New large language models continue to emerge.However, over time, concerns about whether the development of large language models has peaked are gradually rising.Today, industry leaders believe that the development of artificial intelligence is shifting from the research phase of foundational large models to the application development phase of large models.Although the direction of development is clear, applications such as question answering and dialogue have begun to feel tedious, and the future path remains full of uncertainties, making it hard to imagine.
Image: Logo of MetaGPT
However, the case of multi-agent collaboration in generating complex software showcased by MetaGPT gives us a different feeling: the development of AI applications will not be a simple linear path.
Image: Software Engineering SOP, MetaGPT creatively incorporates processes into its AI system, linking roles supported by large language models to collaboratively complete complex software development projects.
MetaGPT works like a virtual software development company. It plays multiple roles such as product manager, software architect, software engineer, and tester, referred to as “multi-agents.” Users only need to describe a software development project in natural language, such as: “Create a ‘2048’ moving block puzzle game,” and MetaGPT’s multi-agents can collaborate based on standardized software engineering operational procedures (SOP) to generate product design documents, architecture design documents, program code, and runnable software.
Image: “2048” (left), Snake (middle), Flappy Bird (right) are all famous video games. They can be generated with zero coding through MetaGPT, successfully running.
Currently, users can easily generate games like “2048”, “Snake”, “Flappy Bird”, and even enterprise software like CRM and ERP through MetaGPT. The generation of these software only requires a few minutes and a small amount of large model API call costs, possibly just a few dollars. This proves the effectiveness of the “multi-agent collaboration” method supported by the current foundational large language models.
Image: Stanford AI Multi-Agent Town SmallVille
From AutoGPT breaking down user needs into small tasks and completing them, to Stanford University building a virtual town called “Smallville” where 25 AI agents live, to MetaGPT’s multi-agent collaboration completing software engineering projects, we can see that over time, the construction and collaborative work of “agents” and “multi-agents” are becoming more in-depth and refined.
The foundational capabilities of these agents, such as natural language understanding, reasoning, and generation, are built on the “black box” foundational large models. However, people are also beginning to provide “white box knowledge” to intelligent systems and agents in various ways. One of the main ideas of MetaGPT is to integrate standardized operational processes of software companies into intelligent systems, filling the gaps of large models.
The case of multi-agent collaboration in MetaGPT heralds a new possibility for the application development of large language models. It will not be a simple continuation of linear technological development, but a sophisticated combination of AI’s black box knowledge and human’s white box knowledge, sedimenting, mining, and leveraging the capabilities of large language models to form breakthrough applications that bring us greater surprises and shocks.
Please note that the multi-agent framework of MetaGPT is an open-source project, and you can find the software on GitHub[1] and learn more about its technology in the paper “METAGPT: META PROGRAMMING FOR MULTI-AGENT COLLABORATIVE FRAMEWORK”[2].
References
[1]. https://github.com/geekan/MetaGPT
[2] “METAGPT: META PROGRAMMING FOR MULTI-AGENT COLLABORATIVE FRAMEWORK,” https://arxiv.org/pdf/2308.00352.pdf