Discover China’s biggest AI Model
GPT-3 caused a stir a few months ago when it impressed the world with its capabilities for generating text content. Its 175 Billion parameters made it look like a true AI giant. Now a new Deep Learning model is making an appearance at an AI event in Beijing. This model, referred to as Wudao, ‘natural law understanding’, is the world’s largest new model with 1.75 Trillion parameters.
The initial version was presented in March but has been quickly replaced by this larger version. The goal was to present the world’s largest model because according to Dr. Zhan Hongjiang, president of BAAI (Beijing Academy of Artificial Intelligence):
“The way to general artificial intelligence is big models and big computers.”
General Artificial Intelligence is the hypothetical ability of a machine to learn any task in the way that a human can. GPT-3 has the ability to do Zero-Shot learning, which is learning other tasks without anyone telling it that this is its task. It can do tasks such as translate, perform arithmetic operations, sort words, and write programming code.
The people in charge of BAAI (Beijing Academy of Artificial Intelligence) imagine a large-scale model with immense computing power that will make it possible to generate general artificial intelligence. A sort of central base to provide artificial intelligence services while feeding in more data and knowledge.
What makes this (pre-trained) model so special is that it uses its 1.75 Trillion parameters to simulate conversations, understand images (making it a multimodal model because it can process both images and text), write poems, and create recipes.
Parameters are defined in ML models and as they evolve and learn, these parameters are updated and improved to find the correct output over time. Models that contain many parameters are usually more sophisticated, but this comes with the disadvantage that more time and money have to be invested for their development.
The Wudao 2.0 model has been trained with 1.2 TB of Chinese text and 1.2 TB of English text and can understand both. It has also been trained with 4.9 TB of images and text in total. Right now it has 22 partners being companies like Xiaomi among them.