Baidu Wenxin starts the internal test experience by benchmarking ChatGPT

ChatGPT, a blockbuster in the scientific and technological world, has changed the stereotype that AI is "artificial intelligence disabled" before. It can realize many operations of human thinking understanding, has the ability of context understanding, and can answer some questions according to the needs of users. For example, generate a piece of code that can run, and improve the code according to the requirements to make it more suitable for the needs. The future direction may be to combine with various office software and application software to assist in office work, generate AI videos, and so on.

At present, Wenxin Yiyan has opened the internal test application, and Baidu said that Wenxin Yiyan cloud service will be launched on March 27. As the first intelligent AI all-around assistant released first in China, Baidu has gained great attention, but whether it can break out of the world and gain public recognition is still a question worth considering.

Wenxinyiyan applied for public beta address:
https://yiyan.baidu.com

Wenxin Yiyan Enterprise Version Application Test Address:
https://cloud.baidu.com/survey_summit/wenxin.html

On March 16, Baidu held a press conference in Beijing. Robin Li, the founder and CEO of Baidu, personally introduced Wenxin on the platform, which shows the importance he attaches to it. The development of AI technology can break the barriers of skills and knowledge and bring a new model of human-computer co creation. In the future, "Wenxin Yiyan" will provide external services through Baidu Smart Cloud, which is a new initiative of Baidu Smart Cloud. At present, more than 400 leading enterprises have joined Baidu Wenxin Yiyan ecosystem, including Internet, media, finance, insurance, automobile, enterprise software and other industries. " The technology of Wenxin Yiyan ", a general AI product, can be widely used in content creation, customer service, digital human scenes, etc. after continuous iteration and cost reduction. Its big language model of knowledge enhancement is based on the fly propeller deep learning platform and Wenxin knowledge enhancement big model, which can continue to integrate learning from massive data and large-scale knowledge, and has knowledge enhancement Technical features of retrieval enhancement and dialogue enhancement.

From the current test, both ChatGPT's and Baidu's Wenxin's words have the problem of making up nonsense, that is, the positive tone that seems to have something in mind is actually made up now. Some of them failed to understand the questions asked, while others pretended to understand, which made people confused about what was true and what was false. Their credibility was greatly reduced. For some subjective problems, only inferring based on the existing model knowledge, there will indeed be some intellectual errors. There may also be problems in the training data. Some data are outdated and inappropriate, but the model has not been rapidly improved according to the changes of the times. The various data sources of the training model may be contradictory, so when some objective questions are asked, the answers given by the intelligent assistant are different each time.

As an intelligent AI assistant that outputs text, voice, video and other content in real time, how to implement content review is a complex and tedious problem. The output content should be positive, consistent with the correct values, and positive. When wrong content occurs, how to make corresponding improvements is also a complex problem, Although AI assistants have the ability to correct errors according to prompts, they are not always so lucky. For example, in 2016, Microsoft chat robot Tay was trained as a racist just 24 hours after it was launched on Twitter, and Microsoft had to make some adjustments to it.

We can make some restrictions on the output of the model, such as embedding it into specific scenes for use according to actual needs, and outputting content in specific directions. A few days ago, Microsoft released a new AI function supported by GPT-4 platform, called Copilot, which will be applicable to Word PowerPoint, Excel, Outlook, these popular Microsoft 365 commercial software of Microsoft, can complete writing documents, automatically generate PPT based on text, generate charts based on data, etc. Microsoft claims that Copilot's function is more powerful than simply "embedding OpenAI's ChatGPT into Microsoft 365". We can look forward to Copilot's performance.

ChatGPT has not opened its registration to users in mainland China and Hong Kong, China, which may be a business decision made by OpenAI after integrating various factors. When a technology is not yet mature, it should be tested within an appropriate range. When it meets the applicable requirements, it will be promoted to a wider range of countries and regions, respecting the laws and regulations of various countries and regions, and abiding by local regulations and restrictions.

AI can indeed solve many problems, but it is not omnipotent, nor can we rely on it too much. It can play its best role only when it is in the right place. It is a promising technology itself. Our approach to it is certainly not to stop it from developing, but how to develop it in an orderly manner and make it benefit all mankind within a controllable range.