Yann LeCun and Randall Balestriero proved that AI models can learn matching world representations across text and images without paired training data. Their method matched images and text using 42 times fewer samples than standard language models.
FireBrief
FireBrief