AI Concept

Notions

There are notions that are important to understand in relation with AI technologies. Here is some of them briefly summarize.

Responsible AI

Responsible AI is a set of principles to ensure AI systems without causing harm :

• Fairness (no discrimination) ;
• Reliability & Safety (minimize risks) ;
• Privacy & Security (protect user data) ;
• Inclusiveness (accessible to everyone) ;
• Transparency (decisions must be understandable) ;
• Accountability (clear ownership when something goes wrong).

Large Language Model (LLM)

A Large Language Model (LLM) is the engine behind an AI application such as ChatGPT. In this case, the engine powering ChatGPT is GPT-4 (or GPT-4o, previously), which is the LLM used by the application.

Azure AI Foundry is a service that allows you to choose which Large Language Model (LLM) you want to use.

Small Language Model (SLM)

A Small Language Model (SLM) is a lightweight version of an LLM, trained for a specific topic or context. It has fewer parameters, making it faster, more portable, and less expensive to fine-tune than an LLM.

Multi-Modal Models

A Multi-Modal Model is a model capable of processing multiple types of data together, such as text and images. It is divided into foundation models (large pre-trained, versatile) and adaptive models (fine-tuned for a specific task).

Tokenization

Tokenization is the process of breaking down text into smaller units called tokens (words or fragments), each converted into a unique number that the model can process.

Embeddings

Embeddings are a numerical representation of tokens in a vector space. They allow the model to understand semantic similarities between words based on their position relative to each other.

Fine-tuning

Fine-tuning is the process of adapting a pre-trained foundation model to a specific use case by re-training it on a targeted dataset, without starting from scratch.

Machine Learning (ML)

Machine Learning is an approach where a model learns from historical data to make predictions. It has two phases :

1. Training : the model learns relationships between features and labels ;
2. Inferencing : the model predicts on new data using the learned function y = f(x)). There are two main types : Supervised Learning (with labeled data, for regression or classification) and Unsupervised Learning (without labels, for clustering.

Deep Learning (DNN)

Deep Learning is an advanced form of ML that mimics human neurons through artificial neural networks. A Deep Neural Network (DNN) consists of multiple layers that progressively detect more complex patterns. Each neuron receives an input, processes it via a weighted sum, and an activation function decides whether to pass the signal to the next layer.

Natural Language Processing (NLP)

Natural Language Processing (NLP) is the ability of a system to understand, interpret and respond to human language. It covers tasks such as sentiment analysis, key phrase extraction, translation, summarization, and conversational AI.

Computer Vision (CNN)

Computer Vision is the ability of a model to interpret and analyze visual data from images or videos. It covers tasks such as image classification, object detection, captioning, tagging, and spatial analysis. It relies on CNN (Convolutional Neural Networks) which apply filters to detect edges, textures and patterns progressively.

Retrieval-Augmented Generation (RAG)

Retrieval-Augmented Generation (RAG) is a technique that combines an LLM with an external knowledge source (documents, database) to generate more accurate and contextualised responses, without re-training the model. Otherwise, it will be a fine-tuning.

Anomaly Detection

Anomaly Detection is the ability of a model to identify unusual behaviors or patterns in data that deviate from the norm. It is particularly useful in AIOps to automatically detect infrastructure issues such as abnormal CPU spikes, unusual latency, or unexpected traffic patterns, before they become incidents.

Agent AI

An Agent AI is an AI system capable of autonomously making decisions and executing actions without human intervention at each step. It perceives its environment, plans a sequence of actions, and uses tools or external systems to achieve a goal. It is closely related to MCP, which standardizes how an Agent AI communicates with those external systems.

Example with Agent AI on N8n, where you can connect external tool such as a model, an MCP Server, etc.

Model Context Protocol (MCP)

A Model Context Protocol (MCP) is a protocol that standardizes communication between Large Language Models (LLMs) trough an Agent AI and external systems , such as ITSM tools (like ServiceNow), Kubernetes clusters, and more.

You can use an MCP client , for example, Continue.dev in your IDE (like VS Code) and then configure MCP servers, such as your Kubernetes cluster, to enable your LLM to interact with these systems.

Low-Code / No-Code

Low-Code / No-Code refers to platforms that allow building applications or AI workflows with little to no programming knowledge, through visual interfaces and drag-and-drop components. Low-Code still requires some coding for advanced customization, while No-Code requires none. Example with n8n.

Prompt Engineering

Prompt engineering is the practice of crafting clear, structured instructions to guide AI models toward producing optimal outputs. Based on my experience :

• Rule 1 : DON'T write too much text at once; if possible, break the work into sequences. Otherwise, use functions of the app.
• Rule 2 : DON'T ask the AI to write prompt engineering instructions for another AI, it creates an infinite loop and wastes time.
• Rule 3 : 🎯 Use emojis to clarify and sequence your prompt, helping the AI recognise when there's a new instruction.

AIOps / MLOps

AIOps and MLOps are methodologies like DevOps, oriented for AI :

• AIOps : application of AI and ML to automate and enhance IT operations, such as anomaly detection, event correlation, root cause analysis, and incident remediation.
• MLOps : set of practices that combines ML and DevOps to streamline the deployment, monitoring, and maintenance of ML models in production. It ensures models remain accurate and reliable over time by automating retraining, versioning, and performance tracking.