#AI22 is a series of articles highlighting what we believe to be 10 developments that will be impacting AI this year.
This series is co-written by Dr. Johannes Otterbach, Dr. Rasmus Rothe and Henry Schröder.
A remarkable portion of humans' everyday tasks are based on text: from reading to comprehending and generating it. According to a McKinsey study the average American employee spends 28% of their working hours on reading and responding to emails. Naturally, the ability to streamline and automate these processes opens up infinite possibilities for optimizing resources and thereby creating value for the user. While there has been significant progress made in these spaces, widespread business applications, especially of generative models is yet to come. Natural language processing models (NLP) will be productized, commercialized and begin to be more widely implemented.
Since the publication of a new technology known as Transformers, initially published by Google scientists in 2017 the advances in the field have been supercharged. With the application of these technologies in extensive NLP models by the big (mostly American) tech companies, such as GPT-3 (OpenAI, 2020), Gopher (DeepMind/Google, 2021), or Megatron (Nvidia, 2021) the commercialization of NLP products will experience significant application in everyday business environments.
These large models can complete increasingly complex assignments: from simple tasks such as generating email subject lines to communicating with customers through chatbots, turning text into code, or even interacting with interviewers. The industry will experience a shift of focus from pure generative models - models creating output that is drawn from the original data - towards models incorporating structured knowledge. With the implementation of structural knowledge, models will start to become more than solely an assistant in automated tasks but a creative, cognitive aid. In the next step AI models will be able to extract information from whatever data they are provided - text, images, audio - without depending on cautiously labeled data. Until recently the need for labeled data was the most time-consuming aspect in training a model but as these newer models now need no specific input data, this considerable barrier is starting to disappear. These self-supervised models will accelerate their move from the early industry adoption phase to noteworthy assistants in the coming years, as the market for NLP products is expected to grow c. 250% within the next four years.
The application of NLP models will begin to affect all aspects of our lives in the years to come: from business to entertainment, to assistance, to education and more - and consequently thereby an explosion in AI-language and content innovation. Through the automation of simple, repetitive tasks by AI, the working life of a large number of jobs will change dramatically.