Listen to the Podcast:
Meta, which is the company that owns Facebook, has made a new artificial intelligence model for making images available to the public.
I-JEPA, Meta’s new “human-like” model, comes at a time when image generators like Midjourney are making fast progress in making photo-like pictures from scratch, even though they often make mistakes, especially with human hands.
Meta says that the new I-JEPA model is more accurate than current models because it has better ways to analyze and finish unfinished pictures.
At first, the technology is likely to be used mostly by AI picture generation experts and fans. However, it’s likely that this technology will lead to a number of AI-based features being added to Meta platforms like Instagram.
I-JEPA is different from other generative AI models because it doesn’t just use the pixels around it. Instead, it uses “an enormous amount of background knowledge about the world.”
The AI services that Silicon Valley giant Meta, which owns Facebook, Instagram, and WhatsApp, has so far mostly been limited to an AI technology called LLaMA, which can power online robots.
Mark Zuckerberg, the CEO of Facebook, made a surprising decision in February to make the AI language model available to the public under an open-source license. This means that anyone can use the technology behind the model for free. The original code for the new I-JEPA model will also be made public.
The platform is based on the ideas of Yann LeCun, Meta’s top AI scientist, who wants AI systems to be able to think “like humans.” I-JEPA wants to fix common mistakes in pictures made by AI by adding this kind of thinking.
“AI researchers have tried to come up with learning algorithms that can take common sense background knowledge about the world and turn it into a digital representation that the algorithm can access later,” Meta says.
In particular, this means that, unlike older models, the system is not set up to look at each pixel of a picture. Instead, the AI model pays attention to the most important parts.
For example, instead of looking at a picture of a person pixel by pixel, I-JEPA has been set up to recognize central areas. This lets it figure out that a person who should only have five fingers on one hand is shown in the photo.
