Meta, which is owned by Facebook, is focusing on artificial intelligence research to create worlds through speech, improve how people communicate with voice assistants, and translate between languages, according to CEO Mark Zuckerberg.
“AI is the key to unlocking a lot of these advances,” he stated at the company’s “Inside the Lab” event, which was live-streamed.
Meta is working on a new class of generative AI models, according to Zuckerberg, that will allow anyone to define a world and generate elements of it. Zuckerberg demonstrated an AI concept called Builder Bot in a taped demo, in which he appeared as a legless 3D avatar on an island and delivered spoken commands to create a beach, clouds, trees, and even a picnic blanket.
“As this technology advances, you’ll be able to create nuanced worlds to explore and share experiences with others using only your voice,” Zuckerberg added. He didn’t provide a date for these breakthroughs or any further information on how Builder Bot operates.
He claimed that Meta was focusing on AI research to allow users to have more natural interactions with voice assistants, a step toward how people will interact with AI in the metaverse. “A fully end-to-end neural model for producing on-device assistants,” he stated of the company’s Project CAIRaoke.
The voice assistant chimed in to advise that salt had already been put to the pot in a demonstration of the Project CAIRaoke gadget, which featured a family using it to help make a stew. The assistant also observed that they were low on salt and placed an order for additional.
Meta said it was using the model in its Portal video-calling gadget and hoped to include it into augmented reality (AR) and virtual reality (VR) devices in the future (VR). Jérôme Pesenti, Meta’s vice president of AI, told Reuters that the company was restricting the responses of its new CAIRaoke-based assistant until it could assure that the system didn’t generate foul language.
“These language models are quite powerful,” Pesenti said, “so we’re putting in a lot of effort to be able to regulate them.”
In addition, Zuckerberg revealed that Meta was working on a universal speech translator that would enable instant speech-to-speech translation in any language. Prior to this, the business established a goal for its AI system to be able to interpret all written languages.
The social network business, which recently lost a third of its market value following a disappointing earnings report, has put a lot of money into its new focus on establishing the metaverse and changed its name to reflect this goal. Meta announced this month that its Reality Labs, the company’s augmented and virtual reality division, will lose $10.2 billion in fiscal year 2021.
Meta’s AI chief, Pesenti, told media that the company is looking into how artificial intelligence may be used to monitor material and activity in the metaverse.
“We utilise a lot of AI for moderation on our primary platforms,” Pesenti said. “The metaverse is a bit of a different beast, it’s a lot more real-time,” he added, adding that this was “developing work” and that Meta was still sorting out metaverse policies.
At the AI event, Zuckerberg said Meta was focusing on “self-supervised learning,” in which AI is given raw data rather than being trained on tonnes of labelled data, to help AI analyse and forecast the types of interactions that will occur in the metaverse.
Meta was also working on egocentric data, which includes viewing worlds from a first-person perspective, according to Zuckerberg. He said the project has brought together a global coalition of 13 universities and labs to work on Ego4D, the world’s largest egocentric dataset.
Meta also stated that it would extend free education efforts targeted at attracting more racial minorities to the field of technology, which academics believe is necessary in order to develop AI systems that are devoid of bias. According to a recent survey conducted by recruiter Heidrick & Struggles in the United States and Europe, about 80 percent of data analytics and AI leaders identify as men and 65 percent as white.
Meta wants to open source the recommendations library TorchRec, which it uses to customise products like Facebook’s news feed, as a nod to transparency, according to Pesenti in another event session. According to him, the business will also release a feed ranking prototype to demonstrate how its algorithms prioritise which material is displayed to Instagram users.
Some of the projects unveiled by Meta on Wednesday, such as Project CAIRaoke and the algorithm transparency initiative, are comparable to those disclosed by competitors such as Alphabet Inc’s (GOOGL.O) Google in recent years.
Meta has just stated that its research team has completed work on a new artificial intelligence supercomputer that it expects to be the fastest in the world when it is completed in mid-2022.