M

Maximum Marginal Relevance (MMR)

• Category: Retrieval & generation
• What it actually means: A selection rule that balances relevance to the query with novelty relative to items already selected.
• Why it matters: It can reduce redundant chunks so a limited context budget covers more distinct evidence.
• In practice: Retrieve a candidate pool, select the next item using a documented relevance-diversity weight, and evaluate both answer quality and source coverage.
• Common confusion: MMR diversifies an existing candidate set; it does not retrieve missing evidence or prove that selected passages are correct.
• Related terms: Reranker, Chunking, RAG (Retrieval-Augmented Generation), Grounding
• Sources: The Use of MMR, Diversity-Based Reranking for Reordering Documents and Producing Summaries

MCP (Model Context Protocol)

• Category: Agents & tools
• What people say: A standard way for AI applications to connect to tools and context.
• What it actually means: An open JSON-RPC protocol for a host to connect to servers that expose tools, resources, prompts, and extensions through defined request, result, discovery, and transport contracts. In revision 2026-07-28, every request carries its protocol version and client capabilities instead of relying on an initialization handshake or protocol session.
• Common confusion: MCP standardizes discovery and exchange. It does not decide which tool is safe to call, grant permission, or forbid an application from using explicit state handles.
• Learn it: Model Context Protocol
• Sources: MCP 2026-07-28 key changes
• Related terms: Stateless MCP, Multi Round-Trip Request (MRTR), Function Calling, Tool Contract, Least Privilege

Membership Inference

• Category: Security & governance
• What it actually means: An attack that estimates whether a particular record or example was included in a model's training data by observing model outputs or other accessible signals.
• Why it matters: Even when the model does not reproduce a record verbatim, distinguishable behavior can reveal information about participation in a sensitive dataset.
• In practice: Test representative members and non-members under the real query interface, limit unnecessary confidence signals, reduce data exposure, and evaluate privacy defenses against utility requirements.
• Common confusion: Membership inference asks whether a record participated in training. Model extraction tries to reproduce model behavior, while direct memorization tests whether content can be recovered.
• Learn it: Differential Privacy for LLMs
• Related terms: Data Leakage, Data Minimization, Eval Set, Data Classification
• Sources: Membership Inference Attacks Against Machine Learning Models

Mixed Precision

• Category: Math & training
• What people say: Using lower-precision arithmetic for speed and memory savings.
• What it actually means: A numerical strategy that uses different data types for different operations, often lower precision for many matrix operations and higher precision for values that need more range or stability.
• Common confusion: Speed, memory, and accuracy effects depend on hardware, data type, scaling method, kernels, and model. They are not a fixed multiplier.
• Related terms: Tensor, CUDA, NaN (Not a Number), Quantization

Modality

• Category: Multimodal systems
• What it actually means: A form of information with its own structure and acquisition process, such as text, image, audio, video, depth, or sensor measurements.
• Why it matters: Different modalities have different sampling rates, noise, spatial or temporal structure, and missing-data behavior, so one preprocessing assumption rarely fits all of them.
• In practice: Document each modality's source, units, resolution, timing, preprocessing, and missing-value policy before designing alignment or fusion.
• Common confusion: A modality is not merely a file extension or feature column. Several encodings can represent one modality, and one sample can contain several modalities.
• Learn it: MIO Any-to-Any Streaming
• Related terms: Multimodal Model, Token, Tensor, Embedding
• Sources: ImageBind: One Embedding Space To Bind Them All; Multimodal Machine Learning: A Survey and Taxonomy

Modality Alignment

• Category: Multimodal systems
• What it actually means: Learning or establishing correspondences between representations from different modalities so semantically or temporally related items can be matched.
• Why it matters: Fusion and cross-modal retrieval fail when the system cannot connect the same event, object, or concept across differently structured inputs.
• In practice: Define positive and negative pairs, preserve time or spatial metadata, evaluate mismatched examples, and measure alignment separately from downstream task accuracy.
• Common confusion: Alignment makes representations comparable or corresponding. It does not require them to become identical or erase modality-specific information.
• Learn it: Projection Layer Modality Alignment
• Related terms: Shared Embedding Space, Contrastive Learning, Grounding, Multimodal Fusion
• Sources: Learning Transferable Visual Models From Natural Language Supervision

Model Card

• Category: Evaluation & safety
• What it actually means: A structured report describing a model's intended uses, evaluation conditions, performance characteristics, limitations, and relevant ethical or safety considerations.
• Why it matters: It gives downstream builders context for deciding whether reported evidence applies to their users and deployment conditions.
• In practice: Document model version, training and evaluation scope, subgroup results, known failure modes, prohibited uses, and the date of each claim.
• Common confusion: A model card communicates evidence and limitations; it is not a certification, warranty, system threat model, or substitute for deployment-specific evaluation.
• Related terms: Eval Set, Dataset Split, Distribution Shift, Alignment
• Sources: Model Cards for Model Reporting

Model Router

• Category: AI-native development
• What it actually means: A component that selects a model or provider for a request using requirements such as capability, latency, cost, context size, policy, and current availability.
• Why it matters: Different tasks and failure conditions justify different models, and routing can improve outcome quality without sending every request to the largest option.
• In practice: Send low-risk extraction to a fast model, complex code review to a stronger model, and fail over only to providers that satisfy the same data policy.
• Common confusion: Routing is a policy decision. Random load balancing only distributes traffic.
• Related terms: Evaluation (Eval), Circuit Breaker, Rate Limit, Cost per Successful Task

Model Serving

• Category: Infrastructure & serving
• What it actually means: The runtime and API layer that loads versioned model artifacts, accepts inference requests, schedules execution, manages resources, and returns results under an operational contract.
• Why it matters: A capable model can still produce an unreliable product when queueing, batching, placement, versioning, cancellation, and response boundaries are not engineered explicitly.
• In practice: Pin model and tokenizer versions, validate request limits, expose readiness and latency signals, control concurrency, and test rollback before routing production traffic.
• Common confusion: Model serving is broader than calling inference once and narrower than the complete application, which may also include retrieval, tools, policy, and user state.
• Learn it: Self-Hosted Serving Selection
• Related terms: Inference, Model Router, Autoscaling, Observability
• Sources: Clipper

MoE (Mixture of Experts)

• Category: Models & inference
• What people say: A large model that activates only part of its parameters for each token.
• What it actually means: An architecture with multiple expert subnetworks and a learned router that selects a subset for each input unit, often each token. Sparse activation can increase total parameter capacity without using every expert on every forward pass.
• Why it matters: Compute, memory, communication, routing balance, and quality depend on the specific architecture and serving system.
• Common confusion: Product names do not prove an MoE architecture unless the model developer discloses it.
• Learn it: Mixture of Experts
• Related terms: Transformer, Model Router, Parameter

Multimodal Fusion

• Category: Multimodal systems
• What it actually means: Combining evidence or learned representations from more than one modality to produce a joint representation, prediction, or generated output.
• Why it matters: Modalities can supply complementary evidence, but naïve combination can amplify noise, timing errors, or one dominant stream.
• In practice: Establish single-modality baselines, specify the fusion point and masks, test missing and contradictory inputs, and report which modalities drive each evaluated slice.
• Common confusion: Fusion is the combination operation. Alignment establishes correspondence, and merely placing two modalities in one request does not prove either occurred successfully.
• Learn it: Cross-Attention Fusion
• Related terms: Early Fusion, Late Fusion, Cross-Attention, Modality Alignment
• Sources: Multimodal Deep Learning; Multimodal Machine Learning: A Survey and Taxonomy

Multimodal Model

• Category: Multimodal systems
• What it actually means: A model that learns from, relates, or generates more than one modality through representation, alignment, fusion, translation, or coordinated prediction.
• Why it matters: Multimodal capability depends on how modalities interact, not simply on accepting several input types, and failures can occur at each representation boundary.
• In practice: Document supported input and output combinations, evaluate each modality alone and together, test missing or conflicting inputs, and track preprocessing versions with the model.
• Common confusion: A pipeline with separate image and text models is multimodal at the system level, but it is not necessarily one jointly trained multimodal model.
• Learn it: MIO Any-to-Any Streaming
• Related terms: Modality, Vision-Language Model (VLM), Multimodal Fusion, Transformer
• Sources: Flamingo: a Visual Language Model for Few-Shot Learning; Multimodal Machine Learning: A Survey and Taxonomy

Multi Round-Trip Request (MRTR)

• Category: Agents & tools
• Aliases: MRTR
• What it actually means: An MCP request pattern in which an operation returns resultType: input_required with one or more inputRequests, then the client retries the original method with inputResponses and the exact returned requestState.
• Why it matters: It lets a stateless server request user, model, or root input without opening a server-initiated JSON-RPC exchange or storing protocol session state.
• In practice: Return an input request from tools/call, collect the authorized response in the host, and retry that same tool call with a new JSON-RPC id.
• Common confusion: requestState is untrusted round-trip data. Integrity-protect it before using it for authorization or business decisions, and do not treat it as a server-side session identifier.
• Learn it: MCP Roots and Elicitation
• Related terms: Stateless MCP, MCP (Model Context Protocol), Human-in-the-Loop (HITL), Tool Contract
• Sources: MCP Multi Round-Trip Requests