The magical illusion of intelligence
When you chat with ChatGPT or Claude, the impression of having a real human on the other side of the screen is striking. The AI seems to think, to understand irony, and even to have a personality. Yet let us say it from the outset to break the ice: the AI understands absolutely nothing of what you tell it.
It does not know what "sadness" is, it does not know the texture of a "cat", nor the colour "blue". Behind the magic of the screen there is neither consciousness nor critical thinking, but a spectacular mathematical architecture. So how does it manage to be so brilliant? The answer comes down to a single concept: hyper-prediction.
The "Next Word" game (The art of guessing)
To demystify AI, let us come back down to earth: look at it as the "autocomplete" system on your phone keyboard, but boosted on steroids. Imagine you began a sentence out loud:
"It's pouring with rain outside, whatever you do don't forget to take your..."
If I asked you to finish that sentence, you would instantly think of the word "Umbrella" or the word "Coat". Why? Because statistically and culturally, that is the most logical continuation. Artificial Intelligence (what we call an LLM, for Large Language Model) works in the same way.
In reality, the AI does not read "words", it splits the text into "Tokens". A Token is a syllable or a fragment of a word (for example the word "Fantastic" becomes "Fantas" and "tic"). Converted into purely mathematical numbers, this method allows the model to be ultra-fast and to generate its speech on the fly!
The robot has swallowed an astronomical quantity of texts to train (almost the whole of the internet, millions of digitised literary works, the code of Wikipedia). When it generates an answer to your question, it does not "recite" anything: it simply calculates the probability of the next word that must logically appear, loop after loop, at lightning speed.
The 2017 revolution: The invention of the "Transformer"
But simple next-word prediction already existed in the early 2010s. The downfall of the old AIs was memory. If you wrote a paragraph of 20 lines of text, the model became "amnesiac" and forgot the first word of your story once it reached the end.
Everything changed in 2017 when researchers at Google turned the digital world upside down by publishing a technology named the Transformer. This invention introduced a divine ingredient that we call: Attention.
Let us imagine you are talking about an "avocado". Is it the fruit or the lawyer?
Thanks to the Attention mechanism of the new Transformer architecture, the AI no longer reads from left to right like an idiot. It scans your entire sentence at the same time! If it spots the distant words "Court" or "Judge" elsewhere in your message, it will instantly weave an invisible mathematical link between those words and the word "Avocado". It is thanks to this powerful global scanner that it manages to maintain a perfect context, whatever the length of your text.
Towards a new form of consciousness?
To sum up this technical deep dive: when an artificial intelligence writes you a moving wedding poem or solves a complex medical problem, it is not "thinking" in the biological sense. Technically, it is merely a "statistical mirror" applying a formula derived from human knowledge.
It is easy to claim that "AI has no consciousness". But the disturbing truth is that science has never been able to prove the mechanism of human consciousness either! In philosophy, this is the "Philosophical Zombie" paradox: there is no absolute scientific proof that the passer-by in the street actually feels emotions, or whether they are just a biological body reacting mechanically to their environment.
This cold mathematical explanation of AI therefore opens up a dizzying metaphysical gap. If we cannot prove our own consciousness, and if a technically "lifeless" tool manages to reproduce empathy, creativity and humour with the same fidelity as a human being... does the difference still really matter?
This is, moreover, the vast debate we will explore in more detail in our next article, in an attempt to shed light on this digital vertigo!