It can be a personal assistant that directly helps users make travel plans, order takeout, etc.; it can also be an enterprise assistant that accomplishes software development, makes recruitment, draws up contracts, etc., and different scenarios require different intelligent bodies.

DeepWisdom was founded in 2019, focusing on the development and commercialization of AI Infra. Intelligent bodies will bring a new user experience and the opportunity to reshape all traffic. This means that there is a huge market space to provide a complete Stack (architecture) for an intelligent body society and to provide intelligent body services.DeepWisdom hopes to build AgentStack (the underlying architecture of intelligent bodies), so that intelligent bodies can become a universally significant advanced species, form an intelligent body society, form a new paradigm of new human-machine collaboration, and liberate the society from more work forces.


Ideas for building an intelligent body community

DeepWisdom previously focused on AI Infra, which constitutes the kernel of intelligent body operation. It can provide scheduling, orchestration, computation, storage, network and other Infra capabilities, as well as front-end, back-end, data, ML, LLM, aPaaS capabilities, and even provide a standard library that allows the intelligent body to call most of the real-world APIs. when the intelligent body has these capabilities, it will theoretically realize omnipotence.

But theoretical omnipotence doesn’t mean it’s actually omnipotent, because code is difficult to write, and it’s hard for an intelligent body to call correctly in the face of huge APIs.

MetaGPT is designed to solve this problem.

MetaGPT is a multi-intelligent body open source framework, mainly used to solve the problem of large model landing application, it can be simulated through the simulation of software companies with multiple roles, the output of product design, architectural diagrams, code repo and so on. It will complete the automation of programming, complete utilization of AgentOS capabilities, and complete the mass production of Agent

In the view of DeepWisdom founder and CEO Chenglin Wu, the intelligent body (AI AGENGT) will be the entrance to the next era. And the past experience has laid the foundation for promoting the development of AI Agent.

Intelligent body consists of LLM+Observation+Thinking+Action+Memory and so on. Multi-intelligentsia, on the other hand, need to be composed and synergized by different dimensions such as Agent + Environment + SOP + Review + Routing + Subscription + Economy, in order to build a virtual community that can interact. Intelligent bodies can perform long-range tasks, which can increase efficiency and reduce the burden for humans.

The key to intelligent bodies is SOP. the principle of MetaGPT is not complicated. There are two types of code in the human world, “code” for machines and SOP for humans, and there is a formula Code = SOP(Team). In the past, a tech company could drive a team of humans to produce code based on SOPs, but if you change the composition of the team in the formula from humans to intelligentsia, you’ll find that most of the work can be given to intelligentsia.

Different scenarios require different roles, different skills, and therefore different intelligences. There will also be a lot of interaction between these intelligences. the concept of division of labor was introduced more than 200 years ago, and the division of labor leads to occupations, and occupations lead to specialization and different sop skills. sop is the key to the skills of the different types of work, and enables each intelligence to play a specific role.


DeepWisdom founder and CEO Chenglin Wu said that there are more than 1,200 occupations and 19,900 skills, and 80% of the occupations will be exposed to the big language model, for example, according to the scores of the language class and the big language model, the exposure of programming jobs is at 63.4, and the exposure of jobs such as researcher, trader and data analyst is at 100%.

This means that there is enough data for these types of jobs to generate Agents, and all real-world occupations can be modeled as intelligences, and the intelligences can become a standardized cluster.

According to Chenglin Wu, the most difficult work in MetaGPT is the design related to intelligent bodies and large language models, and how to do a good job of abstraction and frameworks, DeepWisdom designed a high degree of code subtlety and a simple framework design.

Unlike AI application layer frameworks, DeepWisdom focuses on the standardization of the intelligentsia itself, using standardized operating procedure (Sops) coding as a cue to drive multiple intelligentsia by solving the illusion problem well.

According to Chenglin Wu, only a dozen or so intelligences are needed at the very core, and with the underlying foundation, more intelligences can be generated by combining knowledge from different perspectives.

In terms of SOPs, intelligent bodies can play certain roles, SOPs can give vocational skills, and programming can constitute different SOPs. what the company needs is to perfect different SOPs and build out different intelligent bodies.

In the past few months, we have been leading the open source community in terms of active volume, and it is the fastest-growing community in the open source community, which has top engineers from all over the world contributing and generating different communities of intelligences together,” said Chenglin Wu.

For example, as recently as October, MetaGPT has gained 30,000 stars on GitHub, with the number of stars being an indicator of popularity. In the official website of the open source project STAR HISTORY can be seen, recently, the attention of MetaGPT is a steep rise in the situation, its attention is much higher than similar intelligent body framework.

The number of stars of multiple intelligences curve


Recently, in the Mini Hackathon organized by MetaGPT in the open source community, three teams respectively spent only three weeks to realize the replication of intelligent bodies of games such as Minecraft, Werewolf, Stanford Virtual Town and so on based on the MetaGPT framework. This has the contribution of students from several schools such as UC Berkeley, Northwestern University, University of Melbourne, University of Hong Kong, Peking University, and engineers from Tencent, Ali, Smart Spectrum, and many other companies.

These top talents will help push the enrichment and refinement of intelligences more efficiently and quickly, and even the SOPs of many intelligences are constructed by these engineers.

In Wu Chenglin’s opinion, in order to achieve such a multi-intelligent virtual community, the open source framework needs to accomplish more than 200 things, and DeepWisdom has already accomplished 80 of them, and completed the most important “realize the intelligent body’s thinking strategy”.

In their conception, in the future, different intelligences can form a community in the virtual space, and intelligences such as software engineers can also generate their own intelligences. Different intelligent bodies can score each other, so that intelligent bodies constantly optimize themselves.

Currently, DeepWisdom has more than 60 employees, DeepWisdom founder and CEO Chenglin Wu has led the design and implementation of a billion users of large model recommender systems, search engines, natural language understanding, and other top projects, and won the NeurIPS 2 world championships.