We live in a time when a machine can learn from books.
Realistis' Artificial Intelligence's neural networks is currently learning from the 7 volumes of "À la recherche du temps perdu" (Proust, published from 1913 to 1927).
Every 1000 steps (over 116050), you can ask it what it learnt by generating some Proust like text from its global memory (more below).
Proust is famous for his sentence length and his main theme: involuntary memory.
100 years later, a machine applies involuntary memory to Proust's books.
Today, the machine tries to decipher what Proust's words could mean.
A machine can "learn" word by word, and character by character.
The later is the most creative one.
After 1000 steps, the machine delivers a very-machine like text:
At 2000 steps, it learned what words were but it is very creative in building them
-Ext trop nouvelle d'un dîner charmant, de son brat et sont profond.
Il s'arrêtait nous nous fichait de cette idée et aux "focuttes encore qu'il y eût promemer à la personne qui était dû ma chambre d
After all, if "bras" and "rat" are words, why not "brat"?
At step 4000, it improved on word building:
Oh! ça qu'elle figurait de se connaître que j'avais pu l'empêcher de l'ébrouver les amies qui trouvaient d'abord en la peu de courses, non pas au moment de ce que l'arrêté et un coup d'intelligent
At step 11000, it's getting clearer:
jeune que les faveurs étaient au milieu de son intelligence, tant de présence pour moi, en arrivant à la grande instillance de celle-ci comme j'avais promés de tout sens du plaisir à nous aimer de fai
And so on (it takes days to achieve all the 116050 steps).
The most impressive is that the meaning is not always present but the style and the rythme are here.
AI will be more present in the artistic process. It opens new paths for artists. Not to replace artists, but as a new brush with its technique.
This example is just for text, but look at what AI can do from a photo:
You might think the second image is a photo filter, the third image was produced from the original, not from the second image.
Artistic perception is a filter to reality. This "filter" interprets what we see through a set of free creative expressions and abilities.
The image feeded to the machine is a base to express art, the same way what we call reality impresses our retina and we interpret what we partially see as lines and colors through our perception, technique and tools.
Feeding the last generated result above as the original extracts the essence of the artistic perception. If you feed the machine with its own generated productions, as if it paints what it painted, you're getting more of the core artistic perception:
Is it still perceived as a photo filter or is it abstract?
But dimensions of "windows" on the right, composition and perspective is still relative to the original, nothing upside-down, or disformed. This is AI, it keeps the main structure as much as the feeded original is structured. It doesn't brake perceived rules.
Human is in control for now with these AI because we're in control of what feeds them and we ask them to generate innocent productions. When we see the random creativity it expresses with data, maybe there's risk with AI manipulating sensitive data or feeding another AI producing something harmful, the same way an animal trained to hurt adults may kill a child. Ultimately, it's always human's responsibility.
PS: Only the French will get the pun in the title, sorry, I couldn't resist.