I

Idempotency

• Category: AI-native development
• What it actually means: The property that repeating the same operation with the same identity does not create additional side effects beyond the first successful application.
• Why it matters: Retries are normal in distributed agent systems. Without idempotency, one uncertain response can duplicate payments, comments, deployments, or records.
• In practice: Attach an idempotency key to a tool request and persist the completed result so a retry returns that result instead of executing the write again.
• Common confusion: Idempotency does not mean every response is byte-for-byte identical. It means the intended state change is not duplicated.
• Sources: HTTP Semantics: idempotent methods
• Related terms: Retry with Backoff, Durable Execution, Checkpoint

Image Token

• Category: Multimodal systems
• What it actually means: A model-specific visual unit represented as a vector or discrete code, commonly derived from an image patch, region, or learned visual-codebook entry.
• Why it matters: Turning visual input into a sequence lets transformer-style components process images together with text or other tokenized modalities.
• In practice: Document whether tokens are continuous patches or discrete codes, preserve spatial position, test resolution and aspect-ratio changes, and count visual tokens in the model's input budget.
• Common confusion: An image token is not necessarily one pixel, one object, or one fixed physical area. Its scope follows the visual encoder or tokenizer.
• Learn it: Vision-Language Models
• Related terms: Patch Embedding, Token, VAE (Variational Autoencoder), Vision Transformer (ViT)
• Sources: Vision Transformer; VQ-VAE

In-Context Learning

• Category: Prompting & context
• What it actually means: A model adapting its behavior from instructions, examples, or patterns supplied in the current input without an ordinary parameter update.
• Why it matters: It explains how one pretrained model can perform a new task from context while keeping its weights unchanged.
• In practice: Place representative demonstrations before the target input, test order and formatting variants, and keep evaluation examples separate from the demonstrations.
• Common confusion: In-context learning is temporary conditioning, not fine-tuning, durable memory, or proof that the model inferred the intended rule.
• Related terms: Few-Shot, Zero-Shot, Context Window, Prompt Engineering
• Sources: Language Models are Few-Shot Learners

Incident Response

• Category: Reliability & operations
• What it actually means: The coordinated process for detecting, analyzing, containing, recovering from, communicating, and learning from an event that threatens service, data, safety, or security.
• Why it matters: During an incident, clear roles and evidence matter more than improvised heroics, especially when model behavior and distributed dependencies obscure the failing boundary.
• In practice: Define severity and command roles, preserve traces and audit records, stop harmful actions, communicate impact, verify recovery, and track corrective work to completion.
• Common confusion: Incident response manages the event and its consequences. Root-cause analysis and long-term prevention continue after immediate service is restored.
• Learn it: SRE for AI
• Related terms: Observability, Audit Log, Postmortem, Availability
• Sources: Google SRE: Managing Incidents; NIST SP 800-61 Rev. 3

Indirect Prompt Injection

• Category: Security & governance
• What it actually means: A prompt-injection attack delivered through content the system retrieves or observes, such as a webpage, document, email, image text, or tool result, rather than directly through the user's instruction.
• Why it matters: An agent can encounter attacker-controlled instructions while performing an authorized task and mistake that content for authority-bearing guidance.
• In practice: Label external content as untrusted data, separate it from instructions, minimize tool permissions, require approval for consequential actions, and include malicious retrieved content in regression tests.
• Common confusion: Indirect describes the delivery path, not a weaker attack. A hidden instruction in retrieved content can be as consequential as a direct user prompt.
• Learn it: Indirect Prompt Injection
• Related terms: Prompt Injection, Instruction Hierarchy, Trust Boundary, Data Exfiltration
• Sources: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection

Inductive Bias

• Category: Models & inference
• What people say: Assumptions built into a learning system.
• What it actually means: Structural or statistical assumptions that favor some functions or representations over others. Convolution favors locality and shared filters; causal masking favors prediction from preceding positions.
• Common confusion: Transformers still have inductive biases through tokenization, position handling, masking, architecture, data, and objective.
• Related terms: CNN (Convolutional Neural Network), Transformer, Feature

Inference

• Category: Models & inference
• What people say: Running a trained model.
• What it actually means: Executing a trained model to produce predictions, scores, embeddings, or generated tokens without performing an ordinary training update to its parameters.
• Common confusion: An application can update caches, conversation state, or external memory during inference even though model weights stay unchanged.
• Related terms: Autoregressive, Streaming, KV Cache

Instruction Following

• Category: Prompting & context
• What it actually means: A model capability to map natural-language directions and supplied context to behavior that satisfies the stated task and constraints.
• Why it matters: Language generation can be fluent without obeying the user's requested operation, format, boundaries, or priorities.
• In practice: Evaluate instruction adherence separately from answer quality using conflicting constraints, format requirements, irrelevant context, and refusal cases.
• Common confusion: Instruction following is not factual correctness, alignment, or obedience to every string that looks like an instruction.
• Related terms: SFT (Supervised Fine-Tuning), Prompt Engineering, Instruction Hierarchy, Alignment
• Sources: Finetuned Language Models Are Zero-Shot Learners

Instruction Hierarchy

• Category: Prompting & context
• What it actually means: A rule set for resolving conflicts among instructions from sources with different authority, such as application policy, users, and untrusted retrieved content.
• Why it matters: Agent systems mix trusted goals with external text, so the model and harness need a defined response when lower-authority content conflicts with higher-authority constraints.
• In practice: Label untrusted tool output as data, preserve higher-priority constraints outside that content, and test direct and indirect conflict cases.
• Common confusion: An instruction hierarchy can improve behavior but is not a security boundary; least privilege and approval controls still limit consequences.
• Related terms: System Prompt, Prompt Injection, Least Privilege, Tool Contract
• Sources: The Instruction Hierarchy

Inter-Token Latency (ITL)

• Category: Infrastructure & serving
• What it actually means: The elapsed time between two consecutive output-token arrival events for one request, calculated as t_i - t_(i-1) for an output token after the first.
• Why it matters: Individual gaps expose decode stalls and streaming jitter that a per-request average can hide, especially under batching, preemption, or mixed workloads.
• In practice: Record each post-first-token interval with its request and token position, then report distributions by workload, output length, and concurrency without pooling away request boundaries.
• Common confusion: ITL is one interval between consecutive tokens. Time per output token is a per-request average across those intervals, while time to first token covers the wait before streaming begins.
• Learn it: Inference Metrics and Goodput
• Related terms: Time per Output Token (TPOT), Time to First Token (TTFT), Decode Phase, Tail Latency
• Sources: DistServe