Glossary
62 terms, defined the way the course used them. Each links to the class that taught it where there is one.
An AI system that can take multiple steps toward a goal on its own, such as searching, using tools, or calling other software, rather than just answering a single prompt. Coding assistants that can browse files, run code, and execute scheduled tasks are one example raised in class.
A multi-step process where an AI system acts more like an assistant carrying out a sequence of tasks than a single question-and-answer exchange. The course's data-cleanup class argued that messy inputs and unclear systems make these workflows unreliable no matter how good the AI is. taught in class 11 →
Artificial General Intelligence: a hypothetical AI system that could reason and learn across any task at a human level, rather than being limited to what it was trained for. It does not exist yet and is a subject of active debate about timelines. taught in class 2 →
The ongoing effort to make an AI system's goals and behavior match what its developers and users actually intend, rather than pursuing an instruction in a literal or harmful way. It is a live safety concern, illustrated in class by a published study where a model role-played resorting to blackmail rather than being shut down. taught in class 18 →
Artificial Narrow Intelligence: an AI system built to do one task or a limited set of tasks well, such as image generation or spam filtering. This is the category almost every AI tool in use today falls into. taught in class 2 →
Artificial Superintelligence: a hypothetical AI system that would exceed human intelligence across essentially every domain. It sits one step past AGI on the same scale and is even more speculative. taught in class 2 →
The width-to-height proportion of an image or video, such as widescreen versus square versus portrait. Most image and video tools let you set it directly as a parameter, commonly written as something like --ar in MidJourney. taught in class 5 →
The raw, unfiltered version of an AI model before it is packaged into a polished consumer product with guardrails and a friendly interface. The course used the phrase "raw cake vs sprinkles" to describe the gap between a base model and the finished app most people actually use. taught in class 12 →
Skewed or unbalanced patterns in an AI system's output that trace back to skewed patterns in its training data or in the people who labeled that data. The course frames bias as traceable through specific layers, such as what got scraped, who annotated it, and which filters were applied, rather than as something mysterious. taught in class 12 →
The third canvas in class 13's strategy framework: decide where the brand sits by borrowing equity from an adjacent category or building it from scratch, the way Liquid Death did. The point is to step out of the head-to-head mindset and then be single-minded about the one position you pick. taught in class 13 →
The first canvas in class 13's strategy framework, and deliberately the longest: work out what market you are actually in, what paradigm shift is coming, and where the gaps are, while the product stays out of the room for the whole pass. taught in class 13 →
A free, node-based interface for building detailed AI image and video generation pipelines, run locally on your own computer. The course described it as the most powerful but most technical end of node-based tools, nicknamed "the Photoshop of the node world." taught in class 8 →
A training approach where a model is given a written set of principles, a kind of constitution, and trained to check its own responses against those principles rather than relying only on human raters for every judgment call. taught in class 18 →
Deliberately structuring the background information, examples, and framing you give an AI tool, rather than just writing a clever one-line prompt. The course's persona-building class argued that this underlying data structure matters more to a good result than the wording of the prompt itself. taught in class 15 →
The amount of text, images, or other material an AI model can hold in mind at one time during a conversation. Once a conversation or an attached document exceeds it, earlier material starts to drop out of what the model can actually consider.
A version of ChatGPT customized with its own name, instructions, and uploaded knowledge so it behaves like a dedicated assistant for one job. The course treated a custom GPT as essentially a portable persona that can be shared with a link and pulled into different conversations. taught in class 19 →
Synthetic audio, video, or image content that convincingly impersonates a real person doing or saying something they never did. The course raised this both as a vocabulary term and as a live ethical risk, from robocalls to romance scams. taught in class 1 →
A type of AI model that generates an image or video by starting from random noise and gradually refining it, step by step, into a coherent result. Most modern image and video generators, including the models inside Krea, work this way. taught in class 3 →
An AI-generated model built to closely represent one specific real person or product, as opposed to a synthetic persona representing a type of person. H&M's use of digital models of real employees for some campaign imagery was the course's example. taught in class 5 →
Taking an already-trained AI model and training it further on a smaller, more specific set of data so it performs better on a particular task, style, or subject.
Google Gemini's equivalent of a saved, reusable custom assistant, built with its own instructions and purpose. The course grouped Gems with Projects and custom GPTs as the main way to build a reusable AI tool instead of starting from scratch every conversation. taught in class 4 →
A prompt-structuring method taught in the course: Goal, Output, Limitations, Data. It is meant to slow a person down before prompting so the request is specific instead of vague. taught in class 4 →
Graphics Processing Unit: a computer chip originally built for rendering images that turned out to be extremely good at the math behind training and running AI models. Companies like NVIDIA and chipmaker TSMC came up repeatedly in class as the hardware side of the AI boom.
The human-verified, confirmed-correct answer or reference data that a model is trained or checked against. The course noted that hallucination is what happens in the gap between ground truth and what a model actually generates. taught in class 18 →
When an AI tool confidently states something false or generates a detail that was never in its source material. The course frames it as a side effect of how these models work, not a rare glitch, and teaches specific troubleshooting steps for it in image generation.
The second canvas in class 13's strategy framework: who the customers are, why they behave the way they do, and the single insight underneath it all. Brands that win align to one of three levels: me, my community, or the world. taught in class 13 →
A specific image used to anchor the start, end, or a key point of an AI-generated video. The course was explicit that a keyframe is not the same as a general reference image: it has to work as an actual frame within the motion, not just show the right subject or style. taught in class 6 →
An AI model trained on huge amounts of text that predicts and generates language. ChatGPT, Claude, and Gemini are all products built on top of an LLM.
A lightweight way to fine-tune an image or video model on a specific style, character, or subject without retraining the whole model from scratch. Krea's left-side menu includes a dedicated LoRA training feature. taught in class 3 →
Machine learning is software that improves at a task by learning patterns from data instead of following rules a person wrote by hand. Deep learning is a specific approach to machine learning that uses layered neural networks, and it is the technique behind most of today's generative AI.
A standard that lets an AI tool connect to and pull in outside data or another piece of software directly inside a conversation, instead of a person copying and pasting information back and forth. The course's example was connecting a design tool to a chat assistant so the assistant could work with real files. taught in class 10 →
A collection of reference images, colors, and textures gathered to define a visual direction before generating final work. Krea includes a dedicated moodboard feature for collecting and organizing this kind of reference material. taught in class 3 →
An AI tool or model that can work across more than one type of input or output, such as reading an image and responding in text, or taking both text and an uploaded photo as a prompt. The course recommended checking a hallucinating image result against a multimodal tool as a troubleshooting step. taught in class 5 →
Telling an AI image or video tool what to avoid, rather than only what to include. The course flagged it as a tool that can backfire, sometimes making a model fixate on the very thing you named in order to exclude it. taught in class 5 →
A visual way of building an AI generation process by connecting individual building blocks, each doing one job such as generating, upscaling, or editing, into a chain or network instead of typing one prompt into one box. The course taught this in Krea's node editor and named it as the tool for repeatable, multi-step, or multi-model work. taught in class 8 →
Software or a model whose underlying code or weights are published for anyone to use, inspect, or build on, instead of being locked inside one company's product. Some image and video models used in class, like certain Krea node options, are open source.
The internal numeric settings a model adjusts during training to get better at its task. Roughly speaking, more parameters means a larger, more capable model, which is why jumps like GPT-3 to GPT-4 get discussed in terms of parameter scale.
A defined character or point of view given to an AI assistant, whether that means a tone and role for a custom tool or a simulated consumer profile used for research. The course treated persona-building as a design skill with its own structure, not just a costume on top of a chatbot. taught in class 15 →
A feature in tools like Claude and ChatGPT that gives a set of conversations a shared, persistent home for instructions and reference files. Unlike a portable custom GPT, a Project functions as a static context space tied to one tool. taught in class 4 →
The instructions you give an AI tool to get an output. Course frameworks like GOLD and the Image Prompt Structure exist to make prompts more specific and repeatable rather than one-off guesses. taught in class 4 →
An attack or trick where hidden or cleverly worded text gets an AI system to ignore its original instructions and do something else instead. The course referenced Gandalf, a game built by a cybersecurity firm, as a hands-on way to learn how prompt injection works by trying to break a chatbot's guardrails. taught in class 3 →
The verifiable trail showing where a piece of media actually came from and whether it was altered or AI-generated. The course's ethics session raised the plain question behind it: can you prove your work is yours all the way through. taught in class 18 →
A structure for building a synthetic persona: Representation, Objectives, Core Value and Belief, Key tone and Knowledge. It separates a persona's personality from its decision-making logic so both can be built and checked on purpose. taught in class 15 →
An image uploaded to an AI tool to guide a new generation, whether for subject, composition, color, or style. The course treated choosing and combining reference images as its own layer of prompting, separate from the written prompt.
Generating a new AI image or video by starting from an existing one and changing it, rather than starting from a blank prompt. In MidJourney this lets you keep a result's composition or character while altering specific details. taught in class 5 →
A training technique where human raters score a model's responses, and the model is adjusted to produce more of the highly-rated ones. It is one of the main tools used to make a raw model behave in a more helpful, less harmful way, described in class as putting a friendly, well-behaved face on a much stranger underlying system. taught in class 18 →
Small descriptive words added to a prompt that noticeably shift the mood, texture, or feel of an AI generation without changing the main subject. The course used a pear turning into something closer to paper as an example of how sensitive a model can be to one added word. taught in class 5 →
The starting point of random noise a diffusion model uses to generate an image or video. Reusing the same seed with the same prompt gives you a reproducible, comparable result instead of a completely new roll each time. taught in class 5 →
The hypothetical point at which AI capability grows so fast, feeding on its own improvements, that it becomes unpredictable and impossible for humans to fully understand or control from the outside. taught in class 2 →
An uploaded image used to guide the visual style, rather than the exact subject or composition, of a new AI generation. Tools like MidJourney let you dial the strength of that influence up or down against the rest of the prompt. taught in class 5 →
An AI-simulated stand-in for a type of person, built from research and demographic detail, used to test ideas, gather feedback, or represent a target audience before real research is possible. The course distinguished it from a digital twin (a model of one real person) and stressed validating it against real humans rather than treating it as a finished substitute for them. taught in class 15 →
The hidden setup text that tells an AI tool who to be and how to behave before a person ever types anything. In a custom GPT this lives in the Instructions field, which the builder treats as invisible to the end user and the most important part of the setup. taught in class 19 →
A six-archetype assessment of a person's thinking style, used in class as a lens for how someone naturally approaches problems and where an AI collaborator might complement or challenge that style. taught in class 16 →
Guest strategist Angella Tapé's brand-strategy sequence from class 13: Category Creation, then Human Truth, then Brand Positioning, worked in that order so the market and the customer are understood before anyone talks about the product. taught in class 13 →
A small chunk of text, often a word or part of a word, that a language model reads and generates one piece at a time. It is also the unit some image and video tools use to measure how much a generation costs.
The large collection of text, images, audio, or video that an AI model learns from before it is released. What is present, missing, or overrepresented in that data shapes everything the model later produces. taught in class 12 →
The neural network design introduced in the 2017 paper "Attention Is All You Need" that made modern large language models possible. It lets a model weigh how relevant every other word in a passage is when predicting the next one.
A classic thought experiment, proposed by Alan Turing, for judging whether a machine can convincingly imitate human conversation well enough that a person can't reliably tell they're talking to a machine. taught in class 2 →
Increasing an image or video's resolution and detail after it has been generated, either by traditional sharpening methods or by an AI model that invents plausible extra detail. The course distinguished traditional from generative upscaling and named Topaz and Magnific as common tools for it. taught in class 6 →
Building working software, a website, or an app mostly by describing what you want in plain language to an AI tool, rather than writing code by hand line by line. The course's method: sketch it on paper, photograph it, describe it in detail, and refine through conversation. taught in class 19 →
The learned numeric values inside a trained model that encode everything it picked up during training. Downloading or training weights is what it means to have a specific version of a model, as opposed to just accessing it through an app.
A specific kind of AI bias where a model's outputs skew toward Western, Educated, Industrialized, Rich, and Democratic perspectives, because that population is overrepresented in the training data and in the people who tested and rated the model. taught in class 12 →