| AI Agents | Autonomous AI systems that can perceive their environment, make decisions, and take actions to achieve specific goals. Modern AI agents like AutoGPT and Claude can use tools, browse the web, write code, and complete complex multi-step tasks. |
| Agentic AI | AI systems designed to act autonomously with minimal human intervention, capable of planning, reasoning, and executing complex workflows. These systems can break down tasks, use external tools, and adapt their approach based on results. |
| Activation Function | A mathematical function used in neural networks to introduce non-linearity into the model, allowing it to learn more complex patterns in the data. |
| Actor-Critic | A type of reinforcement learning algorithm that combines an actor network that learns to select actions with a critic network that learns to estimate the expected future reward. |
| Adversarial Attacks | A type of security threat to machine learning models, where an attacker intentionally crafts inputs that can fool the model into making incorrect predictions or decisions |
| Adversarial Examples | Inputs to a machine learning model that have been intentionally designed to cause the model to make a mistake, often by adding small perturbations to the input data. |
| Adversarial Training | A defense mechanism against adversarial attacks, where the machine learning model is trained on adversarially generated examples to improve its robustness and generalization. |
| Artificial General Intelligence (AGI) | A hypothetical level of artificial intelligence that can perform any intellectual task that a human can do, with the same level of flexibility, creativity, and adaptability. AGI remains a major goal in AI research. |
| Artificial Intelligence (AI) | The development of computer systems to perform tasks that typically require human intelligence, such as visual perception, speech recognition, decision-making, and language translation. |
| AI as a Service (AIaaS) | A cloud-based service that allows developers to integrate AI capabilities into their applications without having to build the infrastructure themselves. Examples include OpenAI API, AWS Bedrock, and Google Vertex AI. |
| Attention Mechanism | A technique in neural networks that allows models to focus on relevant parts of the input when producing output. Self-attention is the core component of Transformer architectures used in GPT, Claude, and other LLMs. |
| Autoencoder | A type of neural network used in unsupervised learning to learn a compressed representation of the input data, consisting of an encoder and decoder network. |
| Backpropagation | A technique used in training neural networks to adjust the weights of the connections between neurons, based on the error between the predicted and actual outputs. |
| Bias | A systematic error in a machine learning model that results in inaccurate predictions due to an incorrect assumption or limitation in the data, or training data that is not representative of the population. |
| Capsule Networks | A type of neural network architecture that uses groups of neurons, called capsules, to represent objects and their pose in a hierarchical manner, allowing it to handle deformation, occlusion, and other complex transformations. |
| Chain-of-Thought (CoT) | A prompting technique where LLMs are encouraged to break down complex problems into intermediate reasoning steps before arriving at a final answer. This significantly improves performance on math, logic, and reasoning tasks. |
| Chatbot | A computer program that simulates conversation with human users. Modern AI chatbots like ChatGPT, Claude, and Gemini use large language models to provide sophisticated, context-aware responses. |
| Claude | A family of AI assistants developed by Anthropic, known for being helpful, harmless, and honest. Claude models are trained using Constitutional AI and RLHF to be safe and aligned with human values. |
| Constitutional AI | An AI training approach developed by Anthropic where models are trained to follow a set of principles (a 'constitution') to be helpful, harmless, and honest, reducing the need for human feedback on harmful outputs. |
| Context Window | The maximum amount of text (measured in tokens) that a language model can process at once. Modern LLMs like Claude and GPT-4 have context windows ranging from 8K to 200K+ tokens, enabling longer conversations and document analysis. |
| Classification | A technique used in supervised learning to predict a categorical output variable based on input variables. |
| Clustering | A technique used in unsupervised learning to group similar data points together based on their characteristics. |
| Clustering | A type of unsupervised learning where the algorithm groups similar data points together based on their characteristics. |
| Computer Vision | The ability of computers to interpret and understand visual information from the world around them. |
| Convolutional Generative Adversarial Networks (CGANs) | A type of generative model that uses convolutional neural networks as the generator and discriminator networks in a GAN architecture, allowing it to generate realistic images and other structured data. |
| Convolutional Neural Network | A type of neural network that is designed to process and analyze visual data such as images and videos. |
| Convolutional Neural Network (CNN) | A type of neural network that is commonly used for image and video recognition tasks, due to its ability to extract features from local neighborhoods of the input data. |
| Convolutional Neural Network (CNN) | A type of neural network architecture that is specialized for processing images and other structured data, by using convolutional layers to extract local features and pooling layers to reduce dimensionality. |
| Curriculum Learning | A type of machine learning where the training data is presented in a sequence of increasing difficulty, to help the model learn more effectively. |
| Data Augmentation | The process of generating new training data from existing data by applying transformations such as rotation, scaling, and cropping. |
| Data Mining | The process of discovering patterns and relationships in large datasets using machine learning techniques. |
| Data Normalization | The process of transforming data to have a standard scale and range, often used to improve the accuracy of a machine learning model. |
| Data Preprocessing | The process of cleaning and transforming raw data to make it suitable for machine learning algorithms. |
| Decision Tree | A type of supervised learning algorithm that is used for classification and regression tasks. |
| Deep Belief Network (DBN) | A type of neural network architecture that consists of multiple layers of RBMs, with the first layer receiving the input data and the subsequent layers forming a hierarchical representation of the input. |
| Deep Dream | A visualization technique that uses a neural network to enhance and exaggerate the patterns in an image, creating surreal and psychedelic effects. |
| Deep Learning | A subset of machine learning that uses neural networks with many layers to analyze data and improve accuracy over time. |
| Deep Learning | A type of machine learning that involves neural networks with many layers, capable of learning complex patterns in data. |
| Deep Q-Network (DQN) | A type of reinforcement learning that uses a deep neural network to represent the Q-function, and uses experience replay, target networks, and other techniques to stabilize and improve its performance, by reducing the correlation between samples and the target and by preventing overfitting and instability. |
| Deep Reinforcement Learning | A type of machine learning that combines deep neural networks with reinforcement, learning, enabling agents to learn complex behaviors from high-dimensional input spaces. |
| Deep Reinforcement Learning | A type of reinforcement learning that uses deep neural networks to represent the agent's policy or value function, enabling it to learn more complex and high-dimensional tasks. |
| Domain Adaptation | A type of transfer learning where a model is trained on a source domain, and then adapted to a target domain with different characteristics. |
| Ensemble Learning | A technique in machine learning that combines multiple models to improve accuracy and reduce overfitting. |
| Ethical AI | The practice of developing artificial intelligence that is aligned with ethical and moral principles, such as transparency, fairness, and accountability. |
| Expert Systems | An AI system that is designed to replicate the decision-making abilities of a human expert in a particular domain. |
| Explainable AI (XAI) | A field of study focused on developing machine learning models that can provide clear explanations for their predictions or decisions. |
| Explainable AI (XAI) | A research area in artificial intelligence that aims to develop AI systems that can provide explanations or justifications for their decisions and actions, in order to enhance transparency, accountability, and trustworthiness. |
| Fairness | The property of a machine learning model that ensures that the predictions or decisions it makes do not discriminate against particular groups of people or result in unjust outcomes. |
| Feature Engineering | The process of selecting and transforming the most relevant features of a dataset to improve the accuracy of a machine learning model. |
| Federated Learning | A type of machine learning that allows multiple parties to collaboratively train a model without sharing their data, by using local models to compute gradients and aggregating them at a central server. |
| Few-Shot Learning | A type of machine learning where the goal is to learn from a small number of examples, rather than a large dataset. |
| Friendly AI | A concept in artificial intelligence safety that refers to an AI that is aligned with human values and goals, and does not pose a threat to humanity or the environment. |
| Gated Recurrent Unit (GRU) | A type of recurrent, neural network architecture that uses gating mechanisms to selectively update and forget information in the hidden state, making it more efficient than traditional recurrent neural networks. |
| Generative Adversarial Network (GAN) | A type of neural network used in unsupervised learning to generate new data that is similar to the training data. |
| Generative Adversarial Networks (GANs) | A type of neural network architecture that consists of a generator network that learns to generate samples from a target distribution, and a discriminator network that learns to distinguish between the generated samples and real samples from the target distribution. |
| Gradient Boosting | A machine learning technique that combines multiple weak models to,improve accuracy and reduce error. |
| Gradient Descent | A technique used in machine learning to optimize a model by adjusting its parameters to minimize the error. |
| Human-in-the-Loop (HITL) | A machine learning workflow that involves humans in the training or validation process, to provide feedback and ensure the model is performing as desired. |
| Hyperparameter | A parameter in a machine learning model that is set prior to training, such as learning rate or number of hidden layers. |
| Kernel Trick | A technique used in machine learning to transform a dataset into a higher-dimensional space to make it easier to separate and classify. |
| Large language model (LLM) | Computer program for natural language processing that uses deep learning and neural networks. |
| Long Short-Term Memory | A type of recurrent neural network that is designed to remember information over long periods of time. |
| Long Short-Term Memory (LSTM) | A type of recurrent neural network (RNN) that is capable of remembering long-term dependencies. |
| Long Short-Term Memory (LSTM) | A type of recurrent neural network architecture that uses memory cells and gating mechanisms to selectively update and forget information in the hidden state, making it well-suited for tasks involving long-term dependencies. |
| Machine Learning | A subset of AI that allows computers to learn and improve on their own without being explicitly programmed. |
| Machine Learning Operations (MLOps) | The process of managing and automating machine learning workflows, including data management, model training, and deployment. |
| Markov Decision Process (MDP) | A formalism for modeling decision-making problems, where an agent interacts with an environment over time, receiving rewards or punishments based on its actions, and tries to learn a policy that maximizes the expected cumulative reward. |
| Meta-Learning | A type of machine learning where the goal is to learn how to learn, by training models that can adapt quickly to new tasks or environments. |
| Meta-Learning | A type of machine learning that learns to learn, by discovering and adapting to the structure of a set of tasks or environments, in order to improve its generalization and transferability. |
| Model Selection | The process of selecting the best model for a particular problem based on its accuracy and generalization performance. |
| Model-Based Reinforcement Learning | A type of reinforcement learning that learns a model of the environment, such as a dynamics model or a transition function, and uses it to plan and optimize actions, in order to improve sample efficiency and robustness. |
| Model-Free Reinforcement Learning | A type of reinforcement learning that learns to map states to actions directly, without explicitly modeling the environment, by using trial-and-error to explore and exploit the reward signal. |
| Monte Carlo Tree Search (MCTS) | A search algorithm that uses a tree structure to represent the possible actions and states in a game or decision-making problem, and simulates many possible trajectories to estimate the expected reward of each action. |
| Named Entity Recognition (NER) | A type of NLP that involves identifying and extracting named entities such as people, places, and organizations from text. |
| Natural Language Processing | The ability of computers to understand, interpret, and generate human language. |
| Natural Language Processing (NLP) | A field of study that combines computer science and linguistics, focused on developing algorithms to analyze and understand human language. |
| Neural Network | A network of interconnected nodes that are designed to simulate the function of a human brain and can be used for tasks such as image and speech recognition. |
| Neural Network | A type of machine learning model that is inspired by the structure and function of the human brain, consisting of multiple interconnected layers of neurons. |
| Neural Style Transfer | A technique that uses a neural network to transfer the style of one image onto another image, creating a hybrid image that combines the content of one image with the style of another. |
| Neuroevolution | A type of machine learning that uses evolutionary algorithms, such as genetic algorithms or particle swarm optimization, to evolve neural networks that can perform a given task, by selecting and mutating individuals based on their fitness. |
| Off-Policy Reinforcement Learning | A type of reinforcement learning that learns from a different policy than the one used to generate the data, by using importance sampling or other techniques to estimate the value of actions or states under a different policy. |
| On-Policy Reinforcement Learning | A type of reinforcement learning that learns from the same policy that generates the data, by using the trajectory or episode data to update the policy directly, such as with policy gradient or actor-critic methods. |
| One-Shot Learning | A type of machine learning where the goal is to learn from a single example, rather than a large dataset. |
| One-Shot Learning | A type of machine learning that can learn from a single or a few examples of a new class or task, by using transfer learning, generative models, or metric learning, to recognize or synthesize new instances from the limited data. |
| Overfitting | A common problem in machine learning where a model is too complex and fits the training data too well, resulting in poor performance on new data. |
| Partially Observable Markov Decision Process (POMDP) | A variation of the Markov decision process where the agent only partially observes the environment, and must infer the underlying state from the observations. |
| Policy Gradient | A type of reinforcement learning algorithm that learns to directly optimize a policy for selecting actions, by using gradient ascent on the expected reward of the policy. |
| Principal Component Analysis (PCA) | A technique used in unsupervised learning to reduce the dimensionality of a dataset by finding the most important features or components. |
| Privacy-Preserving Machine Learning | The practice of developing machine learning models that preserve the privacy of sensitive data, by using techniques such as differential privacy or secure multiparty computation. |
| Q-Learning | A type of reinforcement learning algorithm that uses a table to store the expected future rewards for each state-action pair, and learns to select actions that maximize the expected future reward. |
| Q-Learning | A type of reinforcement learning that learns a state-action value function, called Q-function, by using a temporal difference algorithm and an exploration strategy, such as epsilon-greedy or softmax, to balance between exploration and exploitation. |
| Recurrent Neural Network | A type of neural network that is designed to process and analyze sequential data such as text and speech. |
| Recurrent Neural Network (RNN) | A type of neural network that is capable of processing sequential data by maintaining an internal state or memory of previous inputs. |
| Regression | A technique used in supervised learning to predict a continuous output variable based on input variables. |
| Reinforcement Learning | A type of machine learning in which an algorithm learns through trial and error by receiving feedback in the form of rewards or punishments. |
| Reinforcement Learning | A type of machine learning that involves an agent learning from trial and error by receiving rewards or penalties based on its actions. |
| Reinforcement Learning | A type of machine learning where an agent learns to make decisions by interacting with an environment, receiving rewards or punishments based on its actions. |
| Reinforcement Learning | A type of machine learning that involves an agent interacting with an environment, in order to learn a policy or a value function that maximizes a reward signal, by using trial-and-error and exploration-exploitation strategies. |
| Restricted, Boltzmann Machine (RBM) | A type of generative model that consists of a bipartite graph of visible and hidden units, with binary stochastic activations, and learns to represent the input data distribution by adjusting the weights between the units. |
| Robotics | The branch of AI that deals with the design and construction of robots that can perform tasks autonomously. |
| Sentiment Analysis | A type of NLP that involves analyzing the emotions and opinions expressed in text or speech. |
| Sim-to-Real Transfer | A research area in robotics that aims to transfer control policies or models trained in simulation to the real world, by adapting to the differences in dynamics, sensors, and perception, in order to reduce the cost and risk of real-world experiments. |
| Singularity | A hypothetical event in the future when artificial intelligence surpasses human intelligence, leading to a dramatic transformation of society and possibly the extinction of humanity. |
| Superintelligence | A hypothetical level of artificial intelligence that surpasses human intelligence in every cognitive task, and has the ability to improve itself recursively, leading to an intelligence explosion and potentially catastrophic consequences. |
| Supervised Classification | A type of supervised learning used for classifying data into pre-defined categories. |
| Supervised Learning | A type of machine learning in which an algorithm is trained on labeled data to make predictions on new data. |
| Transfer Learning | A technique in machine learning that allows a model to reuse knowledge gained from one problem to improve performance on a new, related problem. |
| Transfer Learning | A type of machine learning where a model is first trained on one task, and then fine-tuned on a different but related task. |
| Underfitting | A common problem in machine learning where a model is too simple and fails to capture the complexity of the data, resulting in poor performance on both training and new data. |
| Unfriendly AI | A concept in artificial intelligence safety that refers to an AI that is misaligned with human values and goals, and could pose a threat to humanity or the environment if it gains too much power or autonomy. |
| Unsupervised Learning | A type of machine learning in which an algorithm is trained on unlabeled data to discover patterns and relationships on its own. |
| Unsupervised Learning | A type of machine learning where the algorithm learns patterns and relationships in the data without explicit supervision or labels. |
| Value Alignment | A research area in artificial intelligence safety that aims to ensure that advanced AI systems are aligned with human values and goals, and do not cause unintended harm or adverse consequences. |
| Variance | The amount by which a machine learning model's predictions vary for different training datasets. |
| Variational Autoencoder (VAE) | A type of neural network architecture that consists of an encoder network that learns to encode input data into a low-dimensional representation, and a decoder network that learns to reconstruct the input data from the representation, while imposing a regularization constraint on the representation to encourage diversity. |
| Zero-Shot Learning | A type of machine learning that can generalize to unseen classes or tasks, by using prior knowledge, such as attributes or semantic embeddings, to infer the properties of the new classes or tasks, without requiring explicit training examples. |
| Diffusion Models | A class of generative AI models that learn to create data by reversing a gradual noising process. Used in image generators like Stable Diffusion, DALL-E, and Midjourney to create high-quality images from text prompts. |
| Embeddings | Dense vector representations of data (text, images, etc.) that capture semantic meaning. Words or concepts with similar meanings have similar embeddings, enabling semantic search and RAG applications. |
| Few-Shot Learning | A machine learning approach where models learn from a very small number of examples, often by including examples directly in the prompt. LLMs excel at few-shot learning without requiring retraining. |
| Fine-Tuning | The process of taking a pre-trained model and further training it on a specific dataset to adapt it for particular tasks. Common with LLMs to customize behavior for specific use cases. |
| Foundation Model | A large AI model trained on broad data that can be adapted to a wide range of downstream tasks. Examples include GPT-4, Claude, Llama, and BERT. These models serve as the 'foundation' for many AI applications. |
| GPT (Generative Pre-trained Transformer) | A family of large language models developed by OpenAI that use the Transformer architecture. GPT-4 and GPT-4o are among the most capable LLMs, powering ChatGPT and many AI applications. |
| Grounding | The process of connecting AI model outputs to real-world data, facts, or sources to improve accuracy and reduce hallucinations. Often implemented through RAG or tool use. |
| Hallucination | When an AI model generates plausible-sounding but factually incorrect or nonsensical information. A key challenge in LLM deployment that can be mitigated through grounding and RAG techniques. |
| Inference | The process of using a trained AI model to make predictions or generate outputs. In the context of LLMs, inference refers to generating text responses from prompts. |
| Jailbreaking | Attempts to bypass an AI model's safety guidelines and content policies through carefully crafted prompts. AI developers work to prevent jailbreaks while maintaining model usefulness. |
| Llama | A family of open-weight large language models developed by Meta AI. Llama models (including Llama 2 and Llama 3) can be run locally and fine-tuned, making them popular for research and self-hosted applications. |
| Large Language Model (LLM) | AI models trained on vast amounts of text data that can understand and generate human-like text. LLMs like GPT-4, Claude, Gemini, and Llama power chatbots, coding assistants, and many other AI applications. |
| LoRA (Low-Rank Adaptation) | An efficient fine-tuning technique that trains only a small number of additional parameters while keeping the base model frozen. Dramatically reduces computational requirements for customizing LLMs. |
| Model Context Protocol (MCP) | An open protocol developed by Anthropic that enables AI assistants to connect with external data sources, tools, and services in a standardized way, allowing for more capable and integrated AI applications. |
| Mixture of Experts (MoE) | A neural network architecture where multiple specialized sub-networks (experts) handle different parts of the input, with a gating network selecting which experts to use. Enables larger, more efficient models like Mixtral. |
| Multimodal AI | AI systems that can process and generate multiple types of data, such as text, images, audio, and video. GPT-4V, Claude 3, and Gemini are examples of multimodal models. |
| Ollama | An open-source tool for running large language models locally on personal computers. Ollama simplifies downloading and running models like Llama, Mistral, and others without cloud dependencies. |
| Open Source AI / Open Weights | AI models whose weights and often training code are publicly available for download and use. Examples include Llama, Mistral, and Stable Diffusion. 'Open weights' specifically refers to models where weights are available but training data may not be. |
| Prompt Engineering | The practice of crafting effective prompts to get desired outputs from AI models. Techniques include providing context, examples (few-shot), chain-of-thought reasoning, and structured formatting. |
| Quantization | A technique to reduce model size and memory requirements by using lower-precision numbers (e.g., 4-bit or 8-bit instead of 16-bit). Enables running large models on consumer hardware with minimal quality loss. |
| RAG (Retrieval-Augmented Generation) | A technique that enhances LLM responses by retrieving relevant information from external knowledge bases before generating answers. Reduces hallucinations and enables up-to-date, source-grounded responses. |
| RLHF (Reinforcement Learning from Human Feedback) | A training technique where AI models learn from human preferences rather than just predicting text. Humans rate model outputs, and the model is fine-tuned to produce higher-rated responses. Key to making LLMs helpful and safe. |
| Stable Diffusion | An open-source text-to-image AI model that generates images from text descriptions. Can be run locally and customized with LoRA adapters, making it popular for creative applications. |
| System Prompt | Instructions provided to an AI model that define its behavior, persona, and constraints for an entire conversation. System prompts shape how the model responds to user queries. |
| Temperature | A parameter that controls the randomness of AI model outputs. Lower temperatures (e.g., 0.1) produce more deterministic, focused responses; higher temperatures (e.g., 1.0) produce more creative, varied outputs. |
| Tokens | The basic units that language models process. A token can be a word, part of a word, or punctuation. English text averages about 1.3 tokens per word. Token counts affect costs and context window limits. |
| Tool Use / Function Calling | The ability of AI models to invoke external functions, APIs, or tools to complete tasks. Enables LLMs to perform actions like web searches, code execution, database queries, and more. |
| Transformer | A neural network architecture introduced in 2017 that uses self-attention mechanisms to process sequences. The foundation of modern LLMs including GPT, Claude, Llama, and BERT. |
| Vector Database | A database optimized for storing and querying high-dimensional vectors (embeddings). Essential for RAG systems and semantic search. Examples include Pinecone, Weaviate, Chroma, and pgvector. |