Google's New AI Assistant Shows Machines Can Now Ask For Information On Their Own

One key limitation in building modern AI systems has been the lack of high quality domain-specific training data, given deep learning's voracious appetite for data. Researchers have responded by getting creative, from less data-intensive training methods to harnessing massive armies of Amazon Mechanical Turkers and teaching algorithms how to watch YouTube videos. Yet, Google's announcement yesterday of its new generation of AI assistant that can engage in eerily lifelike conversations that border on Turing Test material suggests that perhaps in the very near future AI assistants can not only take on growing roles as research assistants and reference librarians, but even go as far as to begin collecting their own training data. While limited in the conversational and task domains it can currently address, Google's new system is particularly notable for the way in which it can fluently interact with real world humans so naturally that they never suspect they are speaking with a human. As with generations of text-only chat bots, Google's system demonstrates that within the confines of a specific task or domain, it is not that difficult to pass the Turing Test.