Generative AI for humans: basic terms to understand it.

Generative Artificial Intelligence (Gen AI) is transforming the way we create, work, and communicate. But to be able to speak appropriately about this technology—and, much more importantly, to be able to make informed decisions about its use—we need to understand how it works. And that starts, like everything else, with language. This glossary was created with a very clear objective: to make the basic terms surrounding Gen AI accessible to everyone, because it's not about learning for the sake of learning, but rather about offering concrete tools so more people can actively participate in the conversations that are shaping the immediate future.

Unlike many traditional glossaries that organize concepts alphabetically, at Mottum we decided to organize them by thematic categories. This structure will allow you to understand the terms as parts of a larger system, not as isolated definitions. Welcome, then, to this tour of the key concepts of Gen AI. They are explained clearly, with examples, and without unnecessary technical jargon. Because we believe that understanding the technology is the first step to using it well.

Fundamentals

• LLM (Large Language Model): This is a language model, trained with enormous amounts of text, capable of understanding and generating natural language text. They use architectures such as transformers, which allow the model to predict the next word, thanks to the attention mechanism. Transformers analyze all the words in a text at once and understand the relationship between them, regardless of the distance between them.
• SLM (Small Language Model): This is a language model capable of understanding and generating natural language text. They are smaller in size and context than LLMs, making them more compact and computationally efficient.
• Prompt: This is the natural language instruction given to a generative model to guide its output. There are different methodologies and frameworks to get the most out of prompts, such as RICCE (Role, Instructions, Context, Constraints, Examples).
• Token: The smallest unit of text (word, part of a word, or symbol) processed by AI models. These are measured as input tokens (the tokens the model processes based on our prompt) and output tokens (the tokens the model generates when responding to us).
• Embedding: Embedding converts a token or text into a numerical vector that captures its semantic meaning. By vectorizing it, we can capture semantic and contextual information about each token.
• Assistant: This is software powered by artificial intelligence that encompasses functionalities such as: Machine learning, natural language processing, reasoning, designed to assist the user in tasks such as answering questions, managing schedules, searching for information, controlling devices, etc. Example: OpenAI's ChatGPT is an example of an assistant. Not to be confused with LLMs or SLMs, which are the models used to generate text, an assistant is the application that is born from these capabilities, allowing us to engage in conversation.
• Corpus: In the field of artificial intelligence, it is a structured collection of data used to train and evaluate learning models.

Control Parameters

• Temperature: Temperature adjusts the degree of randomness in token selection. This allows us to control how creative or conservative the model's responses are. Example: At a low temperature (e.g., 0.2), the model will tend to repeat the most likely option; at a high temperature (e.g., 1.0), it will explore more varied options.
• Top-k Sampling: The top-k sampling technique limits generation to the k most likely candidates at each step (in the generation of each word). This directly helps improve response consistency.

Key Techniques

• Training: This is the process by which an algorithm "learns" to solve a task such as classification, regression, or text generation, based on data we provide to the model. In short, training a model is teaching it, through examples and numerical adjustments, to detect patterns in data and then predict or generate useful results in new situations.
• Fine-Tuning: Fine-tuning consists of using a previously trained LLM and retraining it with a more specific dataset, allowing the model to specialize in a specific topic or style. In this way, we are modifying its behavior so that it is better able to respond in specific contexts. Example: A medical company may need to fine-tune a model, as in the case of GPT-3.5, with medical texts (medical records, clinical guidelines, protocols). This allows them to ensure that the model speaks more accurately in the healthcare field; that is, it is less likely to hallucinate and is better able to understand medical terminology.
• RAG (Retrieval-Augmented Generation): Retrieval-Augmented Generation consists of a two-phase process when answering a LLM:

1. Retrieval: Based on the user's prompt, external documents are consulted using dense or keyword-based searches to find relevant texts. When a user asks a question, information is searched in the external document database to find the most relevant texts to answer the user's query. This search is not based on word matching, but rather understands the meaning of what the user is asking and compares it with the content of the stored documents. These documents are prepared in advance by extracting their numerical representations; this makes it easier to quickly find what most closely relates to the question.
2. Generation: The LLM receives the prompt along with the retrieved fragments and generates a final answer, based on that external context.

• TAG (Table-Augmented Generation): Specifically uses structured data from tables to guide the generation process. In contrast to RAG, this method is more optimal when precision and accuracy are required, such as when working with numerical or categorical data organized in tabular format. This process consists of three phases:

1. Table extraction: When the user sends a prompt, the system identifies the relevant structured data sources and extracts the rows, columns, or cells that meet the request.
2. Data fusion: Once the relevant information has been extracted from the tables, the model receives the raw data and integrates it into the generation process. It does not limit itself to displaying values, but uses them to construct a coherent text that explains the information.
3. Generation step: Finally, the generative model produces a fluid and cohesive text based on the structured data. The model ensures that the generated response corresponds exactly to the original data, maintaining consistency in the numbers extracted from the table.

Cognitive Capabilities

• Chain-of-Thought: CoT is an artificial intelligence technique that helps language models (LLMs) solve complex problems by structuring their reasoning into logical steps. Instead of directly providing an answer, the model explains its intermediate steps to reach the final conclusion, so that the model itself elaborates on the answer as it generates it.
• Inference: Inference in machine learning is the process by which a previously trained model applies what it has learned to new data to generate predictions or content.
• AI Agents: These are LLM-based systems that execute autonomous workflows, capable of invoking external tools to execute an action. The difference between a conventional generative AI chatbot and an agent is that the latter is capable of executing actions beyond responding to a prompt. By integrating with other platforms, they are capable of executing complete workflows that would require human intervention if a conventional chatbot were used. An agent uses LLMs, but it isn't necessarily a chatbot. We could create an agent that, based on a photo of an invoice, fills in a row in an Excel spreadsheet with the invoice data. Example: An agent who, upon receiving a "Book a cheap flight to Barcelona on Friday morning," consults airline APIs, compares prices with the user, and completes the reservation.
• Multi-Agents: Multi-agents consist of systems that combine several agents working collectively to complete a task for a user or another system. All agents are under the command of an orchestrator, which is responsible for calling each agent in the appropriate order to complete the task specified by the user or system. Example: A smart home could have the following agents working together to meet the user's needs: a Lighting agent, which controls smart light bulbs to turn on or off, and regulate intensity and color; a HVAC agent, which adjusts the heating or air conditioning to the desired temperature; a Security agent, which monitors door and window sensors and manages the burglar alarm; and an Entertainment agent, which manages music, TV, and connected speakers. All of these agents would be under the command of the orchestrator; the orchestrator receives the overall command from the user and calls each agent in the appropriate order.

Risks and Limits

• Hallucination: A hallucination occurs when a model generates plausible information that is false or unfounded in the training data.
• Context Window: The context window is the number of tokens that an LLM can "see" or take into account when generating each response. When the total number of tokens in the input exceeds the maximum allowed, the model discards the oldest tokens until the count falls within the limit. This is known as token truncation.

Understanding Generative Artificial Intelligence is not just a technical issue, but a strategic necessity for any organization that wants to innovate wisely. This glossary aims to be a clear and accessible gateway to the world of Generative Artificial Intelligence, not only to understand it, but also to begin using it consciously, creatively, and in alignment with business objectives. At Mottum, we support teams and companies on this journey through our AI Training programs: practical, tailor-made training so that your employees not only know what AI is, but also how to apply it in their daily work in an ethical, efficient, and productive manner. Do you want your team to speak the language of AI and use it with real impact? Let's talk.