A

Activation Checkpointing

• Category: Math & training
• What it actually means: A training-memory technique that saves only selected forward-pass activations and recomputes the omitted ones during backpropagation.
• Why it matters: It lets you train larger models or sequences within a fixed memory budget by trading additional computation for lower activation storage.
• In practice: Checkpoint the memory-heavy transformer blocks, measure the extra step time, and keep recovery checkpoints separate from activation-recomputation settings.
• Common confusion: Activation checkpointing is not a durable training checkpoint. It helps one forward and backward pass fit in memory but cannot resume a crashed run.
• Related terms: Autograd, Backpropagation, Checkpoint, Mixed Precision
• Sources: Training Deep Nets with Sublinear Memory Cost

Activation Function

• Category: Math & training
• What people say: The nonlinear operation between layers.
• What it actually means: A function applied after a linear or affine layer that introduces nonlinearity. Without it, composing layers with weights and biases collapses to one affine transformation. ReLU, GELU, and SiLU are common choices. The choice directly affects whether gradients flow during training.
• Learn it: Activation Functions
• Related terms: ReLU, Gradient, Backpropagation

Adam (Optimizer)

• Category: Math & training
• What people say: The optimizer you use without thinking about it.
• What it actually means: Adaptive Moment Estimation. It combines an exponential average of gradients with an exponential average of squared gradients, applies bias correction, and adapts the update scale per parameter. It is a useful baseline, but it still needs a suitable learning rate and schedule.
• Common confusion: Adam is a strong baseline, not a universal best optimizer.
• Sources: Adam paper
• Related terms: AdamW, Optimizer, Learning Rate

AdamW

• Category: Math & training
• What people say: Adam with weight decay fixed.
• What it actually means: An Adam variant that decouples weight decay from the gradient-based parameter update. That makes the shrinkage behavior easier to reason about than adding an L2 penalty inside Adam's adaptively scaled gradient.
• Common confusion: Decoupled weight decay does not make AdamW universally optimal. Model, data, and training scale still determine the best optimizer and schedule.
• Sources: Decoupled Weight Decay Regularization
• Related terms: Adam (Optimizer), Weight Decay, Optimizer

Admission Control

• Category: Reliability & operations
• What it actually means: A pre-acceptance gate that decides whether a request may enter a bounded queue or service under the system's current capacity, priority, and policy.
• Why it matters: Rejecting excess work at a controlled boundary protects admitted requests from queue growth, timeout cascades, and resource exhaustion.
• In practice: Estimate the request's cost, check tenant and system capacity, reserve the required budget atomically, and identify the overloaded scope when rejecting. Give retry guidance only when the condition is transient and the caller's retry budget permits another attempt.
• Common confusion: Admission control acts before acceptance. Load shedding can reject or remove work at ingress, in queues, at dependencies, or at other overload boundaries.
• Related terms: Load Shedding, Backpressure, Rate Limit, Saturation
• Sources: Google SRE: Handling Overload

Agent

• Category: Agents & tools
• What people say: An autonomous model that thinks and acts alone.
• What it actually means: A software system that lets a model select actions toward a goal, observe tool or environment results, and continue under an orchestration policy. An agent may use a loop, a state machine, a workflow engine, or human approvals. The model is one component, not the entire system.
• Why it matters: Reliability comes from the harness, tool contracts, state, permissions, and verification around the model.
• In practice: A coding agent reads repository context, proposes a patch, runs tests in a sandbox, and stops for approval before deployment.
• Common confusion: Autonomy is a degree of delegated authority, not a required property of every agent.
• Learn it: The Agent Loop
• Related terms: Agent Harness, Agent State, Tool Contract, Human-in-the-Loop (HITL)

Agent Harness

• Category: Agents & tools
• What it actually means: The runtime around a model that assembles context, exposes tools, manages state, enforces limits, records traces, and decides when the agent should continue, retry, ask, or stop.
• Why it matters: Two systems using the same model can perform very differently because their harnesses provide different context, tools, feedback, and safety boundaries.
• In practice: Your harness can limit an agent to five tool calls, persist a checkpoint after each accepted patch, and require a passing test command before completion.
• Common confusion: A harness is broader than a prompt template and narrower than the complete product.
• Learn it: Minimal Agent Workbench
• Related terms: Agent, Tool Contract, Agent State, Verification Gate, Sandbox

Agent Memory

• Category: Agents & tools
• What it actually means: Information stored outside the model and selected for use in later agent steps, such as prior decisions, user preferences, task episodes, or verified facts.
• Why it matters: It gives an agent continuity beyond one context window without forcing every past event into every prompt.
• In practice: Store a compact task outcome with provenance, retrieve it only when relevant, and let the user inspect or correct durable personal information.
• Common confusion: Agent memory is not the same as agent state. State tracks the current run; memory preserves selected information for possible future runs.
• Related terms: Agent State, Context Engineering, Checkpoint, Semantic Cache
• Sources: Generative Agents

Agent State

• Category: Agents & tools
• What it actually means: The explicit data an agent carries across steps, such as the current objective, completed actions, tool results, open questions, budgets, approvals, and artifact references.
• Why it matters: Explicit state makes long tasks resumable, inspectable, and less dependent on the model reconstructing progress from a transcript.
• In practice: Store the selected issue, changed files, latest test result, and remaining checks in a typed object that is updated after each action.
• Common confusion: State is not the same as conversation history. A transcript is evidence; state is the compact operational record used to decide what happens next.
• Learn it: Repository Memory and State
• Related terms: Checkpoint, Durable Execution, Context Engineering, Handoff

Agent Skill

• Category: Agents & tools
• What it actually means: A discoverable directory of procedural instructions whose entry point is SKILL.md, with optional references, scripts, and assets that a compatible runtime can load in stages.
• Why it matters: It packages reusable task knowledge separately from one conversation while keeping deeper context and deterministic helpers available on demand.
• In practice: Publish a compact name and routing description, load the workflow only after activation, and read branch-specific references when the task reaches them.
• Common confusion: Activating a skill supplies context. It does not expose a tool, grant permission, create a sandbox, or prove that the resulting work is correct.
• Learn it: Agent Skills: Portable Contract and Runtime Boundary
• Related terms: Skill Bundle, Skill Catalog, Skill Invocation, Progressive Disclosure, MCP (Model Context Protocol)
• Sources: Agent Skills specification

AI Risk Assessment

• Category: Security & governance
• What it actually means: A documented analysis of how an AI system can affect people, organizations, and environments, including context, hazards, likelihood, impact, controls, residual risk, and monitoring responsibilities.
• Why it matters: Model capability alone does not determine risk. Deployment context, affected groups, human authority, data, and system integrations change both the harms and the controls required.
• In practice: Define the intended use and affected parties, identify credible failure and misuse scenarios, assign owners to controls, record residual risk, and set review triggers for material changes.
• Common confusion: A risk assessment supports a decision under stated assumptions. It is not a one-time safety certificate or proof that every hazard has been found.
• Related terms: Threat Model, Guardrails, Human-in-the-Loop (HITL), Data Classification
• Sources: NIST AI Risk Management Framework

Alignment

• Category: Evaluation & safety
• What people say: Making AI safe.
• What it actually means: The effort to make a model or AI system behave in ways that match intended goals, constraints, and human preferences across both expected and adversarial situations.
• Why it matters: A system can optimize the stated metric while violating the user's real intent, so alignment requires evaluation, oversight, and system controls as well as model training.
• Related terms: Guardrails, Evaluation (Eval), Human-in-the-Loop (HITL)

Approval Gate

• Category: Agents & tools
• What it actually means: A control point that blocks a consequential action until an authorized person or policy grants permission.
• Why it matters: It limits the blast radius of uncertain model decisions while preserving automation for reversible work.
• In practice: Let an agent draft a database migration and run it against a disposable database, but require an owner to approve any production execution.
• Common confusion: An approval gate asks whether an action is authorized. A verification gate asks whether evidence shows the action is correct.
• Learn it: Verification Gates
• Related terms: Human-in-the-Loop (HITL), Verification Gate, Least Privilege

Approximate Nearest Neighbor (ANN)

• Category: Retrieval & generation
• What it actually means: A search method that returns vectors likely to be among the nearest to a query without exhaustively comparing the query with every stored vector.
• Why it matters: Approximation makes large vector indexes practical, but it introduces a measurable tradeoff between search speed, memory, and retrieval recall.
• In practice: Tune index and search parameters against a held-out query set, then report latency together with Recall@K instead of assuming every true neighbor is found.
• Common confusion: ANN describes a search objective and tradeoff, while HNSW is one particular index algorithm that can implement it.
• Related terms: Vector Database, HNSW, Cosine Similarity, Recall@K
• Sources: Efficient and Robust Approximate Nearest Neighbor Search Using HNSW

Attention

• Category: Models & inference
• What people say: How a model focuses on important tokens.
• What it actually means: A mechanism that forms contextual representations by comparing query vectors with key vectors, normalizing the resulting scores, and using them to combine value vectors. Masks, position rules, or sparse patterns can restrict which positions participate.
• Why it matters: Attention lets a model route information between sequence positions, but it does not by itself explain or prove what the model understood.
• Common confusion: Attention weights are computation coefficients, not a faithful explanation of model reasoning.
• Learn it: Self-Attention from Scratch
• Sources: Attention Is All You Need
• Related terms: Self-Attention, Transformer, KV Cache

Audio Token

• Category: Multimodal systems
• What it actually means: A discrete identifier produced by an audio codec or tokenizer for a short segment or feature of an audio signal, sometimes across several codebooks.
• Why it matters: Discrete audio representations let sequence models process, predict, store, or generate sound using token-oriented architectures.
• In practice: Version the codec with the model, preserve sample-rate and codebook metadata, measure reconstruction quality, and distinguish semantic audio tokens from waveform-compression tokens.
• Common confusion: An audio token is not a fixed duration, phoneme, or word. Its meaning and time span depend on the tokenizer and codebook design.
• Learn it: Neural Audio Codecs
• Related terms: Token, Embedding, Automatic Speech Recognition (ASR), Multimodal Model
• Sources: SoundStream

Audit Log

• Category: Security & governance
• What it actually means: A durable, access-controlled record of security- or accountability-relevant events, including who or what acted, what changed, when it happened, and the resulting status.
• Why it matters: Consequential agent actions need evidence that supports investigation, policy review, and responsibility beyond performance debugging.
• In practice: Record tool authorization, approval decisions, external writes, policy versions, and artifact identifiers while redacting secrets and restricting log access.
• Common confusion: A trace helps diagnose one execution path. An audit log preserves events required for accountability across executions and over time.
• Related terms: Trace, Observability, Approval Gate, Provenance Attestation
• Sources: NIST SP 800-92

Autograd

• Category: Math & training
• What people say: Automatic gradients.
• What it actually means: A system that records or transforms tensor operations so it can compute derivatives, usually with reverse-mode automatic differentiation. You write the forward computation and the framework derives the gradients needed for backpropagation.
• Learn it: Chain Rule and Automatic Differentiation
• Related terms: Backpropagation, Gradient, Tensor

Automatic Speech Recognition (ASR)

• Category: Multimodal systems
• What it actually means: The task and system pipeline that maps a speech signal to a transcription, often with optional token or segment timing and confidence information.
• Why it matters: Speech interfaces depend on more than language modeling. Acoustic variation, segmentation, decoding, vocabulary, and domain conditions all affect the final transcript.
• In practice: Evaluate word or character errors by language, speaker, noise, and domain, retain timestamps when downstream grounding needs them, and test the exact audio preprocessing used in production.
• Common confusion: ASR transcribes what was said. Determining who spoke requires diarization or speaker recognition, while translation and intent understanding are separate tasks.
• Learn it: Speech Recognition and ASR
• Related terms: Audio Token, Encoder, Tokenization, Multimodal Model
• Sources: Connectionist Temporal Classification

Autoregressive

• Category: Models & inference
• What people say: The model generates one word at a time.
• What it actually means: A factorization in which each output token is predicted from the tokens that precede it. During generation, the selected token is appended to the sequence and becomes part of the next prediction's context.
• Common confusion: The unit is a token, not necessarily a word, and generation can use decoding methods other than always selecting the highest-probability token.
• Related terms: Token, Temperature, KV Cache

Autoscaling

• Category: Infrastructure & serving
• What it actually means: A control loop that changes the number or capacity of serving workers from observed demand, resource use, or application metrics within configured bounds.
• Why it matters: AI workloads can change faster than manual provisioning, but scaling decisions must account for model-load time, accelerator availability, queueing, and request cost.
• In practice: Scale from a demand signal tied to useful work, set minimum warm capacity, bound scale-down churn, and verify that new replicas pass readiness checks before receiving traffic.
• Common confusion: Autoscaling adds or removes capacity. It does not make an overloaded dependency faster or guarantee that enough hardware can be acquired in time.
• Learn it: GPU Autoscaling on Kubernetes
• Related terms: Model Serving, Saturation, Readiness Probe, Backpressure
• Sources: Kubernetes Horizontal Pod Autoscaling

Availability

• Category: Reliability & operations
• What it actually means: The proportion of eligible service interactions or time windows in which users can obtain the defined acceptable service under a stated measurement boundary.
• Why it matters: A service can be running while users still cannot complete useful requests, so availability must be tied to user-visible success rather than process uptime alone.
• In practice: Define eligible events and acceptable outcomes, exclude only documented cases, calculate the indicator over a fixed window, and investigate both total failures and prolonged partial degradation.
• Common confusion: Availability is one reliability outcome. It does not describe latency, correctness, safety, or the experience of every user segment.
• Related terms: Service Level Indicator (SLI), Service Level Objective (SLO), Error Budget, Incident Response
• Sources: Google SRE: Service Level Objectives