AI Glossary.
Plain definitions of AI terms, how Mangrove handles the problems they name, and the sources behind each one.
A
Agent harness
An agent harness is the software around an AI model that runs its loop, supplies its context and tools, and enforces limits the model cannot change.
Guardrails and control
Agentic AI
Agentic AI is a system's ability to pursue a goal over several steps, choosing actions, calling tools and acting on results with limited human direction.
Agents and tools
AI agent
An AI agent is a system in which an AI model chooses its own next steps, calls tools and acts on their results until a task is done or a limit is reached.
Agents and tools
Artificial intelligence
Artificial intelligence (AI) is the field of building computer systems that turn inputs into predictions, content, recommendations, decisions or actions.
Foundations and history
Artificial neuron
An artificial neuron is a simple computing unit, loosely modeled on a nerve cell, that combines several inputs into one output using a threshold.
Foundations and history
B
Backpropagation
Backpropagation is the method that works out how much each weight in a neural network contributed to its error, working backward from the output.
Neural networks and deep learning
C
Chain of thought
Chain of thought is the series of intermediate reasoning steps that a language model writes out before it gives its final answer.
Language models
Circuit breaker
A circuit breaker is an automatic stop that halts trading or another activity when a measured value, such as a price fall or a loss, passes a set limit.
Guardrails and control
Computer vision
Computer vision is the field of AI that gets computers to interpret images and video, such as recognizing which objects an image shows and where.
Neural networks and deep learning
Confabulation
Confabulation is when an AI model fills missing information with an invented account and presents it as fact, such as an agent misreporting what it did.
Agent failure modes
Context
Context is everything a language model reads when it produces a response, such as its instructions, the conversation so far, documents and tool results.
Language models
Context engineering
Context engineering is the practice of choosing what goes into a language model's context at each step, so it holds what the task needs and little else.
Language models
Context management
Context management is the set of mechanisms that keep a long-running agent's context under its size limit and focused on what the task still needs.
Agents and tools
Context poisoning
Context poisoning is when a false statement enters an AI model's context and the model keeps treating it as true in later steps.
Agent failure modes
Context rot
Context rot is when a language model makes more mistakes on the same task as its context grows, missing information it uses correctly in a shorter one.
Agent failure modes
Context window
A context window is the maximum number of tokens a language model can work with in one request, including its instructions, inputs and its own reply.
Language models
Convolution
Convolution is the operation of sliding a small set of weights, called a filter, across an input and computing a weighted sum at each position.
Neural networks and deep learning
F
Fine-tuning
Fine-tuning is further training of an already trained model on a smaller, task-specific dataset, so that it adapts to a particular task or domain.
Language models
G
Generative AI
Generative AI is AI that produces new content, such as text, images, audio or code, modeled on patterns in the data it was trained on.
Language models
Guardrails
Guardrails are mechanisms that enforce policies and user intentions outside the access and control surface of an AI model.
Guardrails and control
H
Hallucination
A hallucination is content an AI model generates and presents as fact that is false or is not supported by the information the model was given.
Agent failure modes
Hooks
Hooks are commands an agent tool runs automatically at set points in an agent's work; one that runs before an action, such as a tool call, can block it.
Guardrails and control
Human in the loop
Human in the loop is a design in which a person reviews or approves an AI system's actions before they take effect, or can step in to stop them.
Guardrails and control
K
Knowledge graph
A knowledge graph is a store of entities and the typed relationships between them, so that facts can be queried, traced to their source and updated.
Agents and tools
L
Large language model
A large language model is a neural network trained on large amounts of text to predict the next token, which it uses to generate text.
Language models
LoRA
LoRA (low-rank adaptation) is a fine-tuning method that freezes a model's weights and trains a small low-rank update for selected weight matrices.
Language models
M
Machine learning
Machine learning is a way of building software that learns how to do a task from examples or experience, instead of following rules a programmer wrote.
Machine learning
Memory poisoning
Memory poisoning is when false content, planted by an attacker or made by the agent itself, is saved to an AI agent's memory and later treated as fact.
Agent failure modes
Model Context Protocol
Model Context Protocol (MCP) is an open standard for connecting AI applications to outside tools and data, so one connector works with many AI clients.
Agents and tools
Model risk management
Model risk management is the framework a bank uses to control the risk of losses from decisions based on models that are wrong or misused.
Guardrails and control
P
Perceptron
A perceptron is an early learning machine that sorts inputs into two classes with a weighted sum and a threshold, and changes its weights after mistakes.
Foundations and history
Predictive AI
Predictive AI is AI that estimates an unknown value or a future outcome from patterns in past data, such as the chance that a borrower defaults.
Machine learning
Prompt injection
Prompt injection is an attack in which text an AI model reads carries instructions that override the ones its developer or user gave it.
Agent failure modes
Q
QLoRA
QLoRA is a fine-tuning method that applies LoRA (low-rank adaptation) to a quantized model, one whose frozen weights are stored as 4-bit numbers.
Language models
R
Reasoning model
A reasoning model is a language model trained to write a long chain of thought before answering, using more computation than a direct reply.
Frontier and research ideas
Recursive self-improvement
Recursive self-improvement is a hypothesized process in which an AI system improves its own design, each improvement making the next one easier.
Frontier and research ideas
RLHF
RLHF (reinforcement learning from human feedback) is training in which people compare a model's outputs and the model learns to produce the preferred ones.
Language models
RLVR
RLVR (reinforcement learning with verifiable rewards) is a training method that rewards a model only when an automatic check confirms its answer.
Language models
Rules
Rules are instruction files, such as AGENTS.md, CLAUDE.md or Cursor rules, that an agent tool adds to the model's context, usually when a session starts.
Guardrails and control
T
Token
A token is the unit of text a language model reads and writes: a word, a piece of a word, a single character or a byte.
Language models
Tool use
Tool use is when an AI model requests a call to an outside program, such as a search or a trading API, and software runs it and returns the result.
Agents and tools
Transformer
A transformer is a neural network design that processes a sequence with attention, so each position can draw directly on every other position.
Neural networks and deep learning
Looking for trading terms? The trading glossary in the Mangrove app covers them.