Self-aware AI redefines social and gaming

On July 23, 2023, Beijing held the Ultrasound Club AI Open Day event, attracting the participation of many top AI entrepreneurs and well-known investors. Ultrasound Club member and Force Wave co-founder Mou Qi gave a wonderful speech. The following is the summary of Mou Qi’s speech:

Sharing guest: Mou Qi, co-founder of Gravitational Wave

Sharing topic: Self-aware AI redefines social interaction and gaming

In the past, most content in the Web 2.0 era was created by users. From this, many IPs with personalities and characteristics were born, which resulted in a large number of content-based social connections. However, in fact, users are more From a distance, it is impossible to establish interactive connections with these IPs.

So in the AI ​​era, how to use AI technology to create these entertainment contents? Think about it, if we can create an AI with its own life in the virtual world, it will have its own sense of autonomy, live like a human every day, have its own plans and thoughts, and be able to live in this world. Take the initiative to make some friends, and have a social media account, and be able to post on Twitter, Weibo, and Moments. If this is possible, you should be interested in finding him or her in that parallel world.

Rewritten content: Force Wave is a company focused on developing 3D autonomous artificial intelligence, aiming to establish a deep connection between people and artificial intelligence, so that everyone can create their own Western World

The so-called autonomous consciousness AI is to produce responses similar to real people, have a memory system for its own lifestyle, and can be continuously updated based on interactions with other entities (whether human or other AI). All these AI updates and interactions can be displayed in a parallel 3D digital world.

The following is our Auto 3D GPT automatic update system for self-aware AI.

To build the above-mentioned automatic generation system of AI Agent, it is necessary to simulate the human-like working system model based on brain neurology. Our Auto 3D GPT is divided into the following systems:

First: Signal collection system (nervous system)

This system simulates the human nervous system and is used to collect various information and signals from virtual scenes. These signals may be emitted by items in the scene, or by other AI Agents or ordinary users. The AI ​​Agent first collects these signals and then performs filtering and pre-processing to ensure that only relevant and useful information is passed to the next stage. This process mimics the way the human nervous system works to ensure the accuracy and timeliness of information

Second: Thinking and Memory System (Human Brain)

This system simulates the working mechanism of the human brain, performs human-like in-depth processing and feedback on the collected signals, and is able to continuously update based on interactions with other entities (whether human or artificial intelligence)

Among them, we adopt various modules to ensure that the agent exhibits human-like behavior in the application. Our AI Agent can utilize server prompts to assist in processing and storing observed information.

AI Agent's daily schedule is generated based on user settings and background, reflecting human routines. This schedule is dynamic and can change based on circumstances, taking into account factors such as relationships, assessing the importance of events, and potentially adjusting its schedule accordingly.

The memory system of the AI ​​agent is divided into two parts, short-term and long-term, which are similar to human cognitive abilities. Short-term memory ensures logical behavior based on recent events and transfers important content to reflective modules. After evaluation, events with greater impact are stored in long-term memory to influence future behavior

The memory and reflection system also allows AI Agents to update relationships through interaction. Positive interactions allow the agent to consider the user a friend, while negative behavior ensures that the agent keeps its distance in the future. Skills are also stored in long-term memory. For example, an agent with a martial arts master setting can demonstrate martial arts, but if this skill is not recorded, the agent may refuse to play the flute. We expect that AI Agents will eventually learn from their experiences.

Our AI Agent thinking system surpasses the traditional artificial intelligence NPC. They interact with others based on memory and context, responding to messages in various ways. An introverted AI agent may refuse a hug from a stranger, while an extroverted AI agent may comply. The AI ​​Agent's thinking process is visible, providing insight into its decision-making. We also ensured he was situationally aware and able to recognize the user's approach and react appropriately

Third: Expression feedback system (muscle expression system)

The rewritten version of this content is: The system uses a self-developed 3D translation model, which can expand and translate statements describing behaviors into executable three-dimensional instructions. Then, through gamified three-dimensional representation, the AI ​​agent can generate behavioral feedback that users can perceive and understand. This process simulates how the human muscular system translates thoughts into physical actions, making the AI ​​agent's responses not only accurate and coherent, but also intuitive and engaging

The goal of the 3D translation large model is to translate a task into a series of executable fine-grained instructions through a large language model.

For example, the input task is "Go to the gym to exercise", and the language model automatically generates a series of instructions: "1. Approach, the target is the gym. 2. Change, the target is clothes. 3. Exercise, the target is the treadmill."

We need to rebuild an interactive instruction set suitable for 3D environments. We used some methods to generate a batch of initial task data, combined with the business user data in our business system, and after data cleaning, formed a data set for the translation model

We are based on the Meta Llama-2 open source 7B and 13B large language models and fine-tune our instruction translation data set. We use parameter adaptation and 8-bit model training strategies to greatly reduce the demand for GPU memory. General Nvidia 3090/4090 graphics cards can meet the training needs.

The translation models of SpaceU-7B and SpaceU-13B achieved perplexity scores of 1.14 and 1.12 respectively in the verification set (the ideal minimum value is 1).

The SpaceU model reasoning framework uses Controller and multiple model workers to implement load balancing and support the translation needs of a large number of queries.

These systems work together to form a complete virtual AI Agent system. By simulating the human nervous system, brain and muscle expression, they enable AI Agents to interact in a more natural, flexible and intelligent way in virtual environments, providing users with a more realistic and immersive AI experience. This will establish a deep connection between people and AI and provide possible implementation paths.

Our Auto 3D GPT system provides users with the ability to create 3D AI Agents with autonomous consciousness. For example, TA can be a friend who relieves your boredom, or a virtual teacher, or of course a time-travelling robot. Cats from a parallel world! When many AI Agents and people live together in a virtual world, this virtual world has unique culture and business rules, and it becomes a new AI City, which is very similar to Westworld.

Rewritten content: Creators can experience the feeling of becoming a god and create a unique civilization to experience. For example, you can create a feminist society like Barbie, or you can create a brand new animal world, or you can let Jesus and Qin Shihuang live in the same city and see what happens to them. This is a more open-ended gaming and social experience

AI can not only replace traditional game planning to build a large number of more complex, unexpected and reasonable story lines, it can also capture player choices and behaviors in real time, and flexibly adjust story content and direction. This dynamic ensures that every player's gaming experience will be unique and extraordinary. In addition, this ability of AI can also be extended to a multi-player environment. By connecting the content stories of different players, it opens up unlimited possibilities for the game and forms an interactive and rich social network, thereby elevating the game to a new social dimension.

A famous philosopher said: Life is like a game, you can't quit! We hope that with the help of the SpaceU Auto 3D GPT system, everyone can reselect and restart in the virtual parallel world.

You can pay attention to the recently released SpaceU Auto 3D GPT and discuss together what impact artificial intelligence with autonomous consciousness will bring to human society

The above is the detailed content of Self-aware AI redefines social and gaming. For more information, please follow other related articles on the PHP Chinese website!