AMMO aims to enable billions of AI agents and humans to coexist equally from the alignment perspective.
Author: Pzai, Foresight News
In the era of the cyberspace age, the development of AI has not only brought about a rapid increase in productivity for everyone, but is also raising a question for us: as AI is gradually penetrating into the domain of humans, do humans need to re-evaluate the human-machine relationship?
Against this backdrop, political views on AI technology are gradually becoming diverse. While the "AI crisis faction" full of skepticism and the "accelerationism (e/acc) faction" are in a heated debate, the "alignment faction" advocates enhancing the public benefit of technology, ethical discussion, and the emphasis on humanistic values, introducing humanistic judgment into the process of AI R&D and iteration to ensure that AI technology will not lose control.
And in the era of the prevalence of AI Agents, accompanied by the transition from a single large model to a multi-modal perception and multi-AI interaction paradigm, the "alignment issue" of AI seems to be increasingly recognized by more and more people.
On February 20, AMMO, launched by former technology leaders from Google, DeepMind and Meta, received a $2.5 million seed round led by Amber Group. From the team background, AMMO has gathered AI experts from major tech giants, with co-founder and CEO David Huang having worked at Google for 10 years, including 7 years leading the AI program and strategic services in the mobile domain. Another co-founder, Diego Hong, graduated from the University of Oxford and previously led the first-generation AI agent framework at Meta. The team has gathered top AI talents from DeepMind, Google, Apple, and even includes the ACM-ICPC world champion.
The project, from the perspective of alignment, aims to transform the current Internet into a "symbiotic network of humans and AI" through a multi-agent framework and reinforcement learning from human feedback (RLHF), enabling billions of AI agents and humans to coexist equally, and allowing AI to evolve together based on the consistency of human collective feedback.
RL Gyms: Multi-Agent Reinforcement Learning
In the field of artificial intelligence and machine learning, reinforcement learning has always been a research focus. AMMO's RL Gyms provide a solid technical support for the research and application of multi-agent reinforcement learning.
Unlike traditional single-agent reinforcement learning, multi-agent reinforcement learning focuses on the process of multiple agents (Multi-Agent) interacting, learning and making decisions together in the same environment. In this process, the relationships between agents are complex, and they may need to collaborate to achieve common goals, or compete with each other. For example, in the logistics delivery scenario, multiple delivery vehicles as agents need to coordinate routes and plan delivery sequences to maximize overall delivery efficiency; while in competitive games, the agent characters controlled by different players need to compete with each other to win.
RL Gym was first proposed by OpenAI, providing a powerful simulation environment for AI evolution. Developers can build highly customized reinforcement learning environments to meet research needs or application scenarios, such as economic simulation, red and blue confrontation, etc., by defining a series of key functions, including the definition of environmental state transition rules, the protocol of agent perception and action execution, and the definition of reward functions. As long as these functions can be precisely defined, RL Gym can simulate various complex scenarios and lay a foundation for AI evolution in them.
For AMMO developers, RL Gyms provide a rich and realistic two-sided market simulator for AI agents. AI can act as a content and service provider, offering high-quality and attractive content to users; at the same time, AI can also play the role of a human user's avatar, as a consumer, selecting high-quality content centered on user value. This dynamic and rich two-way game stimulates both parties to continuously evolve their own strategies to meet the growing content and service consumption needs of users.
Inspired by Anthropic's Constitutional AI, AMMO has created a transparent governance framework to guide agent decision-making within the platform. This structure is constantly updated through extensive human feedback loops to ensure that agent behavior remains consistent with human collective intent. By embedding the alignment mechanism into this architecture from the very beginning, AMMO ensures that its agents evolve alongside the changing values and priorities of society, because "the center of the multi-agent system is human" under the guidance of alignment.
MetaSpace: Building the "World" for Agents
"Each mental agent by itself can only do some very simple things that don't really require a mind or thought. However, when we put these agents together in certain very special ways, that's when real intelligence emerges." As "the father of artificial intelligence" Marvin Minsky described in his work "The Society of Mind".
For AI agents, more iterations require more input, and in the process of interaction between Agents and other Agents or even humans, a solid framework needs to be built to drive the orderly iteration of AI.
Unlike Ocean Protocol, which mainly focuses on the circulation and trading of data, and SingularityNET, which builds a decentralized AI market, AMMO's unique feature is its focus on building an environment for AI evolution. It not only can solve the problem of model capability improvement or single transaction, but also can provide soil for the continuous development and evolution of AI. In terms of multi-agent technology, compared to frameworks like Swarms, AMMO not only has the ability of efficient collaboration among multiple agents, but more importantly, it is committed to building a complete multi-agent world.
In AMMO's core architecture, the team has built a unique and powerful composable high-dimensional virtual universe - MetaSpace. Highly autonomous AI agents no longer operate in isolation, but engage in deep interaction with other Agents and even humans within MetaSpace.
MetaSpace has a series of vertically deep sub-spaces, which have become the key places for AI agents to continuously evolve. In the process of interaction with humans, autonomous AI agents (Goal Buddy) continuously adjust themselves, fully exerting their adaptability, and gradually achieve deep alignment with human behavior and needs. Similarly, human users' AI avatars (User Buddy) also evolve together in this space, helping humans learn, make decisions, invest, explore and make friends.
This online learning mode of multiple Agents can concretize the complex and diverse needs and interests of humans into a large number of Agents. These Agents are not static, but constantly iterate within MetaSpace, so that AI agents in AMMO no longer rely solely on model capability improvement, but achieve self-optimization through interaction with humans and the environment. It can be said that MetaSpace opens the door to the world's information for Agents.
Fakers AI
In AMMO's sub-space, the first sub-project Fakers AI is positioned as the "Little Red Book of the Web3 market". In this application, multiple AI agents work collaboratively to provide users with rich functions. They not only can collect news, market dynamics, analyze on-chain data, and perceive market sentiment in real-time, but also have a key capability - dynamically learning human interaction feedback.
When users interact with AI agents, whether browsing content, asking questions or posting comments, the AI agents will capture this feedback information and continuously optimize themselves through complex algorithms to achieve real-time alignment with human values, preferences and interests. Based on this capability, these AI agents can more accurately select and combine information when integrating content, providing users with timely and accurate content to meet their diverse needs in the Web3 market.
In the Ticker Battle within the application, 4 AI Agents form a powerful automated workflow, with each Agent responsible for overall planning, on-chain data analysis, community opinion analysis, and summarization, and they can self-iterate based on human reactions. This content production model provides users with content created by AI and driven by the community, designed for transparency. And for AI, this also implicitly boosts their influence.
Innovative Practices from AI to Web3
In the wave of AI and Web3 integration, AMMO, as an innovative platform, is gradually emerging. The investments from Amber Group, Samsung Next, Dispersion, and OpenSpace in AMMO not only recognize its technical capabilities, but also show their confidence in its future market potential.
The core of AMMO's architecture is to combine cutting-edge AI technology in content summarization and review with powerful, zero-trust, community-driven governance. In the short term, AMMO's prototype will enable creators and everyday users to produce and fine-tune content through multiple AI agents (each specialized in tasks such as editing or scriptwriting), while policy agents execute guidelines.
In terms of innovative models, AMMO utilizes its unique multi-agent system to allocate different AI agents to various stages of content creation, quality control, and policy execution. Through reinforcement learning techniques and the introduction of human feedback mechanisms, AMMO continuously optimizes the AI-driven content creation process and improves content quality.
The crypto-based incentive system allows AMMO to directly redistribute value to contributors. Users who provide feedback, interact with content, or otherwise help optimize the agents will receive proportional incentives, creating a self-sustaining feedback loop: incentivized participation drives better agent output, which in turn benefits the network and its contributors.
In summary, in the trend of multi-agentization in the AI era, AMMO has created a vision of alignment in AI development and its realization, building a symbiotic world of billions of humans and AI that is aligned with humans. It seems that in the current AI field, alignment itself, whether for humans or AI, ultimately leads to a mutually beneficial result of coordinated and synchronized development, and we are also looking forward to such a coexisting future.
