The domestic version of ＂ChatGPT＂ is coming! On-line measurement of small ice chain (X-CoTA) drives ＂the next generation action center＂ with logical thinking.

Beijing, February 21 (Reporter Miao Yan) Recently, with the explosion of ChatGPT, domestic giants have also accelerated the layout of generative AI. When will the domestic version of "ChatGPT" go online? How’s it going? Become the focus of people’s attention.

At present, Baidu has officially announced the "ERNIE Bot", saying that it will complete the internal test and open it to the public in March this year; Alibaba dharma ChatGPT products have been in the internal testing stage; Jingdong cloud will launch the industrial version of ChatGPT—ChatJD……2 … On February 20th, Xiao Bing Company’s ChatGPT application "X-Chain of Thought & Action" also opened a small-scale internal test, becoming another China-made ChatGPT for trial.

So, what are the new features of X-CoTA? What is the difference compared with ChatGPT? Li Di, CEO of Xiaobing Company, said in an interview with a reporter from Yangguang. com, "This brief demonstration of the new features of Xiaobing Chain made AI Being not only give you a reply, but present her thinking process in front of you completely and transparently, and uncover the black box of the big model. More importantly, she can really implement some kind of Action. "

Small ice chains are no longer just "chatting"

Can fully demonstrate the thinking process.

After the reporter enters the "Little Ice Chain" internal testing system, the page has an input box similar to a search engine, and its response feedback speed is more agile than that of the mobile phone artificial intelligence assistant. It is worth mentioning that when the answer is fed back, not only the conclusion will be displayed, but also the whole thinking process can be fully displayed.

(Screenshot from the small ice chain internal testing system)

For example, enter the recent news hotspot "How to evaluate the sky-high bride price incident in Jiangxi, should it be accepted as a custom?" After discovering the news reports and information about the sky-high bride price, Xiao Binglian finally concluded: "The sky-high bride price in Jiangxi should not be accepted, and supervision should be strengthened to safeguard social fairness and justice."

(Screenshot from the small ice chain internal testing system)

For another example, enter "How many marathons did you run from Beijing to Suzhou?" The final conclusion is: "It is equivalent to 29.2 marathons." Obviously, the small ice chain (X-CoTA) completed the direct and accurate answer, and showed the thinking process and evidence transparently.

(Screenshot from the small ice chain internal testing system)

Through many tests, the test sentences with good feedback effect at present are: "How to evaluate XXXXX", "What do you think of XXX", "What do you think of XXXXX", "Which is better between XXX and XXX", "XXXXXXXX, what’s your opinion" and so on.

For example, what about the movie "The Shawshank Redemption" entered above? What classic scenes impressed you? The answer given is better; But if you ask, "What about the movie The Shawshank Redemption?" This question is too wide, and the result is not as good as the previous one. It can be seen that the more specific and fresh the problem is, the better the effect will be.

Obviously, the content generation logic of Xiao Binglian is that after you ask a question, she thinks about it and finds that she has to search it, or write a piece of code in real time and actually run it, or decide for herself that she should control a series of devices or vehicles in the physical world to better meet your needs.

Li Di told reporters that the direction represented by the small ice chain is to realize the control center of the next generation by using large model technology. Generally speaking, this makes Xiao Bing no longer just a "chat", but a "next generation action center" driven by "logical thinking", covering the digital and physical worlds. This direction will be the next big model innovation breakthrough with real impact.

Don’t do "carving a boat for a sword" competition

Need to explore big model innovation breakthrough

A few days ago, the "White Paper on the Development of Artificial Intelligence Industry in Beijing in 2022" issued by the Beijing Municipal Bureau of Economy and Information Technology clearly proposed to support head enterprises to build a large model of benchmarking ChatGPT, and strive to build an application ecology of open source framework and general large model; Strengthen the layout of artificial intelligence computing infrastructure; Accelerate the supply of basic data of artificial intelligence.

In fact, the small ice chain is not the only innovation of Xiao Bing in the big model era.

Since 2014, Xiao Bing has been growing with the technical iteration, and has experienced many cycles such as retrieval model, generation model, large model and X-CoTA. Among them, in the field of large models, since 2019, Xiao Bing has formed model training and optimization of different scales, and released them in turn after safety assessment. The small ice chain is just one of them.

Li Di believes that "the safety and ethics of large models are still crucial considerations." Therefore, although the domestic market is very hot, Xiao Bing team will not rashly release all kinds of unsafe products just to show off their muscles. This time, Xiao Bing Chain is the only exception.

(Search for the same question on X-CoTA and ChatGPT respectively)

So, what are the advantages of small ice chains compared with ChatGPT? According to Li Di, the small ice chain obtains information in real time, and ChatGPT summarizes it from training data. The logical thinking process of small ice chains is more transparent and observable, while ChatGPT is a black box. The most essential difference is that the small ice chain has actions, such as going to external search; ChatGPT only talks (dialogue generation), but does not act.

In addition, other problems solved by the small ice chain include: solving the problem that the training data of the large model is not updated in time, improving the accuracy and credibility of the reply, and making the information traceable; Effectively reduce the scale and cost of parameters and promote popularization.

In Li Di’s view, "following ChatGPT to do an arms race is to carve a boat for a sword." Because the big model technology itself is developing rapidly, we should further lay out the future of the next station instead of copying the current ChatGPT. In other words, we should think about what comes after ChatGPT, instead of being China’s ChatGPT.