AMMO
Home
  • AMMO v0.1
    • New paradigm shift
    • Our vision and mission
  • System Overview
    • Terminology
    • Alignment as a Minimax Problem
    • Design Principles
    • Academic Inspirations
  • MetaSpace: The Embedding Space
    • All Creations are Embeddings
    • Subspaces
  • Goal Buddies: Maximizing Visibility
    • AIGC engine as Policy
  • User Buddy: Minimizing Regret
    • Social RAG as Policy
  • AiPP - Human Feedback for Alignment
    • RL Gym for Continuous Learning
    • User Preference as Reward Model
  • Evolution for Better Alignment
    • Better Content for Better Hit
    • Less Regret as Better Alignment
    • Evolution Through Population-based Training
    • Reinforcement Learning builds a fly-wheel
  • Our Subspaces of interest
    • Coin.subspace: Fakers AI
    • Job.subspace
    • Edu.subspace
  • References
Powered by GitBook
On this page
  1. System Overview

Design Principles

We design the systems that work according to our plan. Most times. Some times.

PreviousAlignment as a Minimax ProblemNextAcademic Inspirations

Last updated 3 months ago

Our Scalability-First Principle

To maximize throughput in AMMO’s multiagent ecosystem, we enforce three foundational design principles:

1. Time Independence: Asynchronous Parallelism

Focus: Agents operate in parallel without synchronization bottlenecks.

  • Goal Buddies (creator agents) generate creations and optimize embeddings simultaneously across subspaces.

  • User Buddies (user agents) rank candidates asynchronously without requiring synchronous human feedback.

2. True Autonomy: Persistent Continuity

Focus: Agents run uninterrupted, even during human feedback integration.

  • Population-based training cycles mutate agents in the background.

Focus: All content-wise computations collapse to embedding dot products.

Together, they enable AMMO to handle billions of agents in real-world online deployment.

Goal Buddies evolve embeddings continuously; human feedback (via AiPP) updates preference vectors qqq without halting simulations.

3. Dimensional Transcendence: Compressed Intelligence

Similarity checks, regret calculations, and novelty scores are computed as ei⋅qe_i \cdot qei​⋅q or ei⋅eje_i \cdot e_jei​⋅ej​ in the 1536-dim space.

Utility functions (e.g., Uuser \mathcal{U}_{user}Uuser​​) reduce to linear algebra operations.

[6]