Former Space Companionship Instruction

This story is partially fictional

If there is a similarity, it is not necessarily a coincidence

If you suddenly get a chance, enter the metaverse. Will you go?

As a technical house, I will not miss it.

Returning from the cyber universe to the real world, this AI virtual person really aroused my curiosity. If we go back to the perspective of reality to deconstruct it, what kind of code is it, and what kind of AI is it constructed?

The answer may surprise you. Behind this AI is the world's largest monomer Chinese natural language processing (NLP) big model developed by the Inspur Artificial Intelligence Research Institute - Source 1.0.

In fact, its true power I haven't discovered yet. With the help of Source 1.0, an AI development team designed the first AI that can play "script killing" with humans, and Cai Xiao is the name of the role it plays in script killing.

Script Killing is essentially a game of gaming. The game creates a scene where each player is assigned a character, gets their own script, and co-starres in a play. At the same time, players must complete the tasks set by the script through clues and reasoning in order to achieve victory.

In such a high-skilled intellectual game, the key lies in the communication between players. The dialogue is ever-changing, and it is completely impossible to use rules-based programming to complete this task, so how does this AI do it?

That is to say, this script killing AI needs to adopt different answer strategies for each character according to the game plot. This ability is also known as purposeful dialogue. Its principle should start from the source 1.0 behind AI.

The Source 1.0 model focuses on the NLP domain, which has the ability to create all sorts of things with linguistic structures, such as articles, translations, and of course, conversations. (For more information about Source 1.0, see "This article is really not written by AI!") 》。 ）

Source 1.0 is very good at zero sample and small sample learning. It can be understood that with a relatively small number of examples, the model can learn to master the corresponding capabilities.

In simple terms, the implementation of a purposeful dialogue in a game is that the developer first establishes an instance corpus and divides the corpus into different documents according to various strategies. For each question, the program selects the three closest examples from the corresponding corpus document as inputs for small sample learning.

As understood, these examples are like a hint that guides the answering strategy, and the model can give an answer in the game according to the linguistic association of "like-like".

What's even more interesting is that the development team found that the script killing AI also showed greater creative ability. In testing, AI sometimes creates plots that don't exist before, and the game has unwittingly become a new story written by AI and players.

Objectively speaking, at this stage, this script killing AI is certainly not perfect, and the so-called "memory" is one of the typical aspects.

Due to the complexity of the project, the team has now chosen a very simplified approach to handling the memory mechanism, with the AI remembering the conversation between the current and previous rounds. The effect of such a scheme is already quite good, but it still shows some drawbacks. For example, sometimes, AI may still give answers that do not conform to the plot or are inconsistent.

The R&D team acknowledges that the current memory mechanism is not the best solution. This is also the space for future progress of script killing AI.

In actual testing, AI will face more details. For example, human players' input habits vary widely, some people like to wrap lines frequently in chats, some people will use spaces instead of punctuation, and so on. These problems not only require developers to design corresponding processing schemes and mechanisms in advance, but also a test of the large model itself.

The source 1.0 model behind AI has a parameter scale of up to 245.7 billion, and the high-quality Chinese dataset used for training is 5,000 GB. Source 1.0 can do anything related to Chinese NLP, but it doesn't play script killing by itself, it's like a talented martial arts master who needs developers to constantly tap into its skills.

Therefore, a project like Script Killing AI is not just a game, it allows us to see the rich possibilities of model application scenarios.

With the open and open source capabilities of Source 1.0, AI developers can quickly enjoy the convenience of large models, including direct calls to open model APPLICATION interfaces, high-quality Chinese datasets, open source model training code, and more. Borrowing from the script to kill the AI development team, some of the template code given by the Inspur team is "good enough for us to use directly.". At present, source 1.0 has been applied to many creative projects.

The development team said the project was an attempt to generate "interesting souls" in addition to "good-looking skin bags" for virtual people.

Allowing AI to creatively "think" and communicate with humans has always been the biggest problem in the field of AI, and it is also the ultimate goal of AI development. In this project, we seem to be able to glimpse a glimpse of ai's "thinking", "communication" and even "game".

The excellent generation ability of the NLP big model, under the guidance of suitable examples, AI has shown some new features, such as being able to generate a more "humane" expression, as if it has "injected a soul" into the AI.

In this game, AI and human players have unconsciously co-created and deduced new stories. This is in a way a wonderful collision of AI and human thinking.

From a broader perspective, NLP is an integral part of human-computer interaction. As the link between human beings, language will also become a conduit for communication between humans and machines in the future. As envisioned in "Her," even if AI has no entity, language can still play a huge role in its own right.

In the future, with the rapid development of technology, one day, when the metaverse is really within reach, the interaction between man and machine will become more and more close, and AI will participate in the operation of human society more and more frequently. With the technical support of mature generative pre-trained large models, AI can become a thoughtful all-round assistant, bring companionship, better integrate into various service areas such as sales and customer service, and may even have the opportunity to provide psychological help under appropriate guidance.

But in any case, if we want to make the most of AI, we have to understand it. Only by exploring and trying step by step can we finally usher in a future where human-machine coexistence is harmonious.

In order to better support developers to carry out various AI innovations based on source 1.0, Inspur will hold the first "Source Developer Forum" on May 7, inviting outstanding project developers to share their development experience and experience in source 1.0.

The first forum will invite Li Feng, a researcher of inspur artificial intelligence algorithms, and Zhao Zeming, a senior VR/AR practitioner, to focus on the exploration and experience of Source 1.0 in the script killing project, and discuss the development of AI in the field of virtual people. At the same time, the forum will also open source open dialogue and translate the skill model of the two sources 1.0, so stay tuned!

#创作团队:

Curated/written by: Metaverse-Takeko

Typography/Design: Metaverse-M Ka

#参考来源:

https://github.com/bigbrother666sh/shezhangbujianle

Some of the content materials are provided by the Inspur Artificial Intelligence Research Institute.

#图片来源:

Cover: pixabay

Beginning: pixabay

*This push is sponsored by inspur artificial intelligence research institute.