Meta's Zuckerberg says that with the help of AI, he will create a model that will recognize hundreds of languages and a universal speech translator that will offer instant translation between all languages.
Meta Properties founder and CEO Mark Zuckerberg held a presentation on Wednesday with the company's artificial intelligence group , including Yann LeCun, on "Building the metaverse with AI."
The AI Lab says on its Facebook homepage that it wants to create a world where people can easily share, create, and connect physically and virtually, with anyone, anywhere. Zuckerberg also called AI the most important foundational technology of our time.
Meta unveiled Project CAIRaoke, a fully integrated neural model that will help create assistant devices. Specifically, they are going to combine BlenderBot techniques with the latest conversational AI to achieve the highest possible conversational capabilities. It was also mentioned that the company's focus on artificial intelligence for The Metaverse consists of two areas of artificial intelligence research. The first area concerns egocentric perception and the second is a new class of genetic AI models. Zuckerberg announced that no language was omitted, therefore creating a model that can understand hundreds of languages. A global speech translator was also announced that will have the ability to translate in real time, something that artificial intelligence will offer during our lifetime.
The day’s presentation focused on the topic of “supervised learning,” where a neural network is developed that will gather much more data without having to apply careful tactics. Lecun explained the virtues of self-supervised learning to IEEE Spectrum’s Eliza Strickland in an email published Tuesday.
During the presentation, after Zuckerberg gave an introduction , Jérôme Pesenti, the leader of the AI team, explained that self-supervised models are increasingly approaching the accuracy of neural networks trained and supervised by humans, not only in language models but also in areas such as image recognition.
Project CAIRaoke was introduced in a video, touting itself as a highly intelligent AI assistant that correctly understands spoken commands. The video highlighted the model's ability to exploit common understanding errors in assistant technology. It referred to the latter as Level 2 technology and emphasized that this technology is only useful for single-shot interactions and cannot be used in conversational interactions. It also announced that supercharged assistants are on the way.
See also: DeepMind: We trained an AI to control a nuclear fusion reactor
In the final part of the presentation, LeCun was introduced by Yoshua Bengio to a conversation hosted by Lex Fridman, for a discussion on the path to “human-level intelligence.” More specifically, Bengio mentioned that we are still a long way from human-level AI, noting that humans monitor new situations and need time to reflect on solving a problem.
The conscious processing that human brains perform is an inspiration for creating such an AI. To illustrate, Bengio used the example of a person learning to drive on the left side of the road for the first time when visiting London. Much of the learning is made possible by humans as they have a lot of consistent and familiar information, which means they are able to retrain habits to fit the new knowledge required.
See also: Artificial Intelligence (AI): What is it and how much has it evolved to date?
On the contrary, LeCun, while agreeing with Bengio on the problems that need to be addressed, said that he is more focused on solutions. He referred to human and animal learning, wondering about the type of learning that humans and animals use, which is not easy to reproduce in machines. At the same time, answering himself, LeCun said that the problem lies in the inability of models to learn about the way the world works, having global constants. Referring to the example with left-hand driving, he pointed out that the rules of physics do not change. In conclusion, the main question that needs to be answered is how to create an AI mechanism that is capable of observing these global constants. In conclusion, LeCun argued that the common sense that characterizes humans is a collection of global constant models.
Source: zdnet.com
