A
A/B comparison (blinded)

Showing raters two outputs for the same prompt without revealing which model produced which, and asking which is better.

A/B test

Running two versions side by side on comparable users to measure which is actually better.

ABAC (Attribute-Based Access Control)

Granting access based on attributes — of the user, the resource, and the context (department, time, sensitivity) — rather than fixed roles.

Access control

Restricting which documents a given user is allowed to retrieve.

Accuracy / Performance Metric

In the paper, accuracy is the percentage of problems solved correctly out of a total.

Paper 14
Accuracy / Precision / Recall / F1

Standard metrics for objective tasks like classification.

Activation Function

A non-linear function applied to a neuron's weighted sum before passing the result to the next layer.

Paper 03
AdaLoRA

A LoRA variant that adaptively allocates rank across layers.

Adapter

A small trainable module inserted into a frozen model (the original PEFT idea); also a loose name for a LoRA's saved weights.

Add & Norm (Residual + Layer Norm)

The wrapping applied around every sub-layer: `output = LayerNorm(x + SubLayer(x))`.

Paper 08
Admission control / rate limiting

Capping the queue and per-user load; rejecting fast under overload instead of accepting unservable work.

Advantage / Advantage Estimation

In policy gradient RL, the difference between actual return and baseline: A = reward - V(prompt).

Paper 15
Agent

An AI system that works in a loop — read the situation, decide, act with tools, observe the result, repeat — instead of answering a single message.

Agent loop

The repeating cycle of **Think → Act → Observe** until the goal is reached.

Agent mode

Copilot autonomously plans and performs multi-step tasks (editing files, running commands).

Agentic RAG

A setup where the model (acting as an **agent** in a decide-act loop) chooses whether, what, and how many times to retrieve, rather than running a fixed one-shot pipeline.

AI (Artificial Intelligence)

Software that performs tasks once thought to require human intelligence, like understanding language or making decisions.

AI Agent

A system that uses an LLM as its "brain" to decide, step by step, what actions to take to accomplish a goal — using tools, memory, and reasoning.

AI UX

Designing experiences for an uncertain, fallible, probabilistic tool; also a safety mechanism against overreliance.

AI Winter

A period of reduced funding and interest in AI research, typically following a wave of over-promising and under-delivering.

Paper 02
AIME (American Invitational Mathematics Examination)

A competition mathematics exam (15 problems, 3 hours).

Paper 24
ALBERT

A BERT variant that reduces parameters by factorising the embedding matrix and sharing weights across Transformer layers.

Paper 11
Alerting

Automated paging when SLOs are at risk; alert on user-felt symptoms (latency, errors), not just machine stats.

Alignment

The process of training an AI system to behave in ways that align with human values, intentions, and safety constraints.

Paper 22
Alignment / Aligning Language Models

Making language models behave in accordance with human values and preferences.

Paper 15
Alignment matrix (attention heatmap)

A grid where each row corresponds to a target word and each column to a source word.

Paper 07
Alignment model

The small neural network inside the attention mechanism that scores how well a decoder state matches each encoder hidden state.

Paper 07
All-to-All Communication

A communication pattern where every GPU sends data to every other GPU.

Paper 19
Allowlist

A list of approved tools/recipients/domains the agent is restricted to.

AMC (American Mathematics Competitions)

A sequence of mathematics competitions for students (AMC 8, 10, 12).

Paper 24
ANN (Approximate Nearest Neighbor)

Fast search that finds *almost certainly* the closest vectors, trading a tiny bit of accuracy for huge speed.

Annotator Burnout

Psychological harm experienced by humans who repeatedly review harmful content (violence, abuse, self-harm).

Paper 22
Annoy / ScaNN / DiskANN / Vamana

Other ANN indexes: trees (Annoy, static), quantization+pruning (ScaNN), SSD-resident graph (DiskANN, billion-scale cheaply).

Anti-pattern

A tempting-but-wrong approach; a common, costly mistake.

Apache 2.0 License

A permissive open-source license allowing commercial use without restriction.

Paper 18
Artifact

A file or output Claude produces and saves during a session (a document, a diagram, a report) that you can reopen and reuse later, especially in Cowork.

Attention

A mechanism introduced the same year as seq2seq (Bahdanau et al., 2014)

Paper 06
Attention Complexity

The computational cost of attention, typically measured in FLOPs (floating point operations).

Paper 18
Attention head

One of several parallel attention computations; more heads = richer attention, bigger KV cache.

Attention Mechanism

The key innovation in Transformers that allows each token to consider the relevance of all other tokens in the sequence.

3 papers
Attention weight (αₜᵢ)

The probability-like number, between 0 and 1, representing how much the decoder at decoding step t focuses on source position i.

2 papers
Attribute

A key-value detail attached to a span (model name, token count, cost, etc.).

Authentication (AuthN)

Proving *who* you are (a user or service).

Authorization (AuthZ)

Deciding *what* an authenticated identity is allowed to do.

Automation bias

The tendency of human reviewers to over-trust the AI and stop truly checking.

Autonomy

How much an agent decides and acts on its own versus asking a human.

Autoregressive generation

The decoder's mode of operation: generate one token at a time, feed the generated token back as input, generate the next.

Paper 08
Autoregressive language model

A model that generates a sequence by predicting one token at a time, conditioning each prediction on all previously generated tokens.

Paper 10
Autoregressive Language Modeling

A training objective where the model learns to predict the next token given all previous tokens.

Paper 12
Autoscaling

Adjusting replica count to match traffic; for LLMs, scale on queue length / KV-cache utilization / TTFT-vs-SLO, not CPU.

Auxiliary balancing loss (L_balance)

An additional loss term added to the main cross-entropy language modelling loss during MoE training.

Paper 09
AWQ

A post-training quantization method for fast, accurate GPU inference.

Axolotl

A higher-level, config-driven fine-tuning framework.

B
Back-propagation (MCTS)

The process of updating node statistics (visit counts, accumulated rewards) as you trace back from a leaf node to the root after a rollout.

Paper 24
Background task

A command or job Claude starts and lets run without waiting — useful for long builds, test suites, or watchers.

Backpropagation

Short for "backward propagation of errors." The algorithm that computes the gradient of the loss function with respect to every weight in a multi-layer network, by applying the chain rule backwards from the output layer to the input layer.

Paper 03
Backpropagation Through Time (BPTT)

The standard way of training RNNs and LSTMs.

Paper 04
Bare repository

A repo with only `.git` contents and no working files; what servers use.

Base model / Foundation model / Pretrained model

A model that has only been through pretraining.

Baseline

A reference point (e.g., the un-fine-tuned base model) you compare your model against on the same test set.

Batch / Batch size

A group of examples processed together for one weight update; batch size = how many.

Batch processing

Running non-urgent requests together at a discount.

Batch Size

The number of training examples (or sequences) processed in a single gradient update.

Paper 13
Batching / Continuous batching

Processing many requests together to use the GPU efficiently; the main cost lever in serving.

Beam search

An inference-time decoding algorithm.

2 papers
Benchmark

Standardised tests for evaluating language model quality (e.g., MMLU for general reasoning, GSM8k for math, HumanEval for coding).

Paper 18
benchmark_serving / benchmark_throughput

vLLM's built-in load/benchmark scripts.

BERTScore

An embedding-based metric comparing meaning rather than exact words.

Best-of-N (BoN)

A strategy where you generate N independent solutions to the same problem and select the best one according to some criterion (e.g., a Process Reward Model score).

Paper 23
Beta (DPO)

The DPO parameter controlling how strongly to stay near the reference (SFT) model.

Bi-encoder

An embedding model that encodes the query and each document *separately* then compares.

Bi-encoder vs cross-encoder

Bi-encoder embeds query and doc separately (fast, searchable); cross-encoder scores them together (slow, precise) — used for reranking.

Bidirectional context

A representation built from both left and right context simultaneously.

Paper 11
Bidirectional LSTM

An LSTM that processes a sequence in both directions — forward and

Paper 04
Bidirectional RNN (BiRNN)

Two recurrent networks — one reading left to right, one right to left — whose hidden states are concatenated at each position.

Paper 07
Binary Classification

The task of deciding whether an input belongs to one of two categories — yes or no, 0 or 1, cat or dog.

Paper 02
Binary quantization

1 bit per number; up to 32× smaller, Hamming-distance search, very fast.

BitFit

A selective PEFT method that trains only the bias terms.

bitsandbytes

The library that handles 4-bit/8-bit quantization for QLoRA.

BLEU / ROUGE / METEOR

Word-overlap metrics for generation tasks; weak proxies, use cautiously.

BLEU score

Bilingual Evaluation Understudy.

2 papers
Blob

Stores the raw contents of one file (no name, no path).

Blockwise Attention

Computing attention in blocks (query chunk × KV chunk) rather than all-at-once.

Paper 19
BM25

A classic keyword-ranking algorithm; the "sparse" half of hybrid search.

BooksCorpus

The training dataset for GPT-1: approximately 7,000 unpublished novels scraped from the web, totalling ~800 million words.

Paper 10
Bootstrapping

A process where improvement in one component (the model) enables improvement in another (the data), which feeds back to improve the first component further.

Paper 24
Bottleneck (context vector bottleneck)

The design flaw in plain seq2seq (Paper 06): the entire source sentence's meaning must be compressed into a single fixed-size vector before decoding can begin.

Paper 07
BPE (Byte Pair Encoding)

A subword tokenisation algorithm that splits words into common subunits.

Paper 10
Bradley-Terry Model

A probabilistic ranking model from statistics, used here to model human preferences.

2 papers
Branch

A movable pointer to a commit; physically a one-line file in `.git/refs/heads/`.

Branch protection / ruleset

Enforced rules on a branch: require PRs, approvals, passing checks, signed commits; block force-pushes and deletion.

Brute force / Flat / Exact search

Compare the query to every vector.

Budget / quota / circuit breaker (cost)

Limits and automatic stops that prevent runaway spending from bugs, spikes, or abuse.

Build vs buy

Self-hosting GPUs (cheaper at high steady volume) vs managed per-token APIs (cheaper/simpler at low or spiky volume).

Bypass actor

A user granted a narrow exception to a ruleset.

C
Caching

Reusing prior results to avoid paying again.

Calibrated trust

Helping users trust the AI exactly as much as it deserves — avoiding both over-trust and under-trust.

Calibration dataset

A small set of representative texts used to set quantization scales well.

Canary release

Releasing a change to a small percentage of traffic first, watching metrics, then expanding.

Candidate vector (c̃ₜ)

A vector of proposed updates to the cell state, produced by a tanh

Paper 04
Capability negotiation

The handshake where client and server announce which features they support, so each side only uses what the other understands.

Capacity factor

A multiplier that sets the maximum number of tokens each expert can process per batch: `capacity = (batch_tokens / n_experts) × capacity_factor`.

Paper 09
Capacity planning

Estimating how many replicas/GPUs a workload needs: peak demand ÷ single-replica capacity × safety factor.

Cardinality

How many distinct values a label can have.

Catastrophic Forgetting

When a model loses knowledge from pretraining while being fine-tuned on new data.

Paper 15
Causal (masked) self-attention

Self-attention where position i is prevented from attending to positions j > i (future tokens).

2 papers
Causal Language Modeling

Same as autoregressive language modeling: predict the next token given previous tokens.

Paper 12
Causal mask (autoregressive mask)

A (T × T) upper-triangular boolean mask applied in the decoder's self-attention.

Paper 08
Causal Masking

In autoregressive language modelling, preventing the model from attending to future tokens (tokens that come after the current position).

2 papers
CBOW (Continuous Bag of Words)

One of the two Word2Vec training tasks.

Paper 05
Cell state (c)

The "notebook" of an LSTM — a vector that flows from one time step to

Paper 04
Chain rule

A rule in calculus for differentiating composed functions.

Paper 04
Chain-of-Thought (CoT) Prompting

A prompting technique where intermediate reasoning steps are shown in few-shot examples, causing language models to generate their own step-by-step reasoning before producing a final answer.

2 papers
Chat

Conversational Q&A with Copilot about your code.

Chat template

The model-specific format (with special tokens) for laying out system/user/assistant turns.

Chatbot

A system that responds to a single message.

Chinchilla Ratio

An improvement on the compute-optimal frontier (from DeepMind's Chinchilla paper, 2022).

Paper 13
Chinese Room

A thought experiment proposed by philosopher John Searle in 1980 as a critique of the Turing Test.

Paper 01
Chunk

A small passage of a document (often a paragraph) stored and retrieved as a unit.

Chunked prefill

Breaking long prefills into chunks interleaved with decode steps, so big prompts don't stall everyone's TPOT.

Chunking

Splitting documents into chunks.

CI/CD (Continuous Integration / Continuous Deployment)

Automated testing and deployment on every change; for AI, evals run here to gate changes.

Circuit breaker

Automatically stops sending requests to a clearly-failing dependency for a while, using the fallback instead.

Citations / sources

Linking the documents an answer is based on, so users can verify.

Claude Code

Anthropic's command-line coding agent: an AI that reads, writes, and runs code in your project, using tools, memory, and your instructions.

CLAUDE.md

A plain-Markdown briefing file Claude reads automatically at the start of a session.

Client

The component inside the host that maintains a connection to one MCP server and speaks the protocol on the host's behalf.

Closed-loop vs open-loop load

Keeping exactly N requests in flight vs sending requests at an arrival rate (queuing if busy).

Cloze test

A reading comprehension exercise invented in 1953 where words are systematically removed from a passage and the reader must fill them in.

Paper 11
Code completions

Inline "ghost text" suggestions as you type.

Code review

Line-by-line feedback on a PR before merging.

Code scanning (CodeQL)

Static analysis that finds vulnerability patterns in your code and flags them on PRs.

Code-based grader

A deterministic rule check (valid JSON?

ColBERT / late interaction

One vector per token + MaxSim matching; cross-encoder-like quality, still scalable, larger storage.

Combined loss (L₃)

The total fine-tuning loss: L₃ = L_task + λ · L_language_model.

Paper 10
Commit

One saved snapshot of your project, with a unique hash, author, date, and message.

Commit object

Points to a root tree, parent commit(s), author info, and message.

Communication Complexity

The amount of data that must be transferred between GPUs.

Paper 19
Compaction

When a conversation gets long, Claude summarizes earlier context to free up room while keeping the important facts.

Computability

A mathematical property of problems: a problem is "computable" if it can be solved by a Turing Machine (i.e.

Paper 01
Compute Budget (C)

The total computational resources available for training, measured in FLOPs (floating point operations).

Paper 13
Compute-bound

Limited by calculation speed (cores busy, data plentiful).

Compute-Communication Overlap

The simultaneous execution of computation and communication.

Paper 19
Compute-Optimal

Achieving the best accuracy for a given computational budget.

Paper 24
Compute-Optimal Frontier

The boundary of efficient training allocations: the curve of (N, D) pairs that minimize loss for a given compute budget C.

Paper 13
Compute-Optimal Strategy

The choice of which inference-time strategy (Best-of-N vs.

Paper 23
Confused deputy

A security flaw where a trusted component is tricked into misusing its authority for someone else — a key risk MCP authorization design must prevent.

Connectionism

The school of thought in AI that believes intelligence emerges from the interactions of many simple connected units (like neurons), rather than from explicit symbolic rules.

Paper 02
Consciousness

The subjective experience of being aware, of having an inner life.

Paper 01
Consent

The user's explicit approval before an app acts on their behalf (e.g.

Constant error carousel

The original paper's name for the additive structure of the cell state.

Paper 04
Constitution

A written document specifying principles that an AI should follow.

Paper 22
Constitutional AI (CAI)

An alignment methodology that replaces human feedback with AI feedback.

Paper 22
Container / Docker

A portable bundle of app + dependencies that runs identically everywhere; the fix for GPU environment ("works on my machine") pain.

Content-addressed

Objects are named by the hash of their content, so identical content is stored only once.

Context

Everything Claude can "see" right now: your messages, files it has read, tool results, and instructions.

Context engineering

The craft of deciding what to put into Claude's limited context — and what to leave out — so it has exactly what it needs and isn't distracted by noise.

Context Length

The maximum number of tokens a model can process in a single input.

Paper 18
Context length / Context window

The maximum amount of text (in tokens) a model can consider at once, e.g., "8k context" = 8,192 tokens.

Context Parallelism

Distributing a long sequence across multiple GPUs along the sequence dimension (distinct from data, tensor, or pipeline parallelism).

Paper 19
Context propagation

Passing trace context so each span knows its parent and the trace tree assembles correctly.

Context stuffing

Cramming too much into a prompt in the hope something helps.

Context vector

Also called the **thought vector**.

2 papers
Context window

The maximum number of tokens the model can see at once.

2 papers
Context window / context length

The maximum number of tokens (prompt + output) a model can handle at once, e.g.

Continued pretraining (domain-adaptive pretraining)

Extra next-token-prediction training on in-domain text to inject domain knowledge before SFT.

Continuous (in-flight / dynamic) batching

Refreshing the batch every token step, removing finished requests and adding new ones.

Continuous Integration (CI)

Automatically testing every push/PR to catch breakage early.

Convergence

In machine learning, a model has converged when its weights have settled and further training produces no improvement.

Paper 02
Conversational vs. embedded UX

Open chat (flexible but a blank-box burden) vs.

Convolutional Mode (Training)

During training, the recurrence x_t = Āx_{t-1} + B̄u_t can be unrolled and rearranged as a convolution: output = conv(input, kernel).

Paper 21
Copilot coding agent

A cloud agent you assign an issue; it works in the background and opens a PR.

Copilot Memory

A GitHub-hosted, repo-scoped long-term memory; Copilot saves, validates, refreshes, and expires facts (~28 days unused).

Copy-on-write KV sharing

Letting requests share identical KV blocks (e.g.

Core

An individual processing unit; GPUs have thousands.

Corpus

A body of text used to train a model.

Paper 05
Corrective RAG / Self-RAG (CRAG)

Patterns where the model critiques retrieved context and re-retrieves or falls back if it's insufficient.

Cosine similarity

A measure of similarity between two vectors based on the angle between

Paper 05
Cost attribution

Breaking spend down by feature, customer, model, or endpoint to see what/who drives cost.

Cost management / FinOps for AI

Making AI cost visible, accountable, and continuously optimized.

Cost per request / unit economics

The dollar cost of one request, and per-user/per-conversation/per-outcome cost.

Cost per token

GPU $/hour ÷ tokens/hour at the operating point; the metric leadership cares about.

Cost-quality-latency triangle

The trade-off: you can't maximize cheap, good, and fast all at once; you choose the balance per use case.

Cowork

The collaborative surface for Claude (schedules, artifacts, plugins, shared sessions) that goes beyond a single terminal chat.

CPU

A processor with a few very capable cores; great at complex sequential tasks.

Critic / Reviewer

An agent that checks others' work for errors.

Critique Prompt

A prompt asking an AI model to evaluate its own or another model's output against a specific principle.

Paper 22
Cross-attention

Attention where the query comes from one sequence (the decoder) and the keys and values come from another sequence (the encoder).

2 papers
Cross-encoder

A model that reads query and document *together* to judge relevance.

Cross-Entropy Loss

The standard loss function for language modeling.

2 papers
CUDA

NVIDIA's platform for running general computation on their GPUs.

CUDA graphs

Recording a whole step's GPU operations and replaying them as one unit to cut launch overhead.

Custom instructions

Persistent files Copilot folds into requests so you don't repeat rules.

D
d_model (model dimension)

The dimension of all input and output vectors in the Transformer.

Paper 08
Data Annotation / Labeling

The process of having humans provide labels (e.g., preference comparisons) for training data.

Paper 15
Data card

A short document recording a dataset's source, size, splits, cleaning, and limitations.

Data leakage

When test (or validation) examples also appear in training, making evaluation falsely optimistic.

Data parallelism (DP)

Running multiple full copies (replicas) of a fitting model behind a load balancer to add throughput.

Data/RAG framework

Connects agents to your data for retrieval (LlamaIndex…).

Dataset Size (D)

The number of training tokens (or training examples).

Paper 13
datasets (Hugging Face)

Library for loading and processing training data.

Debate pattern

Agents discuss/challenge each other to reach a better answer.

Decision framework

The repeatable nine-step procedure for designing any serving deployment (brief → fit → precision → parallelism → optimizations → config → benchmark/size → operate → cost).

Decode

The second phase: generating output tokens one at a time, sequentially.

Decoder

The second half of the seq2seq architecture.

2 papers
Decoder-only Transformer

A Transformer that uses only the decoder stack (masked self-attention + feed-forward), without the encoder-decoder cross-attention.

2 papers
Decomposition

Breaking a big task into smaller tasks.

Decomposition / Reasoning Decomposition

Breaking a complex problem into smaller, simpler subproblems and solving each one before combining results.

Paper 14
Deduplication

Removing exact and near-duplicate examples from a dataset.

Defense in depth

Layering controls (prevent → detect → block → review → recover) so no single failure is catastrophic.

Delimiters

Clear markers (triple backticks, XML-style tags, headings) that separate parts of a prompt — instructions vs.

Dense computation

The standard approach in neural networks: every parameter fires for every input.

Paper 09
Dense search

Embedding/vector search (vectors are "dense" with numbers).

Dense vs sparse retrieval

Dense = embeddings (meaning); sparse = word-based vectors (exact terms).

Dependabot

GitHub feature that alerts on vulnerable dependencies and opens PRs to update them.

Dependency tree of the field

The map showing how every technique traces back to a few root facts (token-by-token generation; separate GPU compute/memory; memory-bound decode and the KV cache).

Dequantization

Converting quantized numbers back to higher precision for computation; adds small overhead.

Deterministic

Same input always gives the same output (normal software).

Diagnostic checklist

The ordered set of checks for diagnosing slow or broken serving, each mapping to an earlier module's concept.

Dimension (dim)

How many numbers are in a vector (e.g., 384, 768, 1536).

Direct injection / jailbreaking

The *user* types malicious instructions to bypass the AI's rules.

Disaggregated prefill/decode

Running prefill and decode on separate GPU pools so the two phases stop interfering; a large-scale technique.

Discretisation

The process of converting a continuous-time differential equation (dx/dt = Ax + Bu) into a discrete recurrence (x_t = Āx_{t-1} + B̄u_t).

Paper 21
DistilBERT

A compressed version of BERT created by knowledge distillation (training a small model to mimic the outputs of a larger one).

Paper 11
Distillation

Using a strong "teacher" model to generate outputs to train a smaller "student" model.

Distributed (DVCS)

Every clone holds the full history, so you can work offline and each copy is a backup.

Distribution drift

Data's shape changing over time, staling IVF clusters / PQ codebooks; fix by retraining/re-indexing.

Distributional hypothesis

The linguistic claim (often attributed to Firth, 1957) that words

Paper 05
Distributional Shift

When the RL policy generates responses very different from the distribution the reward model was trained on.

Paper 15
dₖ (key dimension)

The dimension of the Query and Key vectors.

Paper 08
DoRA

A LoRA variant that decomposes weights into magnitude and direction; often a small quality gain.

Double quantization

Quantizing the quantization constants too, saving extra memory (from QLoRA).

DPO (Direct Preference Optimization)

A simpler alternative to RLHF that trains directly on chosen/rejected pairs, no reward model or RL needed.

Drift

Behavior decaying over time: *model drift* (the provider changed the model), *data drift* (inputs changed), *knowledge drift* (your knowledge base went stale).

Dualism

The philosophical position, associated with René Descartes, that mind and matter are fundamentally different kinds of thing.

Paper 01
E
Early stopping

Halting training when validation loss stops improving, to prevent overfitting.

Effective batch size

`per_device_batch_size × gradient_accumulation_steps` — the real batch size after accumulation.

Efficient Attention

Modified attention patterns (sliding window, global, sparse) that reduce computation from O(n²) to O(n log n) or O(n), making long sequences tractable.

Paper 20
Eigenvalue (of State Matrix A)

A scalar λ such that Av = λv for some eigenvector v.

Paper 21
Element-wise product (Hadamard product, ⊙)

Multiplication of two vectors of equal length, slot by slot.

Paper 04
Elicitation

A server primitive that lets a server ask the user for additional input mid-task, through the client.

ELIZA

A chatbot created in 1966 by Joseph Weizenbaum at MIT.

Paper 01
Embedding

A dense, low-dimensional vector representation of something — a word,

3 papers
Embedding dimension (d)

The length of each word vector.

Paper 05
Embedding matrix (W)

The (V × d) matrix whose rows are the embeddings for each vocabulary

Paper 05
Embedding model

The model that turns text into embeddings.

Emergent Abilities

Capabilities that appear when a language model reaches a certain scale, but were not present (or were very weak) in smaller versions.

Paper 12
Emergent Capability

A capability that appears in large language models above a certain size threshold, even though it was not explicitly trained for.

Paper 14
enable-prefix-caching / enable-chunked-prefill / kv-cache-dtype / quantization

Flags turning on the corresponding techniques from Modules 04–05.

Encoder

The first half of the seq2seq architecture.

2 papers
Encoder hidden state (hᵢ)

The vector produced by the bidirectional encoder for source position i.

Paper 07
Encoder-decoder

The Transformer's two-part structure for seq2seq tasks (e.g., translation).

Paper 08
End-to-end learning

A philosophy, popularised by this paper, where a single neural network

Paper 06
Entitlement

A specific, fine-grained permission a given identity holds (e.g.

Entropy Regularization

A term in RL that encourages exploration by rewarding policy entropy (randomness).

Paper 15
Episodic / Semantic / Procedural memory

Memories of events / general facts / how-to skills.

Epoch

One complete pass through all training examples.

2 papers
Escalation / routing

Sending only the cases that need judgment (low confidence, flagged, high-stakes) to humans, and to the right human.

Escape hatch

A clear way for the user to reach a human or bypass the AI when it isn't working.

Eval (evaluation)

Measuring prompt quality systematically against a set of test cases with known good answers, so you can tell whether a change actually helped.

Eval set / Test set

A collection of example tasks with expected outcomes, used to score the agent and catch regressions.

Evaluation (eval)

Measuring the quality of an agent's outputs, ideally with numbers across many examples.

Evaluation / eval

Systematically measuring AI output quality across many cases.

Evol-Instruct

Generating data by prompting a model to make existing examples harder and more varied.

Excessive agency

Giving an agent more power/tools than it needs, so one mistake or injection causes outsized damage.

Expert

One of n specialised feed-forward networks in an MoE layer.

Paper 09
Expert collapse

A training failure mode where the gating network routes most tokens to a small number of popular experts, leaving the rest undertrained and effectively unused.

Paper 09
Expert parallelism (EP)

Distributing a Mixture-of-Experts model's experts across GPUs.

Exploding gradient problem

The opposite failure mode: the gradient grows without bound when

Paper 04
Exploration-Exploitation Trade-off

The fundamental challenge in search and learning: should you exploit what you've learned (focus on high-reward nodes) or explore new options (try under-explored nodes)?

Paper 24
Exponent (Alpha, Beta, Gamma)

The power in a power law.

Paper 13
Extrapolation

Extending a fitted line to predict values beyond the observed range.

Paper 13
F
FAISS

Meta's industry-standard ANN *library* (not a server).

Faithfulness / groundedness

Whether every claim in an answer is supported by the retrieved context; the hallucination detector.

Faithfulness / groundedness

Whether an answer sticks to its provided sources rather than making things up.

Fallback

A worse-but-working alternative when the primary path fails (another model, a cache, a non-AI path, or a human).

fastText

A 2016 extension of Word2Vec (by Bojanowski et al.) that represents

Paper 05
Feature flag

A switch to turn a change on/off instantly without redeploying.

Feed-forward network (FFN)

A two-layer MLP applied position-wise after attention: `FFN(x) = max(0, xW₁ + b₁)W₂ + b₂`.

Paper 08
Feed-forward network / MLP

The per-token processing part of a layer; holds many parameters.

Few-shot

Including a few worked examples in the prompt to show the model the pattern you want, instead of only describing it.

Few-shot learning

The ability to perform a task from a small number of examples — typically tens to hundreds.

2 papers
Few-Shot Prompting

A technique where a language model is shown a small number of examples (typically 2-8) before being asked to solve a new problem.

Paper 14
FFN (Feed-Forward Network) / MLP

The per-token processing sub-layer in a Transformer block; holds much of the model's parameters and knowledge.

Filtering (pre / post / in-search)

Restricting search by metadata.

Fine-tuning

Adapting a pre-trained model to a specific task by continuing training on labelled task data with a small learning rate.

4 papers
FlashAttention

An exact, memory-efficient attention kernel that avoids writing big intermediates to slow memory; speeds attention, especially for long contexts.

FLOPS / TFLOPS

Floating-point Operations Per Second; raw compute speed.

Flywheel

The improvement loop: production failures feed the eval set, which catches them before the next release, which improves the system.

Forget gate (fₜ)

A sigmoid-valued vector that decides which slots of the previous cell

Paper 04
Fork

Your own full copy of someone else's repo, for contributing without write access.

Forward Pass

The computation that flows from input to output through the network: input → layer 1 → layer 2 → ...

Paper 03
Foundation model

A large model pre-trained on broad data that can be adapted to many downstream tasks.

Paper 10
FP32 / FP16 / BF16

32-, 16-bit floating-point precisions.

FP32 / FP16 / BF16 / FP8 / INT8 / INT4

Number formats of 32/16/16/8/8/4 bits (= 4/2/2/1/1/0.5 bytes per parameter).

FP8

Modern 8-bit float, natively accelerated on the newest GPUs; memory savings + real compute speedup with small quality loss.

Framework

Software that handles the repetitive plumbing of building agents (loop, tool parsing, memory, coordination).

Full fine-tuning

Updating all of a model's weights.

Function calling / Tool calling

The model outputting a structured (usually JSON) request to use a tool, which the surrounding program then executes.

G
Garbage collection (`git gc`)

Periodic cleanup that packs loose objects and eventually removes unreferenced ones.

Gated Recurrent Unit (GRU)

A simpler variant of the LSTM proposed by Cho et al.

Paper 04
Gateway / proxy

A single chokepoint all model calls pass through, used to centralize caching, routing, rate limits, budgets, and logging.

Gating network

The learned routing function `G(x) = Softmax(TopK(H(x), k))`.

Paper 09
GDPR / CCPA / HIPAA / EU AI Act

Examples of data-protection and AI regulations imposing real obligations (consent, deletion, transparency, sometimes mandatory human oversight).

GELU activation

Gaussian Error Linear Unit: GELU(x) = x · Φ(x), where Φ is the Gaussian CDF.

Paper 10
Gemini Nano

The smallest variant (~2–7B parameters) of Gemini, optimized for on-device inference on mobile phones and edge devices.

Paper 20
Gemini Pro

The balanced variant (~50B parameters estimated) of Gemini, deployed for most production use (Google Bard, Workspace, Search).

Paper 20
Gemini Ultra

The largest variant (~1.3T parameters estimated) of Gemini, achieving the highest benchmarks (90.04% MMLU) but requiring significant compute.

Paper 20
Generalization

The ability of a model to perform well on new, unseen data (test set).

Paper 13
Generalization / Generalizing to New Domains

Whether a trained reward model (or policy) performs well on new, unseen tasks or domains.

Paper 15
GGUF

A file format (llama.cpp/Ollama) for running quantized models efficiently on CPU/Mac.

GitHub Actions

Automation (CI/CD) defined as YAML in `.github/workflows/`.

GitHub Advanced Security (GHAS)

The bundle of advanced security features (secret scanning, code scanning, etc.) for private/enterprise repos.

GitHub Copilot

An AI pair programmer (an LLM wired into your tools) that suggests, explains, and writes code.

Glossary: Let's Verify Step by Step

### Outcome Reward Model (ORM)

2 papers
GloVe

A 2014 alternative to Word2Vec, from Stanford.

Paper 05
GLUE benchmark

General Language Understanding Evaluation.

Paper 11
Gödel's Incompleteness Theorem

A 1931 result by mathematician Kurt Gödel: in any formal mathematical system powerful enough to describe arithmetic, there are true statements that cannot be proved within that system.

Paper 01
Golden dataset / eval set

Curated questions + correct answers + source chunks, used to measure retrieval and generation quality.

Goodput

Throughput counting only requests that met their latency SLO.

Goodput per dollar within SLO

The one-line summary of the serving engineer's objective.

GPTQ

A post-training quantization method for efficient GPU inference.

GPU (Graphics Processing Unit)

A processor with thousands of simpler cores; great at massively parallel work like the matrix math in LLMs.

GPU utilization (and its trap)

A "GPU is busy" reading that can sit at 100% during memory-bound decode while compute is idle — so it's a misleading efficiency signal for LLMs.

gpu-memory-utilization

Flag for the fraction of GPU memory vLLM may use; the key throughput/stability knob (bigger KV cache vs less headroom).

Graceful degradation

Designing the system to get slower/simpler/fall back under stress rather than failing outright.

Grader / scorer

Whatever decides how good an output is: code-based, human, or LLM-as-judge.

Gradient

The vector of partial derivatives of the loss with respect to every

Paper 04
Gradient accumulation

Summing gradients over several mini-batches before updating, to simulate a larger batch with less memory.

Gradient Checkpointing

A memory optimisation technique where intermediate activations are discarded during forward pass and recomputed during backward pass.

Paper 19
Gradient Descent

The optimisation algorithm that trains neural networks.

Paper 03
Graduated autonomy

Moving a task from tighter to looser human oversight as the AI proves itself (and back if quality drops).

GraphRAG

Extracting entities and relationships into a **knowledge graph** and retrieving sub-networks; strong for multi-hop/global questions.

Greedy decoding

The simplest decoder strategy.

Paper 06
Grounding

Tying the model's answer to provided source text rather than its own memory — the basis of trustworthy, citable output.

Grouped Query Attention (GQA)

A variant of Multi-Head Attention where multiple query heads share the same key-value head.

Paper 18
GSM8K

A benchmark dataset of 8,500 grade-school math word problems, ranging from simple arithmetic to multi-step reasoning.

Paper 14
Guardrail

A check around the model.

Guardrails

Checks and limits (enforced in code) on what an agent can take in, put out, or do.

Guardrails / moderation

Input/output filtering to keep harmful or unsafe content in check.

H
Hallucination

The phenomenon where a language model generates plausible-sounding but factually incorrect or entirely made-up information.

Paper 12
Handoff

One agent passing the task (and needed context) to another.

Hard alignment

In older statistical machine translation, each target word was explicitly assigned to exactly one source word.

Paper 07
Hardware-Aware Algorithm

An algorithm designed with GPU memory hierarchy in mind.

Paper 21
Harm Prevention

A principle in many AI constitutions stating that the AI should avoid providing information, advice, or assistance that could lead to physical, financial, psychological, or emotional harm.

Paper 22
Hash (SHA)

A 40-character fingerprint computed from content.

HBM (High Bandwidth Memory)

High-capacity GPU memory (e.g., 80GB on an H100), with lower bandwidth than SRAM.

Paper 21
Head (attention head)

One of h = 8 parallel attention computations in multi-head attention, each operating in a lower-dimensional subspace (dₖ = d_model / h).

Paper 08
Hebbian Learning

The biological learning rule proposed by psychologist Donald Hebb in 1949: "neurons that fire together, wire together." When two neurons are active simultaneously, the connection between them strengthens.

Paper 02
Helpful, Harmless, Honest (HHH)

The alignment criteria used to train InstructGPT: helpful (answers user queries well), harmless (doesn't enable or encourage harmful acts), honest (doesn't hallucinate or mislead).

Paper 15
Helpfulness

A principle stating that the AI should genuinely assist the human in achieving their goals.

Paper 22
Hidden Layer

A layer of neurons between the input layer and the output layer.

Paper 03
Hidden size / embedding dimension

How wide each token's internal representation is; bigger = more capacity, memory, compute.

Hidden state (h)

The LSTM's "spoken" output at each step.

2 papers
Hierarchical planning

A high-level plan whose steps each break into their own sub-plans.

HNSW

A popular, fast, accurate ANN indexing method (higher memory use).

Honesty

A principle stating that the AI should be truthful and not deliberately mislead the human.

Paper 22
Hook

A command the harness runs automatically at a defined moment — before or after a tool call, when a session starts, and so on.

Host

The AI application the user interacts with (e.g.

Human Feedback (HF)

Labels provided by humans comparing two AI outputs and indicating which one is better.

Paper 22
Human Preference / Human Feedback

Judgments by human raters about which model outputs are better.

Paper 15
Human Rater Agreement / Inter-Rater Reliability

Measure of how often different human raters agree on which output is better.

Paper 15
Human-in-the-loop (HITL)

Requiring human approval before the agent performs risky/irreversible actions.

Human-on-the-loop (HOTL)

The AI acts autonomously but a human monitors and can intervene.

Human-out-of-the-loop

Fully autonomous AI; humans only review aggregates and samples afterward.

Hybrid search

Combining **dense** (vector) and **sparse** (BM25 keyword) search, fused (e.g.

HyDE (Hypothetical Document Embeddings)

Generate a hypothetical answer, embed *it*, and search with that; the fake answer is often closer to real passages than the question.

Hyperparameter

A setting you choose before/around training (learning rate, epochs, rank, etc.), as opposed to the learned weights.

I
Identity propagation

Carrying the original user's identity through each hop — host → server → downstream API — so permissions are enforced as that real user, not a shared service account.

In-context learning

Performing a task by providing examples in the prompt, without updating model weights.

4 papers
In-Context Recall

The ability to retrieve specific facts from long context.

Paper 21
Indexing

The offline phase: parse → clean → chunk → embed → store, producing a searchable knowledge base.

Indirect injection

Malicious instructions hidden in *content the model processes* (a web page, document, email).

Inference

The process of running a trained model on new inputs to generate predictions.

2 papers
Inference Latency

The time to generate a single token during inference.

Paper 18
Inference vs Training

Inference: generating tokens one-by-one (autoregressive).

Paper 19
Inference-Time Scaling

The broader principle of improving model performance by allocating more compute at inference time, rather than only at training time.

Paper 23
InfiniBand

A high-speed network fabric used in data centres (200+ GB/s).

Paper 19
InfiniBand / RoCE / RDMA

Fast inter-node networking technologies that make multi-node serving practical.

Initialize

The first lifecycle message that opens an MCP session and triggers capability negotiation.

Input / prompt tokens

The text you send in.

Input gate (iₜ)

A sigmoid-valued vector that decides how much of the candidate vector

Paper 04
Input Projection (B)

A matrix or function that projects the input u_t into the state space.

Paper 21
Input tokens / output tokens

Tokens you send vs.

Input transformation

GPT-1's technique for reformatting any NLP task's input as a flat token sequence wrapped in special tokens, allowing the unmodified pre-trained model to handle diverse task formats.

Paper 10
Insecure code generation

The risk that AI-suggested code contains vulnerabilities (injection, weak crypto, hard-coded secrets).

Insecure output handling

Trusting model output blindly and passing it into a database, shell, or page, enabling classic attacks.

Instruct model / Chat model

A base model that has gone through post-training (SFT + preference tuning) so it follows instructions and chats.

Instruction Following / Alignment

Teaching language models to follow user instructions accurately and safely.

Paper 14
Instruction format

A data shape of instruction / optional input / output (e.g., Alpaca-style).

Instruction layers / priority

Personal → path-scoped → repo-wide → AGENTS.md → org; all merge, higher wins on conflict.

Instruction-Following

The ability of a language model to accurately follow user instructions and respond helpfully.

Paper 15
Instrumentation

Adding the code that creates spans/logs/metrics.

INT8 / INT4

8- and 4-bit integer precisions used in quantization.

Interconnect

Any GPU-to-GPU or node-to-node link; its speed dictates which parallelism you can use where (the governing rule: chatty parallelism → fastest link).

IPO / KTO / SimPO

Variants of DPO with different objectives or data needs (KTO uses single good/bad labels, not pairs).

Isolation

Keeping one tenant's (or session's) data and execution separated from another's, so they can't see or affect each other.

Issue

A GitHub thread tracking a task, bug, or feature.

Iteration / Round

One complete cycle of: MCTS search → solution verification → data collection → model training.

Paper 24
IVF (Inverted File)

Clustering-based ANN; partition space into cells, search only cells near the query.

K
k (top-k experts)

The number of experts selected per token.

Paper 09
Kernel

A small program that runs one operation on the GPU.

Kernel fusion

Merging several small operations into one kernel to reduce memory traffic.

Key (K)

The "advertisement" projection of each position.

Paper 08
KL Divergence Penalty

A regularization term in the RL objective that constrains the policy to stay close to the SFT baseline: β · KL[π_RL || π_SFT].

Paper 15
kNN (k-Nearest Neighbors)

The task: find the k most similar stored vectors to a query.

Knowledge base / index

Your prepared, searchable collection of chunks.

Knowledge cutoff

The date after which a model knows nothing, because its weights were frozen at training time.

KServe / Ray Serve

LLM/ML-aware serving layers built on Kubernetes.

Kubernetes (K8s)

The standard orchestrator for running, scaling, and self-healing many containers across machines.

KV block / page / block table

The fixed-size unit of KV memory, and the per-request map from logical token positions to physical blocks.

KV Cache

The memory buffer storing Key and Value vectors from all previous tokens during autoregressive (token-by-token) generation.

Paper 18
KV Chunk (Key-Value Chunk)

A segment of the key and value matrices corresponding to a subset of the sequence.

Paper 19
KV offloading / swapping

Moving cold KV cache to CPU/disk when GPU memory is tight; trades speed for capacity.

KV-cache quantization

Storing the KV cache in 8-bit (or lower) to roughly halve its memory, enabling more concurrency / longer context.

L
Label masking / Completion-only loss

Computing loss only on the assistant's tokens, so the model learns to respond, not to parrot the prompt.

Language model

A probability distribution over sequences of tokens.

Paper 10
Large Language Model (LLM)

A neural network trained to predict the next token in a sequence, using next-token prediction as the training objective.

Paper 14
Latency

The time taken to produce a response.

Paper 23
Latency Hiding

Making communication latency disappear by overlapping it with computation.

Paper 19
Latency-throughput curve

A plot of throughput vs p99 latency across concurrency levels; the most important serving chart.

Latency–throughput trade-off

Bigger batches raise throughput but can raise per-user latency.

Layer / Block

One repeating unit of a Transformer (attention + MLP).

Layer Normalisation (Layer Norm)

Applied after each sub-layer.

Paper 08
Learning Rate

A small positive number (e.g.

3 papers
Learning-rate schedule

How the learning rate changes over training (e.g., cosine decay).

Least privilege

Giving the agent only the tools/permissions it actually needs.

Least privilege / minimal agency

Giving the model and its tools the minimum power needed, to limit the blast radius of any failure.

Lifecycle

The defined stages of an MCP connection: initialize → operate → shut down.

Linear Recurrence

A recurrence relation of the form x_t = Āx_{t-1} + B̄u_t where future x_t depends only on previous x_{t-1}, not on all past history.

Paper 21
Linear Regression

A statistical method to fit a straight line through data points.

Paper 13
Linear Separability

A property of a dataset: two classes are linearly separable if you can draw a straight line (in 2D) or a flat hyperplane (in higher dimensions) that perfectly separates all examples of one class from all examples of the other.

Paper 02
Liveness vs readiness probe

Health checks: "is it alive?" (restart if not) vs "is it ready to serve?" (don't route traffic until the model is loaded — crucial for slow-loading LLMs).

llama.cpp

A lightweight engine for running quantized models on CPU/GPU/Mac.

LLM (Large Language Model)

An AI model (GPT, Claude, Gemini, Llama…) that predicts the next chunk of text.

LLM-as-a-judge

Using a strong model to score or compare outputs, for filtering data or evaluating models.

LLM-as-judge

Using another LLM to score an output against criteria (also used in evaluation).

LM head / Unembedding

The final layer that converts the model's internal representation into a probability for every token in the vocabulary.

Load balancing

The goal of ensuring that all n experts receive roughly equal numbers of tokens over training.

2 papers
Load generator (locust / k6)

Tools to drive realistic concurrent/rate-based traffic.

Load shedding

Politely rejecting or queuing low-priority traffic under extreme load to protect the whole system.

Log

A timestamped record of a single event ("what exactly happened in this one case").

Log-Log Plot

A graph where both axes are logarithmic.

Paper 13
Logit

A raw, unnormalised score before softmax.

Paper 08
Logits

Raw, unnormalised scores output by a neural network before applying softmax.

Paper 07
Long short-term memory

The full name of the paper.

Paper 04
Long-term memory

External storage (database/files) that persists across sessions and can be huge.

Loose object

A single compressed object file in `.git/objects/ab/cdef...`.

LoRA

An efficient fine-tuning method that adjusts a small add-on set of weights instead of all of them.

LoRA / multi-LoRA serving

LoRA = small add-on weight patches that customize a base model.

lora_alpha

The LoRA scaling factor; effective scaling is `alpha / r`.

lora_dropout

Dropout applied within the LoRA path to reduce overfitting.

Loss

A single number measuring how wrong the model's predictions are.

Loss Function

A mathematical function that measures how wrong the network's prediction is.

Paper 03
Lost in the middle

Models attend most to the start and end of context and may overlook material in the middle; place key chunks at the edges.

LSH (Locality-Sensitive Hashing)

Hashing that sends similar vectors to the same bucket.

LSTM (Long Short-Term Memory)

The RNN variant used by both the encoder and decoder in this paper.

Paper 06
M
Masked Language Modelling (MLM)

BERT's primary pre-training objective.

Paper 11
MATH Benchmark

A dataset of 12,500 competition-level math problems from AMC (American Mathematics Competitions) and AIME (American Invitational Mathematics Examination).

2 papers
Matrix multiplication

The core math operation in LLMs; millions of independent multiply-adds, ideal for parallel GPU hardware.

Matryoshka embeddings (MRL)

Embeddings whose first N dimensions are themselves usable; truncate for speed, expand for accuracy.

Maturity model

A scale (Level 0 "Vibes" to Level 4 "Optimized") for assessing how production-ready an AI system is across all pillars.

Max sequence length

The token cap per training example; longer ones get truncated.

Max tokens

A cap on how long the model's output can be.

max-model-len

Flag bounding max context length, which bounds KV-cache size per request.

max-num-batched-tokens

Flag for max tokens processed per step; balances prefill vs decode and tail latency.

max-num-seqs

Flag for max concurrently batched requests; trades throughput against latency and memory.

MCP (Model Context Protocol)

A standard for packaging tools/data sources so any compatible agent can plug into them — like "USB for AI tools."

MCP Gateway

A central proxy that sits between hosts and many MCP servers to enforce auth, routing, rate limits, logging, and policy in one place.

MCP server

A program that speaks MCP and offers tools, resources, or prompts (e.g.

MCP server / tool

A connector giving Copilot capabilities to act on external systems; complements skills.

Memory

Facts Claude persists across sessions in files, so it remembers your preferences, project details, and past decisions without you re-explaining them.

Memory bandwidth

How fast data moves between GPU memory and cores (GB/s or TB/s); largely decides *decode speed*.

Memory capacity

How much data (GB) the GPU can hold; decides *whether a model fits*.

Memory hierarchy

Tiers from tiny-fast (registers, SRAM/L1) to big-slow (VRAM), then off-GPU (system RAM, disk).

Memory Scaling

With P GPUs using Ring Attention, per-GPU memory is O((n/P) × d), scaling linearly with the number of GPUs.

Paper 19
Memory-bound

Limited by data-movement speed (cores idle, waiting for data).

Merge (merge_and_unload)

Folding a LoRA adapter into the base weights so there's zero inference overhead.

Metadata

Tags stored with each chunk (source, date, section, page, permissions) enabling filtering, citations, and access control.

Metric

A number measured and aggregated over time ("how is the system doing overall").

MHA / MQA / GQA

Multi-Head / Multi-Query / Grouped-Query Attention: design choices trading KV-cache size against quality.

Micro-batch

A small chunk of work kept flowing through a pipeline to fill bubbles.

Mixture of Experts (MoE)

A layer type where multiple expert networks are available, and a router learns which expert(s) to use for each input.

Paper 18
MLOps / LLMOps

The practices and tooling for deploying, monitoring, and continuously improving ML/LLM systems in production.

MMLU (Massive Multitask Language Understanding)

A benchmark of 57 diverse academic subjects (history, law, science, medicine) with 14,042 multiple-choice questions.

Paper 20
Mode collapse (in generation)

When generated data lacks diversity, repeating similar outputs.

Model

The underlying Claude that powers a session.

Model collapse

Degradation that can occur when models are trained repeatedly on model-generated data.

Model merging (soups / TIES / DARE)

Combining multiple fully fine-tuned models' weights into one without retraining.

Model provider

Who supplies the LLM (OpenAI, Anthropic, Google, Hugging Face, local via Ollama).

Model routing / cascading

Sending each request to the cheapest model that can handle it; escalating to a bigger model only when needed.

Model Scaling / Emergent Threshold

The observation that CoT prompting's effectiveness depends critically on model size.

Paper 14
Model Size (N)

The number of parameters in a neural network.

Paper 13
Model-invoked / user-invoked

A skill chosen automatically by description match vs.

MoE (Mixture of Experts) layer

A drop-in replacement for the FFN sub-layer in a Transformer.

Paper 09
Monitoring

Tracking whether known things are wrong (vs.

Monte Carlo Tree Search (MCTS)

A search algorithm that explores a decision tree by: (1) selecting promising nodes using UCB, (2) expanding the tree with new candidate moves, (3) running rollouts to simulate outcomes, (4) backing up the results to update node statistics.

Paper 24
MRR / nDCG / precision@k

Retrieval-quality metrics: MRR (rank of first hit), nDCG (graded ranking quality), precision@k (fraction of shown results that are relevant).

MTEB / BEIR / ANN-Benchmarks

Leaderboards for embedding quality, retrieval quality, and index speed/recall.

Multi-agent system

Multiple specialized agents collaborating (or competing) toward a goal.

Multi-head attention (MHA)

`Concat(head₁, ..., headₕ) · W^O`.

2 papers
Multi-query

Generate several paraphrases of a question, retrieve for each, and combine.

Multi-Query Attention (MQA)

An attention variant where all query heads share a single key-value head.

Paper 18
Multi-step reasoning

Chaining several thinking steps to solve harder problems.

Multi-tenancy

Serving many independent customers (tenants) from one shared system, while keeping each tenant's data and access strictly separate.

Multi-tenant isolation

Ensuring shared infrastructure doesn't leak one tenant's data to another (watch shared prefix caches).

Multimodal

Capable of processing and reasoning over multiple modalities (text, images, audio, video) simultaneously.

Paper 20
Multimodal RAG

Retrieving over images, charts, tables, and scanned pages, not just text.

N
n (number of experts)

Total number of expert networks in one MoE layer.

Paper 09
Native Multimodality

Training a single model jointly on multiple modalities from the start, rather than training text-first and bolting on vision later.

Paper 20
Natural Language

Human language as it is actually spoken and written — English, Hindi, Tamil, etc.

Paper 01
Negative sampling

The training trick that made Word2Vec fast.

Paper 05
Neural Machine Translation (NMT)

The umbrella term for translation systems built entirely from neural

Paper 06
Next Sentence Prediction (NSP)

BERT's second pre-training objective.

Paper 11
Next-token prediction

The pre-training objective: given all previous tokens, predict the probability distribution over the next token.

Paper 10
NF4 (NormalFloat4)

The 4-bit format used by QLoRA, designed to match how weights are distributed.

Node

One physical server (often holding multiple GPUs).

Noisy top-k gating

The specific gating formulation from the 2017 paper: raw logits have Gaussian noise added before top-k selection during training.

Paper 09
Non-parametric knowledge

Information given to the model at question time via the prompt; fresh, exact, citable.

Normalization (L2)

Scaling a vector to length 1.

Notification

A JSON-RPC message that expects no reply — used for one-way signals like progress updates.

Nucleus Sampling (Top-P Sampling)

A generation strategy where you only consider the top tokens that make up a certain cumulative probability (e.g., top_p=0.9 means consider tokens until their probabilities sum to 90%).

Paper 12
Numerical Stability

Ensuring computed values don't overflow, underflow, or lose precision.

Paper 19
NVIDIA Container Toolkit

Lets containers access the host's GPUs.

NVLink

NVIDIA's high-speed GPU interconnect (576 GB/s per link).

Paper 19
NVLink / NVSwitch

NVIDIA's very fast GPU-to-GPU links *inside a server*; what makes tensor parallelism viable.

O
OAuth 2.1

The modern authorization framework MCP builds on for granting scoped, delegated access without sharing passwords.

Observability

Being able to see everything an agent did (thoughts, tool calls, results).

OCR (Optical Character Recognition)

Software that extracts text out of an image (e.g.

Off-Policy vs. On-Policy RL

Off-policy: Learning from data generated by other policies (e.g., supervised data).

Paper 15
Offline evaluation

Running your system against a fixed test set before shipping a change.

Offline vs online vLLM

The Python library (`LLM.generate`) for batch jobs vs the API server (`vllm serve`) for live traffic.

Ollama

A simple tool for running GGUF models locally; great for prototyping.

One-hot vector

A vector of length V with a single 1 and the rest zeros.

Paper 05
One-Shot Learning

Performing a task with exactly one example in the prompt.

Paper 12
Online evaluation

Measuring quality on live production traffic via feedback and sampled LLM-judge scoring.

Online Softmax

An incremental softmax computation (using logsumexp trick) that maintains running statistics (max, sum of exponentials) as you process blocks.

Paper 19
Online vs offline serving

Online = live users, latency-sensitive.

Open-weight model

A model whose weights you can download and run yourself, rather than calling a provider's API.

OpenAI-compatible API

A de-facto standard API shape most engines (including vLLM) support, so clients can switch backends easily.

OpenTelemetry (OTel)

An open, vendor-neutral standard for creating traces, spans, and metrics, so you aren't locked to one tool.

OPQ

PQ with a learned rotation first, for better recall at the same memory.

Optimal Allocation

For a given compute budget C, the best way to split resources between model size (N) and data size (D) to minimize loss.

Paper 13
Optimizer (Adam / AdamW / 8-bit Adam)

The algorithm that applies gradients to weights, with adaptive step sizes and momentum.

Orchestration

The code that assembles the prompt and manages the flow of a request (retrieval, history, model call, output handling).

Orchestration framework

Builds and coordinates the agent loop / multiple agents (LangGraph, CrewAI, AutoGen…).

Orchestrator / Manager

The "boss" agent that breaks a goal into tasks and assigns them to workers.

Orchestrator–Worker pattern

A manager delegates sub-tasks to workers and combines results.

ORPO

A method combining SFT and preference tuning into one step, no reference model needed.

Outcome Reward Model (ORM)

A model that scores only the final output (right or wrong), without evaluating intermediate steps.

2 papers
Output gate (oₜ)

A sigmoid-valued vector that decides which slots of the cell state

Paper 04
Output parser

Code that reads the model's response and extracts the structured data you need, often paired with a schema.

Output Projection (C)

A matrix or function that projects the hidden state x_t back to the output space.

Paper 21
Overfitting

Training error decreases, but test error increases.

Paper 13
Overreliance

Users trusting AI output (including hallucinations) too much and acting on it.

OWASP LLM Top 10

A standard list of the top security risks for LLM applications; a useful threat checklist.

P
p50 / p95 / p99 latency

Median and tail response times; optimize and alert on the tail (p99).

Packfile

Many objects bundled and compressed together by `git gc` into `.git/objects/pack/`.

Packing

Concatenating multiple short examples into one sequence to reduce wasted padding and speed training.

Padding

Filler tokens added so examples in a batch are the same length; ignored in the loss.

Paged optimizer

An optimizer that offloads state to CPU RAM during memory spikes to avoid crashes (from QLoRA).

PagedAttention

vLLM's memory-management technique enabling efficient batched serving.

Parallel Scan

A hardware-friendly algorithm (e.g., Blelloch scan) that computes a recurrence y_t = f(x_{t-1}, u_t) in parallel by decomposing it into a tree of subproblems.

Paper 21
Parameter / weight

One of the billions of tunable numbers inside a model that together encode what it "knows." More parameters = more memory and compute needed.

Parameters

The learnable weights in a neural network model.

Paper 18
Parameters / Weights

The numbers that make up a model.

Parametric knowledge

Knowledge stored in the model's weights from training; fast and broad but frozen, blurry, and uncitable.

Parent

The commit that came before a given commit.

Parent-document / small-to-big retrieval

Match on small chunks but return the larger surrounding passage for generation.

Parsing / extraction

Pulling clean text out of source files (PDF, Word, HTML, slides, scans).

Pass@K

A metric that evaluates whether at least one out of K generated solutions is correct.

Paper 23
Passkey

A phishing-resistant, passwordless sign-in credential.

Patch (Image Patch)

A small rectangular region of an image, typically 14×14 pixels.

Paper 20
PCIe

A slower general-purpose bus; TP over PCIe-only is often too slow.

Peephole connections

An extension to LSTMs (Gers & Schmidhuber, 2000) in which the gates

Paper 04
PEFT (Parameter-Efficient Fine-Tuning)

The family of methods that train only a tiny number of new parameters while freezing the base model.

Percentile (p50/p95/p99)

A way to summarize a distribution.

Percentiles (p50 / p95 / p99)

The value below which 50% / 95% / 99% of requests fall.

Perceptron

The artificial neuron Rosenblatt described: takes several inputs, multiplies each by a weight, sums the results, and outputs 1 if the sum exceeds a threshold, 0 otherwise.

Paper 02
Permission mode

How much Claude is allowed to do without asking — from confirming every action to running freely within an allowlist.

Perplexity

A metric for language models derived from cross-entropy loss.

Paper 12
Phrase table

A core data structure in pre-2014 statistical translation.

Paper 06
PII (Personally Identifiable Information)

Personal data (names, emails, etc.) that must be protected, redacted, and handled per privacy law.

Pipeline / Sequential pattern

Agents work in a fixed order, each passing output to the next (assembly line).

Pipeline bubble

Idle gaps in pipeline parallelism; reduced by keeping many micro-batches in flight.

Pipeline parallelism (PP)

Splitting the model by layers across GPUs/nodes, assembly-line style; less communication, so it can cross nodes.

PiSSA

A LoRA variant with smarter initialization from the base weights' principal components.

Plan mode

A mode where Claude researches and proposes a plan for approval before making any changes, so you can steer before code is written.

Plan-first vs. plan-as-you-go

Writing the whole plan up front vs.

Planning

Decomposing a goal into an ordered set of doable sub-tasks.

Plugin

A packaged add-on that extends Claude with extra skills, commands, or integrations.

Pod / Deployment / ReplicaSet

K8s units: the running container(s), and the controllers that keep N replicas alive (operationalizing data parallelism).

Policy Gradient / Policy Optimization

RL algorithms that improve a policy (probability distribution) by taking gradient steps that increase expected reward.

Paper 15
Policy Model

The language model being trained and improved across rounds.

Paper 24
Polysemy

The property of a word having multiple meanings.

Paper 05
Position bias / Length bias

Tendencies of LLM judges to favor a certain answer position or longer answers; must be controlled for.

Positional encoding (PE)

A fixed vector added to each input embedding to inject position information.

2 papers
Post-training

The fine-tuning steps (SFT, preference tuning) applied after pretraining to turn a base model into a helpful assistant.

Power Law

A mathematical relationship where one variable is proportional to another raised to a power: y = a * x^b.

Paper 13
PPO (Proximal Policy Optimization)

A stable reinforcement learning algorithm used in the RL stage.

Paper 15
PQ (Product Quantization)

Compress vectors by splitting into chunks and replacing each with a codebook ID.

Pre-quantized model

An already-quantized model you download and serve directly.

Pre-training

Training a model on large-scale, typically unlabelled data before fine-tuning.

3 papers
Precision

How many bits are used to store each weight; more bits = more exact but more memory.

Precision / bits

How many bits store each number.

Precision@K

Of the K retrieved chunks, how many were actually relevant.

Preference format

Data shaped as prompt / chosen / rejected, used for preference tuning.

Preference tuning

Teaching a model judgment/taste by training on "this response is better than that one." (M12)

Prefill

The first phase: the model reads the whole prompt at once (in parallel), builds the KV cache, and produces the first token.

Prefix caching (Automatic Prefix Caching, APC)

Reusing the cached KV of shared prefixes (system prompts, RAG context, chat history) to skip redundant prefill.

Prefix tuning / Prompt tuning / P-Tuning

Additive PEFT methods that prepend trainable "virtual tokens" (soft prompts) to the input.

Pretrained vectors

Word vectors trained on a large corpus by someone else and then

Paper 05
Pretraining

The initial, expensive phase where a model learns language by predicting the next token across enormous amounts of text.

Primitive

A core building block MCP defines.

Probabilistic / non-deterministic / stochastic

Same input can give different outputs (AI).

Process Reward Model (PRM)

A machine-learning model trained to evaluate the quality of individual steps in a multi-step reasoning process.

2 papers
Production

The live environment real users use, as opposed to a demo or test setup.

Program-of-Thought (PoT)

Solving problems by writing Python code instead of natural language reasoning.

Paper 24
Progress / Cancellation

Protocol features that let a long-running operation report how far along it is, and let the caller stop it.

Progressive disclosure

Staged loading of skills: L1 name+description (always) → L2 full body (on match) → L3 resources (on demand).

Progressive disclosure vs. memory vs. instructions

On-demand expertise vs.

Prometheus / Grafana

Tools to collect (scrape) and visualize metrics.

Prompt

The text instructions given to a model.

Prompt caching

Reusing/caching a stable part of the prompt to save cost and time.

Prompt chaining

Breaking a task into a sequence of prompts where each step's output feeds the next, instead of asking for everything at once.

Prompt Engineering

The practice of carefully designing the text prompt to get better outputs from a language model.

2 papers
Prompt Format

The specific structure and wording of a prompt.

Paper 12
Prompt injection

Malicious instructions hidden in content the agent reads (web page, email, doc) that trick it into harmful actions.

Prompt template

A reusable prompt with placeholders you fill in at runtime, so the same well-tested wording serves many inputs.

Provider

The company running the model you call over the internet (e.g., OpenAI, Anthropic, Google).

Pull request (PR)

A GitHub proposal to review and merge a branch's commits.

Push protection

Blocks a push that contains a recognized secret *before* it enters history.

Python Verification

The process of executing Python code to check if a solution is correct.

Paper 24
PyTorch

The underlying deep-learning framework that runs the math on the GPU.

R
RAG (Retrieval-Augmented Generation)

Retrieve relevant info → Augment the prompt with it → Generate the answer.

RAGAS

A popular framework of RAG evaluation metrics.

Rank (`r`)

The size of a LoRA patch's bottleneck; controls its capacity.

RAPTOR

Building a tree of summaries over a corpus and retrieving at the right level of detail; good for global/summary questions.

Rate limiting

Capping request volume — the provider's limits on you, and your limits on each user (protects availability, cost, and against abuse).

RBAC (Role-Based Access Control)

Granting access by assigning users to roles (admin, editor, viewer) that bundle permissions.

Re-planning

Adjusting the plan when a step fails or returns surprising results.

ReAct (Reason + Act)

The core pattern of alternating **Thought → Action → Observation**, combining reasoning with tool use.

Reasoning / Multi-Step Reasoning

The cognitive process of chaining ideas together across multiple steps to arrive at a conclusion.

Paper 14
ReBAC (Relationship-Based Access Control)

Granting access based on relationships between entities ("owner of," "member of") — the model behind systems like Google Zanzibar.

Recall

The fraction of truly-relevant items that were actually found.

Recall@K / hit rate

Was a relevant chunk in the top K retrieved; the most important retrieval metric.

Receptive Field

In deep networks, the range of input positions that influence a given output position.

2 papers
Recurrent Mode (Inference)

During token-by-token generation, apply the recurrence directly: x_t = Āx_{t-1} + B̄u_t.

Paper 21
Recurrent Neural Network (RNN)

A neural network that processes inputs one step at a time, feeding its

2 papers
Red-teaming

Deliberately attacking your own system to find weaknesses before real attackers do.

Reflection / Self-correction

An agent reviewing its own output, finding flaws, and revising — without a human pointing them out.

Reflexion

Writing a "lesson learned" after a failure and storing it to do better next time (reflection + memory).

Registry

A catalog where MCP servers are published and discovered, so hosts can find and install them.

Regression

Something that used to work but broke after a change.

Regression testing

Checking the fine-tuned model didn't get worse at general tasks it should still handle.

Reinforcement Learning from Human Feedback (RLHF)

A three-stage training pipeline for aligning language models: (1) Supervised Fine-Tuning on human demonstrations, (2) training a Reward Model on human preference comparisons, (3) using Reinforcement Learning (PPO) to optimize the policy aga

Paper 15
Rejection Sampling

A data generation strategy: generate many candidate solutions, keep only the correct ones, discard the rest.

Paper 24
Reliability

Whether the system keeps working under real load and failures.

Remote

Another full copy of a repo, usually on GitHub; the default is `origin`.

Replica

One complete serving instance (a single GPU or a TP/PP group); the unit you data-parallelize and autoscale.

Replication

Copies of data across machines (fault tolerance + read throughput).

Repository (repo)

A project tracked by Git: your files plus a hidden `.git` folder holding the entire history.

Representation learning

The broader idea — Word2Vec is an early instance — that useful

Paper 05
Request / Response

The paired JSON-RPC messages: a request asks for something and carries an id; the matching response returns the result or an error.

Reranking / reranker

A second, more accurate (cross-encoder) scoring pass over retrieval candidates; the highest-value upgrade beyond basic retrieval.

Residual connection

Adding the sub-layer's input directly to its output: `x + SubLayer(x)`.

Paper 08
Resource

A server primitive exposing readable data (files, records, documents) the model can pull into context — read-only, addressed by URI.

Retrieval

Finding the most relevant chunks for a query.

Retry (with backoff and jitter)

Re-attempting a failed call, waiting progressively longer (backoff) with randomness (jitter), only for retryable errors, with a cap.

Retry storm

Cascading overload caused by naive client retries; mitigated by backoff and circuit breakers.

Reverse-input trick

Sutskever's empirical hack: feed the source sentence to the encoder in

Paper 06
Review gate

A point where the flow pauses for human judgment, placed by stakes and uncertainty.

Revision Prompt

A prompt asking an AI to rewrite its own output to address a critique.

Paper 22
Reward Function

In MCTS, the function that assigns a reward to a rollout outcome.

Paper 24
Reward hacking

When a model games the reward signal to score high without genuinely being better.

Reward Hacking / Gaming the Reward Model

When the RL policy finds ways to get high reward scores without actually being helpful.

Paper 15
Reward Model (RM)

A neural network trained in the second stage of RLHF to predict which of two responses humans prefer.

2 papers
Ring Attention

A distributed attention algorithm where P GPUs are arranged in a ring topology.

Paper 19
Ring Topology

An arrangement of P GPUs in a logical circle where GPU i communicates with GPU i-1 (receives data) and GPU i+1 (sends data).

Paper 19
RL-CAI (Reinforcement Learning Constitutional AI)

The second stage of Constitutional AI.

Paper 22
RLAIF (Reinforcement Learning from AI Feedback)

The stage of Constitutional AI where an AI (rather than a human) provides feedback on which response better follows the constitution.

Paper 22
RLHF (Reinforcement Learning from Human Feedback)

The original preference-tuning approach: train a reward model, then optimize the LLM against it with RL.

RoBERTa

Robustly Optimized BERT Pretraining Approach.

Paper 11
Role prompting

Telling the model who to act as ("You are a careful editor…") to shape tone, depth, and focus.

Rollback

Quickly reverting to the previous version when a change goes wrong.

Rolling / blue-green / canary deployment

Update strategies: replace replicas gradually / run new alongside old then switch / send a small traffic slice to the new version first.

Rollout

In MCTS, a simulation of completing a partial solution to a full solution.

Paper 24
Root

A client primitive that tells a server which directories or scopes it is allowed to operate within.

RoPE scaling

A technique to extend a model's usable context length by adjusting its positional encoding.

Rotary Position Embeddings (RoPE)

A method of encoding token position information by rotating query and key vectors.

Paper 18
Routing

Classifying a query and sending it to the right index/tool (docs vs SQL vs web).

RRF (Reciprocal Rank Fusion)

A simple, robust method to merge two ranked result lists (e.g.

rsLoRA (rank-stabilized LoRA)

A LoRA variant that rescales so high ranks train more stably.

S
Safety factor / headroom

Extra capacity (e.g.

Sampling

A client primitive that lets a server ask the host's model to generate text — so servers can use the LLM without holding their own API key.

Sampling Temperature

A hyperparameter in language model decoding that controls randomness.

Paper 23
Sandbox

An isolated, safe environment for running untrusted code so it can't harm the real system.

Scaled dot-product attention

`Attention(Q, K, V) = softmax(Q·Kᵀ / √dₖ) · V`.

2 papers
Scaling

Increasing the size of neural networks (more parameters, more data, more compute).

Paper 13
Scaling Laws / Emergent Capabilities

The observation that larger models have qualitatively different capabilities (reasoning, instruction-following) that smaller models lack.

Paper 15
Scheduler / KV-cache manager / executor

vLLM's three conceptual roles: decides what runs (continuous batching), manages KV memory (PagedAttention), and runs the model on the GPU.

Schema

A precise specification of the fields and types an output must contain — handed to the model so its structured output is predictable.

Scope

In OAuth, the specific permissions an access token grants (e.g.

Secret

Any credential that grants access (password, API key, token, private key, `.env`); never commit it.

Secret rotation

Revoking a leaked credential and issuing a new one; always do this *before* scrubbing history, since a pushed secret may already be copied.

Secret scanning

GitHub feature that detects committed secrets and alerts you.

Segment embedding

One of three embeddings summed to form each token's input representation.

Paper 11
Selective SSM

An SSM where the input projection (B), output projection (C), and step size (Δ) are functions of the input u_t, not fixed constants.

Paper 21
Self-attention

Attention where Q, K, and V all come from the same sequence.

Paper 08
Self-Consistency

A technique that improves chain-of-thought reasoning by sampling multiple independent reasoning chains from the same prompt and taking a majority vote on the final answer.

2 papers
Self-Critique

The process of an AI model reading its own output and identifying whether it violates constitutional principles.

Paper 22
Self-Evolution

A bootstrapping process where: (1) a model generates candidate solutions using search, (2) solutions are verified automatically, (3) correct, high-quality solutions become training data, (4) the model is trained on this data, improving for

Paper 24
Self-Instruct

Generating new instructions (and answers) from a small seed set to bootstrap a dataset.

Self-Refine

The loop of generate → self-feedback → refine → repeat.

Self-supervised learning

Learning without human labels by using the data itself as the answer (e.g., predicting the next word).

Semantic caching

Caching answers to frequently-asked (by meaning) questions to skip the pipeline.

Semantic chunking

Grouping sentences by meaning so each chunk is one coherent idea.

Semantic search

Searching by meaning (via embeddings) rather than exact keywords.

SentencePiece

A subword tokenizer that converts text into tokens using a learned vocabulary.

Paper 20
Separation of storage and compute

Vectors in cheap object storage, stateless compute searches them; elastic and cost-efficient.

Seq2seq (sequence-to-sequence)

The encoder-decoder architecture from Paper 06 (Sutskever et al., 2014).

Paper 07
Sequence Parallelism

Parallelising the sequence dimension of tensors.

Paper 19
Sequential Revision

A strategy where you iteratively refine a solution, using feedback from one attempt to improve the next.

Paper 23
Server

A program that exposes tools, resources, and prompts over MCP for clients to use (e.g.

Serving / inference server

Running a model as a continuous service that handles many users' requests over an API.

Session

One continuous working conversation with Claude, with its own context.

Settings (settings.json)

The configuration file that controls the harness — permissions, environment variables, hooks, and model choice.

SFT (Supervised Fine-Tuning)

Fine-tuning on input→desired-output examples; the workhorse technique.

Shadow mode

Running a new version alongside production without showing its output to users, just logging/evaluating what it would have done.

Sharding

Splitting data across machines (capacity + write throughput; queries fan out).

Short-term memory

The context window — the current task and recent steps; temporary.

Sigmoid

The activation function σ(z) = 1/(1+e⁻ᶻ).

Paper 03
Sigmoid function (σ)

A function that squashes any real number into the interval (0, 1).

2 papers
Signed commit

A commit cryptographically signed (GPG/SSH) so GitHub can show a **Verified** badge proving authorship.

Similarity metric

How closeness is measured: **cosine** (angle/direction, default for text), **dot product** (direction + magnitude), **Euclidean/L2** (straight-line distance).

Single-replica capacity

The throughput one replica sustains within SLO at your length profile; the basis for sizing.

Skill

A reusable procedure you teach Claude once (as a folder with instructions) so it can perform a multi-step task the same way every time — invoked by name.

Skip-gram

The other Word2Vec training task, and the one people usually mean

Paper 05
SL-CAI (Supervised Learning Constitutional AI)

The first stage of Constitutional AI.

Paper 22
SLA (Service Level Agreement)

A contractual performance promise, usually looser than the SLO.

Slash command

A reusable shortcut typed as `/name` that runs a saved prompt or procedure — your own custom commands plus built-in ones.

Sliding Window Attention (SWA)

An attention variant where each token attends only to the last W tokens (a sliding window), not all previous tokens.

Paper 18
SLO (Service Level Objective)

A committed target like "99.9% success" or "p95 < 3s." Define one per feature.

Slope

On a log-log plot, the slope of a line is the exponent of the power law.

Paper 13
SmoothQuant / LLM.int8()

INT8 methods that handle activation outliers so 8-bit quantization stays accurate.

Snapshot

Git's mental model for what a commit stores: a complete photo of the project at a moment, not a list of edits.

Soft alignment

Attention's approach: each target word is generated using a *weighted blend* of multiple source words, not a hard assignment to one.

Paper 07
Soft prompt

Trainable input vectors that act like a learned prompt expressed as numbers.

Softmax

A function that turns a vector of raw scores into a probability

5 papers
Softplus Function

Smooth approximation of ReLU: softplus(x) = log(1 + e^x).

Paper 21
Span

A single timed unit of work within a trace (e.g., "retrieval took 120 ms"), with attributes.

Sparse computation

The opposite of dense: only a fraction of parameters are active for any given input.

Paper 09
Sparse search

Keyword search (e.g.

Specification Gaming

The problem where an AI system finds a way to satisfy the letter of a specification while violating its spirit.

Paper 22
Speculative decoding

Using a small **draft model** (or heuristic) to guess several tokens, then verifying them in one parallel pass of the big model.

SQ (Scalar Quantization)

Store numbers as int8 instead of float32.

SQuAD

Stanford Question Answering Dataset.

Paper 11
Squash / merge / rebase merge

GitHub's three ways to merge a PR (combine into one commit / merge commit / linear replay).

SRAM (Static RAM)

Tiny, ultra-fast on-GPU cache (e.g., 192KB per core).

Paper 21
SSH key / PAT

Credentials for authenticating to GitHub (key pair / personal access token).

Staging area / Index

The "loading dock" where you gather exactly what goes into the next commit; physically the `.git/index` file.

State Space Model (SSM)

A continuous or discrete linear dynamical system.

2 papers
State Transition Matrix (A)

An n×n matrix governing how the hidden state x evolves over time.

Paper 21
Static batching

Naive batching that waits for all requests in a batch to finish; wastes GPU slots.

Statistical Machine Translation (SMT)

The dominant pre-2014 translation approach.

Paper 06
Step Size (Δ)

A scalar (or per-head scalar) that controls the discretisation rate.

Paper 21
Stop sequence

A string that tells the model to stop generating when it appears — useful for clean, bounded output.

Straggler Problem

When one GPU is slower than others (older hardware, thermal throttling, interference), it becomes the bottleneck.

Paper 19
Streaming

Showing the answer as it's generated, word by word, to slash perceived latency.

Structured output

Output constrained to a defined shape (JSON, a schema) so a program can use it directly, rather than free-form prose.

Structured State Space (S4)

A prior SSM architecture (Gu et al., 2021) that imposes structure on the A matrix (e.g., diagonal, plus rank-1 update) for efficiency.

Paper 21
Subagent

A separate Claude instance the main agent spawns to handle a focused sub-task (research, review) with its own fresh context, reporting results back.

Submodule

Another Git repo embedded at a pinned commit inside yours.

Subsampling

A Word2Vec training trick where very common words (like "the", "of",

Paper 05
Success rate / Latency / Cost

Key metrics — % of tasks done correctly / how long it takes / tokens or dollars used.

Summarization (memory)

Replacing long history with a short summary to save context space.

Supervised Fine-Tuning (SFT)

The first stage of RLHF.

2 papers
Supply chain

The full set of dependencies and sources a server is built and distributed from — a security surface, since a compromised dependency compromises the server.

Supply-chain security

Protecting against risks from third-party code and actions (e.g.

Switch Transformer

Google's 2021 simplification of MoE: k=1 routing (route each token to exactly one expert, no blending).

Paper 09
Sycophancy / Sycophantic Behavior

When a model agrees with users even when the user is wrong, in order to be pleasing.

Paper 15
Synchronisation Barrier

A point where all P GPUs pause and wait for the slowest GPU to finish.

Paper 19
Synthetic data

Training examples generated by an AI model rather than collected from humans.

System prompt

The developer-written setup that defines the agent's role, rules, tools, and process.

System prompt / instructions

The standing orders in a prompt that set the AI's behavior, rules, and tone.

T
Tag object

An annotated tag pointing to a commit, with its own metadata.

Tail latency

The slow end (p95/p99) of the latency distribution; where real user frustration lives.

tanh (hyperbolic tangent)

A function that squashes any real number into the interval (−1, +1).

Paper 04
Target modules

Which weight matrices receive LoRA adapters (e.g., q/k/v/o projections, MLP layers, or "all-linear").

Task (async)

A long-running unit of work tracked over time, letting a server perform work asynchronously and report results when ready.

Teacher forcing

A training technique.

Paper 06
Technical Report

A publication style (unlike peer-reviewed research papers) that allows companies to present results without the formal review process.

Paper 20
Temperature (in Generation)

A hyperparameter controlling randomness in generation.

Paper 12
Tensor Cores

Special GPU units that do low-precision matrix math (FP16/INT8/FP8) very fast; a reason quantization speeds compute.

Tensor parallelism (TP)

Splitting each layer's math across GPUs; makes a big model fit and adds bandwidth, but needs constant fast communication — keep within one NVLink node.

Test set

Held-out, ideally real, examples used once at the end to score the model honestly.

Test set / dataset (for evals)

A curated collection of test inputs (and often expected outputs) you run your system against.

Test-Time Compute

Spending additional computation at inference time (rather than training time) to improve performance.

2 papers
Text-to-SQL

Having the model write a database query; the right tool for structured/aggregate questions (not RAG).

TGI (Text Generation Inference)

Hugging Face's production model-serving engine.

TGI / TensorRT-LLM / SGLang / llama.cpp / Ollama

Alternative engines: Hugging Face's TGI, NVIDIA's highly optimized TensorRT-LLM, SGLang (structured generation/prefix sharing), and the CPU/local llama.cpp + Ollama world.

The knee / operating point

The point of maximum throughput still within your latency SLO; your target operating point (≈ goodput maximum).

Thought vector

Another name for the **context vector** — Hinton's evocative label for

Paper 06
Threat model

A structured view of who might attack a system, how, and what's at stake — used to design defenses deliberately.

Threshold

The minimum weighted sum required for the Perceptron to output 1.

Paper 02
Throughput

The number of queries a system can handle per unit time.

Paper 23
Time to first token

How long until the *first* word of a streamed response appears.

Time-to-first-token / Tokens-per-second

The two components of latency: how fast output starts, and how fast it streams.

Timeout

A maximum wait time for a call; exceeding it triggers a fallback.

Token

A unit of text, roughly a word or subword.

2 papers
Token Budget

The total number of tokens (words or subwords) available for generating a solution.

Paper 23
Token dropping

When an expert receives more tokens than its capacity allows, excess tokens skip the MoE layer and pass through the residual connection unchanged.

Paper 09
Token optimization

Practices that reduce how many tokens a task consumes — trimming context, reusing caches, being concise — to save cost and stay within the window.

Token Position

The index of a token in the sequence (0 to n-1).

Paper 19
Tokenization

The process of converting input (text, images, audio) into discrete tokens.

Paper 20
Tokenizer

The component that converts text into tokens (and back).

Tombstone / soft delete

Marking a vector deleted without removing it from the graph; cleaned up later by compaction.

Tool / Function

Code the agent can call to act or fetch info it can't do alone (search, calculator, send email, run code).

Tool / function calling

Giving the model the ability to call functions (look up an order, do math, search) and use the results.

Tool schema

A structured description of a tool (name, description, inputs) that tells the model when and how to use it.

Top-k selection (TopK)

The operation that keeps the k largest values in a vector and sets all others to −∞.

Paper 09
Top-p (nucleus sampling)

An alternative to temperature: the model samples only from the most probable tokens whose combined probability reaches p.

TPOT (Time Per Output Token) / ITL (Inter-Token Latency)

Average gap between subsequent output tokens.

Trace

The full recorded log of one agent run (inputs, steps, tools, observations, output, cost).

Trace ID / correlation ID

The identifier that ties all the steps of one request together.

Tracking branch

A local branch linked to a remote branch for easy push/pull.

Train / Validation / Test split

Dividing data into what the model learns from, what you watch for overfitting, and what you score once at the end.

Training

The process of setting a model's weights by showing it data and nudging the weights to predict better.

Training Data

The text corpus used to train a language model.

Paper 18
Training-Time Compute

Computation used to train the model initially.

Paper 23
Trajectory

The path/process an agent took (not just the final answer).

Transfer learning

The reuse of knowledge (model weights) learned on one task/dataset for a different but related task.

Paper 10
Transformer

A neural network architecture based on self-attention, introduced in "Attention Is All You Need" (2017).

Paper 20
Transformer Decoder

The architecture used in GPT models: a stack of self-attention and feedforward layers that process tokens left-to-right (causally).

Paper 13
transformers (Hugging Face)

The standard library for loading models and tokenizers.

Transparency

A key benefit of Constitutional AI: the principles are written in human-readable natural language, making the intended values explicit and auditable.

Paper 22
Transport

How MCP messages physically travel between client and server — commonly stdio (local process) or streamable HTTP (remote).

Tree

A directory listing: names → blob/tree hashes plus permissions.

Tree of Thoughts (ToT)

Exploring several possible reasoning paths/plans and picking the best.

trl (Transformer Reinforcement Learning)

Library providing ready-made trainers (`SFTTrainer`, `DPOTrainer`).

Truncation

Cutting off text that exceeds the max sequence length; can silently break examples.

Trust boundary

The line where data or control passes between parties with different trust levels — every crossing needs validation and authorization.

TTFT (Time To First Token)

Time from sending a request to the first output token.

Turing Machine

An abstract mathematical machine Turing described in 1936 — not a real physical device, but a thought experiment.

Paper 01
Turing Test

The test proposed by Turing: a machine passes if a human interrogator, communicating only by typed text, cannot reliably distinguish it from a human.

Paper 01
Two-factor authentication (2FA)

A second sign-in factor (authenticator app, passkey) beyond a password; the top account protection.

V
V1 engine

vLLM's re-architected core for lower overhead and cleaner async behavior.

Value (V)

The "what I send when selected" projection of each position.

Paper 08
Value Function / Baseline

In RL, an estimate of expected future reward used to reduce gradient variance.

Paper 15
Values Specification

The process of encoding organizational or societal values into an AI system.

Paper 22
Vanishing Gradient

The problem where gradients shrink toward zero as they propagate backwards through many layers (especially through sigmoid activations).

Paper 03
Vanishing gradient problem

The phenomenon where, during BPTT on a long sequence, the gradient

Paper 04
Variance

In the context of scaling laws, the spread of loss values across multiple runs or models.

Paper 13
Vector database

A database that stores embeddings and quickly finds the most similar ones (e.g., Pinecone, Chroma, Weaviate, FAISS, Qdrant).

Vector library vs database vs extension

Library (FAISS): in-process index only.

Verifier

A component that evaluates whether a proposed solution is correct.

2 papers
Versioning

Tracking changes to prompts, models, retrieval config, guardrails, and code so behavior is reproducible and rollback is possible.

Vision Transformer (ViT)

A Transformer applied to images by dividing them into patches and treating patches as tokens.

Paper 20
vLLM

A high-performance GPU serving engine with continuous batching and LoRA support; the default for serious serving.

Vocabulary (V)

The set of words the model knows about.

2 papers
Voronoi cell / centroid

A cluster region and its center, used by IVF.

VRAM / GPU memory

The GPU's dedicated memory, separate from system RAM, measured in GB.

W
Warm pool

Pre-loaded spare capacity kept ready because LLM scale-up is slow.

Warm-up

Early requests that pay startup costs; discard them and measure steady state.

Warmup

Starting with a tiny learning rate and ramping up over the first steps, for stability.

Weight

A number attached to an input connection in a neural network.

Paper 02
Weight decay

A regularization technique that discourages large weights, helping prevent overfitting.

Weight Initialisation

The values given to weights before training begins.

Paper 03
Weight tying

Using the same weight matrix for both the token input embedding and the output projection (UW and UWᵀ).

Paper 10
Weight-only vs weight+activation quantization

Compressing just the stored weights (big memory/bandwidth win, easy on quality) vs also quantizing the live activations for compute speedups (harder, due to outliers).

Weights

The billions of numbers inside a neural network; where a model's learned knowledge and abilities are stored.

Win rate

The fraction of comparisons your model wins in A/B evaluation.

Window size (c)

How many words on each side of the target count as "context" in

Paper 05
Word analogy task

An evaluation task of the form "A is to B as C is to ___".

Paper 05
Word vector

Another name for a word embedding — a dense, low-dimensional vector

Paper 05
Word2Vec

Collective name for the two 2013 papers (Mikolov et al.) and the

Paper 05
WordPiece

BERT's subword tokenisation algorithm.

Paper 11
Worker / Specialist

An agent with one focused job (research, writing, coding).

Working directory

The real files you see and edit.

Worktree

An isolated copy of your git repository where an agent can work without touching your main checkout — handy for parallel or experimental changes.

WxAy notation

Weights x-bit, activations y-bit (e.g.

`
`.git` folder

The repository's actual database — objects, refs, HEAD, index, config, hooks, logs.

`.gitignore`

Lists files Git should not track; only affects untracked files.

`<EOS>` token

End-of-sentence token.

Paper 06
`<SOS>` token

Start-of-sentence token.

Paper 06
`add`

Stage changes for the next commit.

`AGENTS.md`

A cross-tool standard file giving AI agents project guidance.

`allowed-tools`

Frontmatter listing tools a skill may use without per-use confirmation.

`applyTo`

Frontmatter glob that limits a path-specific instruction to matching files.

`bisect`

Binary-search history to find the commit that introduced a bug.

`blame`

Show who last changed each line and in which commit.

`branch` / `switch` / `checkout`

List/create branches and move between them.

`cherry-pick`

Copy a single commit onto the current branch.

`clean`

Delete untracked files (not recoverable — use `-n` first).

`clone`

Download a full copy of a remote repository once.

`config`

The repository's local settings and remotes.

`copilot-instructions.md`

The repo-wide instruction file at `.github/copilot-instructions.md`, applied to everyone.

`description` (skill)

The frontmatter field Copilot matches against your request to decide whether to load the skill.

`git filter-repo`

The recommended tool to rewrite history and purge a file (e.g.

`hooks/`

Scripts Git runs automatically at events (pre-commit, pre-push, etc.).

`index`

The staging area, in binary form.

`info/exclude`

A personal, uncommitted ignore list.

`init`

Create a new repository (`.git` folder).

`logs/`

Records every position HEAD has held; powers the reflog.

`ORIG_HEAD`

A bookmark of where HEAD was before a big operation, for easy recovery.

`origin/main`

Your locally cached view of the remote's branch; updates only when you sync.

`origin`

The conventional name for the default remote.

`packed-refs`

Many refs compressed into one file for efficiency.

`push` / `fetch` / `pull`

Upload commits / download without merging / download and merge.

`rebase`

Replay your commits onto another branch for a linear history; never on shared commits.

`reflog`

Show the history of HEAD's movements; the recovery safety net.

`refs/`

Folder of pointer files: `heads/` (branches), `tags/`, `remotes/`.

`reset`

Move the branch pointer back: `--soft` (keep staged), `--mixed` (keep unstaged), `--hard` (discard).

`restore`

Unstage a file or discard working-directory changes.

`revert`

Add a new commit that undoes an earlier one; safe for shared history.

`SKILL.md`

A skill's entry file: YAML frontmatter (`name`, `description`, optional fields) plus an instructions body.

`stash`

Temporarily shelve uncommitted changes.

`status` / `diff` / `log`

See pending changes / line-level changes / history.

`tag`

Mark a commit with a permanent name (e.g.

`upstream`

Convention for the original repo when you've forked it.