What is Artificial Intelligence?

This is an attempt to explain Artificial Intelligence (commonly called “AI”) in the broad sense. I should point out that I have worked and coded in AI but I am not an expert. I ask for your indulgence if you have any comments on my text and your comments are welcome.

What is Artificial Intelligence?

Artificial intelligence (AI) is fundamentally about finding an exit in a maze.

At each intersection (straight ahead? left? right?) an evaluation function allows you to choose the path that best leads to the exit.

The evaluation function is important in determining the route and which exit to take, because the maze can have multiple exits. It's a human who chooses the evaluation function, not the AI.

Maze exploration is a cognitive problem that generalizes to daily life, war, politics, shopping in a supermarket, navigation in Google Maps, optical reading, writing documents, sport, robotics, etc.

Tools available in AI

To explore the maze, the AI has a number of tools.

Propositional algebra

Propositional algebra (https://en.wikipedia.org/wiki/Propositional_formula) allows you to manipulate logical propositions, as you would do with mathematical formulas. For example, with A = it is Thursday; B = it is not raining; C = Harry Potter practices Quidditch; D = Harry Potter visits Professor Dumbledore, we can write propositions like A and B => C; C => no D etc. and infer conclusions like A and C => B; A and D => not B and not C. Expert systems from the 1980's use this technique and manipulate tens of thousands of rules.

Neural networks

Neural networks work more by intuition and association of ideas, such as "Paris is to France what (...) is to Germany" or "man is to the king what woman is to (…)”. They are used to create huge databases of association of ideas called Large Language Models (LLM). Some have trillions of parameters.

Voice recognition, computer vision, machine translation, robotics, chatbots etc. use neural networks.

Other AI tools

There are other AI tools such as fuzzy logic, probabilistic reasoning or genetic variation algorithms but they are less impressive.

What is currently wrongly called AI consists solely of neural network applications. By contrast, we are reduced to calling “real” AI Artificial General Intelligence (AGI).

Historical general view

Historically, in the 1980s, the Japanese plan for the 5th Generation of Computers had launched an ephemeral fashion for AI with "expert systems" (https://en.wikipedia.org/wiki/History_of_artificial_intelligence#Money_returns:_Fifth_Generation_project) which gave rise to interesting applications in medicine. It was then that I started programming expert systems using different tools.

Recently, chatbots have revived the fashion for AI by imitating human speech, once again giving resounding applications in medicine, chemistry and pharmacy. Nothing has changed since Molière (“the lung” satire, https://www.encyclopedia.com/arts/educational-magazines/imaginary-invalid): medicine and research in general work a lot through the association of ideas.

Limits of the regulatory transparency obligation

The transparency obligation of the European Union's AI Act applies well to modal logic, much less well to applications based on neural networks (how do you explain an intuition?).

At a minimum, it should be declared how the associative and intuitive database was populated and what evaluation function was used for learning. It's might help when you meet a random youngster in Brooklyn, New York, to know if this young man has been educated using the knowledge base of the "Jets" or the knowledge base of the "Sharks" (https://en.wikipedia.org/wiki/West_Side_Story)