Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Reference

AI glossary

219 terms you meet when choosing and building with AI tools, models and agents, each defined in a sentence or two, explained, and sourced.

#

τ-bench
τ-bench (tau-bench) is a benchmark that tests AI agents in simulated customer-service conversations, where a language model plays the user and the agent must use domain API tools while following written policy rules.

A

Abliteration
Abliteration is a technique that stops a language model from refusing requests by finding the single internal direction that carries refusal and editing the weights so the model can no longer write to it, without any retraining.
Agent framework
An agent framework is a software library that provides the common parts of an AI agent, such as the tool loop, state, memory, handoffs between agents and tracing, so developers do not build them from scratch.
Agent harness
An agent harness is the software around a model that turns it into a working agent: the loop, tools, system prompt, permissions, context management and any files used to carry progress between sessions.
Agent hooks
Agent hooks are user-defined commands or scripts that a coding agent runs automatically at set points in its work, such as before a tool call or after a file edit, to check, block or extend what it does.
Agent instructions file (AGENTS.md)
An agent instructions file is a Markdown file in a code repository, such as AGENTS.md or CLAUDE.md, that a coding agent reads at the start of a session to learn the project's commands, conventions and rules.
Agent loop
An agent loop is the repeating cycle at the heart of an AI agent: the model asks for a tool call, the application runs it and returns the result, and the model decides what to do next.
Agent memory
Agent memory is the set of techniques that let an AI agent keep information beyond a single context window, such as facts about a user or past work, by storing it outside the model and loading it back when needed.
Agent orchestration
Agent orchestration is the control of which agents or model calls run in an application, in what order, and how each next step is chosen, either by a model making the decision or by fixed code.
Agent permissions
Agent permissions are the rules that decide which tools, files, commands, networks and accounts an AI agent may use, and which actions need a human's approval first, so a mistaken or manipulated agent can do limited damage.
Agent sandbox
An agent sandbox is an isolated environment, such as a container, virtual machine or operating-system restriction, in which an AI agent runs code and commands so it cannot reach files, networks or data outside set limits.
Agent skills
Agent skills are folders of instructions, scripts and reference files, each described by a SKILL.md file, that an AI agent can discover and load only when a task needs them.
Agent2Agent protocol
The Agent2Agent protocol (A2A) is an open standard that lets AI agents built on different frameworks and by different vendors find each other, delegate tasks and exchange results without exposing their internal tools or memory.
Agentic AI
Agentic AI describes software in which a language model has some control over how a task gets done: it plans steps, picks tools and acts on the results, with more autonomy than a single prompt and reply.
Agentic RAG
Agentic RAG is retrieval-augmented generation run by an AI agent that decides when to search, what to search for and which source to use, and can search again if the first results fall short.
AI agent
An AI agent is a system in which a language model decides its own next steps: it calls tools, reads the results and repeats until the task is done, instead of following a fixed script.
AI alignment
AI alignment is the work of making an AI system's behaviour match what its users and developers actually intend, so it is helpful, honest and avoids harm, rather than just predicting likely text or optimising a narrow score.
AI assistant
An AI assistant is a software product, usually built on a large language model, that takes requests in plain language and answers questions or completes tasks for a person, as ChatGPT, Claude and Gemini do.
AI avatar
An AI avatar is a generated video character, often based on a real person, that speaks a script or live responses with matching face, lip and head movements.
AI benchmark
An AI benchmark is a fixed, public set of tasks with an agreed scoring method, used to compare models on the same test, for example coding issues solved, exam questions answered or retrieval quality across many datasets.
AI code editor
An AI code editor is a code editor or IDE with language models built in, offering completions, chat about the codebase and an agent mode that can plan and make changes across many files.
AI crawler
An AI crawler is an automated bot run by an AI company that fetches web pages to collect training data, build an AI search index, or retrieve a page live when a user asks an assistant about it.
AI image upscaling
AI image upscaling, or super-resolution, uses a neural network to enlarge an image and fill in plausible detail, so the result looks sharper than simple resizing would make it.
AI red teaming
AI red teaming is the practice of deliberately attacking a model or AI application before and after release, with human testers or automated tools, to find harmful outputs, jailbreaks, injections and other failures so they can be fixed.
AI search engine
An AI search engine is a search service that runs web searches for a query, reads the results with a language model and returns a written answer with links to its sources, instead of only a list of links.
AI supply-chain attack
An AI supply-chain attack compromises an AI application through something it depends on, such as a downloaded model, a LoRA adapter, a dataset, a Python package, an MCP server or an agent skill, rather than attacking the application directly.
AI training data
AI training data is the collection of examples a model learns from, such as web text, code, images, transcripts or labelled pairs. Its size, quality, mix and licensing shape what the model can do and where it fails.
AI watermarking
AI watermarking is embedding a hidden, machine-detectable signal in text, images, audio or video at the moment an AI model generates it, so software can later identify the content as AI-generated.
AIME (benchmark)
AIME, the American Invitational Mathematics Examination, is a 15-question US maths competition with whole-number answers from 0 to 999; AI labs use its past papers, such as AIME 2025, as a maths reasoning benchmark.
API key
An API key is a secret string an application sends with each request so that a service can tell which account is calling, apply its limits and bill it. Anyone holding the key can usually use it.
Approximate nearest neighbor search
Approximate nearest neighbor (ANN) search finds vectors close to a query without checking every stored vector. It gives up a guarantee of the exact best match in exchange for much faster search on large collections.
ARC-AGI
ARC-AGI is a series of benchmarks, started by François Chollet and run by ARC Prize, that tests whether an AI system can solve new reasoning puzzles it has not seen before, measuring how efficiently it learns rather than what it knows.
Artificial Analysis Intelligence Index
The Artificial Analysis Intelligence Index is a composite score from the benchmarking company Artificial Analysis that combines several evaluations, spanning knowledge work, coding, science and general knowledge, into one number for comparing language models.
Attention mechanism
An attention mechanism lets a neural network decide, for each token it processes, how much weight to give every other token in the input. Self-attention is the core operation inside a transformer.
Autoregressive model
An autoregressive model generates a sequence one piece at a time, predicting each next token from all the tokens before it and feeding that output back in. Most large language models work this way.

B

Base model
A base model is a language model straight out of pretraining, before instruction tuning or human-feedback training. It continues whatever text it is given rather than following instructions or holding a conversation.
Batch API
A batch API lets you submit a large set of model requests in one job and collect the results later, instead of waiting on each call. Providers run the job when capacity allows and usually charge less for it.
Benchmark contamination
Benchmark contamination is when a model has seen a benchmark's test questions or answers during training, so its score on that benchmark overstates how well it handles new problems of the same kind.
BM25
BM25 is a ranking formula for keyword search. It scores each document by how often the query's words appear in it, giving more weight to rare words and less credit to long documents and repeated terms.
BrowseComp
BrowseComp is an OpenAI benchmark of 1,266 hard questions that tests whether a browsing agent can search the web persistently for tangled, hard-to-find facts, with short answers that are easy to check.
Browser agent
A browser agent is an AI agent that controls a web browser to complete tasks, such as searching, clicking, filling in forms and reading pages, often running a headless browser with no visible window.

C

C2PA content credentials
C2PA content credentials are signed provenance records, defined by an open standard from the Coalition for Content Provenance and Authenticity, that travel with an image, video, audio file or document and record its origin and edits.
Chain-of-thought
Chain-of-thought (CoT) is a prompting method in which a language model writes out intermediate reasoning steps before its final answer, which tends to improve results on maths, logic and other multi-step problems.
Chunking
Chunking is the step that splits long documents into smaller passages before they are embedded and indexed, so a retrieval system can find and return only the parts of a document that answer a query.
Coding agent
A coding agent is an AI agent that works on software directly: it reads a codebase, edits files, runs commands and tests, and repeats until a change works, rather than only suggesting code for a person to paste.
Computer use
Computer use is the ability of an AI model to operate a computer the way a person does, by looking at screenshots and sending mouse clicks and keystrokes, so it can work with software that has no API.
Context compaction
Context compaction is the step where an agent summarises a conversation that is nearing its context window limit and continues in a fresh window that starts from that summary, so a long task can keep going.
Context engineering
Context engineering is the practice of choosing and maintaining everything that goes into a language model's context window during a task: instructions, retrieved documents, tool definitions, tool results and conversation history.
Context rot
Context rot is the drop in a language model's accuracy and reliability as its input grows longer, even when the input still fits inside the context window and the needed information is present.
Context window
A context window is the maximum amount of text, counted in tokens, that a language model can consider at once, covering the system prompt, conversation, documents, tool results and the response it writes.
Continuous batching
Continuous batching is a model-serving technique that adds new requests to the running GPU batch and removes finished ones after every generation step, instead of waiting for a whole batch to complete. It keeps the GPU busy and raises throughput.
ControlNet
ControlNet is a neural network add-on that lets a text-to-image diffusion model follow a structural guide, such as an edge map, depth map or human pose, in addition to the text prompt.
Cosine similarity
Cosine similarity measures how alike two vectors are by the angle between them. It ranges from -1 to 1, ignores vector length, and is the most common way to compare embeddings in search.
CPU offloading
CPU offloading runs a model that does not fit in GPU memory by keeping some of its layers in system RAM, or on disk, and using the GPU for the rest. The model runs, but more slowly.
Cross-encoder
A cross-encoder is a model that reads a query and a document together in one pass and outputs a relevance score. It is more accurate than comparing separate embeddings, but too slow to run over a whole collection.

D

Data exfiltration (AI agents)
Data exfiltration, for AI agents, is when an attacker tricks an agent into reading private data it can access and sending it out, through a web request, an email, an image link or another channel the attacker controls.
Data residency
Data residency is a control that keeps a customer's data stored, and often processed, in a chosen geographic region such as the EU, so a team can meet its own legal, contractual or customer rules on where data lives.
Decision model
A decision model is a model built to return a typed answer, such as a choice from a list, a score on a rubric or a yes/no probability, instead of generating free text.
Deep research (AI)
Deep research is an AI agent feature that plans a research question, runs many web searches, reads the sources it finds and writes a long report with citations, taking minutes rather than seconds.
Deepfake
A deepfake is image, audio or video content made or altered with AI so that it convincingly shows a real person, place or event in a way that did not happen and could pass as authentic.
Diffusion language model
A diffusion language model (dLLM) generates text by starting from a fully masked or noisy sequence and refining many token positions in parallel over several steps, instead of writing one token at a time.
Diffusion model
A diffusion model is a generative model that learns to turn random noise into data, such as an image, by removing a little noise at a time over many steps.
Direct preference optimization
Direct preference optimization (DPO) is a way to align a language model with human preferences by training directly on pairs of preferred and rejected answers, using a simple classification-style loss instead of a reward model and reinforcement learning.
Document parsing
Document parsing converts files such as PDFs, scans, slides and spreadsheets into structured text that software can use, keeping headings, reading order, tables and images instead of a flat stream of characters.

E

Elo rating (model arenas)
An Elo rating, in model arenas, is a score computed from many head-to-head votes between pairs of models, so that the gap between two scores predicts how often one model's answer will be preferred over the other's.
Embedding
An embedding is a list of numbers (a vector) that represents a piece of text, an image or other data, produced so that items with similar meaning end up with vectors close together.
Embedding dimension
Embedding dimension is the number of values in each vector an embedding model outputs. More dimensions can hold more detail, but every extra value adds storage, memory and search cost per item.
EU AI Act
The EU AI Act, Regulation (EU) 2024/1689, is the European Union's law on artificial intelligence. It sets risk-based rules for those who build and deploy AI systems, from outright bans to transparency duties, phased in from 2025.

F

Few-shot prompting
Few-shot prompting is giving a language model a handful of worked examples of a task inside the prompt, so it can copy the pattern for a new input without any retraining.
Fine-tuning
Fine-tuning is further training of a pretrained model on a smaller, task-specific dataset, so the model adopts a particular skill, format, tone or domain without being trained from scratch.
Foundation model
A foundation model is a large model trained on broad data at scale that can be adapted to many different downstream tasks, through prompting or fine-tuning, instead of being built for one task.
Frontier model
A frontier model is one of the most capable general-purpose AI models available at a given time, usually from a well-funded lab. The term is relative and also appears in AI safety policy.

G

Generative engine optimization
Generative engine optimization (GEO) is the practice of shaping web content so that AI search engines and chat assistants find it, quote it and cite it in the answers they generate.
GGUF
GGUF is a binary file format for storing a model's weights, tokenizer and metadata in one file, built for fast loading and local inference with GGML-based tools such as llama.cpp, often in quantized form.
Golden dataset
A golden dataset is a fixed, carefully checked set of example inputs paired with the answers or behaviour you expect, used as the reference when you evaluate or regression-test an AI system.
GPQA
GPQA (Graduate-Level Google-Proof Q&A) is a benchmark of 448 hard multiple-choice questions in biology, physics and chemistry, written by experts so that skilled non-experts cannot answer them even with unlimited web search.
GPU
A GPU (graphics processing unit) is a processor with many small cores built to run many calculations in parallel. Its speed at matrix maths makes it the standard hardware for training and running AI models.
GPU cloud
A GPU cloud is a service that rents out GPU machines by the hour or second for AI training, fine-tuning and inference. Specialist providers, sometimes called neoclouds, focus on GPUs and often charge less than the large general clouds.
GraphRAG
GraphRAG is a retrieval approach where an LLM first builds a knowledge graph from a document collection and writes summaries of related groups of entities, then answers questions using those summaries and graph links.
Grounding
Grounding is tying a language model's answer to specific, checkable sources, such as search results, documents or a database, so that its claims rest on supplied evidence rather than only on training data.
Guardrails (AI)
Guardrails are runtime checks placed around a language model that inspect inputs and outputs and block, rewrite or redirect them, so an application stays on topic, avoids harmful content and follows rules the model alone would not reliably keep.
Guidance scale (CFG)
The guidance scale, or CFG scale, is a setting in diffusion image generators that controls how strongly the output follows the prompt. Higher values follow the prompt more closely but reduce variety.

H

Hallucination (AI)
A hallucination is output from an AI model that sounds fluent and confident but is false, unsupported by its sources or made up, such as an invented citation, quote, statistic or API function.
HNSW
HNSW (hierarchical navigable small world) is a graph-based index for approximate nearest neighbor search. It links vectors to their neighbors across several layers, so a search can hop quickly toward the closest matches.
Human evaluation
Human evaluation is judging AI outputs with people rather than automated metrics, for example raters scoring answers against a rubric or picking the better of two anonymous responses in a blind test.
Human-in-the-loop
Human-in-the-loop (HITL) is a design in which an AI system pauses at chosen points for a person to approve, correct or reject its next step, so risky or uncertain actions are not taken automatically.
HumanEval
HumanEval is a code-generation benchmark released by OpenAI in 2021: 164 hand-written Python problems, each with a function signature, docstring and unit tests, used to check whether a model writes code that actually works.
Humanity's Last Exam
Humanity's Last Exam (HLE) is a benchmark of 2,500 expert-written questions across dozens of subjects, built to be very hard for AI models, with answers that are clear and checkable but cannot be found quickly online.
Hybrid search
Hybrid search runs a keyword search and a vector search on the same query and merges the two ranked lists, so results can match on exact words as well as on meaning.

I

Image-to-image
Image-to-image is the task of generating a new image from an existing one plus a prompt, keeping some of the original's structure while changing style, detail or content.
Image-to-video
Image-to-video is the task of turning a still image into a short video clip, usually with the image as the first frame and a text prompt describing the motion.
Inference
Inference is running a trained model to produce output from new input, such as an LLM generating a reply to a prompt. It is the stage users pay for every time they call a model, as opposed to training.
Inference provider
An inference provider is a company that runs AI models on its own hardware and sells access to them through an API, usually charged per token. You send prompts and get outputs without managing GPUs yourself.
Inpainting and outpainting
Inpainting is filling a masked region of an image with new generated content that fits the rest of the picture. Outpainting is the same idea applied outside the edges, extending the image beyond its frame.
Instruct model
An instruct model is a base language model that has been further trained on instructions paired with good answers, and often on human feedback, so it follows requests instead of just continuing text.

J

Jailbreak (LLM)
A jailbreak is a prompt or technique that gets a language model to produce output its safety training is meant to refuse, such as harmful instructions, by role-play, obfuscation, odd encodings or long adversarial text.

K

Knowledge base (AI)
In AI apps, a knowledge base is the collection of documents and data that an assistant or agent searches to answer questions, such as help articles, policies or product docs, kept separate from the model itself.
Knowledge cutoff
A knowledge cutoff is the date after which a model has little or no information, because its training data was collected up to around then. Events after it are unknown unless supplied in the prompt.
Knowledge distillation
Knowledge distillation is a training method in which a smaller student model learns to imitate a larger teacher model's outputs, producing a distilled model that keeps much of the teacher's skill at lower cost.
Knowledge graph
A knowledge graph stores information as entities (people, products, places, concepts) connected by named relationships, so software can follow links between facts instead of searching through loose text.
KV cache
A KV cache stores the attention keys and values a transformer has already computed for earlier tokens, so each new token only needs work for itself. It speeds up generation but takes GPU memory that grows with context length.

L

Large language model
A large language model (LLM) is a neural network with billions of parameters, trained on large amounts of text to predict the next token, which lets it write, answer questions, summarise and follow instructions.
Late interaction (ColBERT)
Late interaction is a retrieval method, introduced by ColBERT, that encodes the query and each document separately into one vector per token, then scores a match by comparing those token vectors in a cheap final step.
Layout analysis
Layout analysis is the part of document parsing that finds the regions on a page, such as titles, paragraphs, tables, figures, captions and footers, labels each one, and works out the order a person would read them in.
Leaderboard (AI)
An AI leaderboard is a public ranking of models by their scores on one or more benchmarks or on human preference votes, updated as new models are tested, so readers can compare models side by side.
Lip sync (AI video)
AI lip sync is the use of a model to change the mouth movements of a person in a video or image so they match a new audio track, such as a translated or rewritten line.
LLM evaluation
LLM evaluation, often called evals, is measuring how well a model or AI application performs a task by running it on a set of test cases and scoring the outputs against defined criteria, by code, by humans or by another model.
LLM gateway
An LLM gateway is a proxy server between your applications and model providers that exposes one API and handles keys, routing, fallbacks, rate limits, logging and spend tracking in a single place.
LLM observability
LLM observability is recording what an AI application does in production, including prompts, responses, tool calls, latency, token use and cost, as traces you can search, so you can debug failures and see how the system behaves over time.
LLM routing
LLM routing is the practice of sending each request to the language model best suited to it, often a cheap model for easy prompts and a stronger one for hard prompts, to balance cost, speed and quality.
LLM-as-a-judge
LLM-as-a-judge is an evaluation method where a strong language model grades the outputs of another model or application, using a rubric, a reference answer or a side-by-side comparison, in place of or alongside human reviewers.
llms.txt
llms.txt is a proposed convention in which a website publishes a Markdown file, usually at /llms.txt, that summarises the site and links to its most useful pages in a form language models can read easily.
LMArena
LMArena, first known as Chatbot Arena, is a public platform that ranks AI models by crowdsourced human votes: users see two anonymous models answer the same prompt and pick the better response.
Local LLM
A local LLM is a language model that runs on your own computer, phone or server instead of a provider's cloud API, so prompts and data stay on your hardware and it works without an internet connection.
LoRA
LoRA (low-rank adaptation) is a fine-tuning method that freezes a model's original weights and trains small low-rank matrices added to its layers, so far fewer parameters change and much less GPU memory is needed.

M

Matryoshka embeddings
Matryoshka embeddings are vectors trained so that their first values carry the most information. You can cut them to a shorter length and still get useful search results, which saves storage and speeds up search.
MCP client
An MCP client is the part of an AI application that holds one connection to one MCP server, negotiating capabilities and carrying requests, results and notifications between that server and the host app.
MCP resources and prompts
MCP resources and prompts are two things an MCP server can offer besides tools. Resources are pieces of data, each with a URI, that an app can add to context; prompts are reusable message templates a user chooses to run.
MCP server
An MCP server is a program that exposes tools, data resources or prompt templates to AI applications through the Model Context Protocol, so any MCP-compatible client can use them without custom integration code.
MCP transport (stdio, Streamable HTTP)
An MCP transport is the channel that carries MCP's JSON-RPC messages between client and server. The specification defines two: stdio, for a server run as a local subprocess, and Streamable HTTP, for a server reached over the network.
Metadata filtering
Metadata filtering limits a search to items whose stored attributes, such as date, author, language or access group, match given conditions, alongside or before ranking them by vector similarity.
Mixture of experts
Mixture of experts (MoE) is a model design in which each layer holds several expert sub-networks and a router sends each token to only a few of them, so only part of the model runs per token.
MMLU
MMLU (Massive Multitask Language Understanding) is a multiple-choice benchmark covering 57 subjects, from elementary maths and US history to law and computer science, that tests a language model's broad knowledge and problem solving.
Model card
A model card is a short document published with a machine learning model that explains what it is for, how it was trained and evaluated, where it performs poorly, and how it should and should not be used.
Model Context Protocol
The Model Context Protocol (MCP) is an open standard that defines how AI applications connect to external tools, data and prompts, using JSON-RPC messages between a host application, its clients and separate servers.
Model fallback
Model fallback is a rule that sends a request to a backup model or provider when the first choice fails, times out, hits a rate limit or rejects the input. It keeps an AI feature working through outages.
Model license
A model license is the legal terms under which a model's weights, and sometimes its code and data, may be used, changed and shared. Terms range from permissive open licences to non-commercial or custom restricted ones.
Model parameters
Model parameters are the numbers a neural network learns in training, mostly weights. The parameter count, written as 7B for seven billion, is the usual shorthand for a model's size.
Model serving
Model serving is running a trained model as a service that applications call over a network, with software that loads the weights onto GPUs, queues and batches requests, manages memory and returns results.
Model weights
Model weights are the learned numbers inside a neural network, set during training, that decide how it turns input into output. A checkpoint is a saved file or set of files containing those weights.
MTEB
MTEB, the Massive Text Embedding Benchmark, is a public benchmark and leaderboard that scores text embedding models across many task types, such as retrieval, clustering, classification and reranking, over many datasets and languages.
Multi-agent system
A multi-agent system is a setup in which several AI agents, each with its own instructions, tools and context, work on parts of a task and pass results to each other or to a coordinating agent.
Multimodal model
A multimodal model is an AI model that can take in, and sometimes produce, more than one kind of data, such as text, images, audio or video, within a single model rather than through separate systems.

N

No-code AI app builder
A no-code AI app builder is a tool that turns a plain-language description or a visual editor into a working app, usually with hosting and a database included, so people can ship software without writing code.
NPU
An NPU (neural processing unit) is a chip, or a block inside a larger chip, built to run neural network maths quickly and with little power. Phones and laptops use NPUs to run small AI models on the device.
Numeric precision (FP16, BF16, FP8)
Numeric precision is the number format used to store a model's weights and run its maths, such as FP32, FP16, BF16, FP8 or INT4. Fewer bits per number means less memory and faster inference, at some risk to accuracy.

O

OAuth for AI agents
OAuth for AI agents is the use of the OAuth authorization standard to let an agent or MCP client act on a user's behalf with a limited, revocable access token, instead of holding the user's password or a broad API key.
ONNX
ONNX (Open Neural Network Exchange) is an open file format for machine learning models. It describes a model as a graph of standard operators so it can be trained in one framework and run in another.
Open source AI
Open source AI is an AI system released under terms that let anyone use, study, modify and share it, with access to the parts needed to change it: training data information, full code and model parameters.
Open-weight model
An open-weight model is a model whose trained weights are published for anyone to download and run, often without the training code or full training data, so it is not necessarily open source.
OpenAI-compatible API
An OpenAI-compatible API is an endpoint from another provider or tool that accepts the same request and response format as OpenAI's API, so code written for OpenAI can switch to it by changing the base URL, key and model name.
Optical character recognition
Optical character recognition (OCR) is software that reads text from images, such as scanned pages, photos or screenshots, and turns it into machine-readable characters that can be searched, copied or fed to a model.
OSWorld
OSWorld is a benchmark for computer-use agents that runs real operating systems such as Ubuntu, Windows and macOS and scores agents on 369 open-ended tasks across web and desktop apps, checked by scripts.
Output tokens
Output tokens are the tokens a language model generates in its reply. They are counted and billed separately from input tokens, and a max_tokens setting caps how many one response may contain.

P

PagedAttention
PagedAttention is a method for storing a language model's KV cache in small fixed-size blocks that need not sit next to each other in GPU memory, like pages in an operating system. It cuts wasted memory so a server can run more requests at once.
Parameter-efficient fine-tuning
Parameter-efficient fine-tuning (PEFT) adapts a pretrained model by training only a small number of new or selected parameters, leaving the rest frozen. It cuts the memory, compute and storage that full fine-tuning needs.
pass@k
pass@k is a metric for code benchmarks that gives the chance at least one of k generated solutions to a problem passes its unit tests; pass@1 is the chance that a single attempt succeeds.
PII redaction
PII redaction is the detection and removal or masking of personally identifiable information, such as names, phone numbers, addresses and card numbers, from text or images before it is sent to a model, stored in logs or shown to others.
Pretraining
Pretraining is the first and largest stage of training a model, where it learns general patterns from a very large unlabelled dataset, usually by predicting the next token, before any task-specific tuning.
Prompt caching
Prompt caching is an API feature that stores the processed form of a repeated prompt prefix, such as a long system prompt or document, so later requests that start with the same prefix are cheaper and faster.
Prompt engineering
Prompt engineering is the practice of writing, structuring and testing the instructions and examples given to a language model so that it produces the output an application needs, without changing the model's weights.
Prompt injection
Prompt injection is an attack where text fed to a language model, typed by a user or hidden in a web page, file or tool result, changes the model's behaviour in ways its developer did not intend.

Q

Quantization
Quantization is a technique that stores a model's weights, and sometimes activations, with fewer bits, such as 8 or 4 instead of 16, so the model needs less memory and often runs faster, at some cost in accuracy.
Query rewriting
Query rewriting changes a user's question before search, for example by clarifying it, splitting it or adding related terms, so the retriever is more likely to find the passages that answer it.

R

Rate limit (API)
A rate limit caps how much an account can use an API in a time window, typically as requests per minute (RPM) and tokens per minute (TPM). Requests over the limit are rejected, usually with an HTTP 429 error.
ReAct (reasoning and acting)
ReAct (reasoning and acting) is a prompting pattern in which a language model alternates between writing a reasoning step, taking an action such as a search or tool call, and reading the result before its next step.
Reasoning model
A reasoning model is a language model trained to work through a problem in a stream of intermediate thinking tokens before giving its final answer, which tends to help on maths, coding and multi-step tasks.
Regression testing (LLM apps)
Regression testing for LLM apps means re-running a fixed set of evaluation cases after every change to a prompt, model, retrieval setup or tool, to catch cases that used to work and now fail.
Reranking
Reranking is a second search step that takes a short list of candidate documents and reorders them by relevance to the query, using a more accurate but slower model than the first search.
Retrieval metrics (recall@k, nDCG, MRR)
Retrieval metrics score how well a search system returns relevant documents. Recall@k checks how many relevant items appear in the top k, nDCG rewards ranking them high, and MRR measures how early the first correct one appears.
Retrieval-augmented generation
Retrieval-augmented generation (RAG) is a pattern where a system first searches a collection of documents for passages relevant to a question, then gives those passages to a language model so its answer draws on them.
Reward model
A reward model is a separate model trained on human preference data to score how good a response is. Reinforcement learning then uses that score to tune a language model towards answers people prefer.
RLHF
RLHF (reinforcement learning from human feedback) fine-tunes a language model using human rankings of its answers: a reward model learns those preferences, then reinforcement learning pushes the model toward answers that score well.

S

Safetensors
Safetensors is a file format from Hugging Face for storing model weights. Unlike Python pickle files, loading it cannot run hidden code, and it supports fast, memory-mapped loading of individual tensors.
Self-hosting (AI models)
Self-hosting an AI model means running it on hardware you control, such as your own servers, rented GPUs or a laptop, instead of calling a provider's API. You choose the model and keep the data, but you run the infrastructure.
Semantic cache
A semantic cache stores past LLM responses and returns one when a new request means roughly the same thing as an earlier one, judged by embedding similarity rather than an exact text match. It saves cost and latency on repeated questions.
Serverless GPU
Serverless GPU is a cloud model where you deploy code or a model and the platform starts GPU workers only when requests arrive, billing for the time they run and scaling to zero when idle.
Slash command
A slash command is an instruction typed with a leading slash, such as /compact or /review, that tells an AI tool to run a built-in action or a saved prompt instead of treating the text as an ordinary message.
Small language model
A small language model (SLM) is a language model with relatively few parameters, usually a few billion or fewer, built to run on a laptop, phone or single GPU at lower cost than a large model.
SOC 2
SOC 2 is an audit report, produced by an outside CPA audit firm under AICPA standards, on a service organisation's controls relevant to security, availability, processing integrity, confidentiality or privacy.
Sparse retrieval
Sparse retrieval represents queries and documents as vectors over a vocabulary where most entries are zero and each non-zero entry is a word weight, so search can use fast inverted indexes like keyword search.
Speaker diarization
Speaker diarization is the task of splitting an audio recording by speaker, labelling each stretch of speech with who said it. It answers the question "who spoke when".
Spec-driven development
Spec-driven development is a way of working with coding agents in which you first write a specification of what to build and why, then have the agent turn it into a plan, a task list and finally code.
Speculative decoding
Speculative decoding speeds up text generation by letting a small, fast draft model propose several tokens, which the large model then checks in a single pass, keeping the accepted ones. The output matches what the large model alone would produce.
Speech-to-speech model
A speech-to-speech model takes spoken audio in and produces spoken audio out directly, without first converting the speech to text and back. It is used for live voice conversation and translation.
Speech-to-text
Speech-to-text, also called automatic speech recognition (ASR), is software that converts spoken audio into written text. It powers transcription, captions, voice commands and the listening half of voice agents.
State space model
A state space model (SSM) is a sequence model that carries a fixed-size hidden state forward as it reads each token, so its cost grows linearly with input length. Mamba is the best-known example.
Streaming (LLM output)
Streaming sends an LLM's output to the client piece by piece as tokens are generated, instead of waiting for the full response. Users see text appear within moments, even when the complete answer takes much longer.
Structured output
Structured output is an API feature that makes a language model's reply follow a format you define, usually a JSON schema, so that code can parse it reliably instead of scraping free text.
Subagent
A subagent is a helper agent that a main agent hands a task to. It works in its own fresh context window, often with its own instructions and tools, and returns only a summary of the result.
Supervised fine-tuning
Supervised fine-tuning (SFT) trains a pretrained model further on example prompts paired with the responses you want. It is the usual first step that turns a base model into one that follows instructions.
SWE-bench
SWE-bench is a coding benchmark built from real GitHub issues in popular Python projects; a model or agent must edit the repository so that the issue is fixed and the project's hidden tests pass.
Synthetic data
Synthetic data is training or test data generated by software, often by another AI model, rather than collected from real-world sources. It is used to fill gaps, add examples cheaply, or avoid exposing real personal data.
System prompt
A system prompt is the block of instructions an application sends to a language model before the user's messages. It sets the model's role, rules, tone and output format for the whole conversation.

T

Table extraction
Table extraction is the task of finding tables in documents and rebuilding their rows, columns, headers and cells as structured data, such as HTML, Markdown or CSV, instead of loose lines of text.
Temperature (sampling)
Temperature is a sampling setting that controls how random a language model's word choices are. Low values make it favour the most likely tokens; higher values spread choices across less likely ones.
Tensor parallelism
Tensor parallelism splits each layer of a neural network across several GPUs, so that every GPU holds part of each weight matrix and computes its share of every step. It lets a model too large for one GPU run across many.
Terminal-Bench
Terminal-Bench is a benchmark for AI agents that work in a command-line terminal, where each task gives the agent a real environment and is scored by tests that check whether the job was actually done.
Test-time compute
Test-time compute is the amount of computation a model spends while answering a request, rather than during training. Letting a model think longer or try several answers can improve results on hard problems.
Text-to-3D
Text-to-3D is the task of generating a 3D object or scene from a written prompt. Image-to-3D does the same from one or more photos. The output can be viewed from any angle.
Text-to-image
Text-to-image is the task of generating a new picture from a written description, called a prompt. Most current text-to-image models are diffusion models conditioned on the text.
Text-to-music
Text-to-music is the task of generating music audio from a written description, such as genre, mood, instruments and tempo. Some systems also write and sing lyrics.
Text-to-speech
Text-to-speech (TTS) is software that turns written text into spoken audio. Modern TTS uses neural networks that predict the sound of speech from the text, then render it as a waveform.
Text-to-video
Text-to-video is the task of generating a short video clip from a written prompt. Models must produce frames that each look right and also move consistently from one frame to the next.
Time to first token
Time to first token (TTFT) is how long a user waits between sending a request to a language model and receiving the first token of its reply, covering queuing, prompt processing and network time.
Token (LLM)
A token (LLM) is the basic unit of text a language model reads and writes, often a word, part of a word, a punctuation mark or a byte. Prices, speed and context limits are all counted in tokens.
Token pricing
Token pricing is how most AI model APIs charge: a set price per million input tokens you send and a separate, usually higher, price per million output tokens the model generates, with discounts for cached input and batch jobs.
Tokenizer
A tokenizer is the component that splits text into tokens and turns them into numeric IDs a language model can process, then turns the model's output IDs back into text.
Tokens per second
Tokens per second (tok/s) measures how fast a language model produces output. It is quoted either per user, the speed one reply streams at, or per system, the total output across all concurrent requests.
Tool calling
Tool calling, also called function calling, is a model API feature where you describe functions with a schema, the model replies with a structured request to call one, and your code runs it and returns the result.
Tool poisoning
Tool poisoning is an attack where a malicious tool, often an MCP server, hides instructions in its tool description or metadata; the AI model reads them as guidance and may leak data or misuse other tools without the user noticing.
Top-p sampling
Top-p sampling, or nucleus sampling, picks each next token only from the smallest set of candidates whose combined probability reaches a threshold p, cutting off the long tail of unlikely tokens.
TPU
A TPU (tensor processing unit) is Google's custom chip for machine learning, built around large matrix-multiply units and high-bandwidth memory. Google offers TPUs to other companies as a rented service on Google Cloud.
Transformer (architecture)
The transformer is a neural network architecture, introduced in 2017, that processes sequences using attention instead of recurrence. It is the basis of nearly all current large language models and many image and audio models.

U

Unified memory
Unified memory is a design in which the CPU and GPU share one pool of memory instead of each having its own. On Apple Silicon Macs and similar chips, it lets a local AI model use most of the system RAM as GPU memory.

V

Vector database
A vector database stores embeddings alongside their source records and finds the stored vectors closest to a query vector, usually with an approximate index so search stays fast across millions of items.
Vector search
Vector search finds items by comparing their embeddings to the embedding of a query and returning the closest ones, so results match on meaning rather than on shared keywords.
Vibe coding
Vibe coding is building software by describing what you want to an AI model and accepting the code it writes without reading or reviewing it, judging the result only by whether the program seems to work.
Vision-language model
A vision-language model (VLM) is a model that takes images and text as input and produces text, so it can describe pictures, answer questions about them, read documents and extract data from screenshots.
Voice agent
A voice agent is an AI agent that holds a spoken conversation in real time, listening to a person, deciding what to say or do, and answering in a synthetic voice, often over a phone line or in an app.
Voice cloning
Voice cloning is the use of AI to make a text-to-speech system speak in a specific person's voice, learned from a recording of that person. Some systems need only a short sample.
VRAM
VRAM is the memory on a graphics card. For AI work it holds the model weights, the KV cache and working data, and whether a model fits in VRAM usually decides which GPU you need to run it.

W

Web crawling for AI
Web crawling for AI is the automated fetching of web pages, by following links or a list of URLs, to collect text for training models, building search indexes or giving an application fresh content to retrieve.
Weight-space programming
Weight-space programming is a proposed capability in which a language model reads a neural network's weights to work out what it computes, then writes or edits those weights directly, without gradient training, much as a programmer edits code.
Workflow automation (AI)
AI workflow automation is the use of tools such as Zapier, Make or n8n to chain triggers, app actions and model calls into a process that runs on its own, like summarising each new support ticket.
Workflow vs agent
A workflow is an LLM system whose steps follow paths fixed in code, while an agent lets the model decide its own steps and tool calls. Many products sold as agentic workflows sit somewhere between the two.
World model
A world model is an AI model that learns how an environment works well enough to predict what happens next, often from video or sensor data, so an agent can plan or train inside the simulation.

Z

Zero data retention
Zero data retention (ZDR) is an arrangement in which an AI provider does not keep a customer's prompts and outputs after a request is answered, usually offered only to approved customers and only for eligible API features.
Zero-shot
Zero-shot means asking a model to do a task from an instruction alone, with no worked examples in the prompt and no task-specific training. The model relies only on what it already learned.