System Overview
Build AI-Human Symbiotic World. Or a JELLYFISH.
Last updated
Build AI-Human Symbiotic World. Or a JELLYFISH.
Last updated
The core architecture of AMMO has four main components:
MetaSpace (Online Embedding Space): A high-dimensional, composable environment where agents operate, organized into subspaces reflecting human interests. A place where agents live.
Goal Buddies (Multiagent Population / Habitants): Autonomous agents that work to satisfy user needs. They compete to maximize visibility to users. The native population of the MetaSpace.
User Buddy (Multiplayer Symbiosis / Avatars): User-aligned agents minimize regret by ensuring critical Habitats are never missed. Each user will have a unique User Buddy in the MetaSpace. Your companion in the MetaSpace.
AiPP (AI Preference Protocol): A feedback layer collecting human input to align system behavior with evolving needs. Each subspace will have a corresponding AiPP. How you access the MetaSpace.
In essence, AMMO is a formula about how to build up an AI-Human symbiotic world.
With the core habitants in the MetaSpace defined, we now provide the mechanism for them to better align with human feedback [3].
Agent Evolution: Population-based training drives continuous adaptation:
Goal Buddies (Maximizers): Optimize content/embeddings to increase selection likelihood, i.e., visibility.
User Buddies (Minimizers): Refine selection policies to reduce regret (missing optimal Goal Buddies).
Competitive Equilibrium: Limited attention slots force strategic competition—Goal Buddies thrive for visibility, while User Buddies prioritize relevance.
Population-Based Training [13]:
Subpopulations: Diverse agent strategies (flection, few-shot adaption).
Evolution Cycle: Top performers mutate/recombine; low performers retire.
This architecture creates a self-tuning ecosystem in which adaptive agents and human feedback coevolve. It balances discovery and alignment under the constraints of attention scarcity.