AI that produces human-like language for chatbots, video games, semantic search and customer service is getting a lot of attention lately. The latest news comes out of OpenAI, the research company known for its GPT-3 machine-learning model used to automatically generate sophisticated text. The research outfit will now let all its API users create customized versions of GPT-3 that are fine-tuned for specific applications.
And, in keeping with the low-code trend making AI more accessible to people without computer engineering Ph.D.s, it works to train AI models on those application-specific datasets — say, a sampling of a few hundred product reviews — with the proverbial “single line of code.” The idea is to make this form of content-generating AI faster and cheaper because developers don’t have to create and train AI systems from the ground up, which can be a pricey and time-consuming endeavor.
GPT-3 is the third iteration of the GPT natural language model. It employs a form of machine learning called few-shot learning, which can be trained using a "few" examples of natural language sentences. Few-shot learning is also the foundation for tools Meta is testing in its constant battle to detect disinformation and other content that violates policies on Facebook and Instagram.
As OpenAI has advanced its language AI, others including Nvidia and China’s Inspur AI Research have pushed the limits of natural language processing. Meanwhile, Google’s DeepMind unveiled new language AI research last week including a study of its Gopher language model.
OpenAI, DeepMind and others say they are dedicated to producing “beneficial” and “safe” AI. However, making machine-learning models more and more accessible and easy-to-use poses serious risks of even more prolific dissemination of false, sensitive or discriminatory content and information.
DeepMind’s recent research looks into these risks by creating a way to classify them. OpenAI also helps to mitigate the negative impacts of text-generating AI in the wild by reviewing and actively monitoring applications built with GPT-3.
Correction: This story originally identified OpenAI as an open-source company.
Climate TRACE's data show emissions at the facility-level.Image: Climate TRACE