A Student’s Lexicon of AI

As a student reading and learning about artificial intelligence, my biggest challenge has been keeping up with all the unfamiliar terminology. Every week, it seems another technical paper introduces terms that sound completely alien to me - from architectural concepts like LoRA and attention mechanisms to safety debates around jailbreaking and sovereign AI. Rather than letting these unfamiliar concepts slip past, I decided to start documenting them as I went. I compiled this lexicon to serve as my personal working glossary, defining each phrase plainly so that anyone wrestling with AI literature can spend less time figuring out hard words and more time understanding the ideas behind it.

Adversarial Robustness: The capacity of an AI system to resist manipulation, evasion attacks, or performance failure when fed intentionally misleading inputs.

Agentic Workflow: An iterative execution loop where models self-reflect, break down multi-step plans, call APIs, and verify intermediate outputs before providing a final answer.

AI Agent: An autonomous software entity that perceives its environment, reasons through goals, calls external tools, and takes sequential actions to execute tasks.

AI Alignment: The field of research focused on ensuring artificial intelligence systems consistently steer toward human intentions, values, and ethical objectives.

AI Swarm: A decentralized collective of multiple AI agents working cooperatively to solve complex problems through distributed decision-making.

Artificial General Intelligence (AGI): A theoretical milestone where a synthetic system matches or surpasses human cognitive competence across virtually all economically and intellectually valuable tasks.

Attention Mechanism: A mathematical technique allowing a model to dynamically weight the importance of different tokens across an entire context window.

Autoregressive Model: A generative sequence model that predicts the next token or element strictly conditional on all previously generated tokens in the sequence.

Chain-of-Thought (CoT): A prompting or architectural technique that directs a model to break down complex problems into explicit step-by-step intermediate reasoning paths.

Classifier-Free Guidance (CFG): A technique used in text-to-image/video diffusion that balances prompt adherence against generation diversity by interpolating between conditional and unconditional outputs.

CLIP (Contrastive Language-Image Pre-training): A joint embedding architecture trained to map text descriptions and visual images into a shared vector space, foundational for aligning prompts with generated visuals.

Compute Governance: Regulatory and institutional oversight tracking the physical data centers, advanced semiconductor chips, and energy resources required to train frontier models.

Constitutional AI: A training methodology where an AI model uses an explicit set of principles or rules to self-critique, filter, and revise its own outputs without direct human intervention.

Context Window: The maximum number of tokens an AI model can take as input and hold in memory during a single interaction.

Cross-Attention: A transformer mechanism that injects conditioning information (such as textual prompt embeddings) into another modality's network layers (such as a visual denoiser).

Data Poisoning: An adversarial attack where corrupted, misleading, or malicious samples are deliberately injected into a training dataset to compromise the AI model's accuracy or manipulate its behavior.

Deepfake: Synthetic or manipulated digital audio, photo, or video content generated using deep learning to convincingly depict an individual saying or doing something that never occurred.

Diffusion Model: A generative architecture that learns data distributions by iteratively adding Gaussian noise to training data and subsequently learning to reverse the denoising process to synthesize new samples.  

DPO (Direct Preference Optimization): An alignment method that optimizes model weights directly on pairwise preference data without requiring an auxiliary reward model.

Embedding: A dense, high-dimensional vector representation of text, audio, or images that captures semantic relationships in vector space.

Fine-Tuning: The supervised process of adapting a pre-trained base model on a specialized, curated dataset for a specific domain or task.

Function Calling / Tool Use: The capability of an AI model to detect when external data or computation is required and output structured parameters to invoke APIs or databases.

Generative Adversarial Network (GAN): A dual-network framework consisting of a generator producing synthetic data and a discriminator attempting to distinguish real samples from fake ones.  

Generative AI (GenAI): A branch of AI focused on synthesizing novel, high-dimensional content (text, image, audio, video, code, or 3D) that mimics the distribution of its training data.  

Hallucination: An occurrence where a generative AI produces output that is factually incorrect, nonsensical, or ungrounded while presenting it as truth.

In-Context Learning (Few-Shot Prompting): Providing a model with relevant examples directly inside the prompt context to guide its output format and reasoning without altering model weights.

Inference: The runtime execution phase where a trained model processes new input to generate predictions or outputs.

Jailbreaking: The act of using crafted prompts or exploit techniques to bypass the built-in safety guardrails, content filters, and behavioral restrictions of an AI model.

Kill Switch: A physical or software-based failsafe mechanism designed to instantly shut down or neutralize an AI system in the event of dangerous or uncontrollable behavior.

Large Language Model (LLM): A deep learning model trained on vast text corpora using self-supervised learning to predict tokens and generate human-like language.

Latent Diffusion Model (LDM): A diffusion model optimized to execute the denoising process within a compressed latent space rather than pixel space, substantially cutting computational overhead.

Latent Space: An abstract multi-dimensional mathematical space where complex data points (like semantic concepts or visual features) are compressed into continuous vector representations.

LoRA (Low-Rank Adaptation): A PEFT technique that freezes the base weights and injects trainable rank-decomposition matrices into transformer layers, drastically reducing VRAM and storage footprints.

Mechanistic Interpretability: The reverse-engineering of neural networks to understand the specific circuits, neurons, and internal representations responsible for model behavior.

Mixture of Experts (MoE): An architecture that routes specific queries to specialized subsets of neural network parameters ("experts") to improve efficiency and reduce inference costs.

Mode Collapse: A generative failure mode—particularly in GANs—where the model discovers a narrow subset of realistic outputs and repeatedly generates them, losing output diversity.

Model Collapse: The degradation in quality, diversity, and coherence that occurs when new generative models are trained primarily on synthetic data produced by older models.

Multi-Modal AI: Models capable of simultaneously processing, interpreting, and generating multiple data modalities, such as text, code, audio, images, and video.

Negative Prompting: Explicit instructions or tokens provided to a diffusion model that penalize undesirable features, styles, or artifacts during the reverse-denoising pass.

Parameter: The internal weights and biases adjusted during training that determine how the model processes input into output.

Parameter-Efficient Fine-Tuning (PEFT): Methods designed to adapt large pre-trained models to downstream tasks by modifying only a small fraction of parameters while freezing the base model.

Perceptual Loss: A loss metric computed using features extracted from deep layers of a pre-trained network to evaluate whether generated media matches the perceived quality and structure of real data.

Pre-training: The initial, compute-heavy phase where a base model learns general linguistic and conceptual patterns from broad, unlabeled datasets.

Prompt: An input - typically text, code, or an image - provided by a user or system to direct an AI model to produce a specific response or perform a task.

Prompt Engineering: The craft of designing, structuring, and optimizing natural language inputs to elicit specific, reliable outputs from an AI model.

Prompt Injection: A security vulnerability where malicious user input hijacks the system instructions of an application and overrides its intended operational parameters.

QLoRA: An extension of LoRA that quantizes the frozen base model to 4-bit precision while optimizing 16-bit LoRA adapter weights, enabling fine-tuning on consumer-grade hardware.

Quantization: A post-training optimization technique that reduces model weight precision (e.g., from 16-bit float to 4-bit integer) to save memory and increase execution speed.

Recursive Self-Improvement: The hypothetical process where an AI system redesigns and enhances its own code or architecture, triggering rapid, iterative leaps in capability.

Red Teaming: A structured, adversarial stress-testing practice where security specialists intentionally attack an AI system to uncover vulnerabilities, risks, and failure modes before bad actors can exploit them.

Retrieval-Augmented Generation (RAG): A framework that retrieves authoritative external documents from a database to ground the model's generation in factual source material.

RLHF (Reinforcement Learning from Human Feedback): A training methodology that aligns model outputs with human preferences using reward models trained on human ratings.

Sandbox: An isolated, secure computing environment where AI models, code, or applications can be run and tested without risking harm to the broader host system or network.

Soft Prompting / Prompt Tuning: Freezing model weights and prepending continuous, learnable virtual token embeddings to the model's input vector space.

Sovereign AI: A nation's independent capability to build, train, deploy, and govern its own AI infrastructure, models, and data without reliance on foreign technology or providers.

Superalignment: The technical challenge of ensuring artificial systems that vastly exceed human intelligence remain safely steerable and aligned with human intent.

Synthetic Data: Artificially generated datasets created by generative models to train downstream algorithms, augment sparse real-world data, or protect privacy.

Temperature: A parameter that scales the logits before the softmax layer, adjusting the probability distribution to make token selection either more deterministic (near 0) or more random and creative (near 1).  

Token / Tokenization: The process of breaking raw text into discrete linguistic chunks (sub-words, characters, or words) that a neural network can process numerically.

Transformer: A neural network architecture that relies on self-attention mechanisms to process input data in parallel rather than sequentially.

Variational Autoencoder (VAE): A probabilistic architecture that compresses input data into a continuous, low-dimensional latent space distribution before decoding it back to reconstruct or generate variations of the original data.

Vector Database: A specialized database optimized for storing, indexing, and executing fast similarity searches across high-dimensional vector embeddings.

Zero Vulnerability: A theoretical, zero-flaw security state where a system contains no known or unknown software bugs, misconfigurations, or exploitable attack surfaces.

#ArtificialIntelligence
#AILexicon
#MachineLearning
#GenerativeAI
#AIEthics
#TechLiteracy
#DeepLearning
#AIAlignment
#AIEducation
#LearnAI