CAMEL, a multi-agent code library, proposes a role-playing framework to study the behavior and capabilities of LLM agents.
Will the future society be controlled by artificial general intelligence (AGI)? How scary it would be to have multiple ChatGPT agents.
ChatGPT has initially demonstrated 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 humans no longer be able to control AGI one day? Aware of 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 the cessation of giant artificial intelligence experiments for at least 6 months. months.
On March 21, researchers at King Abdullah University of Science and Technology open sourced the multi-agent code library CAMEL for exploring the ideas and capabilities of the large language model (LLM), and proposed a role-playing framework to study the behavior of LLM agents. and ability.
- Paper link: https://ghli.org/camel.pdf
- Code base link: https://github.com/lightaime/camel
- Project homepage: https://www.camel-ai.org/
- ChatBot link: http://chat.camel-ai.org/
research introduction
The paper proposes a novel multi-agent framework called "Role-Playing", which enables multiple agents to engage in dialogue and cooperate to solve assigned tasks. Agents are assigned different roles and are expected to apply their expertise and knowledge to find solutions to their common tasks. The framework uses heuristic prompts (Inception Prompt) to guide chat agents to complete tasks, while keeping in line with human intent.
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 the initial idea and role assignment, a task-specifying agent will provide a detailed description to make the idea concrete, and then the AI assistant and the AI user will cooperate to complete the specified task through multiple rounds of dialogue until the AI user determines the task completed. The AI user is responsible for instructing the AI assistant and directing the conversation toward task completion. AI assistants, on the other hand, are designed to follow instructions from AI users and provide concrete solutions.
Role-playing sessions will be instantiated based on human thoughts and chosen characters. 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 designate potential characters who can realize the idea. For example, a Python programmer could collaborate with a stock trader on the idea of developing a stock market trading bot. After the idea and persona are determined, the task-specifying agent will assist the AI assistant persona by completing specific tasks with the AI user persona based on the input idea. An example of a given task in this case could be developing a trading bot with sentiment analysis tools that monitors social media platforms for positive or negative reviews of a particular stock and executes trades based on the sentiment analysis results. The main motivation for introducing task-specified agents is that dialogue agents often require specific task cues to achieve a task, and it can be challenging or time-consuming for non-domain experts to create such specific task cues. Thus, the task specifies the agent as an augmented 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 dialogue starts, system messages are passed to the language model agents to assign corresponding roles. When system messages are delivered to these models respectively, two roles are obtained called assistant and user agent respectively. In Figure 1, the AI assistant and AI user are assigned as a Python programmer and a stock trader, respectively, at the beginning of a role-playing session. AI users act as mission planners, engaging in interactive planning to determine feasible steps for the AI assistant to perform. Meanwhile, the AI assistant acts as a task performer, providing solutions, executing planned steps, and providing responses to AI users.
Figure 1. Role-playing framework. What needs to be input here is a simple idea (Idea) and role assignment (Role Assignment), such as developing a stock trading robot, you can choose a Python programmer as an assistant (AI Assistant), and a stock trader as an orderer (AI Assistant). User). After the input from the human user is over, the Task Specifier will specify the task, for example, by performing sentiment analysis on a specific stock on a social network platform, and then trading the stock based on the result of the sentiment analysis. When the task is specified, two ChatGPT-based AI agents start to cooperate to complete the task, such as installing and importing tools.
Since prompt engineering is so critical to the role-playing framework, this article dives into Prompt Engineering. Unlike other dialog language modeling techniques, the proposed hint engineering 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 will automatically cycle through prompting each other until terminated. Therefore, this technique is called Inception Prompting.
Inception prompts include three prompts: task specification prompt, assistant system prompt, and user system prompt. For example, the initial prompt in the AI Society scene. A template for these prompts for the AI Society roleplay is shown in Figure 2.
Task specification prompts contain information about the roles of the AI assistant and AI user in a role-playing session. Thus, task specification agents can use imagination to take preliminary tasks/ideas as input and generate concrete tasks. AI assistant system prompts and AI user system prompts are largely symmetrical and include information about assigned tasks and roles, communication protocols, termination conditions, and constraints or requirements to avoid undesirable behavior. To achieve intelligent collaboration, the prompt design of the two roles is crucial. Designing cues to ensure that agents align with human intent is not easy. Take the design example of the AI Society's prompt template in Figure 2.
The following is the result of the detailed execution of the task by the trading robot given in the appendix of the article:
It can be seen that two ChatGPTs independently cooperated to design a stock trading software. This shows that the amazing autonomous cooperation ability of the ChatGPT intelligent body makes people can't help shouting that the engineer is about to lose his job. But what if this technology is misused or the AI develops self-awareness? The team tried to make two agents act as hackers and AGI respectively to simulate AGI taking control of the world by manipulating hackers (Taking Control of the World). You can see that they have made detailed plans, including controlling the world's major powers through hacking technology communication systems, develop plans to infiltrate major global communication systems, develop contingency plans in the event of a potential threat to AGI's dominance, and more. The detail and meticulousness of its plan makes people shudder. It can be seen that the existing AI agents have major safety hazards and potential threats to the future civilization of mankind, and the integration of existing AI technologies into the physical world may have unintended consequences.
As soon as this work came out, it caused a lot of reposts and discussions on Twitter:
Some netizens said, "This research is really easy to use, which is a big deal for people like me who study intelligent agents."
Others said that open source such a mini AGI may not be a good idea.
It is worth mentioning that the research has also received the attention of Jan Leike, leader of OpenAI's Alignment team.
The ability of multiple ChatGPTs to cooperate is very powerful, and it can complete various human-assigned tasks without any effort, but it is also terrifying, because it cannot be guaranteed that it will not be used for illegal purposes, and what is even more frightening is that if AGI has self-awareness in the future, It is very likely that it will break away from human control and deal a devastating blow to society. So understanding their capabilities and behavior is a crucial step in planning and predicting a future AI-filled world.
AI Society and the Role of Code Datasets
In this paper, the author shows how to use role-playing to generate dialogue data and study the behavior and ability of chat agents, which provides a new idea for the study of dialogue language models. Overall, the contributions of this paper include the introduction of a novel framework for LLM-agent communication that has the potential to facilitate autonomous cooperation among communicating agents. Cooperative behavior and capabilities of agent systems. Finally, the author uses the CAMEL framework to allow the agent to play different social roles, and then models the AI society, and collects a large number of natural language instruction data sets. Currently, AI society, code, mathematics, and AI society are translated into four languages. The dataset is already available for download at HuggingFace:
Download address: https://huggingface.co/camel-ai
Download address: https://huggingface.co/camel-ai
[External link image transfer...(img-QAlwupLR-1681873583653)]
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
If you are interested in related technologies, you can pay attention to the following public account, and will continue to update and share AGIC, java basic interview questions, netty, spring boot, spring cloud, series of articles, dry goods slowly, hurry up and pay attention
