model's behaviour.
Inductive bias: Assumptions built into a model's architecture, such as locality and weight sharing in CNNs, that shape what it can learn easily from limited data.
Inertia (WCSS): The sum of squared distances from each point to its assigned cluster centroid, the objective k-means minimizes.
Inference: Using a trained, fixed model to produce outputs for new inputs.
Inference runtime: Software that runs the model: Ollama, LM Studio, llama.cpp, MLX, vLLM and others.
InfoNCE loss: A cross-entropy loss that asks a model to identify the positive pair among a set of negatives using softmax-normalized similarities.
Information gain: The reduction in entropy achieved by a split, weighting each child by its share of samples.
Inpainting: Regenerating only a masked region of an image while keeping the rest consistent.
Input tokens: Everything sent to the model in a request, including instructions, history, documents and tool definitions.
Instance segmentation: Producing a separate pixel mask for each individual object instance.
Instruction tuning: Further training on example instructions and good responses so the model behaves like an assistant.
Interconnect: The link that carries data between GPUs or machines, such as PCIe, NVLink, Thunderbolt, or Ethernet.
Intersection over union (IoU): The area where a predicted region and the true region overlap, divided by the area of their union.
Inverted dropout: Zeroing activations with probability p during training and scaling survivors by 1/(1-p) so no rescaling is needed at inference.
Inverted index: A map from each term to the list of documents (and often positions) containing it, enabling fast lexical retrieval.
Irreducible loss: The asymptotic loss term E that no model size or data amount can remove, reflecting the entropy of the text.
ISO/IEC 42001: An international standard specifying requirements for an AI management system, against which organisations can be certified.
IsoFLOP profile: A curve of final loss versus model size for runs that all use the same compute budget, whose minimum gives the optimal size.
Isolation forest: An ensemble of random trees that scores points as anomalous when they are isolated in few random splits.
IVF (inverted file index): An index that clusters vectors with k-means and searches only the lists of the nprobe centroids closest to the query.
Jacobian: The m-by-n matrix of first partial derivatives of a function from n inputs to m outputs.
Jagged frontier: The uneven and hard-to-predict boundary between tasks an AI system does well and similar-looking tasks it does badly.
Jagged technological frontier: The uneven boundary of AI capability, under which similar-looking tasks may be done well or badly by the same model.
Jailbreak: An attempt to get a model to bypass its own safety training, often through role-play or obfuscated wording.
Jitter: Deliberate randomness added to retry delays so that many clients do not retry at the same moment.
JSON mode: A decoding option that forces output to be syntactically valid JSON without enforcing any particular schema.
JSON Schema: A standard format for describing the structure, types and required fields of JSON data.
JSON-RPC: A lightweight format for remote procedure calls in which requests and responses are JSON objects matched by ID.
Just-in-time retrieval: Keeping references in context and loading full content through tools only when it is needed.
k-means++: An initialization that picks each new starting centroid with probability proportional to its squared distance from the nearest centroid already chosen.
k-nearest neighbours: A classifier that predicts the majority label (or mean target) of the k training points closest to the query.
Kalman filter: A recursive estimator that alternates prediction with a motion model and correction with measurements, weighting each by its uncertainty.
Kernel (filter): A small learned weight tensor spanning all input channels that is slid across the input to produce one output feature map.
Kernel trick: Computing dot products in a high-dimensional feature space through a kernel function, without constructing the feature vectors.
Key and value: Vectors computed for each token; keys describe what a token offers and values carry its information.
KL divergence: A non-negative, asymmetric measure of how much a distribution q differs from p, equal to cross-entropy minus entropy.
KL penalty: A term that penalises divergence between the trained policy and a reference model to prevent drift and reward hacking.
Knowledge acquisition bottleneck: The difficulty, cost and slowness of extracting and encoding human expertise as explicit rules.
Knowledge distillation: Training a smaller student model to match the outputs or output distribution of a larger teacher model.
Knowledge graph: A graph whose nodes are entities and whose typed, directed edges are relations between them, often with properties and provenance.
Knowledge injection: Adding new factual knowledge to a language model system, whether through weights, retrieval or the prompt.
Kolmogorov–Smirnov statistic: The maximum absolute difference between two empirical cumulative distribution functions, used to test whether two samples share a distribution.
KV cache: The model's working record of text it has already read, which grows with context length and uses memory.
Label: The correct answer attached to a training example, such as the actual sale price.
Label delay: The lag between an event and the arrival of its ground-truth label, such as a chargeback weeks later.
Label shift: A change in the class proportions P(Y) while the class-conditional input distributions P(X|Y) stay the same.
Label smoothing: Replacing one-hot training targets with a mixture of the one-hot vector and a uniform distribution to reduce overconfidence.
Labelled property graph: A graph model in which nodes have labels and properties and edges have a type and properties, typically queried with Cypher or GQL.
Laplace smoothing: Adding a pseudo-count to every feature-class count so that unseen features never receive zero probability.
Large language model (LLM): A large neural network trained on text to predict the next token, used to understand and generate language.
Late chunking: Embedding a whole document with a long-context model to obtain contextualized token embeddings, then pooling them within each chunk's span.
Late interaction (MaxSim): Scoring by summing, over query token embeddings, the maximum similarity to any document token embedding.
Latency: The delay before data starts arriving over a link, separate from how much data it can carry.
Latent diffusion: Running the diffusion process in the compressed latent space of a pretrained autoencoder rather than in pixel space.
Layer normalization: Normalizing across the features of each individual example, independent of batch size and identical in training and inference.
Leaderboard: A ranked table of model scores on one or more benchmarks.
Learning rate: The step size used in gradient descent, controlling how far parameters move on each update.
Least privilege: Granting only the minimum access a component needs to do its job.
Least squares: Choosing model parameters to minimise the sum of squared differences between predictions and targets.
Leiden algorithm: A community-detection method that improves on Louvain by guaranteeing well-connected communities, used by GraphRAG to build a community hierarchy.
Lemmatization: Reducing a word to its dictionary form using vocabulary and part-of-speech information, such as 'saw' (verb) to 'see'.
Lethal trifecta: The combination of access to private data, exposure to untrusted content and an outbound communication channel in one AI agent, which together enable data exfiltration.
Liar's dividend: The ability of people to dismiss authentic recordings as fake because convincing fakes are known to exist.
Lidar: A sensor that measures distances by timing reflected laser pulses, producing 3-D point clouds.
Lineage: The recorded chain linking a model version to the run, code, data, configuration and evaluation that produced it.
Linear attention: An attention variant that replaces the softmax with a kernel feature map so causal attention can be computed as a recurrence over a fixed-size state.
Linear probe: Training only a new linear output layer on top of a frozen pretrained network to measure the quality of its features.
Linearly separable: Describes two classes of points that can be divided exactly by a straight line, or by a flat boundary in higher dimensions.
Little's law: In a stable system, average concurrency equals arrival rate times average time in the system.
LLM-as-a-judge: Using a language model with a rubric to grade another model's outputs, which requires validation against human judgements.
LLM-as-judge: Using a language model with a rubric to grade the outputs of another system.
LLMOps: The practices and tooling for versioning, evaluating, deploying, observing and improving applications built on large language models.
Load-balancing loss: An auxiliary training loss that penalises uneven routing so that tokens spread across experts.
Local model: An AI model whose weights are stored and run on your own computer rather than on a provider's servers.
Location-based accounting: Scope 2 method using the average emission factor of the grid where electricity is consumed.
Log-loss: The negative average log-probability a model assigns to the true labels, also called binary cross-entropy.
Log-Mel spectrogram: A time-frequency representation of audio on a perceptually scaled frequency axis with log-compressed energies, used as input to speech models.
Log-odds (logit): The logarithm of p/(1-p); logistic regression models it as a linear function of the features.
Log-sum-exp trick: Subtracting the maximum logit before exponentiating so that log-softmax and related quantities can be computed without overflow.
Logical points: Resolution-independent screen coordinates used by many operating systems, equal to physical pixels divided by the display scale factor.
Logit: The raw, unnormalised score a model assigns to each possible next token before softmax turns the scores into probabilities.
Logits: The unnormalized real-valued scores a model outputs before a softmax or sigmoid converts them to probabilities.
Long-context prompting: Supplying a large portion of the relevant corpus directly in the prompt on each request instead of retrieving selected passages.
Long-term memory: Information persisted outside the model across sessions and selectively retrieved into the context.
Loop closure: Recognising a previously visited place so that accumulated drift in a SLAM estimate can be corrected.
LoRA: Low-Rank Adaptation, a fine-tuning method that freezes pretrained weights and trains a low-rank update BA added to them.
Loss: A single number measuring how wrong a model's predictions are.
Loss masking: Setting the labels of tokens that should not be learned, such as the prompt, to an ignore index so they do not contribute to the loss.
Loss of control: A scenario in which one or more AI systems operate outside human control and humans cannot effectively correct or stop them.
Loss scaling: Multiplying the loss by a large factor before backpropagation in FP16 training so small gradients do not underflow, then dividing it out before the update.
Lost in the middle: The finding that models use information at the start or end of a long context better than information in the middle.
Lovelace objection: The argument, named by Turing after Ada Lovelace's 1843 Notes, that a machine can only do what it has been instructed to do and cannot originate anything.
LSTM: A gated recurrent architecture with a separately maintained cell state updated additively under control of forget, input and output gates.
Machine learning (ML): A kind of AI in which a system learns patterns from example data instead of following only hand-written rules.
Macro-F1: The unweighted mean of per-class F1 scores, giving every class equal weight regardless of its size.
mAP: Average precision averaged over classes and, in the COCO convention, over IoU thresholds from 0.50 to 0.95.
Margin: The distance from a classifier's decision boundary to the closest training points; for a scaled SVM it is 1/||w||, and the full width between the two margin hyperplanes is 2/||w||.
Market-based accounting: Scope 2 method reflecting contractual purchases such as PPAs and renewable energy certificates.
Markov decision process (MDP): A formal model of sequential decisions defined by states, actions, transition probabilities, rewards and a discount factor.
MASE: Mean absolute scaled error: test MAE divided by the in-sample MAE of the seasonal naive method.
Masked language modeling: A pretraining objective in which some input tokens are hidden and the model predicts them from the surrounding context on both sides.
Matrix factorization: A recommendation technique that represents users and items as learned vectors whose dot products approximate observed ratings.
Maximum likelihood estimation: Choosing model parameters that maximize the probability (likelihood) of the observed data.
McCulloch-Pitts neuron: A binary threshold unit that outputs 1 when its weighted (or counted) excitatory inputs reach a threshold and no inhibition blocks it.
McNemar's test: A paired test that compares two classifiers on the same examples using only the examples on which they disagree.
MCP: The Model Context Protocol, an open standard for connecting AI applications to tools and data sources.
MCP client: A component in the host that maintains a connection to one MCP server.
MCP server: A program that exposes tools, resources or prompts to AI applications through MCP.
Mean average precision (mAP): A detection metric that averages precision over recall levels and over object classes.
Mean opinion score (MOS): The average of listener ratings on a 1 to 5 quality scale, used as a subjective measure of audio quality.
Mean reciprocal rank: The average over queries of one divided by the rank of the first relevant result, or zero if none is retrieved.
Mean squared error (MSE): A loss equal to the average of the squared differences between predictions and correct values.
Means-ends analysis: A search strategy that repeatedly picks an action to reduce the most important difference between the current state and the goal, setting subgoals when needed.
Median absolute deviation (MAD): The median of absolute deviations from the median, a robust measure of spread.
Mel scale: A perceptual frequency scale, roughly linear below 1 kHz and logarithmic above, used to space filterbank channels.
Mel spectrogram: A time-frequency representation of audio with the frequency axis warped to the perceptual mel scale.
Membership inference: An attack that determines whether a specific record was part of a model's training data.
Memory bandwidth: How many gigabytes per second the hardware can read from memory, the main limit on generation speed.
Memory bandwidth (GB/s): How many gigabytes per second the chip can read from memory. The single biggest driver of local chat speed.
Memory capacity: How much data the processor's memory can hold, which decides whether a model fits.
Memory poisoning: An attack that plants false or instruction-like content in an agent's persistent memory or knowledge store to influence future behaviour.
Memory-bound: Describes work whose speed is limited by how fast data can be read from memory rather than by arithmetic throughput.
Message passing: The GNN computation in which each node aggregates transformed messages from its neighbours and updates its own representation.
MFCC: Mel-frequency cepstral coefficients: a compact feature vector obtained by applying a discrete cosine transform to log-mel energies.
MinHash deduplication: A locality-sensitive hashing method that estimates Jaccard similarity between documents' shingle sets to find near-duplicates at scale.
Minkowski distance: The p-th root of the sum of absolute coordinate differences raised to the power p; p=2 is Euclidean and p=1 is Manhattan distance.
Mixture of experts (MoE): A model design where each layer has many expert sub-networks and only a few are used per token.
ML engineer: An engineer who builds the pipelines to train, evaluate and serve machine learning models reliably.
MLOps: The practices and platforms for deploying, monitoring and maintaining machine learning models in production.
MLX: Apple's open-source machine-learning framework designed for unified-memory Apple silicon.
Modality: A type of data, such as text, images, audio or video.
Mode collapse: A GAN failure in which the generator produces only a small subset of the data's variety because that subset fools the discriminator.
Model card: Structured documentation of a model's intended use, data, evaluations, limitations and license.
Model cascade: A setup that sends every request to a cheap model first and escalates only the failures to a stronger, more expensive model.
Model collapse: Progressive loss of distributional tails and diversity when models are recursively trained on model-generated data.
Model Context Protocol (MCP): An open standard for connecting AI applications to external tools and data sources.
Model FLOPs utilisation (MFU): The fraction of the hardware's peak floating-point throughput that is spent on the model's useful training FLOPs.
Model FLOPs utilization (MFU): Achieved model FLOP/s divided by the hardware's theoretical peak FLOP/s.
Model inversion: An attack that uses a model's outputs to reconstruct representative features of training data or sensitive attributes of individuals.
Model registry: A versioned catalog of deployable models with lineage, aliases or stages, and approval history.
Model risk: The potential for adverse consequences from decisions based on incorrect or misused model outputs.
Model routing: Automatically choosing which model handles each request based on rules, difficulty, cost, speed and privacy.
MoE (mixture of experts): A model that uses only some of its parameters per token. All experts must fit in memory; only the active ones set the speed.
Momentum: An optimizer technique that accumulates an exponentially decaying sum of past gradients and steps along it.
Multi-head attention: Running several attention operations in parallel on lower-dimensional projections and concatenating their outputs through an output projection.
Multi-head latent attention (MLA): DeepSeek's method of compressing keys and values into a smaller latent vector.
Multi-hop question: A question whose answer requires chaining several facts, typically from different documents.
Multi-query attention (MQA): An attention design where all query heads share a single key and value head, minimising the KV cache.
Multilayer perceptron: A feed-forward network of fully connected layers with non-linear activations between them.
Multimodal: Able to take in or produce more than one kind of data, such as text, images and audio.
Multimodal model: A model that can take in or produce more than one modality.
n-gram language model: A model that estimates the probability of a token from counts of the preceding n-1 tokens.
Naive Bayes: A probabilistic classifier that applies Bayes' theorem assuming features are independent given the class.
Named-entity recognition (NER): Identifying spans of text that refer to entities such as people, organizations, locations, dates and amounts, and labeling their types.
Narrow AI: An AI system designed for a specific task or limited range of tasks.
NDCG: Normalised discounted cumulative gain, a ranking metric that rewards placing relevant items near the top of a list, scaled so a perfect ranking scores 1.
NDCG@k: Discounted cumulative gain of the top-k list divided by that of the ideal ordering, rewarding relevant items placed near the top.
Nested cross-validation: An outer cross-validation loop for performance estimation wrapped around an inner loop used only for hyperparameter tuning.
Neural audio codec: An autoencoder with a quantized bottleneck that compresses audio into discrete codes and reconstructs it, trained with reconstruction and adversarial losses.
Neural network: A function made of layers of simple units connected by adjustable weights.
Neural vocoder: A neural network that converts an acoustic representation such as a mel spectrogram into an audio waveform.
Neuro-symbolic AI: Approaches that combine neural networks with symbolic representations, logic or search in a single system.
NIST AI RMF: The US National Institute of Standards and Technology's voluntary AI Risk Management Framework, organised into the functions Govern, Map, Measure and Manage.
Noise schedule: The sequence of per-step noise variances beta_t, equivalently the curve of cumulative signal levels abar_t, that controls how fast information is destroyed.
Non-consensual intimate imagery (NCII): Intimate or sexual images or videos of a person published without their consent, including AI-generated depictions.
Non-maximum suppression (NMS): A post-processing step that keeps the highest-confidence box and removes other boxes that overlap it above an IoU threshold.
Norm: A function that assigns a non-negative length to a vector, such as the L1, L2 or L-infinity norm.
Normal equations: The linear system X^T X theta = X^T y whose solution gives the least-squares parameters.
Normalizing flow: A generative model built from invertible transformations whose exact likelihood follows from the change-of-variables formula.
North-star metric: The single product metric chosen to represent the user value a feature is meant to create.
Notification: A JSON-RPC message with a method but no ID, sent without expecting any response.
Notified body: An independent organisation designated by an EU member state to assess whether medium- and high-risk medical devices conform to the MDR before CE marking.
NPU: A neural processing unit, a low-power AI accelerator built into many recent chips.
NVLink: NVIDIA's high-bandwidth GPU-to-GPU scale-up interconnect, switched across up to 72 GPUs in NVL72 racks.
OAuth: A widely used standard for granting an application limited access to a service on a user's behalf.
OCR: Optical character recognition, the conversion of images of text into machine-readable characters.
Off-policy learning: Learning about one policy, such as the greedy policy, from data generated by a different behaviour policy.
Offload: Running part of a model from system RAM when it does not fit in GPU memory. Works, but much slower.
Offload fraction: The share of a model's weights held in system RAM rather than GPU memory when a model is split between the two.
On-policy distillation: Distillation in which the student generates sequences and the teacher provides token-level targets on those student-generated sequences.
One-class SVM: A kernel method that learns a boundary enclosing most training data, flagging points outside it.
One-hot encoding: Representing a categorical value as a vector of binary indicators with a single 1 marking the category.
ONNX: An open, framework-neutral format that represents a model as a graph of versioned operators so it can run in many inference engines.
Open Source AI Definition (OSAID): OSI's 2024 definition requiring freedoms to use, study, modify and share an AI system, with data information, code and parameters available.
Open weights: A release in which trained model parameters are downloadable, regardless of the license conditions attached.
Open-weight model: A model whose trained parameters are published so anyone can download and run it, subject to its licence.
OpenAI-compatible API: A local or hosted interface that accepts requests in the same format as OpenAI's API, so the same client software works with it.
OpenRAIL: A family of royalty-free model licenses that include use-based restrictions which must be passed on to derivatives.
Operational design domain (ODD): The specific conditions under which a driving-automation feature is designed to operate.
Optimizer state: Extra per-weight statistics a training algorithm such as Adam keeps between steps.
Orchestrator: An agent that divides a task among sub-agents and combines their results.
Orchestrator–workers: A pattern in which a lead agent decomposes a task, delegates subtasks to worker agents, and synthesizes their results.
Out-of-bag estimate: A validation score computed for each tree on the roughly 37% of training rows absent from its bootstrap sample.
Output tokens: The tokens the model generates in its reply.
Over-refusal: A model or guardrail declining a benign request that policy actually permits.
Over-smoothing: The tendency of node representations in deep GNNs to converge to similar values as repeated averaging spreads information across the graph.
Over-training: Training a model on more tokens than compute-optimal in order to obtain a smaller, cheaper-to-serve model at a given quality.
Overfitting: When a model learns its training data too specifically and performs poorly on new data.
Overlap: The number of tokens shared by consecutive fixed-size chunks, so that the stride equals chunk size minus overlap.
Overthinking: When a reasoning model spends many tokens on a problem that did not need them.
p-value: The probability, assuming the null hypothesis is true, of observing a result at least as extreme as the one obtained.
Padding: Extra border values (usually zeros) added around the input so kernels can be applied at the edges and output size can be controlled.
PagedAttention: A serving technique that stores the KV cache in fixed-size blocks allocated on demand, reducing wasted memory and allowing prefix sharing.
Paired comparison: Comparing two models question by question on the same items, which removes shared difficulty and narrows the uncertainty of the difference.
Parallel tool calls: A single model turn that requests several tool executions at once.
Parameter: A number inside a model, such as a connection weight, that is adjusted during training.
Parameter (weight): One adjustable number inside a neural network that is set during training.
Parameters: The adjustable numbers a model learns during training, usually counted in billions.
Parameters (B): The model's learned weights, in billions. More parameters usually means more capable and more memory.
Parent-child retrieval: Indexing small child chunks for precise matching but returning their larger parent sections to the generator.
Partial derivative: The derivative of a multivariable function with respect to one input while all other inputs are held fixed.
Partition function: The normalizing constant Z in an energy-based model that makes exp(-E(x))/Z integrate to one; usually intractable.
pass@k: The probability that at least one of k independent samples solves a problem.
pass^k: The chance that a model succeeds on all k repeated attempts at the same task.
Patch: A small square block of pixels that a vision model embeds as a single vector, analogous to a text token.
Patent grant: An explicit license from contributors to use their patents embodied in the work, as provided by Apache-2.0.
Payback period: The time needed for cumulative net benefits to recover the up-front investment.
Perception–planning–control loop: The repeated cycle in which a robot senses its environment, decides on actions and executes them.
Perceptron: An early single artificial neuron that learned to classify inputs by adjusting its weights.
Perceptron learning rule: An update that adds or subtracts the input to the weights only when the perceptron's output is wrong.
Permutation importance: The drop in a model's held-out score when one feature's values are randomly shuffled.
Perplexity: The exponential of the average per-token negative log-likelihood, interpretable as an effective number of equally likely choices.
Physical symbol system hypothesis: Newell and Simon's claim that a system manipulating physical symbols has the necessary and sufficient means for general intelligent action.
Pilot purgatory: The state in which AI pilots run indefinitely without success criteria or a decision to scale or stop.
Pinball (quantile) loss: A loss that scores a forecast of a given quantile, penalising under- and over-prediction asymmetrically.
Pipeline: A scikit-learn object that chains preprocessing steps and a model so that all of them are fit together on training data only.
Pipeline parallelism: Splitting a model by layers so each device runs a consecutive group of layers in turn.
Planning: Having the model outline the steps of a task before and during execution.
pLDDT: AlphaFold's per-residue confidence score (0 to 100) estimating how reliable the predicted local structure is.
Point-in-time correctness: Joining each training example only with feature values that were available at that example's timestamp.
Policy: A mapping from states to a probability distribution over actions.
Pooling: A fixed, non-learned downsampling operation, such as taking the maximum or average over each local window of a feature map.
Population Stability Index (PSI): A binned measure of distribution change computed as the sum over bins of (actual − expected) × ln(actual/expected).
Positive predictive value (PPV): The proportion of positive test results that are true positives, which depends on both test accuracy and disease prevalence.
Positive predictive value (precision): The share of a system's positive predictions that are correct: TP / (TP + FP).
Post-training quantization: Quantizing a model after training is finished, without further training.
Posterior collapse: A VAE failure in which the encoder's output matches the prior regardless of input, so the latent carries no information.
Power law: A relationship y = a·x^(−α) that appears as a straight line with slope −α on log-log axes.
Pre-activation: The weighted sum plus bias, z = w·x + b, computed before the activation function.
Pre-filtering: Restricting the search to items that satisfy metadata constraints before or during the similarity search, rather than discarding results afterwards.
Pre-norm: A transformer block layout that applies normalization to the input of each sub-layer inside the residual branch rather than after the residual addition.
Pre-routing: Choosing a model for a request before running any model, based on rules and a prediction of the request's difficulty.
Pre-training: Training a model on a large general corpus, typically to predict the next or masked token, before adapting it to specific uses.
Precision: The fraction of items a model flagged as positive that really are positive.
Predetermined Change Control Plan (PCCP): A plan, authorised with a device, that specifies which future model modifications are intended and how they will be validated, so they can be made without a new submission.
Prediction interval: A range expected to contain a future observation with a stated probability.
Preemption: When a higher level of government's law overrides conflicting laws from a lower level.
Prefill: The phase where the model reads and processes the whole prompt before writing.
Prefill / prompt processing: How fast a model reads your prompt before answering. Mostly limited by compute (TFLOPS).
Prefix caching: Reusing stored KV-cache blocks for prompt prefixes already processed by earlier requests.
Pretraining: The first, largest training stage, where a model learns to predict the next token over a huge body of text.
Principal component: A unit direction of maximal remaining variance, equal to an eigenvector of the data covariance matrix.
Procedural memory: Stored knowledge of how to perform tasks, such as rules, workflows or skills that shape behaviour.
Product quantization (PQ): A compression scheme that splits vectors into sub-vectors and replaces each with the ID of its nearest centroid in a learned per-sub-space codebook.
Productivity: How much useful output is produced per unit of time or effort.
Profile (NIST AI RMF): An application of the AI RMF functions, categories and subcategories to a specific setting, technology or sector, such as generative AI in NIST AI 600-1.
Projection layer: A learned mapping that converts vision-encoder outputs into vectors in the language model's embedding space.
Projector: A small network, such as a linear layer or MLP, that maps image-encoder features into a language model's token embedding space.
Prompt: The text instruction or question you give an AI model.
Prompt caching: Reusing a model's processed state for an identical prompt prefix across calls to cut cost and latency.
Prompt chaining: Splitting a task into a sequence of prompts in which each step's output becomes input to the next.
Prompt eval set: A fixed collection of representative inputs with expected outputs or grading criteria used to compare prompt versions.
Prompt injection: An attack where text the AI reads contains hidden instructions meant to override the user's or developer's instructions.
Prompt template: Reusable prompt text with named variables that are filled in at run time.
Promptable segmentation: Segmentation in which the user specifies the target with a prompt such as a click, box, mask, text phrase or exemplar.
Provider pinning: A rule that a task must use a specific model provider (for example, local only) and must never silently fall back to another.
Proxy: An easy-to-measure stand-in for the thing you actually want to predict, such as spending instead of health need.
Proxy label: An early-available signal that correlates with the true outcome and is used to estimate performance before real labels arrive.
Pseudonymisation: Replacing identifiers with placeholders while keeping a separate key that can map them back.
Pseudonymization: Replacing identifiers with consistent tokens, such as keyed hashes, that cannot be reversed without separately held information.
PUE: Power usage effectiveness: total facility energy divided by IT equipment energy, with 1.0 as the ideal.
Pushdown automaton: A finite-state machine augmented with a stack, able to recognize nested structures such as balanced JSON brackets.
Pydantic: A Python library that parses, coerces and validates data against typed model classes and can emit their JSON Schema.
QLoRA: LoRA training on top of a base model stored in 4-bit precision to save memory.
Quantization: Storing a trained model's weights with fewer bits to save memory and speed it up.
Quantization (Q4, Q8): Storing weights with fewer bits. Q4 cuts a model to roughly a quarter of its 16-bit size.
Quantization-aware training (QAT): Training or fine-tuning a model while simulating low precision so it holds up better when quantized.
Quasi-identifier: A detail, such as a postcode or date of birth, that is not identifying alone but can identify a person in combination with others.
Query decomposition: Splitting a complex question into simpler sub-questions that can be retrieved and answered separately.
Query, key, value: Three learned linear projections of each token: queries are compared with keys to produce attention weights that are applied to values.
R-squared: One minus the ratio of the model's squared error to the squared error of predicting the mean.
RAG: Retrieval-augmented generation: retrieving relevant text and including it in the prompt so the model answers from it.
Random forest: An ensemble of decision trees trained on bootstrap samples with random feature subsets at each split, whose predictions are averaged.
Random variable: A function that assigns a numerical value to each outcome of a random process.
Rank: The number of linearly independent columns (equivalently rows) of a matrix, i.e. the dimension of its column space.
Rank (LoRA): The inner dimension r of LoRA's two adapter matrices, which bounds the rank of the learned weight update and sets its size.
Rate limit: A cap on how many requests or tokens an account may send in a given period.
Rational agent: A system that perceives its environment and chooses the action expected to achieve the best outcome by a given measure of success.
RDF: The W3C Resource Description Framework, a graph data model of IRI-identified triples queried with SPARQL.
RDMA: Remote direct memory access, which lets one machine write into another's memory with minimal operating-system overhead.
Readiness probe: A Kubernetes check that tells the cluster a container is ready to receive traffic, typically only after its model has loaded.
Real-time factor: Processing time divided by the duration of the audio produced; values below 1 mean faster than real time.
Real-time factor (RTF): Processing time divided by audio duration; values below 1 mean faster than real time.
Realisation rate: The share of a theoretical benefit, such as time saved, that actually turns into financial or operational value.
Reason code: A standardised explanation identifying a factor that most reduced an applicant's score or outcome.
Reasoning effort: A setting that tells a model roughly how much thinking to do before answering.
Reasoning model: A language model trained to work through intermediate steps before giving its final answer.
Recall: The fraction of actual positives the model detects, TP / (TP + FN), also called sensitivity or true positive rate.
Recall (sensitivity): The fraction of truly positive items that the model correctly flagged.
Recall@k: The fraction of relevant items that appear among the top k retrieved results.
Recall@k (index): The fraction of the exact top-k neighbours that an approximate search also returns in its top k.
Receptive field: The region of the input image that can influence a given unit in a deeper layer.
Reciprocal rank fusion: A method that merges several ranked lists by summing 1/(k + rank) for each document across the lists.
Reciprocal rank fusion (RRF): A fusion method scoring each document by the sum over ranked lists of 1/(k + rank), with k typically 60.
Reconstruction error: The difference between an input and an autoencoder's reconstruction of it, used as an anomaly score.
Recursive splitting: Splitting on the coarsest natural separator first and recursing to finer separators only for pieces that remain too large.
Red-teaming: Deliberately attacking a system to find ways it can be made to fail or misbehave.
Reference monitor: A component that mediates every access by a subject to an object, is protected from tampering, and is simple enough to verify.
Reflection: Having the model review its own output against the goal or external checks and correct it.
Reflection tokens: Special tokens in Self-RAG that let the model signal retrieval need, passage relevance, support and usefulness.
Regression testing: Re-running existing tests after a change to make sure previously working behaviour still works.
Regularisation: A penalty on model complexity, such as the size of the weights, added to the loss to reduce overfitting.
Reinforcement learning: Training a model by rewarding outputs that achieve a goal, such as a correct answer.
Rejection sampling (data): Generating several candidate outputs and keeping only those that pass a verifier or score highest under a judge.
Release bundle: The versioned combination of model snapshot, prompts, tool definitions, retrieval settings and decoding parameters that determines an LLM app's behaviour.
Reparameterization trick: Writing a random latent as a deterministic function of model outputs and independent noise (z = mu + sigma * epsilon) so gradients can flow through sampling.
Reranker: A model that re-scores retrieved candidates by reading the query and each chunk together.
Research engineer: An engineer who implements and scales new methods and experiments, often for large-model training.
Residual connection: A skip path that adds a block's input to its output, so the block learns a correction and gradients flow easily through deep networks.
Residual vector quantization (RVQ): Quantization with a sequence of codebooks in which each stage encodes the residual error left by the previous stages.
Responsibility: In a Gaussian mixture, the posterior probability that a given point was generated by a given component.
Retrieval evaluator: In Corrective RAG, a lightweight model that scores retrieved documents and triggers keep, discard-and-search, or combine actions.
Return: The discounted sum of future rewards from a time step onward.
Return on investment (ROI): Net benefit divided by cost, expressing how much value an investment returns per unit spent.
Reverse-mode automatic differentiation: An algorithm that computes vector-Jacobian products by propagating adjoints from outputs back to inputs; backpropagation is its use for neural networks.
Reward hacking: An agent achieving high reward through behaviour that exploits flaws in the reward signal rather than doing the intended task.
Reward model: A network that outputs a scalar score for a prompt-response pair, trained on human or AI preference comparisons.
Ridge point: The arithmetic intensity, peak compute divided by peak bandwidth, at which work switches from memory-bound to compute-bound.
Risk tiering: Classifying AI systems into levels of governance scrutiny according to the potential severity, scale and likelihood of harm and their legal classification.
RLAIF: Reinforcement learning from AI feedback, where preference labels come from a model rather than humans.
RLHF: Reinforcement learning from human feedback: training a model toward responses that people ranked as better.
ROC-AUC: The area under the curve of true positive rate versus false positive rate, equal to the probability a random positive outranks a random negative.
RoCE: RDMA over Converged Ethernet: a protocol that lets Ethernet networks perform remote direct memory access with low CPU overhead.
Rolling-origin evaluation: Backtesting by repeatedly training on data up to an origin, forecasting the next h steps and scoring, for several origins.
Roofline model: A performance model in which attainable throughput is the minimum of peak compute and bandwidth multiplied by arithmetic intensity.
Rotary position embedding (RoPE): A positional scheme that rotates pairs of query and key dimensions by position-dependent angles so their dot product depends on relative position.
ROUGE: A family of recall-oriented summarization metrics based on n-gram or longest-common-subsequence overlap with reference summaries.
Round-to-nearest (RTN): The simplest quantization method, which rounds each weight independently to the closest representable value.
Router: The component that chooses which knowledge source or tool should handle a query or sub-query.
Router (gate): A small network that scores the experts and picks which ones process each token.
Runtime (inference engine): The software that loads model weights and generates text from them.
Saddle point: A point with zero gradient where the function curves upward in some directions and downward in others.
safetensors: A tensor serialisation format that stores only data and metadata, so loading it cannot execute code.
Safety case: A structured, evidence-backed argument that a system is acceptably safe for a defined use and environment.
Sampling: Choosing the next token at random according to the model's predicted probabilities.
Sampling rate: The number of audio samples recorded per second, which caps the highest representable frequency at half its value.
Sandbox: An isolated environment where code can run without reaching the rest of the system or its secrets.
Saturation: The regime where an activation's derivative is near zero, so little gradient flows through the unit.
Saved memory: Short notes a chat app stores about you and adds to the context of new conversations.
Scalable oversight: Methods that allow humans or weaker systems to reliably supervise AI systems whose capabilities exceed their own on the task.
Scale: A shared multiplier stored for a block of weights so they can be represented with few bits.
Scaled dot-product attention: The operation softmax(QK^T / sqrt(d_k) + M) V, where M is an optional mask.
Scaling law: An empirical power-law relationship between a model's loss and its size, training data or compute.
Schema-constrained decoding: Generation in which a compiled grammar or schema masks, at each step, every token that would make the output an invalid prefix.
Scope 2 emissions: Indirect emissions from purchased electricity, heat, steam or cooling under the GHG Protocol.
Score: The gradient of the log-density with respect to the data, grad_x log p(x), which does not depend on the normalizing constant.
Seasonal naive forecast: A forecast that repeats the value from the same point in the last observed season.
Seasonality: A repeating pattern with a fixed, known period m, such as weekly or yearly cycles.
Secure aggregation: A cryptographic protocol that lets a server learn the sum of many clients' updates without learning any individual update.
Selection rate: The share of a group that a system approves, advances or selects.
Selective SSM: An SSM whose discretization step and input and output matrices are computed from the current input, letting it choose what to keep or forget.
Self-consistency: Sampling several independent reasoning paths and returning the most frequent final answer.
Self-consistency (majority voting): Sampling several independent reasoning paths and returning the most common final answer.
Self-consistency check: Asking the same question several times independently and treating disagreement among the answers as a warning sign.
Self-information: The surprise of an outcome, equal to the negative logarithm of its probability.
Self-Instruct: A method in which a language model generates and filters new instruction-following examples from a small seed set.
Self-Route: A hybrid strategy in which the model answers from retrieved chunks when it judges them sufficient and otherwise falls back to long-context prompting.
Self-supervised learning: Learning in which the training labels are derived automatically from the data itself, such as predicting the next word of a text.
Semantic chunking: Placing chunk boundaries where the embedding similarity between consecutive sentences drops, as a proxy for a topic change.
Semantic memory: Distilled, general facts about the user, domain or world, independent of when they were learned.
Semantic segmentation: Assigning a class label to every pixel without distinguishing separate objects of the same class.
Sensitivity: The maximum change in a query's output, measured in a chosen norm, between any two neighboring datasets.
Sequence-level distillation: Training a student with ordinary cross-entropy on complete sequences generated by the teacher.
Set-of-Mark prompting: Overlaying numbered marks on image regions or UI elements so a vision model can refer to them by label instead of by coordinates.
Shadow deployment: Running a new model on a copy of live traffic while discarding its outputs, so its behaviour can be measured without affecting users.
Short-term memory: The information in the model's current context window, which is lost when the session ends.
Sigmoid: The function 1/(1+e^(-z)) that maps any real number to a probability between 0 and 1.
Silhouette coefficient: For a point, (b - a) / max(a, b), where a is its mean distance to its own cluster and b its mean distance to the nearest other cluster.
Singular value decomposition: The factorization of any matrix as U Σ Vᵀ with orthogonal U and V and non-negative singular values on the diagonal of Σ.
SIP: The Session Initiation Protocol, used to set up, modify and terminate voice and video calls over IP networks.
SLAM: Simultaneous localisation and mapping: estimating a robot's trajectory and a map of its environment at the same time.
Sliding-window attention: Attention that only looks back over a fixed number of recent tokens.
SMOTE: An oversampling method that creates synthetic minority-class examples by interpolating between a minority example and its minority-class neighbours.
Soft binding: A watermark or perceptual fingerprint used to recover a content credential after metadata has been stripped.
Soft targets: The teacher's full probability distribution over outputs, used as training targets instead of one-hot labels.
Softmax: A function that exponentiates a list of scores and divides by their sum, producing positive probabilities that add up to 1.
Software as a Medical Device (SaMD): Software intended for a medical purpose, such as diagnosis or treatment, that performs that purpose without being part of a hardware medical device.
Spacetime latent: A compressed representation of a video in which an autoencoder has reduced both spatial and temporal resolution, later split into patches for a transformer.
Span: A timed, attributed record of one operation within a distributed trace, linked to its parent operation.
Sparse autoencoder (SAE): An autoencoder with an overcomplete, sparsity-penalized hidden layer trained on model activations to decompose them into interpretable feature directions.
Sparse model: A model that uses only a subset of its parameters for each token, such as a mixture of experts.
Speaker diarization: The task of segmenting audio by speaker to determine who spoke when, without prior knowledge of the speakers.
Specification gaming: Behavior that satisfies the literal specification of an objective without achieving the outcome its designers intended.
Spectrogram: A grid showing how much energy an audio signal has at each frequency over time, used as the input to many speech models.
Speculative decoding: A lossless technique in which a cheap draft proposes several tokens that the target model verifies in one forward pass.
Speech-to-speech model: A single model that takes speech as input and produces speech as output, without an explicit text hand-off between separate ASR, language and TTS systems.
Speech-to-text: Technology that converts spoken audio into written text, also called automatic speech recognition.
SQNR: Signal-to-quantization-noise ratio: the power of the original values divided by the power of the rounding error, usually in decibels.
Standardization: Rescaling a feature to zero mean and unit standard deviation using statistics from the training data.
State-space model (SSM): A sequence model defined by a linear recurrence h_t = A-bar h_{t-1} + B-bar x_t with readout y_t = C h_t, obtained by discretizing a continuous linear system.
Stationarity: The property that a series' mean, variance and autocovariance do not change over time.
Statistical reproducibility: Obtaining results within known run-to-run variance when re-running an experiment, rather than bit-identical outputs.
stdio: Standard input and output, the text streams a parent process uses to talk to a child process.
Stochastic gradient descent: Gradient descent that estimates the gradient on a random mini-batch of examples at each step.
Stratified k-fold: Cross-validation that preserves each class's proportion in every fold.
Streaming: Receiving the model's reply token by token as it is generated instead of all at once.
Streaming multiprocessor (SM): The repeated compute block of an NVIDIA GPU that schedules warps and contains general-purpose units, tensor cores, registers and shared memory.
Stride: The number of input positions the kernel moves between successive outputs; a stride of 2 roughly halves each spatial dimension.
Structured output: A mode in which the model's response is constrained to match a supplied schema.
Sub-agent: A separate model instance given a focused task in its own clean context window that returns a condensed result to a coordinating agent.
Superposition: The representation of more features than a layer has dimensions by encoding sparse features along nearly orthogonal directions.
Supervised fine-tuning (SFT): Continued training on prompt-response pairs with next-token loss applied only to response tokens.
Supply-chain risk: Danger that comes from third-party components, such as downloaded models, datasets or plugins, being malicious or compromised.
Support vector machine: A classifier that finds the maximum-margin separating boundary, optionally in a kernel-induced feature space.
Support vectors: The training points on or inside the margin that alone determine an SVM's decision boundary.
Symbolic AI: AI that represents knowledge as explicit, human-readable symbols and rules and draws conclusions by logical inference.
System prompt: Standing instructions, usually set by an app or in settings, that apply to every message in a conversation.
T-shaped skills: Broad working knowledge across a field combined with deep expertise in one area.
Target encoding: Replacing each category with a smoothed average of the target variable for that category.
Task horizon: The length, in human-expert working time, of tasks an AI system can complete at a stated reliability such as 50%.
Teacher forcing: Training a decoder by feeding it the ground-truth previous token rather than its own prediction.
Technology-assisted review (TAR): Using a trained classifier to rank or select documents for production in litigation, validated by statistical sampling.
Temperature: A setting that controls how random the choice of the next token is.
Temporal consistency: The property that objects, identities and textures remain stable and move plausibly across video frames.
Temporal-difference error: The difference between the bootstrapped target r + gamma V(s') and the current estimate V(s).
Tensor / pipeline parallelism: Ways to split one model across several GPUs or machines.
Tensor core: A specialized hardware unit that performs small matrix multiply-accumulate operations in low-precision formats and supplies most of a GPU's AI FLOPS.
Tensor parallelism: Splitting each layer's weights across devices so they all work on the same token simultaneously.
Test set: Held-out data used once at the end to give an unbiased estimate of performance on new data.
Test set (for prompts): A fixed collection of realistic inputs with pass/fail rules, used to compare prompt versions objectively.
Test-time compute: Computation spent during inference, such as longer reasoning, sampling multiple answers, or search, to improve output quality.
Text-to-speech: Technology that converts written text into spoken audio.
Text-to-SQL: Translating a natural-language question into an executable database query whose result answers it.
TF-IDF: A text weighting that multiplies a term's frequency in a document by the log-inverse of how many documents contain it.
TFLOPS: Trillions of floating-point operations per second, a measure of raw compute.
TFLOPS / PFLOPS: Trillions / quadrillions of math operations per second at a stated precision (FP4, FP8, FP16).
Thinking tokens (chain of thought): The intermediate working-out a reasoning model generates before its answer.
Threat model: A structured description of what you are protecting, from whom, and the paths through which it could be exposed.
Tied embeddings: Using the same matrix for the input token embeddings and the output prediction head, which saves V × d parameters.
Time to first token: The delay between sending a request and receiving the first generated token, dominated by queueing, network time and processing the input.
Time to first token (TTFT): The delay between sending a prompt and seeing the first token of the reply, mostly prefill time.
Time-series foundation model: A model pretrained on large collections of series that can forecast new series zero-shot.
Time-weighted exposure: The share of a person's working hours spent on tasks that AI could speed up or assist.
Token: A chunk of text, often a word or part of a word, that a language model reads and writes.
Token budget: An explicit allocation of the context window's tokens among the components of a prompt and the expected output.
Token passthrough: Accepting a client's access token and forwarding it unchanged to downstream APIs, which MCP's security guidance forbids.
Tokenization: Splitting text into the units (words, subwords or characters) that a model processes.
Tokenizer: The component that splits text into tokens and converts them to numbers.
Tokens/sec (decode): How fast a model writes its answer for one user. Mostly limited by memory bandwidth.
Tool annotations: Optional MCP metadata such as read-only or destructive hints that describe a tool's behaviour to clients.
Tool loop: The repeated cycle of calling the model, executing requested tools and returning results until the model gives a final answer.
Tool use / function calling: A model capability for requesting that the application run a named function with specified arguments.
Top-5 error: The fraction of test examples for which the correct label is not among the model's five highest-ranked predictions.
Top-k routing: Sending each token only to the k experts with the highest router scores and weighting their outputs by renormalised gate values.
TOPS: Trillions of integer operations per second, usually INT8; the usual NPU rating.
torch.export: A PyTorch API that captures a model ahead of time as a full graph (an ExportedProgram) for compilation and deployment without Python-level tracing at run time.
Total cost of ownership (TCO): The full cost of a system over its useful life, here hardware depreciation plus energy, expressed per year.
Total parameters: Every weight in the model, all of which must be stored in memory.
TPOT: Time per output token: the average interval between consecutive generated tokens after the first.
Training: The phase in which a model learns by adjusting its parameters to reduce errors on example data.
Training cutoff: The date after which a model has no information, because its training data stops there.
Training opt-out: A setting or agreement that stops your conversations being used to train future models.
Training set: The portion of data used to fit a model's parameters.
Training-serving skew: A mismatch between features or data processing at training time and at inference time.
Training–serving skew: A difference between how inputs or features are computed during training and during serving that degrades a model's live performance.
Training/serving skew: A mismatch between how features are computed during training and in production, causing degraded live performance.
Transfer learning: Reusing a model trained on one task or dataset as the starting point for a different but related task.
Transformer: The neural network design behind most modern language models, built around attention.
Transport: The mechanism that carries protocol messages between client and server, such as stdio or HTTP.
Triple: The atomic fact (subject, predicate, object), such as (Nimbus Labs, ACQUIRED, Kestrel AI).
True positive rate (recall): The share of actually positive cases that a system correctly identifies: TP / (TP + FN).
trust_remote_code: A Hugging Face loading flag that downloads and executes model-specific Python code from the repository.
TTFT: Time to first token: the delay from request arrival to the first generated token, including queueing and prefill.
TTL (time to live): An expiry time after which a stored memory is automatically removed or ignored.
TTS: Text-to-speech, the synthesis of spoken audio from text.
Turing machine: An idealized computing device with an unbounded symbol tape, a read/write head and a finite rule table, used to define what is computable.
Turing test: Alan Turing's 1950 proposal to judge machine intelligence by whether a person conversing by text can tell the machine from a human.
Two-tower model: A retrieval model with separate user and item encoders whose embeddings are compared by a dot product, allowing item embeddings to be precomputed and indexed.
Unified memory: A single pool of memory shared by the CPU and GPU on the same chip package.
Universal machine: A Turing machine that can simulate any other Turing machine when given that machine's description as input, the theoretical basis of the stored-program computer.
Usable AI memory: The part of memory a model runtime can actually use, after the OS and any vendor GPU-allocation limit.
Usable memory: The memory actually available to run a model: VRAM on a discrete GPU, or on unified memory total minus an operating-system reserve (the AI Index's measure, which assumes a Mac's default GPU cap has been raised).
User prompt: The individual message a person types in each turn of a conversation.
v-prediction: A parameterization in which the network predicts v = sqrt(abar_t) epsilon - sqrt(1 - abar_t) x_0 instead of the noise or the clean sample.
Validation set: Held-out data used repeatedly to choose models, features and hyperparameters.
Value × feasibility matrix: A 2×2 prioritisation grid that scores candidate use cases on business value and on how practical they are to deliver.
Vanishing gradient: Gradients shrinking exponentially as they propagate backwards through many layers, so early layers barely learn.
Variance: The expected squared deviation of a random variable from its mean.
Vector: An ordered list of n real numbers, treated as a point or arrow in n-dimensional space.
Vision Transformer (ViT): A transformer that processes an image as a sequence of embedded patches with position information.
Vision-language model (VLM): A model that accepts images, and usually text, as input and produces text grounded in both.
Vision-language-action model: A model, typically built on a vision-language model, that maps images and language instructions to robot actions.
Visual instruction tuning: Fine-tuning a vision-language model on instruction-and-response examples grounded in images.
Vocoder: A model that converts an intermediate acoustic representation, such as a mel spectrogram, into a waveform.
Voice cloning: Using a short audio sample to make a text-to-speech system imitate a particular person's voice.
Voice-to-voice latency: The time from the end of the user's speech to the first audible sound of the agent's reply.
VRAM: The dedicated memory on a discrete graphics card.
Warmup: A schedule phase that increases the learning rate gradually from near zero at the start of training.
Watermark: An imperceptible signal embedded in content at creation that an algorithm with the right key or model can detect.
WebRTC: A set of browser and mobile standards for encrypted real-time audio, video and data with built-in NAT traversal.
Weights: The learned numbers inside a model that determine its outputs.
Wilson score interval: A confidence interval for a binomial proportion that behaves well for small samples and proportions near 0 or 1.
Word embedding: A dense vector representation of a word learned so that words used in similar contexts have similar vectors.
Word error rate (WER): The number of word substitutions, deletions and insertions needed to turn a transcript into the reference, divided by the number of reference words.
Workflow: A sequence of steps fixed in code, possibly including model calls, where the model does not choose the path.
Workflow (vs agent): A system whose control flow is fixed in code, as opposed to an agent whose model chooses the next step.
Workstation GPU: A professional graphics card, often with much more VRAM than consumer cards, aimed at heavy compute work.
WUE: Water usage effectiveness: liters of site water consumed per kilowatt-hour of IT energy.
ZeRO / FSDP: Data-parallel training that shards optimizer state, gradients and optionally parameters across workers instead of replicating them.
Zero point: The integer that represents the real value 0 in asymmetric (affine) quantization.
Zero-shot classification: Classifying into categories supplied at inference time, typically by comparing image embeddings to text-prompt embeddings, without training on those categories.
Zero-shot voice cloning: Generating speech in a new speaker's voice from a short reference recording without training on that speaker.

Put it to work

See which machines run which models, at what speed, in the AI Hardware & Model Index — built on the same methods this program teaches.

Open the AI Index →