Where generative AI fits in the AI landscape¶

Purpose of this document¶

This document situates generative AI within the wider field of artificial intelligence. Its goal is to make one idea clear: generative AI is not "artificial intelligence" as a whole, but one branch among several, with its own mechanisms, strengths and limits.

This framing matters in a professional setting. Confusing generative AI with the entire field leads to two recurring mistakes: expecting a text generator to behave like a recognition system or a calculator, and attributing to a model capabilities it does not actually have. Understanding where each technology sits helps choose the right tool for a given problem.

This page is the entry point to the section. The next pages then look more closely at generative AI itself and at the architectures used to deploy it in an organisation.

1. Artificial intelligence is a field, not a single technology¶

Artificial intelligence is a broad field made up of sub-domains whose structures and objectives differ significantly. Supervised learning, unsupervised learning, generative models, reinforcement learning and hybrid statistical methods all answer distinct goals and should not be confused.

An algorithm designed to generate text does not work at all like a face-recognition model or a product-recommendation system. This diversity means we should avoid grouping everything under a single definition, and above all avoid attributing to a model capabilities it does not possess. Speaking of "AI" without saying which kind is often a source of misunderstanding.

2. A useful way to picture it: nested levels¶

A simple way to place generative AI is to see the field as a set of nested levels, from the most general to the most specific:

Artificial intelligence (AI)
└─ Machine learning
   └─ Deep learning
      └─ Generative AI

Artificial intelligence is the overall goal: getting a machine to carry out tasks that, for a human, call on analysis, recognition, decision-making or language.
Machine learning is a family of methods in which a system learns patterns from data instead of being programmed with explicit rules.
Deep learning is a branch of machine learning based on neural networks with several successive layers. Its distinctive feature is that it learns to extract the relevant characteristics directly from the raw data, whereas in classic machine learning those characteristics often have to be defined by an expert.
Generative AI is a specific deep-learning approach, dedicated to producing content rather than only classifying or predicting.

Each level is a special case of the one above it. Generative AI is therefore deep learning, which is machine learning, which is artificial intelligence, but the reverse is not true: most AI is not generative.

3. How a neural network learns¶

Deep learning relies on artificial neural networks made up of several successive layers. Each layer gradually transforms the input data into more and more abstract representations, which lets the model capture complex patterns and make precise predictions.

In practice, training rests on an iterative optimisation process. The model is first given a set of annotated data, for example images of cats and dogs with their corresponding label. The network makes an initial prediction, often wrong at the start, and compares its result with the ground truth. Through an algorithm such as gradient descent, the weights of the connections between neurons are adjusted little by little so as to reduce the error. Repeated thousands or even millions of times, this process lets the model refine its parameters and generalise its predictions to new data.

4. Different architectures for different problems¶

The architecture of a neural network varies according to the need.

Convolutional networks (CNNs) are optimised for image processing, detecting visual patterns such as edges and textures.
Recurrent networks (RNNs, LSTMs) read a sequence in a fairly linear way and are suited to data such as text or time series.
Transformers are a more recent architecture. Unlike recurrent networks, they do not read a sequence step by step: they rely on a mechanism called attention, which weighs the relative importance of the elements against one another. This architecture has largely replaced recurrent networks for language and now underpins most generative models.

These architectures are effective, but they often demand significant computing power and large volumes of data to be trained well. This is why optimisations such as transfer learning (reusing a pre-trained model) or distributed training on GPUs or TPUs are widely used in industry to make these models more accessible and more efficient.

5. What makes generative AI distinct¶

With the arrival of generative AI, a new paradigm emerged. Unlike classic models, which only classify or predict, these systems are able to create content: text, images, audio or even code.

They rely on advanced architectures such as Transformers, which let them analyse and generate content by drawing on very large bodies of data and sophisticated probability models. These models do not merely give back what they have learned: they produce plausible responses according to the context, which gives an impression of creativity and intelligence. A notable aspect of generative AI is that it turns the "meaning" of the input into a vector, or a set of vectors, in order to respond in a relevant way.

This is the key point of this page: generative AI is a specific architecture of AI that uses a deep-learning approach. It is powerful, but it remains one branch of the field.

6. Patterns built on top of generative models¶

Several approaches have been developed to extend generative models and adapt them to precise uses. They are not competing models but patterns layered on top of a generative model, and they can be combined.

RAG (Retrieval-Augmented Generation) combines an external knowledge base with a generative AI, letting it look up up-to-date information instead of relying only on what it memorised during training.
ReAct (Reasoning + Acting), used for advanced conversational agents, lets a model reason and interact with tools according to the instructions it receives.
MoE (Mixture of Experts) activates only part of the network depending on the problem to be handled, which improves efficiency and reduces resource consumption.
Chain of Thought (CoT) helps a model structure its reasoning by detailing the steps of its reflection, making its responses more reliable and easier to explain.

These approaches are not exclusive and can be combined to optimise a model's performance according to the need.

7. You do not always need AI¶

It is worth remembering that it is not always necessary to use AI to solve a problem. In many cases, simpler statistical approaches are enough to obtain precise and interpretable results.

For example, analysing the distribution of data often rests on the assumption that the data follows a normal distribution, also called a Gaussian distribution. Checking this assumption simplifies many predictive analyses. A method such as the Henry line tests whether a variable follows a normal distribution by comparing the observed data with a theoretical straight line on a probability plot. These statistical principles are everywhere in concrete applications, whether for detecting anomalies in series of measurements, evaluating the performance of a system or managing financial risk.

8. AI is an extension of mathematics and statistics¶

Ultimately, AI is a powerful tool, but it is an extension of classic mathematics and statistics. Behind its sometimes mystified image, it rests on logical principles and well-established probabilistic models. Contrary to the common idea that AI works as an autonomous, independent "black box", it is above all an assembly of techniques adapted to specific problems.

That said, some AI systems are genuinely black boxes, in the sense that for many models, particularly in deep learning, it is difficult to explain precisely why a given decision was made. This stems from the probabilistic nature of how they work.

Conclusion¶

The pragmatic approach is therefore to understand what AI can do, but also what it cannot do, and to choose the appropriate tools according to the problem to be solved.

Generative AI is one of the most visible branches of this field today, but it remains a branch among several. Keeping this map in mind, AI as a whole, machine learning, deep learning, then generative AI and the patterns built on top of it, helps set realistic expectations and make sound technology choices.