Xiaoice Company, Microsoft’s independent R&D team of AI products based in China, started the internal testing of its ChatGPT-like application “X-CoTA” on February 21.
The interface of the X-CoTA is similar to the search box of ChatGPT and the chatbot will make a logically integrated answer according to the user’s input. Unlike ChatGPT, it splits the process of calling data into several steps, and calculates the information input by users step by step, and then outputs the answers.
Facing the input of proposition composition, news commentary and objective questions, the understanding ability of X-CoTA is not inferior to ChatGPT, and answers are time-sensitive. However, in the face of complicated arithmetic and context, X-CoTA has common problems with large language models, such as low accuracy and incomplete answers, which need to be guided by clarifying previous questions to improve the answer.
According to Xiaoice, X-CoTA has the following characteristics. First, it can obtain information in real time from the world and save the data. The step-by-step logic is transparent, which makes it easier to find and correct any mistakes.
Secondly, X-CoTA can split the logic while reducing the scale of large model parameters, reducing the dependence on computing power. Therefore, the cost of training for developers can be reduced, thus promoting its application. Finally, X-CoTA can operate driven by logical thinking.
Data security and ethical issues in the process of model training and debugging have attracted some attention and discussion lately. Xiaoice responded that the question records kept by X-CoTA will not be used for training or other purposes. At the same time, for the sake of safety, Xiaoice makes a trade-off between interest and length. For example, X-CoTA will not generate homework, summary or speech for users.
There will be high service fees in the process of calling large models. In a report by 36Kr, Li Di, CEO of Xiaoice, gave an example. “If the ChatGPT method is used, the daily cost of X-CoTA will be as high as 300 million yuan ($43 million). The OpenAI model is not necessary for most companies. Almost all digital behaviors can be disassembled into three steps: search, programming and calculation. By splitting the large model into several small models and running them step by step, the training and operation costs can be effectively reduced.”
In the report of Chinese Entrepreneurs, as for the advantages of Chinese tech giants such as Baidu, Alibaba, Tencent and ByteDance, Li said, “Big models need tons of technical accumulations. All these companies have this ability. The real advantage of OpenAI is its focus on adjusting the model.”