Do you love playing script kill?
The good news is that you don't have to worry about not having enough teammates anymore; the bad news is that you can't get through or see through your teammates anymore — because someone has trained an AI that can play a role in script killing with the help of the source 1.0 model launched by Inspur Information.
"The first time I played script killing, I didn't expect it to be with AI, I felt that I was not very into the drama, but this Cai Xiao (AI) was more into the drama than me. In the plot, my role is a 'wall grass', belonging to the ticket that was won, this Cai Xiao is too active, let me be a little overwhelmed. Frequently showing me overtures, and then talking to me like a little girl about her dreams, I was embarrassed to refuse. A real player ,"Kong Mo", said so.
The application of the "pre-trained big model" AI to the script killing this set scene has brought surprises and surprises to developers and players. There is no perfect technology, but technology opens the door to a new world.
Will this door lead to the future metaverse?
Script killing training
The idea of teaching AI to play script killing came from 6 programming enthusiasts isolated in Shanghai, from Shanghai Jiao Tong University, East China University of Science and Technology, and industry insiders engaged in VRAR.
After all, it is training AI, the script has been specially adapted, and the miniature online script with five characters has been selected to kill. Each game gathers four live players and, without their knowledge, the AI plays the role that assigns the remaining characters.
"In line with the principle of pulling the details full, we also prepared a WeChat account for the AI, and carefully set a nickname and avatar for her, and even before each game, we will closely follow current events and prepare nearly three days of circle of friends content for her, and after the game, we will continue to send three consecutive days of circle of friends content to provide extended plots." The developer said.
Such AI accompaniment can be described as full of sincerity, and players are a little surprised when they learn that they and AI have played a script kill. After all, when the player makes a confession about the AI's character, her performance is definitely a loyal girlfriend, without hesitation to distribute the "good guy card": we are best friends.
Although the development cycle is a short period of about a month, iterations are carried out every day. The developers used Inspur Source 1.0, a generative pre-trained model and the world's largest single Chinese NLP model. Source 1.0 excels at Zero-Shot and Few-Shot learning, not finetune, which more models are currently good at.
If you use a string of professional data to describe, the source 1.0 open API uses the Language Model (LM) model structure, the parameter scale is as high as 245.7 billion, and the Chinese dataset used for training is 5000 GB, compared with the GPT-3 model 175 billion parameter volume and 570GB training dataset, the "source 1.0" parameter scale is 40% ahead, and the training dataset scale is nearly 10 times ahead.
The advantage of a small sample is that under the demonstration of 2 to 3 or even 1 suitable example, the model can well understand the "dialogue strategy" that the developer hopes to achieve, as if it has the ability to "learn from one example". The developer compares AI to the talented and intelligent Zhang Wuji, but if he encounters not a high-ranking person in the world, but a common man, what he demonstrates every day is how to go to work to touch fish and slip around in class. "The same is true of the Source 1.0 model, although it carries 5.02TB of Chinese data, which is almost equivalent to more than 5 million books, but it has not played script killing, and all it can do is simulate and learn from the sample."
The AI needs to use different answer strategies for different characters according to the plot, the "example corpus" is divided into 4 TXT files, and the program will select the example source according to the questioner. In a real game, though, the wording of the player's possible questions is endless. Developers also use the pre-trained model released by Baidu Flying Propeller - simnet_bow, which can automatically calculate phrase similarity, and the operation speed is fast, and the video memory occupation is low.
It can be seen that training a successful AI also requires eclectic strengths. Deciding which open source model to use is also influenced by factors such as model features and whether the API is convenient enough. "Inspur's engineering support is still in place, so the development threshold is actually not so high." The relevant interfaces of the API are very complete. I am actually not a professional programmer, a little exploration can be used, I think I can further make a PaaS platform, serve more people. The developer told First Finance that the team gave a high evaluation of the template code given by Inspur for prompt generation, request submission and reply query, saying that it was "good enough that we can use it directly."
"Infusing the Soul" into the Metaverse
AI Play Script Killing is an application that closely combines natural language generation and natural language understanding in a limited scene, and it is also an attempt by virtual people to generate "interesting souls" outside of "good-looking skin bags".
The NLP field has made great strides in generative pre-training of large models, but intelligent customer service is still difficult to say intelligent, and companion service robots such as smart speaker products will happily talk back to the owner.
"There are two kinds of dialogue, one is open context, that is, completely unrestricted. This minimum requirement for AI is to speak human words, but there is no requirement for logic, understanding of context and so-called targeting, which is difficult to have practical value, and if it is a targeted dialogue, the traditional approach is not very good at meeting the needs. ”
The developers of the AI script wanted to make a difference by introducing an "end-to-end" approach. "We also found that if ai does not memorize its previous answers, the results generated by the ai will be relatively lack of continuity, and even give inconsistent answers." These problems forced us to decide to add 'multi-round conversation memory mechanism'. Under the guidance of the appropriate example, it can even reflect a certain amount of emotional intelligence, increasing the expression of emotion.
The developer provides an example
"It's not that it's going to be like this every time AI, it's random. Take this example to illustrate that under the generation mechanism of the big model, it will not only speak human words, but also bring a little human flavor, which is a new feature, a very interesting new feature. The developer told CBN.
In addition, just as humans do not remember everything in their daily lives, the "memory mechanism" without an "attention mechanism" is actually more harmful than beneficial to dialogue AI. However, if you want to introduce a "attention mechanism", you need to add more complex NLU (natural language understanding) algorithms. Fortunately, the actual application scenario of script killing pays more attention to the dialogue of the current round, and for the situation that requires the content of the distant round dialogue to answer, the AI can say "forget".
"For me personally, the research and application of AI is now completely interested, and then we may have some research cooperation with some professionals in universities and entertainment industries." Ai can certainly play script killing with humans, but there are many more possibilities waiting to be explored.