Google Search is an arguably comprehensive product which is used quite widely around the web and on mobile devices. Google regularly adds more features to it and redesigns it from time to time as well. Although it handles basic and somewhat advanced search queries quite well, Google says that it is now working to handle even more complex questions.
This endeavor utilizes a new technology called Multitask Unified Model (MUM) to fetch answers to more complex search queries. It is built on Google's Transformer architecture but is reportedly 1,000 times more powerful than BERT. MUM has been trained on 75 languages and various tasks at once which means that it has a stronger contextual world knowledge compared to existing models. An additional benefit is that MUM can handle different modalities. As of now, it can understand text and images but will be able to handle video and audio in the future too.
Google has highlighted a case to demonstrate the usefulness of MUM. If you've hiked Mount Adams and are now planning to hike Mount Fuji, you would want to know how to prepare differently for your next endeavor. Right now, you'd probably be able to get your answer from Google Search but it would require complex search queries and time. If instead, you asked the same question from a professional hiker, they would quickly be able to provide you an answer. MUM aims to build on this example even further by utilizing world knowledge to recommend you training techniques, top-rated gear, and other helpful subtopics with a single and simple search query like "I've hiked Mt. Adams and now want to hike Mt. Fuji, what should I do differently to prepare?"
The fact that MUM is trained on dozens of languages also means that it can efficiently transfer knowledge from one language to the other and show you recommendations in your preferred one. A future possibility could also be uploading a photo of your hiking boots for MUM to determine whether they are reliable for hiking Mount Fuji, and recommend you alternatives if they're not.
Google says that avoiding bias in new AI systems and reducing their carbon footprint is essential so it will follow a rigorous evaluation process similar to BERT to ensure that MUM ticks all the right boxes. Features powered by MUM will be coming to Google Search eventually as the project is in early exploration phases right now.