Presentation of DALL-E
The January 5, 2021, the openai.com website put forward an AI able to translate a word or a sentence into a picture.
Say like this, it seems easy to do... But it is a really challenge !
For exemple, if you give this sentence :
"An armchair in the shape of an avocado."
You can have this beautiful pictures :
Source : Website of openai.com
This extraordinary AI is called DALL-E, a contraction of the name of the surrealist artist Salvador Dali and the robot WALL-E.
Github official repo is here
DALL-E is based on the GPT-3 model, which had a bad buzz this year because it suggested to a patient to commit suicide when used as a medical chatbot by Nabla.
The GPT-3 for Generative Pre-trained Transformer 3 is an autoregressive language model that uses deep learning to produce human-like text.
It is the successor of GPT-2, and he the third-generation language prediction model who was created OpenAI.
But openai.com had warned Nabla on the use of the GPT-3 model in medical, and had say to Nabla :
"Health is in the high-stakes category because people rely on accurate medical information to make life-and-death decisions, and mistakes in this area could result in serious harm."
But with DALL-E, the risks would not be of this magnitude because the use here of GPT-3 is different.
The GPT-3 Model
We already know that with this model, OpenAI is for example able to generate code from a sentence in English.
OpenAI has shown it for example with the presentation of Codex accessible in private beta since August 2021.
He has been trained and is actually able to program in many langage like CSS, JSX, Python and others.
Incredibly powerful, this model has also managed to write a dissertation in less than twenty minutes against three days for a student.
Here is an example of many use of GPT-3 if you want to see more
But as we have already seen, this model is not without its flaws...
Indeed it can be dangerous, and can for example develop conspiracy theories on Reddit.
But this is normal because the AI has been trained on hundreds of billions and he necessarily used resources present on the web, which tends to amplify the racist and sexist bias of the model. But of course researchers are already working on this problem.
With this model, many technological advances can be made in the field of AI. And maybe one day machines will really be able to understand everything we say...
Top comments (0)