Artificial Intelligence Lexicon

Artificial Intelligence (AI)

Artificial Intelligence encompasses the overarching field concerned with the theory and development of computational systems capable of performing tasks traditionally requiring human intelligence. These tasks include, but are not limited to, visual perception, natural language understanding and generation, speech recognition, decision-making, and language translation. AI research integrates diverse methodologies to enable machines to exhibit intelligent behavior across a wide range of domains.

Machine Learning (ML)

Machine Learning represents a subfield of AI that focuses on the development of algorithms capable of learning from data, rather than relying on explicitly programmed instructions for each specific task. ML systems infer patterns and relationships from empirical data and adjust their internal parameters accordingly, thereby improving their performance over time. Common applications include classification, regression, and clustering, where models generalize from examples to make predictions or uncover structure in new data.

Deep Learning (DL)

Deep Learning is a specialized branch within machine learning that employs artificial neural networks with multiple hierarchical layers, commonly referred to as deep neural networks. These architectures are designed to automatically learn increasingly abstract and complex feature representations from data, enabling increased performance in tasks such as image recognition, natural language processing, and speech synthesis. Deep learning models typically require massive datasets and substantial computational resources due to their depth and complexity.

Relationship Between AI, ML, and DL

Artificial Intelligence serves as the broad conceptual framework under which Machine Learning is situated. ML constitutes the data-driven approach to achieving AI, where systems learn to perform tasks through pattern recognition rather than explicit programming. Deep Learning, in turn, is a subset of ML that utilizes multilayered neural network architectures to model high-level abstractions in data. While all deep learning is machine learning, and all machine learning is a form of AI, the reverse is not necessarily true.

Artificial General Intelligence (AGI)

A theoretical class of artificial intelligence systems that possess the capacity to understand, learn, and apply knowledge across a wide array of tasks and domains, exhibiting cognitive flexibility comparable to that of a human being. Unlike narrow AI systems, which are designed to perform specific tasks within constrained domains, AGI would demonstrate generalized problem-solving capabilities, including reasoning, abstraction, transfer learning, and adaptability to novel and unforeseen situations. Achieving AGI would represent a major milestone in AI research, signifying the development of machines capable of autonomous intellectual performance across diverse contexts without task-specific programming.

Large Language Model (LLM)

A class of artificial intelligence systems designed to process and generate human language by leveraging deep learning architectures, typically based on transformer neural networks. These models are characterized by their extensive parameterization and their training on vast collection of textual data, enabling them to learn complex linguistic patterns, semantic relationships, and contextual dependencies. The "large" designation refers both to the scale of the model training dataset and the complexity of the model architecture. This enables them to perform a wide range of natural language processing tasks, including text generation, translation, summarization, and question answering, with high degrees of fluency and coherence.

Token

In the context of large language models, a token is a fundamental unit of text used during the model’s input and output processing. Tokens may represent complete words, subword units, or even individual characters, depending on the tokenization strategy employed. These discrete units serve as the model’s basic building blocks for understanding and generating language, allowing it to operate on variable-length sequences of text in a structured manner.

Token Sequence

An ordered list of discrete units (tokens) derived from raw text data during the tokenization process. Tokens may represent characters, subword units, or whole words, depending on the tokenizer's design. In language models, token sequences serve as the fundamental input and output structure for training and inference, allowing the model to learn and generate text in a structured, sequential manner. The sequence order is critical for preserving syntactic and semantic dependencies.

Tokenization

The computational process of segmenting continuous text into smaller units known as tokens. This transformation converts unstructured natural language input into a structured format that can be ingested and interpreted by language models. Tokenization plays a critical role in enabling LLMs to handle complex linguistic structures by standardizing input into a consistent, model-compatible representation. Techniques vary by model and may involve rule-based, statistical, or learned approaches to optimize performance across diverse languages and contexts.

Inference

The process by which a trained machine learning model is applied to new, previously unseen data in order to generate predictions, classifications, or other outputs. This stage follows the model's training phase and represents its practical deployment, wherein the learned parameters are used to generalize from past examples to novel inputs. In the context of large language models and other AI systems, inference involves leveraging the model's internal representations and learned knowledge to perform tasks such as text generation, language understanding, decision-making, or pattern recognition in real-world applications.

Pre-training

The foundational stage in the development of large language models (LLMs), during which the model is exposed to a vast collection of unlabeled textual data. The objective of this phase is to enable the model to acquire general-purpose linguistic competence, including syntactic structure, semantic relationships, and discourse-level coherence. Learning during pre-training is typically unsupervised or self-supervised, with common training objectives such as masked language modeling (MLM) or autoregressive next-token prediction.

The pre-training collection is often drawn from large-scale internet text, encompassing a wide range of domains, genres, and styles. While this diversity contributes to the model’s broad generalization capabilities, it also introduces challenges related to factual reliability, epistemic uncertainty, and value alignment, particularly when the data comes from unverifiable or low-quality sources. In contrast, models trained or evaluated primarily on verifiable domains, such as scientific literature, code repositories, or structured encyclopedic content, exhibit more predictable and auditable behavior, especially in tasks requiring factual accuracy.

Post-training

The suite of methodologies applied to a pre-trained model to adapt and specialize its behavior for particular downstream tasks, application contexts, or user-aligned objectives. This phase typically involves supervised fine-tuning, instruction tuning, or reinforcement learning, each contributing to the alignment and refinement of the model's outputs.

Post-training is especially critical when deploying language models in safety-sensitive or high-stakes environments, where mere linguistic fluency is insufficient. It is in this stage that distinctions between verifiable and unverifiable domains become pronounced: verifiable domains allow for more objective evaluation and calibration of model outputs, whereas unverifiable domains, for example those involving subjective opinions, open-ended advice, or speculative reasoning, pose unique challenges for ensuring truthful and reliable responses.

Supervised Fine Tuning (SFT)

A process by which a pre-trained model is adapted to specific tasks through continued training on labeled datasets consisting of input-output pairs. This technique uses standard supervised learning to minimize loss functions that capture task-specific correctness, such as cross-entropy loss on paired prompts and desired responses.

SFT is particularly effective in grounding models in verifiable domains, where labeled data can be curated with high fidelity. In contrast, applying SFT in unverifiable or ambiguous domains may risk reinforcing dataset biases or subjective conventions, especially when ground-truth labels are difficult to define or contestable.

Reinforcement Learning (RL)

A learning paradigm wherein an agent improves its behavior through interaction with an environment and feedback in the form of scalar rewards or penalties. When applied to language models, RL facilitates the optimization of generated text based on higher-order objectives, such as informativeness, coherence, or adherence to stylistic norms, beyond what can be captured by likelihood-based training alone.

In verifiable domains, reward signals can sometimes be constructed from measurable task performance or consistency with external ground truth. However, in unverifiable domains, the design of reward functions is less straightforward and often requires subjective judgment, increasing the risk of misalignment or reward hacking.

Reinforcement Learning from Human Feedback (RLHF)

A specialized form of RL designed to align the outputs of large language models with human preferences and values. The approach involves collecting human judgments on model responses to train a reward model, which is then used to guide further optimization of the language model via reinforcement learning algorithms such as Proximal Policy Optimization (PPO).

RLHF has become a cornerstone in the alignment of LLMs for interactive, safety-critical, or socially sensitive use cases. Its efficacy depends heavily on the clarity and consistency of human feedback. In verifiable domains, evaluators can ground their judgments in objective reference materials. Conversely, in unverifiable domains, such as moral reasoning, aesthetic preference, or interpersonal advice, evaluative signals may be inherently noisy or culturally contingent, complicating alignment and generalization.

Synthetic Data

Artificially generated information that replicates the statistical characteristics, distributions, and structural patterns of real-world datasets. It is produced through algorithmic processes, such as simulation, generative models, or rule-based systems, with the goal of providing data that is useful for training, validating, or testing machine learning models. Synthetic data is particularly valuable in scenarios where real data is scarce, sensitive, or expensive to collect, and it plays a growing role in enhancing model robustness, ensuring privacy, and enabling experimentation at scale.

Parameters

In a machine learning model, parameters refer to the internal variables that are learned from data during training. In neural networks, these typically include the weights and biases of each layer. Parameters define the model’s functional mapping from inputs to outputs, and their values are iteratively adjusted through optimization algorithms (for example, stochastic gradient descent) to minimize a specified loss function.

Weights

A subset of a model's parameters that determine the strength or influence of connections between neurons in adjacent layers of a neural network. During training, weights are adjusted to capture statistical patterns in the data. Their configuration directly affects the model’s ability to generalize from training data to unseen inputs. In deep learning, large-scale models may contain billions of individual weights.

Biases

In the context of neural networks, biases are trainable scalar parameters added to the weighted sum of inputs before the application of activation functions. These parameters allow the model to shift activation thresholds independently of input values, enhancing the network’s capacity to fit complex patterns by enabling flexibility beyond zero-centered data.

Separately, biases also denote systematic errors or prejudices that can arise within machine learning models or datasets. Such biases manifest as distortions in model predictions or behaviors caused by imbalanced, incomplete, or unrepresentative training data, flawed labeling, or design assumptions. These statistical and social biases are critical considerations in the deployment of AI systems, as they may lead to unfair, inaccurate, or discriminatory outcomes, particularly in socially sensitive applications.

Evaluations

The systematic assessment of a model's performance on specific tasks, typically using quantitative metrics such as accuracy, F1 score, BLEU, perplexity, or mean squared error. Evaluations may be conducted on held-out validation or test datasets to measure generalization, and can be domain-specific (factuality in verifiable domains) or subjective (coherence or helpfulness in unverifiable domains).

Data Labeler

A human or automated system responsible for assigning structured annotations (labels) to raw data, such as classifying images, categorizing text, or ranking responses. Labeled data is essential for supervised learning, enabling models to learn input-output mappings. Human data labelers are particularly important in nuanced or ambiguous tasks, though consistency and subjectivity must be carefully managed, especially in unverifiable domains.

Neural Network

A computational architecture composed of layers of interconnected nodes (neurons) designed to approximate complex, non-linear functions. Inspired by the structure of biological neural systems, neural networks are the foundation of modern deep learning models. Each neuron computes a weighted sum of its inputs, applies an activation function, and passes the result to the next layer. Through back-propagation and gradient descent, neural networks learn to adjust their parameters to minimize task-specific error.

System Prompt

A specially designed input instruction or set of directives provided to a language model to guide its overall behavior, tone, and output style during interaction. Unlike user prompts, which typically contain specific queries or tasks, system prompts establish contextual parameters or constraints that influence the model’s responses consistently across a session or application. System prompts are integral in shaping model alignment, ensuring outputs adhere to desired safety, ethical, or functional guidelines, and are often employed in fine-tuning or real-time prompting strategies to control generative behavior.

User Prompt

The input text or query submitted by an end-user to a language model, requesting information, assistance, or task completion. User prompts typically contain specific instructions, questions, or context that guide the model’s generation of relevant and coherent responses. Unlike system prompts, which set overarching behavioral parameters, user prompts are dynamic and varied, reflecting the immediate intent and needs of the user during an interaction. Effective interpretation of user prompts is crucial for achieving task-specific accuracy and user satisfaction in conversational AI systems.

Agent

A program acting on behalf of a person or organization to perform complex tasks that require planning or the use of external tools.

Overfitting

In machine learning, when a model learns the training data too well including its noise and outliers, leading to poor generalization on new, unseen data. Essentially, the model becomes too tailored to the training set and fails to capture the underlying patterns that would allow it to make accurate predictions on data it hasn't seen before. While the model performs excellently on training data, its performance on validation or test data is often significantly worse.

Context Window

The amount of text a model can process at once measured in tokens.

Catastrophic Forgetting

The abrupt loss of previously acquired knowledge in a neural network when it is trained on new data or through the process of fine-tuning.