The Internet contains billions of images, freely available online and constitutes a dense sampling of the visual world. In the first part of the talk, we will explore this world with a large dataset of 80 million images collected from the Internet. Using non-parametric methods, we show how object recognition can be performed in spite of the noisy labels attached to each image. For certain classes that are particularly prevalent in the dataset, such as people, we are able to demonstrate a performance comparable to class-specific Viola-Jones style detectors. In the second part of the talk, we will describe methods for efficiently searching Internet-sized databases. Using machine learning techniques, we represent each image with a compact binary code, at most a few hundred *bits* in length, which preserves the original neighborhood structure of images in the database. Our scheme is able to perform real-time search on millions of images using a standard PC, obtaining a retrieval performance comparable with that of more complex descriptors, despite being many orders of magnitude faster. Joint work with: Antonio Torralba (MIT), Yair Weiss (Hebrew University) and William T. Freeman (MIT).