Glossary of AI Terms and Definitions

Latest Technology Awareness Date: June 2024
(Incorporates concepts from GPT-4, GPT-4-turbo, GPT-4o, LLaMA 3, Gemini, Claude, Mistral, and more as of mid-2024.)

---

A. Core AI Concepts

Artificial Intelligence (AI)

The field of computer science focused on creating systems capable of tasks that typically require human intelligence (e.g. perception, reasoning, learning, planning, natural language understanding).

Machine Learning (ML)

A subset of AI where computers learn from data without explicit programming; algorithms improve performance on a task over time as they are exposed to more data.

Deep Learning

A subfield of ML that uses neural networks with many layers (“deep” architectures) to learn hierarchical representations of data, powering applications like image recognition and language modeling.

Neural Network

A computational model inspired by biological neurons; consists of interconnected nodes (neurons) that transform input data through weighted connections to produce output.

Supervised Learning

An ML approach where the model learns from labeled data (input-output pairs) to predict outputs for new, unseen inputs.

Unsupervised Learning

Learning patterns or structures from data without explicit labels, such as clustering or dimensionality reduction.

Reinforcement Learning (RL)

An ML paradigm where an agent learns to take actions in an environment to maximize cumulative reward through trial and error.

Transfer Learning

Applying knowledge gained from one task or dataset to improve learning on a related but different task.

Generative AI

Systems that can produce new content—text, images, audio, or code—by learning the underlying distribution of the data.

Foundation Model

A very large, general-purpose model trained on broad data (e.g. GPT, PaLM, LLaMA) that can be adapted to many downstream tasks.

---

B. NLP & Language Models

Natural Language Processing (NLP)

The branch of AI focused on enabling computers to understand, interpret, and generate human language.

Large Language Model (LLM)

A neural network trained on vast text data to predict and generate human-like text, often with billions or trillions of parameters.

Transformer

A neural network architecture introduced in 2017 (Vaswani et al.) that uses self-attention mechanisms to model dependencies in sequential data, enabling scalable training of LLMs.

Prompt

An input given to an AI model (especially an LLM) to steer its response toward a desired output.

Prompt Engineering

Crafting prompts strategically to guide AI model outputs toward specific tasks or styles.

Chain-of-Thought (CoT)

A technique where the model generates intermediate reasoning steps to improve complex problem-solving.

Instruction Tuning

Fine-tuning a model on datasets where instructions and expected outputs are paired, improving its ability to follow human directions.

Few-Shot Learning

The ability of a model to perform new tasks with very few examples provided at inference time.

Zero-Shot Learning

The ability of a model to perform tasks it was not explicitly trained for, relying on general knowledge encoded during training.

Context Window

The amount of input text (tokens) a model can “see” at once; recent models have context windows of tens or hundreds of thousands of tokens.

---

C. Technical Terms

Parameter

A learned weight in a neural network; LLMs often have billions or trillions of parameters.

Embedding

A dense vector representation of data (words, sentences, images) capturing semantic or contextual meaning.

Attention Mechanism

A method in neural networks that dynamically focuses on different parts of the input when producing output, core to the transformer architecture.

Loss Function

A metric optimized during training to measure the difference between predicted and actual outputs.

Backpropagation

The algorithm for computing gradients of the loss with respect to model parameters, enabling learning through gradient descent.

Gradient Descent

An optimization technique that iteratively adjusts parameters to minimize the loss function.

Overfitting

When a model learns the training data too well, including noise, and fails to generalize to new data.

Regularization

Techniques to prevent overfitting by penalizing model complexity.

Fine-Tuning

Continuing the training of a pretrained model on a new, often smaller, dataset to specialize it for a particular task.

---

D. Deployment & Safety

Inference

The process of using a trained model to make predictions on new data.

Serving

Deploying a model so it can respond to user requests in real-time.

Latency

The delay between a user request and the model’s response.

Alignment

Efforts to ensure AI systems behave in ways consistent with human values and intentions.

Hallucination

When an AI model generates text that is syntactically correct but factually false or misleading.

Red Teaming

Systematic testing of AI systems to discover vulnerabilities or harmful outputs.

Guardrails

Techniques and policies to constrain AI behavior to safe and appropriate outputs.

RLHF (Reinforcement Learning from Human Feedback)

Training technique where human preferences are used to shape a model’s outputs via reward signals.

---

E. Emerging & Industry Terms

MoE (Mixture of Experts)

A neural network design where only a subset of the model’s parameters (experts) are active per input, increasing efficiency and scalability.

Multimodal Model

An AI model that can process and generate multiple types of data, such as text, images, audio, and video.

Vision-Language Model (VLM)

A model that jointly understands images and text, such as CLIP or GPT-4 with vision.

Retrieval-Augmented Generation (RAG)

Combining retrieval systems with generative models to improve factual accuracy by grounding responses in external documents.

Synthetic Data

Artificially generated data used to augment training datasets.

Open-Weight Model

A model whose weights are publicly released (e.g. Meta’s LLaMA 2/3).

Closed-Weight Model

A model whose weights remain proprietary (e.g. OpenAI’s GPT-4).

---

📅 Technology Awareness Date

Definitions current as of June 2024.
Includes trends and terminology reflecting the state of the field up through mid-2024.