


Targeting the intelligent society, MetaGPT joins hands with Jürgen Schmidhuber team
In the past few months, MetaGPT [1]’s Agents software company examples have been impressive. It quickly gained 30k stars on GitHub and dozens of global Professional media and big V reports. But agent software companies are just a microcosm of Agent Society. An intelligent society may have software companies, e-commerce companies, and game companies, as well as a large number of independent agents to provide productivity. Jürgen Schmidhuber, the father of modern artificial intelligence, also strongly recognizes the concept of an intelligent society. He and his team have made significant contributions to MetaGPT and are included in the list of MetaGPT authors.
As early as 1986, Marvin Minsky led an ideological revolution in the field of artificial intelligence with his work "Society of Mind (SOM)" [2]. He proposed a very original theory: the mind does not need to be composed of individual intelligent components, but rather a complex system composed of the interaction of a series of simple components. It is this collection that gave birth to what we know. Intelligence and consciousness. This concept has had an immeasurable and far-reaching impact on the construction of autonomous agents and their subsequent development.
With the leap forward in artificial intelligence technology to 2023, we can now imagine how if each tiny component itself possesses a certain degree of intelligence, how they will interact and what level of collective intelligence they will produce. Among the research papers on Natural Language Agent Society (NLSOM, Language Agent Society) in the first half of 2023 [3], from well-known research institutions such as King Abdullah University of Science and Technology, Swiss Artificial Intelligence Laboratory, Oxford University and ETH Zurich scientists jointly explored the possibility of intelligent agent communities.
They proposed to build a language-driven community of agents that can collaboratively complete tasks that a single agent cannot or is difficult to complete independently. The study proposes a series of experimental ideas that are more than just proofs of concept. They are seen as a precursor to a society containing trillions of intelligent agents, which may also include human members.
https://arxiv.org/pdf/2305.17066.pdf
At the 2023 CogX Festival, Jürgen showed the audience his vision for large-scale Insights from Language Models (LLMs). When discussing topics related to agents, he mentioned various ways to build self-improving systems, including Universal Turing Machines [4] and Gödel machines [5]. He pointed out that the current large language model provides us with a new way of thinking - by using a universal symbolic language (such as natural language or programming code) as an interface to connect different models. These models can communicate with other language models to build a natural language society of mind (NLSOM) paradigm.
Professor Jürgen Schmidhuber is Scientific Director of the Swiss Artificial Intelligence Laboratory (IDSIA) and Director of the Artificial Intelligence Center at King Abdullah University of Science and Technology (AI Initiative, KAUST) . His work has a profound impact on important artificial intelligence directions such as reinforcement learning (Reinforcement Learning), meta learning (Meta Learning), and neural networks (Neural Network).
So far, Professor Schmidhuber’s Google Scholar citations are 210,000, of which the long short-term memory (LSTM) paper as a co-inventor has been cited 90,000 times. At the age of 15, he hoped to develop an artificial intelligence that was smarter and capable of self-improvement so that he could retire. Two of the four people who founded DeepMind, as well as their first AI Ph.D. recruit, came from Jürgen Schmidhuber’s lab.
In the society Jürgen envisions, all communication is transparent and easily interpretable. He mentioned a concept called "Mindstorm", that is, when a problem is given, this natural language mind society can cooperate to solve the problem.
In this process, each member of society may have different ideas and perspectives, and they will collect and integrate these different ideas to make collective decisions.
This method is particularly suitable for solving problems that cannot be effectively solved by a single agent. Jürgen further explained that such problems can be of a programming nature, such as using the Python language to solve a specific programming problem. Through this synergy, the collective intelligence of the intelligent agent society will be able to achieve solutions that transcend individual capabilities.
This iteration of the MetaGPT project received direct guidance from Jürgen, and his team also provided a lot of support in coding, writing, and engineering.
Next, this article will analyze the updated content of the MetaGPT paper in detail so that readers can understand its details more deeply.
1. Paper and Framework Update
Paper Section 3.1 Update: Expounds the concept of role specialization design and role division of labor in the MetaGPT framework, and illustrates the role of a single agent in Behavioral patterns in MetaGPT and organization under SOPs.
Paper section 3.2 update: Introducing the communication mechanism in the MetaGPT framework, including structured communication interface design and publish-subscribe mechanism.
Update to Section 3.3 of the paper: An executable feedback mechanism is introduced, which is a mechanism for continuous iteration and self-correction during code execution.
Fig.2. Example of communication protocol (left) and example of iterative programming with on-the-fly executable feedback (right). Left: Agents use a shared message pool to publish structured messages. They can also subscribe to relevant messages based on their configuration. Right picture: After generating the initial code, the engineer agent can execute the code and check whether errors are reported during operation. If an error occurs, the Agent will check the execution results and compare them with PRD, system design and code files to rewrite and optimize the code.
1.1. Agent communication protocol
Currently, most multi-agent agents complete collaboration through dialogue based on natural language, but this is not suitable for solving specific tasks. The best way.
Natural language output without constraints and specific requirements may lead to distortion of information content or shift in semantic focus.
Therefore, structured communication content and interface forms help agents to quickly and accurately understand task requirements, and are also conducive to maximizing the retention of information content. Referring to the role requirements for different positions in human SOPs, we set output specifications for each role that are in line with human experts in the corresponding position, requiring the agent to convert the original natural language information into a more structured expression (as shown in the figure below). Such as data structure, API design and sequence diagram.
Fig.3 MetaGPT software development process diagram shows that structured SOPs can bring better results. For a more detailed demonstration, see Appendix B
In subsequent experiments, we compared the actual solution effects of MetaGPT and ChatDev (using a chat-form communication and collaboration mechanism) to perform this complex task of software development. The results illustrate the structure. Specialized communication interface design can bring significant effects to intelligent agent collaboration.
Publish-Subscribe Mechanism
In the communication process of multi-agent, relying only on the 1v1 single-point communication method will not only increase the complexity of the communication topology, but also lead to collaboration The low efficiency will also sharply increase development costs. Therefore, we implemented a shared message pool and interest-based subscription within the framework through the [publish-subscribe] messaging mechanism.
Specifically, the environment provides a shared message pool from which agents can obtain information directly without asking other agents one by one. At the same time, the agent can filter and filter messages based on its own interests/concerned behaviors, thereby reducing message/memory overload. As shown in Figure 3, the architect mainly pays attention to the product manager's PRD document output, but pays less attention to the test engineer's document.
1.2. Executable iterative feedback design
Debugging and execution feedback play an important role in daily programming tasks. However, existing methods often lack self-correction mechanisms and only conduct code feasibility assessment through code review and review mechanisms. To further reduce the hallucination problem of LLM in generating code, we introduce an executable feedback mechanism to iteratively improve the code. Through automatic code execution test result feedback, code feasibility assessment and judgment are carried out, and LLM is promoted to carry out self-iteration and optimization. As shown in Figure 2, engineers can continuously update the code based on the code execution results and iteratively test until the test passes or exits after a maximum of N retries.
2. Experiment update
In the experimental part, we have added an exploratory experiment on the effect of introducing a multi-agent framework into SOPs, and code brought by executable iterative feedback Quality improvement experiments. On the dataset:
- Effect evaluation on code quality: We use two public benchmark data sets: HumanEval and MBPP.
1) HumanEval includes 164 handwritten programming tasks. These tasks include functional specifications, descriptions, reference code, and tests.
2) MBPP contains 427 Python tasks. These tasks cover core concepts and standard library functionality and include instructions, reference code, and automated tests. - We propose SoftwareDev, a benchmark dataset for more challenging software development tasks: Our SoftwareDev dataset collects 70 representative software development task instances, each with its own task prompt (See Table 5 of the paper). The range of these tasks is diverse (see Figure 5 of the paper), such as mini-games, image processing algorithms, data visualization, etc. They provide a powerful testing platform for real development tasks. Unlike previous datasets, SoftwareDev focuses on engineering aspects. In the comparison, we randomly selected seven representative tasks for evaluation.
2.1. Executable iterative feedback design
Figure 4 shows that MetaGPT performs well in both the HumanEval and MBPP benchmarks. Outperforms all previous methods. When MetaGPT (using GPT-4 as the base model), its Pass@1 on the HumanEval benchmark is significantly improved compared to GPT-4. It achieved 85.9% and 87.7% in these two public benchmarks (considering the experimental cost, the numerical results of some models are directly used the results provided by Dong et al. (2023). [6]).
2.2. Software development task data set & evaluation indicators
For SoftwareDev, we prioritize the actual usability of build projects and evaluate performance through manual evaluation (A, E) or statistical analysis (B, C, D). We demonstrate MetaGPT through visual examples The ability to generate independent software (Figure 5 of the paper). Additional experiments and analysis can be found in Appendix C of the paper: (A) Executability: This metric rates the generated code from 1 (failed/non-functional) to 4 (defect-free). 1 means non-functional, 2 means working but not perfect, 3 means nearly perfect, and 4 means no defects. (B) Cost: The cost evaluation here includes (1) project running time (2) Token consumption and (3) actual costs. (C) Code statistics: including (1) the number of code files (2) the average number of lines of code per file and (3) the total number of lines of code. (D) Production efficiency: The basic definition is the Token usage divided by the number of lines of code, that is, the Token consumed by each line of code. The smaller the value, the higher the code production efficiency. (E) Manual revision cost: Quantified by the number of revision rounds required to ensure that the code runs smoothly, which represents the frequency of manual intervention, such as debugging or importing dependencies and other revisions.2.3, SOPs vs ChatChain
In the scenario of solving specific tasks, in order to explore the effect of SOPs on multi-agent collaboration ,We selected ChatDev, an agent framework that supports ,software development tasks in open source work, as the ,experimental comparison object. ChatDev is a framework for agent organization and collaboration based on the role division of ChatChain and software development waterfall flow. We selected 7 tasks from SoftwareDev for comparison and compared the above-mentioned relevant indicators to illustrate the differences.#3. Acknowledgments
Thanks to Sarah Salhi, executive secretary of KAUST AI Center, postdoctoral fellow Wang Yuhui, and doctoral student Wang Wenyi for this The paper provides suggestions and help. [1] https://arxiv.org/pdf/2308.00352.pdf[2] https://en.wikipedia.org/wiki/Society_of_Mind
[3] https://arxiv.org/pdf/2305.17066.pdf
[4] https://en.wikipedia.org/wiki/Universal_Turing_machine
[5] https://en.wikipedia.org/wiki/Gödel_machine
[6] https://arxiv.org/abs/2304.07590
The above is the detailed content of Targeting the intelligent society, MetaGPT joins hands with Jürgen Schmidhuber team. For more information, please follow other related articles on the PHP Chinese website!

Hot AI Tools

Undresser.AI Undress
AI-powered app for creating realistic nude photos

AI Clothes Remover
Online AI tool for removing clothes from photos.

Undress AI Tool
Undress images for free

Clothoff.io
AI clothes remover

Video Face Swap
Swap faces in any video effortlessly with our completely free AI face swap tool!

Hot Article

Hot Tools

Notepad++7.3.1
Easy-to-use and free code editor

SublimeText3 Chinese version
Chinese version, very easy to use

Zend Studio 13.0.1
Powerful PHP integrated development environment

Dreamweaver CS6
Visual web development tools

SublimeText3 Mac version
God-level code editing software (SublimeText3)

Hot Topics

But maybe he can’t defeat the old man in the park? The Paris Olympic Games are in full swing, and table tennis has attracted much attention. At the same time, robots have also made new breakthroughs in playing table tennis. Just now, DeepMind proposed the first learning robot agent that can reach the level of human amateur players in competitive table tennis. Paper address: https://arxiv.org/pdf/2408.03906 How good is the DeepMind robot at playing table tennis? Probably on par with human amateur players: both forehand and backhand: the opponent uses a variety of playing styles, and the robot can also withstand: receiving serves with different spins: However, the intensity of the game does not seem to be as intense as the old man in the park. For robots, table tennis

On August 21, the 2024 World Robot Conference was grandly held in Beijing. SenseTime's home robot brand "Yuanluobot SenseRobot" has unveiled its entire family of products, and recently released the Yuanluobot AI chess-playing robot - Chess Professional Edition (hereinafter referred to as "Yuanluobot SenseRobot"), becoming the world's first A chess robot for the home. As the third chess-playing robot product of Yuanluobo, the new Guoxiang robot has undergone a large number of special technical upgrades and innovations in AI and engineering machinery. For the first time, it has realized the ability to pick up three-dimensional chess pieces through mechanical claws on a home robot, and perform human-machine Functions such as chess playing, everyone playing chess, notation review, etc.

The start of school is about to begin, and it’s not just the students who are about to start the new semester who should take care of themselves, but also the large AI models. Some time ago, Reddit was filled with netizens complaining that Claude was getting lazy. "Its level has dropped a lot, it often pauses, and even the output becomes very short. In the first week of release, it could translate a full 4-page document at once, but now it can't even output half a page!" https:// www.reddit.com/r/ClaudeAI/comments/1by8rw8/something_just_feels_wrong_with_claude_in_the/ in a post titled "Totally disappointed with Claude", full of

At the World Robot Conference being held in Beijing, the display of humanoid robots has become the absolute focus of the scene. At the Stardust Intelligent booth, the AI robot assistant S1 performed three major performances of dulcimer, martial arts, and calligraphy in one exhibition area, capable of both literary and martial arts. , attracted a large number of professional audiences and media. The elegant playing on the elastic strings allows the S1 to demonstrate fine operation and absolute control with speed, strength and precision. CCTV News conducted a special report on the imitation learning and intelligent control behind "Calligraphy". Company founder Lai Jie explained that behind the silky movements, the hardware side pursues the best force control and the most human-like body indicators (speed, load) etc.), but on the AI side, the real movement data of people is collected, allowing the robot to become stronger when it encounters a strong situation and learn to evolve quickly. And agile

At this ACL conference, contributors have gained a lot. The six-day ACL2024 is being held in Bangkok, Thailand. ACL is the top international conference in the field of computational linguistics and natural language processing. It is organized by the International Association for Computational Linguistics and is held annually. ACL has always ranked first in academic influence in the field of NLP, and it is also a CCF-A recommended conference. This year's ACL conference is the 62nd and has received more than 400 cutting-edge works in the field of NLP. Yesterday afternoon, the conference announced the best paper and other awards. This time, there are 7 Best Paper Awards (two unpublished), 1 Best Theme Paper Award, and 35 Outstanding Paper Awards. The conference also awarded 3 Resource Paper Awards (ResourceAward) and Social Impact Award (

Deep integration of vision and robot learning. When two robot hands work together smoothly to fold clothes, pour tea, and pack shoes, coupled with the 1X humanoid robot NEO that has been making headlines recently, you may have a feeling: we seem to be entering the age of robots. In fact, these silky movements are the product of advanced robotic technology + exquisite frame design + multi-modal large models. We know that useful robots often require complex and exquisite interactions with the environment, and the environment can be represented as constraints in the spatial and temporal domains. For example, if you want a robot to pour tea, the robot first needs to grasp the handle of the teapot and keep it upright without spilling the tea, then move it smoothly until the mouth of the pot is aligned with the mouth of the cup, and then tilt the teapot at a certain angle. . this

This afternoon, Hongmeng Zhixing officially welcomed new brands and new cars. On August 6, Huawei held the Hongmeng Smart Xingxing S9 and Huawei full-scenario new product launch conference, bringing the panoramic smart flagship sedan Xiangjie S9, the new M7Pro and Huawei novaFlip, MatePad Pro 12.2 inches, the new MatePad Air, Huawei Bisheng With many new all-scenario smart products including the laser printer X1 series, FreeBuds6i, WATCHFIT3 and smart screen S5Pro, from smart travel, smart office to smart wear, Huawei continues to build a full-scenario smart ecosystem to bring consumers a smart experience of the Internet of Everything. Hongmeng Zhixing: In-depth empowerment to promote the upgrading of the smart car industry Huawei joins hands with Chinese automotive industry partners to provide

Conference Introduction With the rapid development of science and technology, artificial intelligence has become an important force in promoting social progress. In this era, we are fortunate to witness and participate in the innovation and application of Distributed Artificial Intelligence (DAI). Distributed artificial intelligence is an important branch of the field of artificial intelligence, which has attracted more and more attention in recent years. Agents based on large language models (LLM) have suddenly emerged. By combining the powerful language understanding and generation capabilities of large models, they have shown great potential in natural language interaction, knowledge reasoning, task planning, etc. AIAgent is taking over the big language model and has become a hot topic in the current AI circle. Au
