On December 25, the beta version of Kunlun Wanwei AI Agents development platform "Tiangong SkyAgents" was officially opened for testing. Users can experience it immediately at https://model-platform.tiangong.cn/ .
Kunlun Wanwei's "Tiangong SkyAgents" AI Agents development platform is built based on Kunlun Wanwei's "Tiangong Large Model" and has independent learning and independent thinking capabilities from perception to decision-making, from decision-making to execution. Users can build their own single or multiple "personal assistants" through natural language, and can modularize different tasks and implement execution through operating system modules, including question presets, designated responses, knowledge base creation and retrieval, and intent recognition. , text extraction, http request and other tasks.
At a time when large model technology is developing rapidly and the application of AI Agents continues to advance, Kunlun Wanwei's "Tiangong SkyAgents" is our exploration and attempt in the field of intelligent agents. This platform may not be perfect, but we hope to work with developers to build and grow together, and continue to expand the application boundaries of artificial intelligence technology. Imperfection now is for the sake of perfection in the future. We have always believed in and dared to make breakthroughs in the process of technological pursuit.
Work together to explore and cooperate to create
In the era of large models, interactive AI is expected to become the mainstream implementation direction of large model technology in the future. History tells us that the evolution of emerging things will always find a stable term to describe this carrier, and AI Agents (intelligent agents) have shown great potential.
At present, the world's attention to agents is extremely enthusiastic. OpenAI pays great attention to the field of agents and released custom GPTs and Assistance API at the OpenAI dev day conference; the co-founder of DeepMind also recently mentioned the next generation of artificial intelligence technology. The direction of development will be interactive AI, not generative AI. This kind of interactive AI is largely consistent with the description of an agent. Users can ask the agent to complete various tasks, and the agent can operate software or collaborate with humans to complete work in complex scenarios.
In terms of technology paradigm, Kunlun Wanwei is also constantly thinking about the underlying technologies and architecture that drive the rapid development of agent technology. We also clearly realize that even with the support of language interaction capabilities of large models, we are still far away from an intelligent agent that can fully automatically make decisions and perform tasks.
Today, Kunlun Wanwei officially opened the beta version of "Tiangong SkyAgents" as an exploration of our technical capabilities and application capabilities of AI Agents . We hope that through this exploration, more and more users and developers will be able to apply large model technology to their work and life, creating exclusive AI Agents that meet daily needs and inspire innovation. We also hope that more friends who are interested in AI Agents can join us to create together. We welcome suggestions and comments from our partners.
What are AI Agents?
Agent is generally translated as "intelligent body" or "agent". The concept was first proposed by Marvin Minsky, one of the founders of the MIT Artificial Intelligence Laboratory (MIT AI Lab), in his book "The Society of Thinking" published in 1986. propose. It is introduced into computing systems from the concepts of society and social behavior, and refers to computing entities that can continue to function autonomously in a certain environment.
AI Agents refer to intelligent entities driven by artificial intelligence technology that can perceive the environment, make decisions and perform actions.
AI Agents is not an emerging concept. Since the establishment of the artificial intelligence technology discipline, research around AI Agents has emerged one after another. After the rise of the deep neural network wave in 2012, an academic faction was born that uses reinforcement learning to train AI Agents. The world-famous Go robot AlphaGo can be regarded as the research result of this school. However, this type of AI Agents is more suitable for confrontational game scenarios and is difficult to implement in the real world.
However, the emergence of large models changed all this.
In 2023, with the breakthroughs of large model technology in the fields of natural language understanding, engineering capabilities, data capabilities, storage capabilities, etc., a large number of conversational interaction "GPT" will emerge. AI Agents driven by large model technology will be more advanced in versatility, practicality, Implementability and other aspects have been developed rapidly, setting off another wave of AI Agents craze around the world.
Traditional large model applications are mostly implemented based on prompts (user prompt words). The quality of prompts will directly affect the answer effect of the large model. Ordinary users who lack prompt word engineering capabilities will find it difficult to maximize the true capabilities of the large model . AI Agents only require the user to give a work goal, and then they can gradually complete the task through independent thinking and calling tools, which greatly reduces the threshold for the application of large model technology.
AI Agents three core modules: brain, perception, and execution
According to the Fudan University paper "The Rise and Potential of Large Language Model Based Agents: A Survey", AI Agents can be divided into three major modular capabilities: brain (Brain), perception (Perception), and execution (Action).
（图片来源：《The Rise and Potential of Large Language Model Based Agents: A Survey》）
1. Brain
The brain is the "core information processing center" of AI Agents. It has the ability to understand the current environment and form "Memory", and it also has the ability to store and retrieve long-term memory. The "brain" can perform logical reasoning based on "memory" and currently received information, and decompose complex problems into achievable sub-tasks to cope with complex scenario tasks. At the same time, through RAG (Retrieval Augmented Generation) technology, AI Agent can make further decisions based on the current scenario and the goals set by the user, achieving independent thinking, planning (Planning) and reasoning (Reasoning).
2. Perception
The perception module allows AI Agents to obtain sufficient information based on the current environment and scenarios, which is what distinguishes it from traditional RPA systems. RPA systems cannot work when faced with a large amount of unknown information and unpredictable environments. AI Agents can perceive, understand and autonomously explore the world by sensing information and making corresponding thoughts and actions.
3. Execution (Action)
The execution module gives AI Agents the authority and ability to perform tasks. After receiving user task instructions, AI Agents combine the current scene information collected by the perception module, summarize and reason through the brain, and output it to the execution module, so that AI Agents can complete instructions according to user needs. At the same time, AI Agents have the ability to call and use tools. These tools can help Agents complete complex tasks more efficiently, while also improving their credibility and flexibility in certain specific scenarios. Related application scenarios This includes allowing AI Agents to purchase airplane tickets, order takeout, and complete corporate IT/customer service/legal tasks, etc.
SkyAgents
"Tiangong SkyAgents" is built based on Kunlun Wanwei's "Tiangong Large Model" and has exclusive "brain", "perception" and "execution" modules.
Individual users/developers can perform natural language and simple operations through "Tiangong SkyAgents". Without coding skills, they can deploy their own AI Agents in a few minutes to complete industry research reports, document filling, trademark design, and even Fitness plans, travel flight bookings and many other customized needs.
Enterprise users/developers can assemble the many capabilities of "Tiangong SkyAgents" on demand into many personalized applications such as enterprise IT, intelligent customer service, corporate training, HR, legal consultants, etc., and support one-click service deployment to ensure their Seamless access in different business systems.
Behind the AI capabilities of "Tiangong SkyAgents" is the accumulation of capabilities of Kunlun Wanwei AI Agents technology in the fields of modular task components, intelligent knowledge base construction, third-party tool invocation, and one-click sharing of personalized AI Agents.
1. Modular task components, zero code to create exclusive AI Agents
Currently, most users have neither code development experience nor the ability to train large model prompt engineering. It is difficult to quickly realize many practical needs of daily life through dialogue and question-and-answer formats, and cannot maximize the capabilities of large models. . In order to solve this problem, "Tiangong SkyAgents" has modularized a large number of task components and integrated capabilities such as intelligent dialogue, information processing, information extraction, information classification, third-party data acquisition, and vector retrieval.
- Intelligent dialogue: The intelligent dialogue module uses AI capabilities to process the content sent by the user through a large language model and reply to the user's specified content.
- Information processing: Using preset prompt words (Prompt), the large model processes specific information input to obtain content that meets needs.
- Information extraction: Through the understanding of semantics by large models, target information can be extracted from input information
- Information classification: With the help of intelligent analysis of large models, user problems are classified and different operations are performed for different types of problems to facilitate personalized processing;
- Third-party data acquisition: Third-party data access will carry relevant parameters. The system sends a POST request to the specified address and receives a response. While carrying relevant parameters, the system can realize data interconnection and interoperability with other application services. Based on third-party data acquisition modules, the capabilities of AI Agents can be greatly expanded, opening up more scenarios such as database operations and online searches.
- Vector retrieval: For common user questions, the system can add questions to the knowledge base for easy search and retrieval. For the "knowledge base" module, users can enter questions, and the system will search for relevant questions and answers in the knowledge base and output them in natural language.
2. Intelligent knowledge base construction to support large-scale knowledge import
Although large models are powerful, they also have their inherent weaknesses. On the one hand, the knowledge obtained by large models through parameter training can only stay at a certain point in time, and the update cost is very high; on the other hand, the training data of large models are usually based on general knowledge, and data in subdivided fields are often lacking. In order to solve this problem, "Tiangong SkyAgents" supports the import of more formats and larger scale data and knowledge, adding a "knowledge base brain" to large models.
- Supports a variety of data import forms: text, files, websites, Q&A pairs, online documents, etc. to easily import existing knowledge
- Knowledge base Embedding: Represent content elements in the knowledge base as low-dimensional vectors, making it easier for elements in the knowledge base to perform mathematical operations such as calculating similarities and finding adjacent entities, thereby improving the knowledge base in AI Agents operability in.
- Free linking of knowledge bases: Each AI Agents can freely link to its own knowledge base and be provided with content by multiple knowledge bases at the same time. Each knowledge base can also link to multiple AI Agents at the same time. Knowledge base content can be enabled and disabled as needed. Enable more flexible knowledge content management.
3. Call third-party tools to handle multiple scenarios as you wish
- Third-party tool calling: The tool calling capability is one of the core capabilities that distinguishes AI Agents from a large number of conversational GPTs. For example, in the air ticket booking scenario, in addition to analyzing and judging user needs and flight information, AI Agents also need to call ticketing Different tools such as platforms and electronic payments. Therefore, in addition to basic modules, "Tiangong SkyAgents" also supports the call of various third-party tools. Users can develop tools according to their own needs, giving them more flexibility when building AI Agents.
4. Personalized AI Agents share with one click
- One-click sharing: In order to give back to users and developers and make it easier to create and use AI Agents, "Tiangong SkyAgents" launched an exclusive New Year event and launched three official New Year templates: Ideal Partner, Opportunity Partner, and Warm Home. The distribution and usage process is completely simplified. AI Agents designed by users based on their own creativity can be shared with more people through links. Users only need to click on the link to gain access to the AI Agents. Click to create: https://model-platform.tiangong.cn/
The capabilities are more comprehensive, the applications are smarter, sharing is more convenient, and the platform is easier to use. The officially opened content of the beta version of "Tiangong SkyAgents" will further promote the universalization of large model technology, help individuals and small and medium-sized enterprises who lack code development capabilities to actively embrace large model technology, and help large models enter thousands of households. Contribute to the ecological development of artificial intelligence.
Scan the QR code to enter the "Tiangong Open Platform" and quickly build AI Agent
