Today, when we talk about AI, we are most often talking about generative AI. This AI, which is capable of creating original content based on what it has learned, has become very popular with the general public thanks to GPT Chat, and it is certainly the most accessible form of AI for the general public, with obvious use cases.
But AI is a rich and fairly old field: The first work on the subject dates back to the 1950s, but it has remained a complicated subject reserved for specialists and a little abstruse for the average person, even though ten years ago IBM Watson, a platform for hybrid artificial intelligence that combines NLP, machine learning, automation and symbolic reasoning (What is cognitive computing ? IBM Watson as an example), attracted attention.
But for AI to become a social issue, it will logically have to enter people’s homes (thank you ChatGPT), but there is a downside to this: by dint of only seeing this type of tool, we forget that there are many other forms of AI, to such an extent that some fear that if there is a Generative AI bubble, its explosion could hurt the entire sector despite the fact that there are mature types of AI with an established business that have been proving themselves for a long time (AI heading for an economic dead end?)
Generative AI can’t do everything, and that’s a good thing, it’s not designed to do everything. The other day I was making a list of potential uses of AI for business governance, and it was an opportunity to realize that each case corresponded to a type of AI that had its own mode of operation.
Not being an expert in the technologies in the “deep” sense of the term, it was an opportunity to clarify things and improve my knowledge in the field. But I also realized that many decision-makers and users used the generic term AI without really knowing what lay behind it, the variety and extent of the field of possibilities, what they could do with it…and even without knowing that they had sometimes been using it for years without knowing it. A bit like talking about cutlery when you have in mind a fork when what you need is a knife…
So I thought to myself that if I have to have this job for me, I might as well share it with those who are asking themselves the same questions. A popular science article that experts will no doubt find not thorough enough, where they will find details that are open to discussion but which I hope will be useful to all those who are primarily concerned with use cases and want to understand what is behind it all.
If we want to understand what AI does (and what it doesn’t do), we need to take a step back and lay the foundations, which is what we’re going to do here.
In short :
- Artificial intelligence encompasses a variety of forms and purposes, far beyond the sole generative AI popularized by tools such as ChatGPT: each type of AI has a distinct objective, whether it is to predict, describe, recommend, interact or act autonomously.
- AI operates using different learning methods (supervised, unsupervised, reinforcement, symbolic or deep learning), each adapted to specific needs and determining the type of task that the AI can accomplish.
- In addition to these methods, there are complementary technological building blocks (NLP, vision, OCR, RPA, GANs, diffusion models, etc.) that make it possible to process varied data and build concrete use cases in all sectors of activity.
- Some concepts need to be clarified: a tool like ChatGPT can understand instructions without examples or adapt if it is given a few, and there are several ways for a machine to produce a text or an image.
- To understand AI in a relevant way, it is essential to distinguish its objectives, learning methods and tools, in order to evaluate the real uses and concrete added value of the solutions that claim to use it.
The main types of AI: what AI seeks to do
We can start by classifying AI according to its intended purpose. What does it seek to accomplish? What can we use it for?
A table is better for this than big words.
| Type of AI | Purpose | Examples |
|---|---|---|
| Predictive AI | Predict what will happen | Sales forecasting, fraud detection |
| Descriptive AI | Explain what happened | Sentiment analysis, customer clustering |
| Generative AI | Create content | ChatGPT, DALL·E, Midjourney |
| Prescriptive AI | Recommend an action | Product recommendation, GPS |
| Conversational AI | Engage in dialog with humans | Chatbots, voice assistants |
| Cognitive AI | Simulate human reasoning | Expert systems, rule engines |
| Explainable AI | Make decisions understandable | SHAP, LIME, model audits |
| Adaptive AI | Learn in real time | Autonomous agents, embedded AI |
| Embedded AI | Operate locally, without connection | Connected objects, intelligent cameras |
| Agentic AI | Act autonomously in an environment | AutoGPT, professional co-pilots, proactive assistants |
Well yes, when you use a GPS to find your way there is AI in it…
A little clarification on artificial intelligence since it will surely be the one that will quickly transform our daily lives and that we tend to associate too much with advanced generative AI.
In fact, these AIs are not themselves generative models, but they rely heavily on these models (such as GPT) to understand, reason, generate and interact. In other words: generative models are a key building block that makes agents more intelligent, more adaptive and more autonomous, but these AIs cannot be reduced to these models.
What about AGI, artificial general intelligence? It doesn’t exist yet, so we’ll see about that later.
Now the question is how they work to achieve this, which explains why each has its own playing field and limitations.
The main learning methods: how AI learns
I’m not keen on the term “intelligence” because what distinguishes AI above all is its ability to learn. And it does so in a variety of ways.
| Method | Idea | Example |
| Supervised learning | AI learns with labeled data | Churn prediction, image recognition, GPT, LLaMA, Copilot |
| Unsupervised learning | AI discovers hidden groups or patterns | Customer segmentation, clustering |
| Reinforcement learning | AI learns by trial and error AI learns by trial and error | Chess game, robotics |
| Symbolic learning | AI applies logical rules | Expert systems, inference engines |
| Deep learning | AI learns via deep neural networks | Speech recognition, vision, GPT, LLaMA, Copilot |
| Few-shot / Zero-shot | AI generalizes with few or no examples | GPT, LLaMA, Copilot |
It may be surprising to see Generative AI in the “few shots” category (at least it was for me), but it is important to understand that they go through several stages.
A tool like ChatGPT has been pre-trained on billions of texts (self-supervised learning, via deep learning): this is called pretraining: it learns the language, the structures, the concepts.
But then, once trained, it can be used in few-shot or zero-shot.
Zero-shot because if you ask it a question without an example, it understands and responds directly, and few-shotbecause if you give it a few examples, it will adapt its response to the context. This is the case, for example, when you “train” an AI to adopt a given style.
It is therefore in use that the model is few-shot/zero-shot, not in its training itself.
What AI manipulates: the technological building blocks
We have talked about the end goal, about how AI acquires the potential (knowledge) necessary to achieve it, but there remains the question of how it uses this potential.
It will mobilize it in technological building blocks, each of which has its own purpose, its own functioning, and its own business use cases.
| Technology | Main use | Example | Business use case |
| NLP (natural language processing) | Understanding texts | Comment analysis | Customer support, e-reputation, monitoring |
| OCR (Optical Character Recognition) | Reading text in an image | Scanning an invoice | Document processing, archiving |
| Computer vision | Recognize objects and shapes | Defect detection | Quality control, security, in-store counting |
| LLM (Large Language Models) | Generate and understand natural language | ChatGPT, Copilot | Writing assistance, intelligent agents |
| RPA (Robotic Process Automation) | Automate repetitive tasks | Data entry | Back office, HR, finance |
| GANs (Generative Adversarial Networks) | Generate images | Artistic creation | Advertising, design, prototyping |
| Diffusion models | Create realistic visuals | Midjourney, DALL·E | Marketing illustration, design assistance |
| Speech-to-text | Transcribe voice | Automatic subtitling | Media, healthcare, legal transcription |
| Chatbots | Automatic dialog | Automated FAQ | Customer service, internal support |
| Expert systems | Logical reasoning via rules | Technical diagnostics | Maintenance, taxation, legal |
| Recommendation systems | Suggest a choice | Movie suggestions | E-commerce, HR, online education |
| Explainability tools | Explaining AI decisions | SHAP, LIME | Compliance, model auditing, insurance |
| Embedded AI | Operating locally in real time | Smart camera | Drones, connected objects, automotive |
| NLG (Natural Language Generation) | Generating text automatically | Automatic reports | Finance, insurance, sports |
If you are wondering what differentiates LLM from NLG, know that I have also asked myself the question.
Let’s just say that LLM such as GPT 4 can do content generation (fairly obvious to understand) but that you can do NLG without LLM using simpler systems (rules, statistical models, etc.).
Let’s take the example of an automatically generated weather report. We will use meteorological data contained in a database and then a rules engine will reposition them in a sentence to give something like “This morning, the sky will be partially cloudy with 18°C. In the afternoon, temperatures will rise to 23°C with a moderate southwest wind.”
The same goes for GANs and diffusion models. So, if I understood correctly, these are two technologies that allow AI to generate images, but they work in very different ways.
GANs are a bit like two artists competing with each other: the first one invents an image, the second one has to say whether it is true or false, the first one tries to fool the second one, and they improve with each round, resulting in increasingly realistic images.
Diffusion models, on the other hand, start from a random scribble (like noise) and gradually remove this noise until a clear image appears. It is slower, but often more stable and more accurate.
Today, diffusion models dominate the AI world (DALL·E, Midjourney, etc.), but ChatGPT’s new image generator has recently changed the game by using “an integrated multimodal predictive approach, generating content via a single neural network trained simultaneously on text, images, and audio” that is totally different from the iterative denoising process inherent to diffusion models.
Bottom line (without the jargon)
When we think of a “product”, many AI solutions will in fact integrate different types of AI, and therefore different learning models serving several building blocks to meet a given use case, especially if it is complex.
There is no such thing as a single AI, but several types, with different objectives, and what each one does depends on how it learns.
Each technology is therefore a specific building block, serving a specific purpose.
Just because a business “does AI” doesn’t necessarily mean it’s innovative. What matters is what it uses it for, how it uses it, and what concrete added value it brings.
Next time you hear about AI, you will (hopefully) know how to ask the right question: what type of AI is it, and what is it for?
Illustration: generated by AI via ChatGPT / DALL·E
