Microsoft's artificial intelligence research department says it has managed to develop the largest language model in deep learning. With 17 billion parameters, Turing Natural Language Generation (T-NLG) surpasses models like Nvidia's Megatron, which has 8 billion parameters, or OpenAI's GTPT-2, with 1.5 billion.
According to Microsoft researchers, T-NLG is based on Transformer language models. The models are capable of organizing sequences of data in tasks such as, for example, automatic translation or text synthesis. T-NLG can generate words to complete incomplete sentences or tasks based on texts, as well as providing direct answers to questions and summarizing documents.
To train the AI model and make it able to respond in a way that resembles humans, the Microsoft team used tools such as the DeepSpeed library and ZeRo optimizer.
With regard to future uses, Microsoft says that T-NLG can open up the range of possibilities for the company's customers. In addition to saving users time when summarizing documents and emails, T-NLG could improve their Microsoft Office experiences, researchers say.
Experts reveal that in the Micrososft productivity suite, the language model can respond to users' requests for help and questions. The technology giant's research department indicates that T-NLG may also bring new possibilities when it comes to chatbots or digital assistants, since they may have a more fluent and possibly less "robotic" language.
For now, the tool is not available for public use. Microsoft indicates that it will launch an experimental demonstration of T-NLG exclusively for a community of academic users.