I know, we were all distracted by something else: in January 2021 OpenAI, the company founded by Elon Musk and backed by Microsoft, has launched its most ambitious project to date: a machine learning system called GIVE HER.
What does this artificial intelligence do? Simply put: it produces images starting from a simple text description. An example? You write “A camel with a book for a hat”, or “A tie made of forks” and he draws, creates, in short: he comes up with an image.
Now, Open AI has unveiled the second version of DALL-E, and things have gotten pretty damn serious: the resolution has gotten a lot better and the production time a lot lower. A typical exponential dynamic to which artificial intelligences have accustomed us, and above all will accustom us. Also because, if you remember, the head of Open AI just said that these machines maybe (maybe) they are becoming sentient.
A bit Wall-E, a bit Dali
DALL-E version one (a portmanteau of “Dali,” after the artist, and “WALL-E,” after the animated Disney character) was able to generate images and stitch together various photos into a collage, offering perspectives editable, and deducing elements of an image yourself, such as shadow effects.
A super good designer, with all the images in the world at his disposal, who interprets the requests and draws instantly.
“Unlike a 3D rendering engine, which requires precise input, DALL-E is often able to 'fill in the blanks' when the caption implies that the image should include a certain element is not explicitly stated,” wrote the team last year on the Open AI blog.
How does AI that draws on demand work?
DALL-E was not intended to be a commercial product, and the OpenAI team decided it would be for the best. The researchers "prevented" him from creating sexual or political images, for example, to prevent the system from being used to generate misinformation. The same goes for specific names: it is impossible to ask him, for example, “draw Frank Sinatra playing tennis with a polar bear”.
DALL-E 1, essentially, takes advantage of the incredible GPT-3 language interpreter (here I talk about it more in depth, if you're interested) and draws what he understands plus what he "intuits".
Unlike the first version, which was available to everyone on the OpenAI website, this is only available for testing by a limited number of partners: I await a few days for a response from Open AI to join the group.