Glossary of AI Terms

Prompt

A prompt is a command we send to a generative model (e.g. ChatGPT, Midjourney, Suno) to tell it what to create – it can be a question, a scene description, a command or a set of instructions. The quality and precision of the prompt have a huge impact on the outcome – it’s a bit like asking a question to an expert: the better you phrase it, the better answer you’ll get. In this context, it is important to understand prompting techniques, which I devote a lot of space to in my training courses on the use of AI.

Negative prompt

A negative prompt is usually a special part of a prompt (command) that tells the AI model what to avoid in the response. It’s like adding a list of forbidden things to better control the end result. For example, we can indicate that we don’t want suggestions from a certain category among the suggestions for solving a problem.

In some tools, a separate field is designated for the negative prompt.

Classic automation with AI

The classic automation process uses clearly defined rules and procedures – for example, “if the customer has not paid for 14 days, send a reminder.” Such solutions work well in predictable situations. Thus, also the place within the process and the task of artificial intelligence should be clearly defined, i.e. AI always performs a very similar task using the same prompt.

AI agents

These are small, intelligent programs using (most often) large language models that can independently plan and execute sequences of tasks. What differentiates AI agents from classic automation processes is that they can be given the right to autonomy (decision-making), and they can operate in a non-linear manner (we distinguish different topologies, as I discuss in training courses on automation with AI). For example, agents can plan a trip, book tickets, write an itinerary and email it.

Specialized models

Specialized models are versions of large language models that have been tailored to specific tasks, industries or users – e.g. GPTs in ChatGPT adapted to help learn English, create prompts, etc. They work more accurately in their area, because they have “built-in” instructions and can use included knowledge (in the form of files).

This notion is my proposal to systematize the reality, where manufacturers call similar or the same concepts differently: in ChatGPT it’s GPT (or GPT) Models, in Copilot: Agents, and in Gemini: Gems.

Fine-tuning

Fine-tuning is the process of further “teaching” a large AI model on specialized data (e.g., medical, legal, programming code) to suit a specific application. It’s like training a general employee to be an expert in a particular field – so he or she better understands the specific needs and language of the user. Because fine-tuning requires data collection and preparation, as well as a training process, it is often an expensive solution.

RAG (Retrieval-Augmented Generation).

RAG is a technique that combines text generation with a search of external knowledge sources (mainly vector databases), which allows the model to give more accurate answers because it is based on a specific source. In this way, the LLM can be better suited to provide answers using company or domain knowledge. This is often a simpler solution than fine-tuning, and also results in specialization, where, in addition, knowledge can be updated.

Modele dyfuzyjne

Modele dyfuzyjne to jeden z typów modeli generatywnych, które generują treść wychodząc od losowego szumu/zamglenia, stopniowo ujawniając finalny rezultat ich pracy. Można powiedzieć, że obrazy generowane przez modele dyfuzyjne są jak Buka w Muminkach, która wychodzi z ciemności.

Modele dyfuzyjne są najczęściej wykorzystywane do generowania grafiki, ale można je także spotkać w kontekście generowania dźwięku, a nawet tekstu (Large Language Diffusion Models).