AI Glossary

Artificial General Intelligence is the idea of a machine that can understand, learn, and apply knowledge across any task at a human-like level of flexibility — not just the narrow tasks today's systems are trained for. No existing system meets this bar; it remains a research goal and a subject of intense debate about timelines, definitions, and risks.

Why it matters: AGI shapes long-term policy, investment, and safety discussions. When headlines claim a model is a step toward AGI, it helps to know that the term has no agreed test — which is exactly why such claims deserve careful reading.

Example: When a lab announces a model that reasons across many domains, commentators debate whether it counts as progress toward AGI — and what would even prove it.

An AI agent is a system that pursues a goal through multiple steps: it plans, uses tools (search, code, APIs), observes results, and adjusts — instead of answering a single prompt once. The 'agentic' part is the loop of action and feedback, not any single model call.

Why it matters: Agents turn chatbots into workers that can complete tasks, but they also multiply failure modes: a wrong step early can cascade. Understanding agents helps you judge claims about AI 'doing things' autonomously.

Example: An agent asked to 'plan a team offsite' might search venues, compare prices, draft an email, and refine the plan — calling tools repeatedly until the goal is met.

An Application Programming Interface is a defined way for software to talk to other software: one system makes a structured request, the other returns a structured response. AI providers expose their models through APIs so developers can build applications on top of them.

Why it matters: APIs are how AI capabilities move from a lab's demo into real products. When a company 'integrates AI', it almost always means calling a model provider's API.

Example: AI Panic's market data arrives through a provider API: the server requests end-of-day prices once per trading day and stores the response for all readers.

Artificial intelligence is the broad field of building systems that perform tasks we associate with human thinking: understanding language, recognizing images, planning, and learning from data. Modern AI is dominated by machine learning — systems that improve from examples rather than hand-written rules.

Why it matters: AI is an umbrella term, not a single technology. Precision matters: a spam filter, a chess engine, and a large language model are all 'AI', but they work completely differently.

Example: Grouping similar news stories, scoring market attention, and translating text can all be called AI — yet each uses different methods underneath.

The AI Panic Attention Score is this site's own metric for how much a story is being talked about: it weighs how many independent sources cover the same event, how fresh the coverage is, and how it spreads. It is deliberately not a sentiment measure and not investment advice — it says how loud the room is, not whether the news is good.

Why it matters: Attention separates signal from noise: a genuinely breaking development is covered by many independent outlets at once, while a single rehashed press release is not. Knowing what the score measures — and what it does not — keeps you from misreading it as a recommendation.

Example: A major model release covered by twenty independent sources within hours gets a high attention score; one blog repeating the same announcement does not move it much.

A benchmark is a standardized test used to compare AI systems: fixed tasks with fixed scoring, so results are comparable across models. Popular examples test reasoning, coding, math, or knowledge — but every benchmark measures only what its tasks actually capture.

Why it matters: Benchmark scores drive headlines and marketing. Knowing their limits — saturation, data contamination, narrow scope — helps you read 'state of the art' claims with healthy skepticism.

Example: When two labs both claim the 'best' model, the benchmark tables usually reveal they each chose the test where their model wins.

The context window is the amount of text a language model can consider at once — your prompt, the conversation so far, and any documents you provide, all measured in tokens. Everything outside the window is invisible to the model for that answer.

Why it matters: Context limits explain real behavior: long documents get truncated, early instructions get 'forgotten' in long chats, and retrieval (RAG) exists partly to work around the window's size.

Example: Pasting a 300-page report into a chat may exceed the window; the model only 'sees' what fits, so a summary of the middle chapters could silently be missing.

Deep learning is machine learning with large, multi-layered neural networks. The 'deep' refers to many stacked layers that learn increasingly abstract features — from edges to shapes to objects in images, or from letters to words to meaning in text.

Why it matters: Nearly every modern AI breakthrough — language models, image generation, speech recognition — is deep learning. Understanding the term separates the technique from the marketing around it.

Example: A deep network trained on millions of photos learns to recognize cats not from rules about whiskers, but from statistical patterns discovered across the layers.

An embedding is a numerical fingerprint of a piece of content — a vector of numbers that captures meaning well enough that similar texts land close together. Embeddings let software search, group, and compare content by meaning instead of exact wording.

Why it matters: Embeddings power semantic search, recommendations, and duplicate detection. AI Panic uses this style of similarity when grouping stories that report the same event in different words.

Example: “OpenAI unveils new model” and “GPT-maker announces next release” share no keywords but have similar embeddings — so they can be grouped as one story.

Fine-tuning is additional training that adapts an existing model to a narrower purpose: a company's support tone, medical terminology, or a specific task format. The base model's general abilities are kept; the extra training shifts its habits.

Why it matters: Fine-tuning explains how one base model becomes many specialized products. It also has limits — it adjusts behavior more than it adds fresh knowledge.

Example: A law firm fine-tunes a general model on its document style so drafts come out in house format instead of generic legalese.

A foundation model is a large, general-purpose model trained on broad data that serves as the base for many applications — via prompting, fine-tuning, or tool integration. GPT-class and Claude-class models are the canonical examples.

Why it matters: The foundation-model layer is where capability, cost, and concentration of power live. Most AI products you use are applications on top of a handful of foundation models.

Example: A writing assistant, a coding helper, and a customer-support bot can all be built on the same foundation model with different instructions and data.

Generative AI produces new content — text, images, audio, code, video — rather than only classifying or ranking existing content. It learns the statistical shape of its training data and samples new outputs that fit the prompt.

Why it matters: Generative systems are why AI suddenly touches creative and knowledge work directly. They can also produce fluent nonsense, which is why verification habits matter more than ever.

Example: Asking a model to draft a product announcement, generate a logo concept, or write a database migration are all generative AI tasks.

A hallucination is when an AI system states something false with full confidence — inventing sources, dates, functions, or facts. It happens because the model generates plausible text, not verified truth.

Why it matters: Hallucinations are the core reliability problem of generative AI. Any workflow that matters needs a verification step — sources, tests, or human review.

Example: A model cites a research paper with a convincing title and author list — but the paper does not exist. The citation sounded right because it fit the pattern.

Inference is using a trained model: you send input, the model computes an output. Training happens once (or rarely) at great cost; inference happens on every single request, which is why its speed and price dominate the economics of AI products.

Why it matters: When news mentions cheaper or faster models, it usually means inference got more efficient — the same capability at lower cost per answer, which is what makes new applications viable.

Example: Every message you send to a chatbot is an inference call: the trained model reads your tokens and generates the reply token by token.

A large language model is a neural network trained on massive text collections to predict and generate language. 'Large' refers to billions of parameters. LLMs power chatbots, writing tools, code assistants, and most current AI agents.

Why it matters: LLMs are the engine of the current AI wave. Understanding that they model language statistically — rather than 'knowing' like a database — explains both their fluency and their failure modes.

Example: The same LLM can summarize a paper, translate it, and answer questions about it — three tasks, one model, driven by different prompts.

Machine learning is the branch of AI where systems learn patterns from data instead of following explicit rules. You show examples; the system adjusts internal parameters until its outputs match the examples — then it generalizes to new cases.

Why it matters: ML is the practical core of modern AI. It explains why data quality decides model quality — and why a system can fail exactly where its training data was thin.

Example: A spam filter learns from millions of labeled emails; it never reads a rule saying 'FREE MONEY = spam', it infers the pattern itself.

Multimodal AI works across types of content at once: text, images, audio, video. A multimodal model can look at a chart and discuss it, or listen to speech and answer in writing — one system, several senses.

Why it matters: Real-world information is multimodal. Models that combine modalities unlock use cases pure text models cannot touch, from document understanding to video analysis.

Example: Upload a photo of a broken error screen and ask what went wrong — a multimodal model reads the screenshot and explains it.

A neural network is a computation model built from layers of simple units that transform inputs into outputs through learned numeric weights. Loosely inspired by neurons, but really just very large, trainable math.

Why it matters: Neural networks are the substrate of deep learning. The term tells you how a system works mechanically — layers and weights — without implying anything brain-like about its 'understanding'.

Example: In an image network, early layers respond to edges and colors, middle layers to shapes, and late layers to whole objects like faces or cars.

An open-source (more precisely: open-weight) model publishes its trained parameters so anyone can download, run, inspect, and often modify them — instead of offering only hosted API access. Licenses vary widely in what they actually permit.

Why it matters: Open models shift power: they enable local running, customization, research scrutiny, and independence from a single provider — while raising their own misuse and support questions.

Example: A company downloads an open-weight model and runs it on its own servers so sensitive data never leaves its infrastructure.

A prompt is the input you give a language model: the question, instructions, context, and examples it should respond to. The same model can give wildly different results depending on how the prompt is written.

Why it matters: Prompting is the user interface of the AI era. Clear instructions, relevant context, and stated constraints are the difference between a generic answer and a useful one.

Example: “Write a summary” gets a generic paragraph; “Summarize this for a busy executive in three bullet points, keeping all numbers” gets a usable result.

Prompt engineering is the deliberate craft of designing prompts: structuring instructions, adding examples, defining output formats, and iterating until results are reliable. At product scale it becomes a disciplined practice closer to programming than to chatting.

Why it matters: Well-engineered prompts turn a demo into a dependable feature. They are also reusable — which is exactly the idea behind curated command libraries like the one in this Academy.

Example: A support team refines one tested prompt that classifies tickets consistently, instead of every agent improvising their own wording each time.

RAG combines search with generation: before answering, the system retrieves relevant documents (usually via embeddings), places them into the prompt, and the model answers grounded in that material instead of memory alone.

Why it matters: RAG is the standard answer to two LLM weaknesses: outdated knowledge and hallucination. Answers anchored in retrieved, citable documents are easier to trust and verify.

Example: Ask a company chatbot about this year's pricing: RAG fetches the current price sheet into the prompt, so the answer reflects today's document, not last year's training data.

In AI Panic, a Signal is a product feature, not a generic AI term: a rule-based market observation derived from stored data — for example an unusual price move or a stock appearing in high-attention news. Every signal lists its evidence and its category, and it is an observation with explicit limits, never a prediction or advice.

Why it matters: Signals show what stands out in the data right now — with the evidence attached, so you can judge it yourself. Knowing they are observations, not forecasts, keeps them in the right place in your thinking.

Example: A signal might read: “NVDA: +4.2% daily move while appearing in three high-attention stories” — with the underlying numbers visible.

A token is the chunk of text a language model actually processes — roughly a word or word-piece. Models read and write token by token; context windows, speed, and API pricing are all measured in tokens.

Why it matters: Tokens are the currency of LLMs. Knowing this explains why long documents cost more, why models sometimes split words oddly, and what '128k context' really bounds.

Example: The sentence “AI moves fast” is about four tokens; a full novel is hundreds of thousands — which is why it cannot fit into one prompt.

The Transformer is the neural network architecture behind modern language models, introduced in 2017. Its core mechanism, self-attention, lets the model weigh which earlier words matter for predicting the next one — across the whole input at once. Note: this architectural 'attention' has nothing to do with the AI Panic Attention Score, which measures news coverage.

Why it matters: Transformers are why AI leapt forward: they train efficiently at massive scale and handle long-range context. Nearly every model in the news — GPT, Claude, Llama, Gemini — is a Transformer.

Example: In “The trophy doesn't fit in the suitcase because it is too big”, self-attention helps the model decide whether 'it' refers to the trophy or the suitcase.