At the beginning of the year, Yann Le Cun explained that ChatGPT “was nothing revolutionary”. But on Tuesday, the Meta group’s artificial intelligence manager estimated that OpenAI’s chatbot would even be obsolete in five years.
Present in Paris on the occasion of the reissue of his work when the machine learns, Yann Le Cunn met Usbek&Rica. The opportunity to return to the weaknesses of the AI model used by ChatGPT. Because if the tool fascinates the majority, it is immediately less revolutionary in the eyes of a specialist, particularly in the use of a corpus to draw their responses.
“A very short shelf life”
The French artificial intelligence researcher regrets that the solution is not controllable, that is, that it is not possible to plan your responses. According to him, that is why the generated long texts have problems to be coherent. He also points to a lack of reliability in the information exchanged. This led to the filing of a complaint with the CNIL by the deputy Eric Bothorel against ChatGPT.
Making AI models controllable is therefore Yann Le Cun’s priority. By giving directives and constraints to solutions, he wants to “develop AIs that reason and plan according to given goals.”
Never dominate “by accident”
However, a higher form of artificial intelligence raises the question of its regulation and its level of autonomy. Defining the criteria to ensure the safety and reliability of AI models is called alignment. But Yann Le Cun tries to reassure: “The desire for domination, and therefore the desire for submission, are specific to social species such as humans, baboons or wolves.”
Finally, Yann Le Cun believes that the artificial intelligence of the future will be open. This is the direction that Meta has chosen to take with LLaMA. This open source artificial intelligence project should help the company design personalized ads for its clients.
“All the Internet technologies that have finally prevailed are based on open source”, concludes Yann Le Cun.
Source: BFM TV
