AI/ML Glossary — 250+ Terms

Fraction of predictions that match the ground truth label, the simplest classification metric.

Small trainable module inserted into a frozen base model for parameter-efficient tuning.

The repeated think/act/observe cycle by which an LLM agent makes progress on a task.

System where an LLM iteratively plans, calls tools, and evaluates progress autonomously.

Human-centric benchmark drawn from SAT, LSAT, GRE and Chinese gaokao exams.

Field studying how to build AI systems that are reliable, controllable and beneficial.

Position-encoding method that adds a fixed linear distance penalty to attention scores.

Engineering models so their behaviour matches human values, intent, and safety policies.

Official Python and TypeScript client for Anthropic's Claude API.

Snowflake's open MoE LLM aimed at enterprise SQL, code and analytics workloads.

Mechanism letting a model weight different parts of its input when producing each output token.

Vaswani et al. 2017 paper that introduced the Transformer architecture.

Network trained to reconstruct its input through a low-dimensional bottleneck.

Generative model that produces output one token at a time, each conditioned on prior tokens.

Cohere For AI's open multilingual model family covering 100+ languages.

Processing many requests together — async, off-peak — at sharply reduced per-token cost.

Number of training examples processed in parallel in one forward+backward pass.

Decoding that keeps the top-N highest-probability partial sequences at each step.

Architecture that encodes query and document independently — fast for retrieval at scale.

Systematic skew in model outputs reflecting unrepresentative data or training choices.

Beyond the Imitation Game — a 200+ task collaborative LLM benchmark suite.

N-gram-overlap metric for machine-translation quality versus reference translations.

Tokenisation algorithm that iteratively merges the most-frequent adjacent token pairs.

Google's filtered 750GB English web-text dataset, used to train T5.

Informal measure of how much a model can learn, set by parameter count and architecture.

State-space-model TTS provider focused on ultra-low-latency real-time voice.

Prompting strategy that asks the model to reason step-by-step before answering.

Lightweight open-source embedding database, popular for prototyping LLM apps.

Splitting documents into smaller passages before embedding for retrieval.

Anthropic's frontier LLM family known for long context, safety, and constitutional AI.

Cosine similarity between CLIP embeddings of a prompt and a generated image.

Microsoft's 330k-image dataset with bounding boxes, segmentation and captions.

Cohere's mid-tier LLM optimised for RAG, tool use and enterprise workflows.

Cohere's flagship 104B-parameter model for high-stakes RAG and agentic applications.

Open repository of web-crawl data, the largest training source for nearly every LLM.

Text generated by an LLM in response to a prompt — the model's output portion.

Anthropic's alignment method using model-generated critiques against a written constitution.

Provider-side caching of long shared contexts (documents, codebases) across calls.

Fine-tuning a model so that previously long system prompts are baked into its weights.

Maximum number of tokens (prompt + completion) a model can attend to in a single request.

Adversarial payload designed to overflow or exploit a model's context window.

Network using sliding convolution filters, dominant for vision before transformers.

Transformer that scores a (query, document) pair jointly — used for reranking.

OpenAI's text-to-image diffusion model family, currently DALL-E 3.

Replicating the model on every GPU and splitting the batch across them.

Documentation describing the contents, provenance and risks of a training dataset.

Databricks' open-weights 132B mixture-of-experts LLM.

Transformer half that generates output tokens autoregressively, attending to prior context.

Transformer variant with only a stack of causal-attention decoder layers, no encoder.

Chinese lab behind DeepSeek-V3 and R1, frontier-quality open MoE models at low cost.

Generative model that learns to reverse a gradual noising process to synthesise data.

Training a small 'student' model to mimic a large 'teacher' model's outputs.

Training a model across many GPUs or nodes via data, model or pipeline parallelism.

Industry-leading neural text-to-speech models, with voice cloning and dubbing.

Skill rating computed from pairwise wins — the basis of Chatbot Arena rankings.

Dense numeric vector that represents text, an image, or other input in a learned semantic space.

Number of floats in an embedding vector — typically 256 to 4096 for modern models.

Local store keyed by content hash so repeated inputs aren't re-embedded — cuts cost.

Network half that compresses an input sequence into contextualised representations.

Transformer with both an encoder reading inputs and a decoder generating outputs.

TII Abu Dhabi's open-weights LLM family — Falcon 7B/40B/180B and Falcon 2.

Solving a task with a handful of demonstrations included directly in the prompt.

Including a small number of solved examples in the prompt to teach the model a pattern.

Distributional distance between generated and real images, lower is better.

Continued training of a pre-trained model on a task-specific dataset to specialise behaviour.

Memory-efficient attention algorithm that fuses operations to avoid materialising the score matrix.

Training objective that learns vector fields connecting noise to data — diffusion's cousin.

Black Forest Labs' state-of-the-art open-weights image-generation diffusion family.

4-bit floating-point format introduced on Blackwell GPUs — double FP8 throughput at inference.

8-bit floating-point format used on H100/B200 to double FP16 throughput at minor quality cost.

Structured-output mode where a model emits JSON describing which tool/function to invoke.

Two-network setup where a generator and a discriminator compete during training.

Google DeepMind's natively-multimodal frontier model family, with Nano, Flash and Pro tiers.

OpenAI's flagship language model family, decoder-only transformers from GPT-1 to GPT-5.

Massively-parallel accelerator that dominates neural-network training and inference.

Summing gradients over multiple micro-batches before applying an optimiser step.

xAI's LLM family — Grok-3, Grok-4 — with real-time X data and reasoning modes.

Grade-school-math benchmark of 8.5k word problems with step-by-step solutions.

Programmatic validators that constrain LLM inputs and outputs against policies.

MiniMax's text-to-video model — Hailuo 02 and beyond.

Confident model output that is factually wrong, fabricated, or unsupported by sources.

Classifier that blocks or modifies model outputs containing disallowed content.

deepset's open-source framework for production-grade RAG and search pipelines.

Commonsense-reasoning benchmark of completing video captions and how-to texts.

Nous Research's instruction-tuned fine-tunes of Llama, Mistral and Qwen base models.

Python library providing thousands of pre-trained transformer models with a unified API.

164-problem coding benchmark from OpenAI testing function generation from docstrings.

Combining lexical (BM25) and vector search scores to get the best of both worlds.

Image-generation model known for accurate in-image typography and posters.

Google DeepMind's text-to-image diffusion model family, currently Imagen 3.

1.4M-image, 1000-class image-classification benchmark that catalysed the deep-learning era.

Training a policy from demonstrations rather than from rewards.

A model's ability to perform new tasks from examples shown in the prompt — no weight updates.

Running a trained model on new inputs to produce predictions or generations.

Pre-processing step that checks user inputs against schemas, PII rules and length limits.

The capability of an LLM to interpret and execute natural-language directions reliably.

Supervised fine-tuning on (instruction, response) pairs to make a base model follow directions.

4-bit integer weight quantisation used widely for fast local LLM inference.

Adversarial prompt that bypasses a model's safety policies or refusal behaviour.

AI21 Labs' hybrid Mamba-Transformer LLM family with long context and high throughput.

Google's high-performance numerical-computing library with NumPy-like API and XLA backend.

API mode that forces a model to emit syntactically-valid JSON output.

Moonshot AI's Chinese LLM with industry-leading 2M-token context window.

Kuaishou's Chinese text-to-video model, known for long, motion-rich clips.

The most-recent date represented in a model's training data — beyond which it has no knowledge.

The classical Hinton-2015 formulation of distillation: minimise student-teacher logit divergence.

Stored attention keys and values from prior tokens, reused so generation does not recompute them.

Quantising attention keys/values to INT8 or INT4 to shrink KV-cache memory.

Open image-text dataset (5B pairs) that trained Stable Diffusion and many CLIP models.

Python/TypeScript framework for composing LLM calls into chains, agents and RAG pipelines.

Diffusion model that operates in a compressed latent space rather than raw pixels.

Learned high-dimensional space where each point encodes the abstract features of an input.

1000-hour English audiobook corpus, the standard speech-recognition benchmark.

Continuously-updated benchmark that swaps tasks monthly to avoid training-data contamination.

Meta's open-weights LLM family — the backbone of most open-source fine-tunes.

Pure-C++ inference of LLMs on CPU, GPU and Apple Silicon — the foundation of local AI.

Data framework for connecting LLMs to private data via retrieval and indexing.

Unnormalised scores a model emits over its vocabulary before softmax converts them to probabilities.

Parameter-efficient fine-tuning that injects small trainable low-rank matrices alongside frozen weights.

Scalar measure of how wrong a model's predictions are; gradient descent minimises it.

Selective state-space model with input-dependent dynamics, competitive with transformers.

12,500-problem competition-math benchmark spanning algebra, geometry and number theory.

Embeddings trained so that prefix slices remain useful — choose any dimension at query time.

Translation metric using stem and synonym matching for better human-correlation than BLEU.

Open-source vector database designed for billion-scale similarity search.

Chinese AI lab behind MiniMax-01 LLMs and Hailuo video model.

French AI lab whose Mistral 7B and Mixtral MoE models popularised sparse open-weights LLMs.

Training in lower-precision floats to double throughput and halve memory versus FP32.

Architecture where a router activates only a few expert sub-networks per token, saving compute.

Vision architecture using only multilayer perceptrons — no attention, no convolutions.

57-subject multiple-choice benchmark spanning STEM, humanities and professional exams.

Documentation describing a model's intended use, performance, limitations and risks.

Anthropic-introduced open standard for LLMs to access tools and data sources uniformly.

Combining the weights of multiple fine-tuned models into a single composite model.

Splitting a model's layers or tensors across multiple GPUs that do not fit one.

System that dispatches each request to the most-appropriate model from a pool.

Algorithm that selects which expert sub-networks process each input token.

Microsoft's COCO dataset — same as COCO, full name often used in academic citations.

Public benchmark and leaderboard scoring embedding models across 56 datasets.

Attention applied in parallel across several projections, each capturing different relations.

Dialogue with state carried across multiple user/assistant exchanges.

Model that accepts or produces more than one input/output type — text, image, audio, video.

NVIDIA's 2020 datacentre GPU — 40/80GB HBM2e, 312 TFLOPS FP16, the workhorse of pre-2024 AI.

NVIDIA's 2024 Blackwell GPU — 192GB HBM3e, ~2.5× H100 performance for FP8 training.

NVIDIA's 2022 Hopper datacentre GPU — 80GB HBM3, ~3× A100 performance.

Open-source transformer policy for robot control trained on Open X-Embodiment.

Command-line + REST tool that makes running local LLMs as simple as `ollama run llama3`.

Model whose trained weights are publicly downloadable, with or without an open-source licence.

Cross-institution open robotics dataset of 1M+ demonstrations from 22 embodiments.

Official Python and TypeScript client for the OpenAI REST API.

Aggregator API that unifies 200+ LLMs behind an OpenAI-compatible interface.

Open-weights vision-language-action model for robot policies, 7B parameters.

Mechanism forcing model output to match a regex, grammar or schema during decoding.

Median per-request latency — half of requests are faster, half slower.

95th-percentile request latency — 5% of requests are slower than this.

Number of learned weights in a neural network — a rough proxy for capacity and cost.

Transformer variant that cross-attends to a fixed-size latent array, decoupling from input size.

Exponential of the average per-token loss; lower means the model finds the data less surprising.

Microsoft's small-but-mighty model family trained on heavily-curated synthetic data.

Physical Intelligence's foundation model for generalist robot manipulation.

Pika Labs' consumer text-to-video model focused on stylised short clips.

Managed serverless vector database for retrieval and semantic search.

Splitting a model's layers into stages, with micro-batches pipelined across GPUs.

Mistral's open-weights vision-language model — Pixtral 12B and Pixtral Large.

Initial unsupervised training of a model on massive corpora to learn general patterns.

Learning trainable key/value prefixes prepended to every transformer layer.

Input text — system, user or assistant — that conditions a model's next response.

Reusing computation from previously-seen prompt prefixes across requests.

Decomposing a task into a sequence of model calls, each consuming the prior output.

Practice of designing prompts that elicit reliable, high-quality model outputs.

Attack where untrusted input overrides the developer's intended model instructions.

Reusable prompt with placeholders that get filled at runtime with dynamic data.

Learning a small set of continuous 'soft prompt' embeddings while keeping the model frozen.

Meta's dynamic deep-learning framework, dominant in research and increasingly in production.

Open-source Rust-based vector database with strong filtering and hybrid search.

LoRA on a 4-bit quantised base model, enabling 65B fine-tunes on a single 24GB GPU.

Compressing model weights to lower-precision integers (INT8/INT4) to save memory and speed inference.

Alibaba's open-weights LLM family covering text, code, vision and audio.

Pattern where a model is grounded on retrieved documents fetched at inference time.

Agent pattern interleaving Thought, Action and Observation steps in a single trace.

A model's ability to chain logical steps, plan, or solve problems beyond pattern recall.

Image-generation models specialised for vector graphics, logos and brand-consistent design.

Sequence model that processes inputs one step at a time, carrying a hidden state.

Systematic adversarial testing of an AI system to find failure modes before deployment.

Open recreation of the Llama training set — 1.2T tokens of curated web, books and code.

TII's 5T-token web-only dataset that powered Falcon — quality from filtering alone.

Model behaviour of declining a request that violates its safety policy.

Second-stage re-scoring of retrieved candidates with a more expensive model for better top-k.

End-to-end pattern combining retrieval with generation to ground answers in fresh data.

Transformer with a retrieval module integrated into its layers, not just its prompt.

Replacing human preference labels in RLHF with judgments from another model.

Training step that aligns a model to human preferences via a reward model and RL.

Position-encoding scheme that rotates query/key vectors in 2D subspaces by their position.

Recall-oriented n-gram and longest-common-subsequence metric for summarisation eval.

Google's Robotic Transformer 2 — a VLM fine-tuned to emit robot actions as text tokens.

Runway ML's commercial video-generation model family — Gen-3 and Gen-4.

Decoding strategy that picks the next token from the model's output distribution.

Empirical relations predicting model loss as a function of parameters, data and compute.

Explicit working-memory section a model writes into before producing its final answer.

Sampling multiple chain-of-thought traces and voting on the majority answer.

Microsoft's SDK for orchestrating LLMs, plugins and memory across .NET, Python and Java.

Search that ranks documents by meaning similarity, computed from embeddings.

Google's language-agnostic tokeniser library implementing BPE and unigram models.

Function that normalises a vector of logits into a probability distribution summing to one.

OpenAI's text-to-video diffusion-transformer model, up to 60-second cinematic clips.

Attention variants that restrict each query to a subset of keys, reducing compute.

Acceleration where a tiny draft model proposes tokens that the main model verifies in parallel.

Models that convert spoken audio into written text — the inverse of TTS.

Open-weights latent-diffusion image model — the spark of the open-source image-gen ecosystem.

Sequence model that uses continuous-time linear dynamics instead of attention.

String at which the model halts generation — used to enforce output boundaries.

Constrained decoding that guarantees output matches a schema (JSON, regex, grammar).

Real-world coding benchmark of 2294 GitHub issues with verified fixes.

Evaluation set generated by an LLM rather than collected from humans.

Training data generated by another model rather than collected from humans or the web.

Highest-priority instruction block that sets a model's persona, rules, and output format.

Sampling parameter that controls randomness — lower is more deterministic, higher more creative.

Splitting weight matrices across GPUs and synchronising on every layer's forward and backward.

Google's graph-based deep-learning framework, dominant in production until PyTorch's rise.

NVIDIA's LLM-specific extension of TensorRT for maximum GPU inference throughput.

Models that synthesise spoken audio from written text — the inverse of STT.

HuggingFace's optimised LLM serving stack used by their Inference Endpoints.

EleutherAI's 825GB diverse text dataset spanning 22 sources, used to train GPT-Neo/J.

Tokens or requests processed per second by an inference system.

OpenAI's fast Rust+Python BPE tokeniser library, the reference for counting GPT tokens.

TTS-specific latency from text-input to first audio packet — the voice analogue of TTFT.

Latency between sending a prompt and receiving the model's first response token.

Application-set ceiling on tokens consumed per request or per session for cost control.

Process of splitting text into the discrete units (tokens) that a model processes.

Per-stream generation speed of a model — the most reported speed number for LLMs.

Letting an LLM invoke external tools — APIs, code execution, search — during generation.

Sampling restricted to the K highest-probability tokens at each step.

Sampling from the smallest token set whose cumulative probability exceeds threshold p.

Google's custom ASIC for neural-network training and inference, available on GCP.

Neural architecture built on stacked self-attention layers, the basis of all modern LLMs.

Reasoning method that explores multiple branching thought paths and backtracks.

Probabilistic autoencoder that learns a smooth latent space with a Gaussian prior.

Database optimised for storing and querying high-dimensional embedding vectors.

Google DeepMind's text-to-video generation model, currently Veo 3.

Transformer that treats image patches as tokens — no convolutions required.

Multimodal model that jointly understands images and text in a shared representation.

Foundation model class that maps images and language directly to robot actions.

High-throughput LLM serving engine with paged-attention KV-cache management.

Fixed set of tokens a tokenizer and model can represent, typically 32k to 200k entries.

Voyage AI's embedding model family — voyage-3, voyage-large-2, voyage-code-2.

Embedding statistical patterns into model outputs so they can be detected as AI-generated.

Open-source vector database with built-in hybrid search and module-based vectorisers.

OpenAI's open-weights speech-to-text model, robust to noise and 100+ languages.

Solving a task with no labelled examples, relying only on a natural-language instruction.

Asking the model to do a task purely from a natural-language description, no examples.