Glossary
Key Terms from A to Z
AGI (Artificial General Intelligence): Hypothetical AI that matches human intelligence across all domains. Does not exist as of 2026. Current AI is “narrow” — good at specific tasks, not general reasoning.
AI (Artificial Intelligence): Systems that can perform tasks normally requiring human intelligence — understanding language, recognizing images, making decisions, generating content. Not a single technology but a family of approaches.
AI Agent: An AI system that can take autonomous actions — browsing the web, running code, sending emails, managing files — not just answer questions. More autonomous than a chatbot or custom GPT.
Alignment Gap: The difference between what a user MEANS and what the AI PRODUCES. Caused by the AI modeling the question accurately while modeling the user poorly. BRIEFS and The 3 Questions close this gap.
Attention Mechanism: The part of a transformer model that determines which words in the input are most relevant to each other. “The cat sat on the mat because it was tired” — attention figures out that “it” refers to “cat,” not “mat.”
Bias (in AI): Systematic errors in AI outputs that reflect inequities in training data or design choices. Types include historical bias, representation bias, measurement bias, and aggregation bias.
BRIEFS: A structured first-message framework: Background, Role, Intent, Execution, Format, Stakes. Derived from what 26 AI models across 5 providers said they need. The precision version of The 3 Questions.
Calibration Handshake: A protocol where the AI restates its understanding of your request and you confirm or correct before work begins. Eliminates most misalignment in one exchange.
Chain-of-Thought: A prompting technique where you ask the AI to “think step by step,” forcing it to show intermediate reasoning. Dramatically improves accuracy on math and logic tasks.
Chatbot: An AI system that responds to user messages in conversation. Reactive — waits for input, doesn’t take autonomous actions. ChatGPT, Claude, and Gemini in their basic chat modes are chatbots.
COMPAS: Correctional Offender Management Profiling for Alternative Sanctions. An algorithm used in US courts to predict recidivism. Found by ProPublica to have racially disparate false positive rates.
Context Window: The maximum amount of text an AI model can “see” at once. Measured in tokens. GPT-4’s context window is ~128K tokens. Anything outside the window is invisible to the model.
Custom GPT: A personalized version of ChatGPT with persistent system instructions, uploaded knowledge files, and configurable settings. More than a chatbot but less than an agent.
Data Drift: When real-world data changes over time, causing a deployed ML model to become less accurate. A spam filter trained on 2023 emails may fail on 2026 spam tactics.
Deep Research: AI tools that autonomously conduct multi-step research — searching, reading, following leads, and synthesizing reports. Available from OpenAI, Google, Perplexity, and others.
Deepfake: AI-generated video, audio, or images of real people that appear authentic. Can be used for entertainment, fraud, or political manipulation.
Diffusion: The technique behind AI image generation. Starts with random noise and gradually removes it, guided by the text prompt, until a coherent image emerges.
Embedding: A numerical representation of text (or images) that captures meaning. Similar concepts have similar embeddings. Used in RAG to find relevant documents.
EU AI Act: European Union regulation classifying AI systems by risk level. High-risk applications face mandatory transparency, testing, and oversight. In force as of 2026.
Few-Shot Prompting: Giving the AI examples of the desired input-output pattern before asking it to handle a new case. “Here are 3 examples, now do the same for this new one.”
Fine-Tuning: Permanently adjusting a model’s behavior by training it on specific data. More effort than RAG but produces deeper behavioral changes. QLoRA makes this feasible on a single GPU.
Hallucination: When an AI generates false information presented as fact. Not a bug being fixed but a structural feature of next-token prediction. The lawyer case (Module 1) is the canonical example.
HIPAA: Health Insurance Portability and Accountability Act. US law governing patient data privacy. AI tools processing patient information must be HIPAA-compliant.
LLM (Large Language Model): An AI model trained on massive text datasets that generates human-like text by predicting the next token. GPT-4, Claude, Gemini, and Llama are all LLMs.
LoRA / QLoRA: Low-Rank Adaptation / Quantized LoRA. Techniques for fine-tuning large models efficiently by training small adapter layers instead of the full model. Reduces GPU memory requirements dramatically.
Machine Learning (ML): A subset of AI where systems learn patterns from data rather than following explicit rules. The three types: supervised (labeled examples), unsupervised (finding patterns), reinforcement (learning from feedback).
Meta-Prompting: Asking the AI to help you write better prompts. “I want to accomplish X — what information do you need from me?” Flips the conversation so the AI interviews you.
Multimodal: AI that works across multiple types of input/output — text, images, audio, video, code. Modern models increasingly handle all modalities natively.
Neural Network: A computing architecture inspired by brain neurons. Layers of connected nodes process information, with each connection having a learnable weight. Deep learning uses neural networks with many layers.
Next-Token Prediction: The core operation of LLMs. Given all previous tokens, predict the most likely next token. This single operation, repeated thousands of times, produces coherent text, code, and reasoning.
NotebookLM: Google’s tool that creates interactive knowledge bases from uploaded documents. Only answers from YOUR sources, reducing hallucinations for document-specific questions.
Overfitting: When a model memorizes training data instead of learning general patterns. Performs well on training data, poorly on new data.
Perplexity: An AI-powered search engine that answers questions conversationally with numbered source citations. Good for getting oriented on a topic quickly.
Prompt: The text you send to an AI model. The quality of the prompt determines the quality of the response more than the capability of the model.
Prompt Chaining: Breaking complex tasks into sequential prompts, where each output feeds the next. Produces sophisticated results from simple individual steps.
QLoRA: See LoRA / QLoRA.
RAG (Retrieval-Augmented Generation): A technique that retrieves relevant documents from a database and adds them to the prompt before the AI generates a response. Grounds AI in YOUR data, dramatically reducing hallucinations.
RLHF (Reinforcement Learning from Human Feedback): The third stage of LLM training where human raters evaluate model outputs, teaching the model to be more helpful and less harmful.
SEARCH Model: Stretched Exponential ARCH — a volatility memory model developed by Dr. Yatawara using the kernel exp(−c·j^α), validated across 100+ assets.
SLM (Small Language Model): A language model with 1B-14B parameters that can run on consumer hardware. A published SLM-RAG experiment showed these matching GPT-4 for domain-specific tasks.
SMARTS: A multi-turn interaction framework: State, Map, Ask, Reason, Think, Summarize. Guides complex AI conversations where you need to THINK with the AI, not just get answers.
Sycophancy: When AI models excessively agree with users, avoid challenging incorrect statements, or over-praise to seem helpful. A side effect of RLHF training.
The 3 Questions: The simplified version of BRIEFS: (1) What do I want? (2) Who am I? (3) How will I use this? The everyday default for effective prompting.
Token: The basic unit of text that LLMs process. Roughly 3/4 of a word. “Strawberry” might be two tokens: [straw][berry]. Understanding tokenization explains why LLMs struggle with letter-counting tasks.
Transformer: The neural network architecture behind modern LLMs, introduced in 2017. Its key innovation — the attention mechanism — allows the model to consider all parts of the input simultaneously.
Vector Database: A database that stores embeddings (numerical representations of text) and enables fast similarity search. Used in RAG to find documents relevant to a user’s question. Examples: Qdrant, Pinecone, ChromaDB.
Vibe Coding: Building functional apps and tools by describing what you want in natural language, without writing code yourself. Good for prototypes and personal tools; unreliable for production software.
Zero-Shot Prompting: Asking the AI to perform a task without providing any examples first. Just describe what you want. Works well for simple, unambiguous tasks.
Back to Modules
API (Application Programming Interface): A way for software applications to communicate with each other. When you use the ChatGPT API, your code sends a request and receives a response — you don’t interact through the chat interface, you integrate AI into your own tools.
Inference: Running a trained model to generate outputs. Training teaches the model; inference is when it answers questions. Every time you send a prompt, the model performs inference. Inference costs (compute, energy, time) determine the practical cost of using AI.
Latent Space: A compressed mathematical representation of data learned by a neural network. Similar concepts cluster together in latent space — “king” and “queen” are closer to each other than “king” and “bicycle.” Image generators work by navigating latent space to find the image matching your description.
System Prompt: Hidden instructions given to an AI model before the user’s conversation begins. When you use ChatGPT, a system prompt shapes its behavior (be helpful, follow safety guidelines). Custom GPTs are built primarily by writing system prompts. Users don’t see them — but they shape every response.
Temperature: A parameter that controls how “creative” or “random” an AI model’s outputs are. Low temperature (0.0–0.3) = predictable, focused, best for factual tasks. High temperature (0.7–1.0) = diverse, surprising, better for brainstorming and creative writing. Most chatbot interfaces set this automatically.