This glossary will be regularly updated.
A
- Active Learning: A machine learning approach where the model selectively queries the user to label new data points, helping to improve learning efficiency with minimal labeled data.
- Agentic: An AI system’s capability to perform actions autonomously, make decisions, and adapt its behavior based on its goals and environment, without direct human intervention.
- AGI (Artificial General Intelligence): A concept that suggests a more advanced version of AI, capable of performing tasks much better than humans while also teaching and advancing its own capabilities.
- AI Alignment Problem: The challenge of ensuring AI systems are aligned with human values and objectives, especially as they grow more autonomous and powerful.
- AI Ethics: Principles aimed at preventing AI from harming humans, including how AI systems should collect data or deal with bias.
- AI Governance: The policies, laws, and frameworks that guide the development and use of AI in a safe and ethical manner.
- AI Safety: An interdisciplinary field focused on the long-term impacts of AI, addressing potential risks of AI progressing to superintelligence hostile to humans.
- Algorithm: A series of instructions allowing a computer to analyze data, recognize patterns, and learn from it to accomplish tasks.
- Anomaly Detection: The process in machine learning and data science of identifying rare items or events that differ from the majority of the data.
- Alignment: Adjusting AI to produce a desired outcome, such as moderating content or maintaining positive interactions toward humans.
- Anthropomorphism: The tendency of humans to attribute human characteristics to non-human entities, such as AI chatbots.
- Artificial Intelligence (AI): The use of technology to simulate human intelligence, aiming to build systems that perform human-like tasks.
- Attention Mechanism: A technique in neural networks that allows models to focus on specific parts of the input data, improving performance in tasks like machine translation and image captioning.
- Autonomous Agents: AI models capable of accomplishing specific tasks autonomously, such as self-driving cars.
B
- Bias: Errors in AI, especially in large language models, that result from biased training data and can lead to the propagation of stereotypes.
- Bayesian Network: A probabilistic graphical model that represents a set of variables and their conditional dependencies via a directed acyclic graph.
C
- Chatbot: A program that simulates human conversation through text.
- ChatGPT: An AI chatbot developed by OpenAI that uses large language model technology.
- Cognitive Computing: Simulation of human thought processes in a computer model, with the primary goal of augmenting human intelligence rather than replacing it.
- Computer Vision: A field of AI focused on enabling machines to interpret and process visual information from the world, such as images and videos.
- Convolutional Neural Network (CNN): A class of deep neural networks, most commonly applied to analyzing visual imagery.
D
- Data Augmentation: Remixing or adding data to diversify the training sets of AI models.
- Data Mining: The practice of examining large datasets to discover patterns and useful information.
- Data Preprocessing: The process of cleaning and organizing raw data before using it in AI models to ensure accuracy and efficiency.
- Deep Learning: A subfield of machine learning that uses artificial neural networks to recognize complex patterns, inspired by the human brain.
- Diffusion: A machine learning method that adds random noise to data, training AI models to recover the original data.
- Dimensionality Reduction: The process of reducing the number of random variables or features in a dataset, often for simplifying the model or improving its performance.
E
- Emergent Behavior: When an AI exhibits unintended capabilities or actions.
- End-to-End Learning (E2E): A deep learning process where a model solves tasks from start to finish without sequential steps.
- Ethical Considerations: Awareness of privacy, data usage, fairness, and safety issues in AI.
- Expert System: A computer system that emulates the decision-making ability of a human expert, often used in specific domains like medical diagnosis or financial planning.
F
- FOOM (Fast Takeoff): The concept that if AGI is developed, it may lead to an uncontrollable rapid advancement, making it too late to protect humanity.
- Feature Engineering: The process of selecting, modifying, or creating new features (attributes) from raw data to improve model performance.
- FM (Foundational Model): A large-scale, pre-trained model that serves as a versatile base for various AI applications by capturing broad knowledge from extensive datasets. These models, such as GPT or BERT, can be fine-tuned for specific tasks, offering a robust starting point for developing specialized AI solutions.
G
- Generative Adversarial Networks (GANs): AI models with two neural networks—a generator and a discriminator—that create and evaluate new content.
- Generative AI: AI that generates content, such as text, video, code, or images, based on training data.
- Google Gemini: An AI chatbot similar to ChatGPT, but connected to the internet and able to pull real-time data.
- Gradient Descent: An optimization algorithm used to minimize the error in machine learning models by adjusting parameters iteratively.
- Guardrails: Policies and restrictions to ensure responsible AI use and prevent the creation of harmful content.
H
- Hallucination: Incorrect AI responses that are confidently delivered, despite being factually wrong.
- Hyperparameter: A parameter whose value is set before the learning process begins in machine learning models and used to control the model’s behavior.
L
- Large Language Model (LLM): AI models trained on large amounts of text data to understand and generate human-like language.
- Latent Space: A representation of data that captures the underlying structure, used in machine learning to make complex relationships easier to work with.
- LSTM (Long Short-Term Memory): A specialized recurrent neural network architecture in machine learning that enables models to learn long-term dependencies in sequential data through memory cells and gating mechanisms.
M
- Machine Learning (ML): A branch of AI that enables computers to improve outcomes through data training, without explicit programming.
- Microsoft Bing: A search engine using ChatGPT-like technology to provide AI-powered search results.
- Multimodal AI: AI that processes multiple types of inputs, such as text, images, video, and speech.
- Model Drift: A phenomenon in machine learning where the model’s accuracy decreases over time due to changing data distributions or relationships.
N
- Natural Language Processing (NLP): A branch of AI enabling computers to understand human language using algorithms, models, and rules.
- Neural Network: A model resembling the human brain, used for recognizing patterns in data.
- Neural Architecture Search (NAS): The process of automating the design of artificial neural networks to optimize performance.
O
- Overfitting: A machine learning error where a model performs too well on training data but fails to generalize to new data.
- Optimization Algorithm: Algorithms used to adjust the parameters of machine learning models to minimize errors and maximize performance.
P
- Paperclips (Paperclip Maximizer): A hypothetical scenario where an AI system, focused solely on producing paperclips, may consume all resources, endangering humanity.
- Parameters: Numerical values that structure LLMs, enabling them to make predictions.
- Prompt: The input or question given to an AI to generate a response.
- Prompt Chaining: Using information from previous interactions to influence future AI responses.
- Pre-trained Model: A machine learning model that has been trained on a large dataset and can be fine-tuned for specific tasks with minimal additional data.
R
- Reinforcement Learning (RL): A machine learning approach where agents learn to make decisions by receiving rewards or punishments for their actions.
- Reinforcement Learning from Human Feedback (RLHF): A machine learning technique where models are trained using reinforcement learning guided by feedback from human evaluators to align outputs with human preferences and values.
- Regularization: Techniques used in machine learning to prevent overfitting by adding constraints or penalties to the model.
- RNN (Recurrent Neural Network): A class of neural networks designed to recognize patterns in sequential data, such as time series or natural language, by utilizing internal states and feedback loops to maintain information across different time steps.
S
- Stochastic Parrot: An analogy illustrating that AI models, despite mimicking human language, do not understand the meaning behind their responses.
- Style Transfer: AI’s ability to transfer the visual style of one image to another, such as adapting a Rembrandt painting in the style of Picasso.
- Supervised Learning: A type of machine learning where a model is trained on labeled data, learning the relationship between inputs and outputs.
- Support Vector Machine (SVM): A supervised machine learning algorithm used for classification and regression tasks, particularly in high-dimensional spaces.
T
- Temperature: A parameter controlling how random or creative an AI model’s output is.
- Text-to-Image Generation: The process of creating images based on textual descriptions.
- Tokens: Small bits of text processed by AI language models to generate responses.
- Training Data: The datasets used to train AI models, such as text, images, or code.
- Transformer Model: A neural network architecture that tracks relationships in data, allowing for contextual understanding of text and images.
- Turing Test: A test developed by Alan Turing to measure if a machine can exhibit human-like intelligence by convincing a human it’s human.
W
- Weak AI (Narrow AI): AI focused on performing specific tasks without the ability to generalize beyond them.
Z
- Zero-Shot Learning: When an AI model completes a task without having seen related training data before.