ChatGPT the computer robot that writes texts… Bluffing and a little disturbing!
Developed by the Californian start-up OpenAI, the ChatGPT software is designed to generate texts. Particularly powerful, managed by artificial intelligence, this computer robot is capable of producing surprisingly coherent and successful writing.
Cooking recipe, poem, scientific concept, political speech, dissertation, children’s story, press article… The performance of ChatGPT, a text generator based on GPT-3 technology, managed by artificial intelligence, is as much fun as they fascinate. This online platform answers your questions and executes your instructions by writing content. Its latest version, launched at the end of November, is particularly successful.
After the success of DALL-E 2, an online software that generates a computer-generated image as realistic as a real photo, based on a few descriptive words, again thanks to artificial intelligence, the Californian start-up OpenAI, co-founded in 2015 by Elon Musk, the boss of Tesla (who has since left the company), launched ChatGPT at the end of November, a conversational robot (chatbot), capable of answering questions and carrying out a real exchange. with a human interlocutor.
This version is currently freely available on the internet. All you have to do is fill in your email address to access it.
Specific training and already prohibitions
In an article, OpenAI explains how ChatGPT works and has been trained. “We trained this model using Reinforcement Learning from Human Feedback (RLHF), using the same methods as InstructGPT (human-in-the-loop models), but with slight differences in the configuration of data collection”. This technique gives more relevant results which pleasantly surprised the testers of the weekend. Note that to train these models, OpenAI uses HPC instances on Microsoft’s Azure cloud.
ChatGPT tries not to fall into the trap of recent chatbot tests that do not filter out bias, hateful, conspiratorial, offensive, discriminatory, etc. (like Meta which recently unplugged its AI Galactica). But if OpenAI recognizes efforts in this area, it emphasizes that the results of the chatbot are not infallible and lacks critical thinking and nuance. Some sites have already announced the prohibition of comments generated by ChatGPT like Stack Overflow deeming the results unreliable. Others already lend this system the vocation of undermining the supremacy of Google in search. There are multiple use cases that will have to confront ethical issues. It should be noted that inexorably these generative AIs are destined to develop and attract a large number of people.
Photo credit: DeltaWorks via Pixabay