As fuzzy as Meta's Metaverse vision still is, artificial intelligence will play a central role in generating the new digital world. The start-up A.I.Reverie will help with artificial data.
Whether the Metaverse becomes primarily a 3D virtual world, a digital twin of reality for augmented reality, or something in between, mixed with traditional computer systems: Artificial intelligence is the key to getting to this digital future faster and managing the complex generation processes cost-effectively through automation. It's unthinkable, for example, that a gigantic real-time VR world like the one in the sci-fi film Ready Player One would be programmed by humans alone.
AI also plays a central role in the development of XR hardware and operating software, such as tracking systems: Facebook's AI chief Yann LeCun called AI fundamental to mass-market AR glasses. On the one hand, for certain functions such as object recognition, on the other hand, so that the computing effort of the hardware and thus its heat generation can be kept in check despite complex computing processes. Nvidia's DLSS is a vivid example of more efficient rendering with AI that saves massive amounts of computing power.
Artificial data can accelerate AI development
But the widespread use of AI systems in many and often very specialized areas (such as hand tracking) requires more and more data for training. In many application scenarios, however, manually obtaining and labeling this data is very costly. Or the data simply does not exist, making the use of AI impossible.
This is precisely where the start-up A.I.Reverie, which was bought by Meta in August for an undisclosed sum, comes in: Reverie generates artificial data for AI training.
To do this, the company creates digital simulations that are as realistic as possible, such as scenes on a street corner (see cover picture) or for the realistic illumination of virtual worlds, from which specific data sets can then be generated for AI training.
According to former Reverie chief Daeil Kim, the company's own software can generate a "nearly infinite supply of annotated images and videos," speeding up computer vision development while reducing AI training costs.
One example: Meta is currently collecting videos filmed from the first-person perspective of a few selected people in an elaborate process. This data will be used to train AI to better understand our daily lives, which in turn is essential for augmented reality glasses. Reverie could generate similar artificial data based on the limited real-world data, thus significantly increasing the data set and thereby speeding up AI training and increasing the quality of the AI system.
Artificial data also carries risks
Reverie provided the U.S. Department of Defense and companies in the agriculture and retail industries, among others, with a platform for generating artificial AI training data before its acquisition by Meta. In January 2021, Reverie entered into a three-year contract with the US Air Force for up to $950 million to develop a battle management system and AI-powered command and control systems. The contract with the Air Force was terminated when Meta acquired the company.
Reverie will be absorbed into Meta's XR division, Reality Labs, and will help develop advanced AI algorithms needed to build the Metaverse, according to Meta. The acquisition amount is not known.
It would be too simple, however, if the data hunger of AI systems could be satisfied with artificial data alone. After all, generating artificial data also carries significant risks, especially in scenarios where AI systems make momentous decisions.
In the aforementioned artificial enhancement of real data, for example, prejudices can be reinforced. In addition, errors that are imperceptible to humans could creep into the artificial data and lead to a malfunction.
