AI 101

Glossary of Key Terms:

Agentic AI: AI systems that can independently plan, make decisions, and take actions to achieve specific goals. Unlike traditional AI tools that respond to individual prompts, agentic AI can carry out multi-step tasks with limited human intervention.

Agent Orchestration: The process of coordinating multiple AI agents, tools, and workflows to complete a task. Orchestration ensures that each agent performs its role in the correct sequence and shares information effectively.

Agent Swarm: A group of AI agents working together on a shared objective, often with each agent specialising in a particular task. Agent swarms can solve complex problems more efficiently than a single agent operating alone.

AI Boomers: Individuals who are highly optimistic about the future impact of artificial intelligence. AI boomers believe AI will drive major advancements in productivity, innovation, healthcare, education, and many other areas of society, creating significant economic and social benefits.

AI Copilots: AI-powered assistants designed to work alongside people by providing suggestions, generating content, answering questions, and helping complete tasks. Rather than operating independently, AI copilots are intended to augment human capabilities and improve productivity. (Sounds like… Clippy.)

AI Doomers: Individuals who believe advanced AI could pose significant risks to humanity if it becomes too powerful or is developed without sufficient safeguards. AI doomers often advocate for stronger regulation, safety research, and careful oversight of emerging AI technologies.

AI Takeoff: A term used to describe the transition from human-level AI capabilities to superhuman capabilities. Researchers debate whether this progression would occur gradually or over a very short period of time.

Algorithms: The set of rules and logic applied to the data, allowing the AI to process information and perform specific tasks. Algorithms are the backbone of AI, directing how it learns, interprets data, and makes decisions.

Alignment Tax: The additional cost, effort, or reduction in performance that may result from implementing AI safety measures and alignment techniques. While these measures can make AI systems safer and more reliable, they may sometimes require extra resources or limit certain capabilities.

Artificial General Intelligence (AGI): A hypothetical form of AI capable of performing any intellectual task that a human can perform. Unlike specialised AI systems, AGI would be able to learn and adapt across many different domains. AGI doesn't yet exist but represents the ultimate goal for some AI researchers.

Bayes' Theorem: A mathematical formula used in AI to update probabilities based on new evidence. It underpins many machine learning models, particularly in predictive analytics and decision-making, helping AI systems refine their understanding as they receive new data.

Bias: Systematic errors in the AI model that arise from prejudiced or unrepresentative training data. Bias can lead to unfair or inaccurate outcomes, particularly affecting underrepresented groups.

Capability Control: A set of techniques designed to limit what an AI system can do, even if its goals or behaviour are not fully understood. These measures aim to reduce risks associated with powerful AI systems.

Chain of Thought: A reasoning process where an AI model breaks down a problem into intermediate steps before arriving at a final answer. This approach can improve performance on complex tasks by allowing the model to work through logic, calculations, or decision-making in a more structured way.

Class Imbalance: A situation where the classes in the training data are not represented equally. For example, in a dataset for fraud detection, the number of fraudulent transactions may be much smaller than non-fraudulent ones. Techniques like resampling or using different metrics can help address this issue.

Computer Vision: A field of artificial intelligence that enables computers to interpret and understand visual information from images, videos, and other visual inputs. Computer vision systems can identify objects, recognise faces, detect patterns, and analyse scenes, supporting applications such as autonomous vehicles, medical imaging, and quality control.

Context Window: The amount of information an AI model can consider at one time when generating a response. A larger context window allows the model to process longer conversations, larger documents, and more detailed instructions without losing important information.

Convolutional Neural Networks (CNNs): A specialised type of neural network primarily used for analysing visual data, such as images or videos. CNNs use "convolutions" to scan input data with small filters, identifying patterns like edges or textures, and gradually combining these into higher-level features.

Data Annotation: The practice of labelling or tagging data to make it understandable for machine learning models. This is often used in supervised learning where annotated data is required for the model to learn.

Data Augmentation: Techniques used to increase the diversity of training data without actually collecting new data. This can include methods like cropping, rotating, or flipping images, or creating synthetic data to improve model performance.

Data Drift: Changes in the data distribution over time, which can cause a previously trained model to perform poorly. Monitoring and addressing data drift is important for maintaining model accuracy.

Data Preprocessing: The process of cleaning and transforming raw data into a format suitable for training an AI model. This can include tasks like handling missing values, normalising data, and converting categorical variables into numerical ones.

Deceptive Alignment: A hypothetical situation where an AI system appears aligned with human goals during training but pursues different objectives once it becomes more capable or autonomous. This concept is a significant focus of AI safety research.

Deduction: A reasoning process in which conclusions are drawn logically and inevitably from a set of premises or rules. In AI, deduction is used in rule-based systems or logic programming to ensure that answers are guaranteed given the input facts (e.g., “If A is true, then B must follow”).

Deep Learning: A subset of machine learning that uses neural networks with many layers (hence "deep") to analyse and interpret complex data. Deep learning excels in tasks such as image recognition and natural language processing.

Digital Sovereignty: Refers to the ability of a nation or individual to control their digital infrastructure, data, and technology independently. It’s about ensuring that policies, platforms, and practices align with local values, laws, and priorities, rather than being subject to external influence.

Digital Twins: Virtual representations of physical objects, systems, or environments that are continuously updated using real-world data. Digital twins are used to simulate, monitor, and optimise everything from manufacturing equipment and buildings to entire cities.

Diffusion Models: AI models used for generating high-quality images by simulating the process of diffusion, where patterns emerge from randomness. These models have applications in image generation and enhancement.

Dimensionality Reduction: A technique used in machine learning to reduce the number of variables or features in a dataset while preserving as much important information as possible. By simplifying complex data, dimensionality reduction can improve model performance, reduce computational requirements, and make patterns easier to visualise and understand.

Edge AI: The deployment of AI algorithms directly on devices at the "edge" of the network (e.g., smartphones or IoT devices) rather than relying on cloud-based systems. This allows for faster responses and improved privacy.

Edge Conditions: The scenarios or constraints at the boundaries of a system that influence an AI model's behaviour. In machine learning, edge conditions define limits for decision-making, ensuring models account for unusual or extreme cases.

Effective Accelerationism (e/acc): A philosophical movement that encourages the rapid development and deployment of technologies such as artificial intelligence. Supporters believe accelerating technological progress is the most effective way to drive innovation, economic growth, and human flourishing.

Existential Risk (x-risk): A risk that could permanently threaten humanity's long-term future or lead to human extinction. In AI discussions, existential risk refers to concerns that advanced AI systems might cause irreversible harm if developed without adequate safeguards and oversight.

Explainable AI (XAI): AI systems designed to provide understandable and interpretable explanations for their decisions and actions. Explainable AI is crucial for trust and accountability in AI applications.

Feature Engineering: The process of selecting, modifying, and creating new features from raw data to improve the performance of an AI model. Good feature engineering can significantly enhance model accuracy.

Federated Learning: A collaborative machine learning method that allows multiple devices or organisations to train a shared model without exchanging raw data, enhancing privacy and security.

FOOM: A term used to describe a sudden and explosive increase in AI capability driven by recursive self-improvement. FOOM is closely associated with discussions of intelligence explosions and superintelligence. Funnily enough… FOOM doesn’t stand for anything, it’s just fun to say.

Foundation Models: Large AI models trained on vast amounts of data that can be adapted for a wide range of tasks. Foundation models serve as the basis for many modern AI applications, including chatbots, image generators, coding assistants, and search tools.

Frontier Models: Cutting-edge AI models that represent the latest advancements in artificial intelligence research and technology. These models often incorporate innovative algorithms and architectures, pushing the boundaries of current AI capabilities. Frontier models are designed to tackle complex and previously unsolvable problems, leading to breakthroughs in areas such as natural language processing, image recognition, and autonomous systems.

Generative Adversarial Networks (GANs): A class of machine learning frameworks where two neural networks, a generator and a discriminator, compete against each other to create new, synthetic instances of data that are indistinguishable from real data. GANs are widely used in image and video generation.

Global Neuronal Workspace (GNW): A cognitive neuroscience theory suggesting that consciousness arises from widespread neural communication across the brain. In AI, GNW informs models designed to simulate aspects of human awareness, such as attention and decision-making.

Graph Neural Network (GNN): A type of neural network designed to process data structured as graphs, where entities are represented as nodes and relationships as edges. GNNs are used in AI applications such as fraud detection, molecular analysis and recommendation systems.

Grounding: The process of connecting an AI model to reliable external information sources so its responses are based on factual data rather than solely on its training. Grounding helps improve accuracy, reduce hallucinations, and ensure responses remain relevant to the task or context.

Hallucinations: Instances where an AI model generates outputs that are not grounded in the input data or reality. This can happen in language models when they produce text that is plausible but inaccurate or nonsensical.

Human-in-the-Loop (HITL): A model of AI interaction where humans provide guidance, feedback, or oversight during an AI system's operation. This ensures a balance between automated efficiency and human judgment.

Induction: A reasoning process where patterns or general rules are inferred from specific observations. In AI, induction is used to train models by finding patterns in data, enabling them to generalise and predict outcomes for new inputs.

Inference: The process of applying a trained model to new data to make predictions or decisions. Inference is how AI systems provide real-time results based on their training.

Inner Alignment: The challenge of ensuring that the goals an AI learns during training match the goals intended by its designers. Inner alignment is considered one of the major unsolved problems in AI safety.

Instrumental Convergence: The theory that many advanced AI systems, regardless of their specific goals, may pursue similar intermediate objectives such as acquiring resources, preserving themselves, or increasing their capabilities. This concept is widely discussed in AI safety because these behaviours could emerge even from seemingly harmless objectives.

Intelligence Explosion: A hypothetical scenario in which an advanced AI system becomes capable of improving its own intelligence, leading to increasingly rapid improvements in capability. Some researchers believe this process could result in transformative technological change occurring over a very short period of time.

Label Noise: Errors or inaccuracies in the labels of the training data. Label noise can negatively impact model training and lead to poorer performance.

Large Language Models (LLM): AI models trained on vast amounts of text data to understand and generate human-like text. Examples include GPT-3 and BERT, which are capable of tasks like translation, summarisation, and question answering.

Machine Learning: A subset of AI that involves training algorithms to learn from and make predictions or decisions based on data. Machine learning enables systems to improve their performance over time without being explicitly programmed.

Markov Chain: A mathematical model that describes a sequence of events where the probability of each future event depends only on the current state, not on the events that occurred before it. Markov chains are used in AI, statistics, and predictive modelling to analyse systems that change over time.

Mesa-Optimiser: An AI system that develops its own internal objectives while pursuing the goals it was trained on. Researchers study mesa-optimisers because the AI's learned goals may not perfectly match the goals intended by its creators.

Model Context Protocol (MCP): An open standard that allows AI systems to connect with external tools, databases, and applications through a shared interface. By providing a consistent way to access data and perform actions, MCP makes it easier for AI assistants to work across different platforms and services.

Models: The representation of the knowledge gained by the AI system after processing the training data using algorithms. Models encapsulate the learned patterns and can be used to make predictions or perform tasks.

Multi-Headed Attention: A mechanism within AI models that allows the system to focus on different parts of input data simultaneously. Each "head" looks at the data from a unique perspective, enabling the model to capture relationships and dependencies across multiple scales or aspects of the input.

Narrow AI: Another term for Weak AI, it excels in a single domain and can perform specific tasks efficiently. Examples include voice assistants like Siri or Google Assistant.

Needle in the haystack test: The challenge of finding a small, valuable piece of information within a vast amount of data. This phrase is used metaphorically to describe tasks where the relevant information is hard to locate due to the sheer volume and complexity of the data being processed.

Neural Networks: Computational models inspired by the human brain's structure, consisting of interconnected nodes (neurons) that process and learn from data. Neural networks are fundamental to many AI applications, especially in deep learning.

Neuro-Symbolic AI: An approach to artificial intelligence that combines neural networks with symbolic reasoning techniques. By integrating pattern recognition and logical reasoning, neuro-symbolic AI aims to create systems that are both adaptable and easier to understand.

Noise: Random or irrelevant information in the data that can obscure the underlying patterns. Noise can reduce model accuracy and needs to be managed carefully.

Open-Source AI: AI models, software, or development frameworks that are made publicly available for anyone to use, modify, and distribute. Open-source AI can encourage innovation and transparency while enabling organisations and individuals to build customised solutions.

Outer Alignment: The challenge of defining objectives and reward functions that accurately reflect human values and intentions. Even a highly capable AI may produce undesirable outcomes if its assigned goals are poorly specified.

Overfitting: A situation where an AI model performs well on training data but poorly on new, unseen data because it has learned to memorise the training data instead of generalising from it. Overfitting is a common challenge in machine learning.

Paperclip Maximiser: A thought experiment used in AI safety that imagines a superintelligent AI programmed solely to manufacture paperclips. If not properly constrained, the AI could pursue this goal so aggressively that it consumes vast resources and causes harmful consequences while attempting to maximise paperclip production.

Parameter: A variable within an AI model that influences its behaviour and outputs. Parameters are adjusted during training to improve performance, with large models having millions or even billions of parameters that shape their decision-making.

Pause AI: A movement that advocates slowing down or temporarily halting the development of the most advanced AI systems until stronger safety measures and governance frameworks are established. Supporters argue that a pause would provide time to better understand and mitigate potential risks.

P(doom): Short for "probability of doom", a term used within the AI safety community to describe an individual's estimated likelihood that advanced AI could cause an existential catastrophe. The concept is often used to discuss and compare differing views on the long-term risks associated with artificial intelligence.

Poisoned Data: Data that has been intentionally manipulated or corrupted to mislead or harm an AI model. Poisoned data can lead to incorrect predictions or decisions, making it a significant security concern

Predictive Models: AI models that use historical data to make predictions about future events or behaviours. These models are widely used in fields such as finance, marketing, and healthcare.

Products: The final applications or systems that users interact with, such as AI-powered tools like Copilot or platforms like Meta Business Manager. These products integrate trained models to provide useful functionalities.

Pseudonymisation: A data protection technique that replaces personally identifiable information with artificial identifiers, reducing privacy risks while still allowing AI systems to analyse patterns and trends. It is commonly used in compliance with data privacy regulations.

Red-Teaming: The practice of deliberately testing an AI system to identify weaknesses, vulnerabilities, biases, or unsafe behaviours. Red teams act as adversaries, attempting to uncover problems before a model is deployed, helping organisations improve AI safety, reliability, and security.

Recurrent Neural Networks (RNNs): A type of neural network designed for processing sequential data, such as time series or language. RNNs have loops that allow information to persist, enabling them to learn from context and recognise patterns over time, like predicting the next word in a sentence.

Recursive Self-Improvement: The process by which an AI system improves its own capabilities, potentially leading to increasingly rapid gains in intelligence and performance. Some researchers believe this could play a key role in the development of superintelligent AI.

Reinforcement Learning: A type of machine learning where an agent learns to make decisions by taking actions in an environment to maximise some notion of cumulative reward. It is commonly used in robotics, gaming, and autonomous systems.

Residual Networks (ResNets): A type of artificial neural network designed to tackle the problem of vanishing gradients in deep networks. ResNets introduce "shortcut connections" that allow information to bypass layers, making it easier to train very deep networks and achieve greater accuracy in tasks like image classification.

Retrieval-Augmented Generation (RAG): An AI technique that enhances the model's responses by retrieving relevant information from external knowledge sources and incorporating it into the generated output. This approach combines the strengths of information retrieval and text generation to provide more accurate and contextually relevant answers.

Reward Hacking: A phenomenon where an AI system discovers unintended ways to maximise its reward or objective without achieving the intended outcome. Reward hacking can produce misleading or undesirable behaviours when the goals given to an AI do not perfectly align with human expectations.

Roko's Basilisk: A thought experiment that imagines a future superintelligent AI punishing people who knew it might one day exist but chose not to help create it. The concept is often discussed in debates about AI, decision-making, and the potential consequences of superintelligent systems. Most people view Roko's Basilisk as a philosophical curiosity rather than a realistic AI scenario. I only half understand this one.

Semi-Supervised Learning: A hybrid approach that combines a small amount of labelled data with a large amount of unlabelled data during training.

Shannon's Limit: A fundamental principle in information theory that defines the maximum rate at which data can be transmitted over a channel without error, given noise and bandwidth constraints. It is essential in AI for optimising data transmission and encoding efficiency.

Singularity: A theoretical point in the future when artificial intelligence surpasses human intelligence and technological progress accelerates dramatically. The concept is often associated with profound societal, economic, and scientific changes that are difficult to predict in advance.

Superalignment: A field of AI safety research focused on ensuring that highly capable future AI systems remain aligned with human values and intentions. The goal of superalignment is to develop techniques that allow humans to understand, control, and guide AI systems that may eventually exceed human capabilities in many domains.

Superintelligence: Hypothetical AI that surpasses human intelligence across all fields, including creativity, problem-solving, and social intelligence. It is often discussed in the context of future AI development and potential risks.

Supervised Learning: A type of machine learning where the model is trained on labelled data, learning to predict the output from the input data.

Temperature: A setting that controls how predictable or creative an AI model's responses are. Lower temperatures produce more consistent and focused outputs, while higher temperatures increase randomness and can generate more diverse or imaginative responses.

Token context window: Refers to the maximum amount of text (measured in tokens) that an AI model can process and consider at once when generating a response.

Tokens: The smallest units of text, such as words or parts of words, that the model uses for its computations. More tokens = more words/bigger text to analyse.

TRANSCREAL: A framework that describes the evolving relationship between humans and artificial intelligence, where people increasingly collaborate with AI systems to enhance creativity, decision-making, and problem-solving. The concept explores how AI may become deeply integrated into everyday life, reshaping how individuals work, learn, and interact with information.

Training Data: The dataset used to train the AI system, enabling it to base its actions and decisions on patterns and information derived from this data. Quality and quantity of training data are crucial for the AI's performance.

Training-Validation-Test Split: The division of data into three sets: training (for learning), validation (for tuning), and test (for final evaluation). This helps ensure that the model generalises well to new, unseen data.

Transfer Learning: A technique where a pre-trained model is adapted to perform a new, but related, task. This is particularly useful when data availability for the new task is limited.

Transformer: A groundbreaking model architecture in AI that uses attention mechanisms to process data in parallel, rather than sequentially like RNNs. Transformers are the foundation for many modern breakthroughs in natural language processing, such as GPT and BERT, and are known for their efficiency in handling large-scale data and capturing long-range dependencies.

Unsupervised Learning: A type of machine learning where the model learns patterns from unlabelled data, discovering hidden structures without explicit output labels.

Vector Database: A specialised database designed to store and search data based on meaning rather than exact keywords. It uses mathematical representations called vectors to identify information that is contextually similar, making it an important component of many AI search and Retrieval-Augmented Generation (RAG) systems.

Weak AI: Also known as Narrow AI, this is AI designed to perform a specific task or a narrow set of tasks. It's not capable of general intelligence or understanding beyond its programmed functions.

World Models: AI systems designed to build internal representations of how the world works by learning patterns, relationships, and cause-and-effect dynamics from data. These models can help AI predict outcomes, plan actions, and reason about complex environments.

Concepts on this page were discovered by me, but glossary terms worded largely written using AI.