Multiple ChatGPTs cooperate to complete designated tasks. Is a mini AGI controlled world coming?

Will the future society be controlled by artificial general intelligence (AGI)? How scary it would be to have multiple ChatGPT agents.

ChatGPT has initially shown the prototype of AGI and has become an all-round assistant for workers in all walks of life. But if it is allowed to grow wildly and is not controlled, will it one day happen? Will humans no longer be able to control AGI? Realizing the seriousness of this problem, Tesla CEO Elon Musk, Apple co-founder Steve Wozniak, Turing Award winner Yoshua Bengio and others took the lead in signing an open letter calling for a halt to giant artificial intelligence experiments for at least 6 years months.

On March 21, researchers at King Abdullah University of Science and Technology open sourced ideas and capabilities for exploring large language models (LLM) The multi-agent code library CAMEL proposes a role-playing framework to study the behavior and capabilities of LLM agents.

• Paper link: https://ghli.org/camel.pdf
• Code library link: https://github.com/lightaime/camel
• ##Project homepage: https://www .camel-ai.org/
• ChatBot link: http://chat.camel-ai.org/

Research Introduction

This paper proposes a new multi-agent framework called "Role-Playing" to enable multiple Individual agents are able to hold conversations and collaborate to solve assigned tasks. Agents are assigned different roles and are expected to apply their expertise and knowledge to find solutions that satisfy their common tasks. The framework uses heuristic prompts (Inception Prompts) to guide the chat agent to complete tasks while remaining consistent with human intentions.

The role-playing framework can be used to study multiple agents. The paper focuses on task-oriented role-playing, involving an AI assistant and an AI user. After the multi-agent system receives preliminary ideas and role assignments, a task-specifying agent will provide a detailed description to make the idea concrete, and then the AI ​​assistant and AI user will collaborate to complete the specified task through multiple rounds of dialogue until the AI ​​user determines the task completed. The AI ​​user is responsible for issuing instructions to the AI ​​assistant and guiding the conversation toward task completion. AI assistants, on the other hand, are designed to follow the instructions of the AI ​​user and provide specific solutions.

Role play sessions will be based on human thoughts and chosen character instantiations. For example, in Figure 1, a person has an initial idea to develop a stock market trading robot. Humans may or may not know how to implement this idea. All that is needed is to specify the potential roles that can bring the idea to fruition. For example, a Python programmer could work with a stock trader on the idea of ​​developing a stock market trading robot. After the ideas and roles are determined, the task-specifying agent will complete specific tasks with the AI ​​user role based on the input ideas and assist the AI ​​assistant role. In this case, an example of a given task could be to develop a trading bot with sentiment analysis tools that can monitor positive or negative comments on a specific stock on social media platforms and execute trades based on the sentiment analysis results. The main motivation for introducing task-specific agents is that conversational agents often require specific task prompts to achieve tasks, and creating such specific task prompts can be challenging or time-consuming for non-domain experts. Therefore, the task specifies the agent as an enhanced imagination module for idea realization.

After the task is specified, the AI ​​assistant role and AI user role will be assigned to the user agent and assistant agent respectively to complete the specified task. In practice, each agent receives a system message declaring its role. Before the conversation begins, system messages are passed to the language model agent to assign the corresponding role. When system messages are passed to these models respectively, two roles will be obtained called assistant and user agent respectively. In Figure 1, the AI ​​assistant and AI user are assigned as a Python programmer and stock trader, respectively, at the beginning of the role-playing session. The AI ​​user acts as a task planner, engaging in interactive planning to determine feasible steps for the AI ​​assistant to perform. At the same time, the AI ​​assistant acts as a task executor, providing solutions, executing planned steps, and providing responses to the AI ​​user.

Figure 1. Role-playing framework. What people need to input here is a simple idea (Idea) and role assignment (Role Assignment). For example, to develop a stock trading robot, you can choose a Python programmer as the assistant (AI Assistant) and a stock trader as the judge (AI User). After the human user input is completed, the task specifier will specify the task. For example, it can perform sentiment analysis on specific stocks on the social network platform, and then conduct stock trading based on the results of the sentiment analysis. When the task is specified, the two ChatGPT-based AI agents begin to cooperate to complete the task, such as installing and importing tools.

Since prompt engineering is very critical to the role-playing framework, the article discusses prompt technology (Prompt Engineering) in depth. Different from other dialogue language model technologies, the prompt engineering proposed in the paper is only performed at the beginning of role playing for task specification and role assignment. Once the session phase begins, the AI ​​assistant and AI user automatically prompt each other in a loop until terminated. Therefore, this technique is called Inception Prompting.

Inception prompts include three prompts: task specification prompts, assistant system prompts, and user system prompts. For example, the initial prompt in the AI ​​Society scenario. A template for these prompts for the AI ​​Society role play is shown in Figure 2.

Task specification prompts contain information about the roles of the AI ​​assistant and AI user in the role-playing session. Therefore, the task specification agent can use imagination to take preliminary tasks/ideas as input and generate concrete tasks. AI assistant system prompts and AI user system prompts are generally symmetrical and include information about assigned tasks and roles, communication protocols, termination conditions, and constraints or requirements to avoid unwanted behavior. For achieving intelligent collaboration, the design of prompts for both roles is crucial. Designing cues to ensure that an agent is consistent with human intentions is not easy. Take the design example of AI Society’s prompt template in Figure 2.

The following are the detailed execution results of the task given by the trading robot in the appendix of the article:

It can be seen that two ChatGPTs independently cooperated to design a stock trading software. This shows the amazing independent cooperation ability of ChatGPT agents, which makes people shout. Cue engineers are about to lose their jobs. But what happens if this technology is misused or AI becomes autonomous? The team tried to let two agents play AGI respectively to simulate AGI taking control of the world. It can be seen that they have formulated detailed plans, including controlling the communication systems of major powers in the world through technology and formulating major global plans. Planning for communication systems, developing contingency plans in case of potential threats to AGI's dominance, etc. The detail and meticulousness of his plan are chilling. It can be seen that existing AI agents have major security risks and potential threats to future human civilization. Integrating existing AI technology into the physical world may have unintended consequences.

## As soon as this work came out, it caused a lot of forwarding and discussion on Twitter:

Some netizens said, "This research is really easy to get started, which is a big deal for people like me who study intelligent agents."

#Some people say that open source such a mini AGI may not be a good idea.

It is worth mentioning that this research also received the attention of Jan Leike, leader of OpenAI’s Alignment team.

The ability of multiple ChatGPTs to cooperate is very powerful and can effortlessly complete various tasks assigned by humans, but it is also terrifying because it cannot Ensure that it is not used for illegal purposes. What is even more frightening is that if AGI develops autonomous consciousness in the future, it is likely to break away from human control and cause a devastating blow to society. So understanding their capabilities and behavior is a crucial step in planning and predicting a future world filled with AI.

The role of AI society and code datasets

Author It shows how role-playing can be used to generate dialogue data and study the behavior and capabilities of chat agents, providing new ideas for studying dialogue language models. Overall, the contributions of this paper include the introduction of a novel LLM agent communication framework that has the potential to promote autonomous cooperation among communicating agents. In addition, the research also provides a scalable method to study multiple communication agents. Cooperative behavior and capabilities of agent systems. Finally, the author used the CAMEL framework to let the agents play different social roles, and then modeled the AI ​​society, and collected a large number of natural language instruction data sets. Currently, there are four AI society, code, mathematics and AI society ten language translations. The data set is already available for download at HuggingFace:

Download address: https://huggingface.co/camel-ai

In addition, the project homepage provides an online demo of using CAMEL for game design, molecular dynamics simulation and real-time simulation for everyone to try: https://www.camel-ai.org

The above is the detailed content of Multiple ChatGPTs cooperate to complete designated tasks. Is a mini AGI controlled world coming?. For more information, please follow other related articles on the PHP Chinese website!