Flower SuperGrid Agents: Scaling AI through Collaborative Networks

Flower SuperGrid Agents: Scaling AI through Collaborative Networks

The Collaborative AI Frontier

Collaborative AI represents a shift from centralized AI—where data is moved to a central computation cluster—to a decentralized approach where computation is moved to the data. This paradigm shift is necessary because the vast majority of the world's high-quality data exists in private silos. While public English web data consists of roughly 15 trillion tokens, an estimated 2,000 trillion tokens remain unused in private silos, meaning less than 1% of available data is currently utilized for foundation model training.

By building collaborative networks rather than larger individual silos, organizations can scale horizontally. This allows non-AI-native companies with unique data to become competitive by training models on resources that no single entity possesses in isolation.

Flower SuperGrid: The Decentralized AI Platform

Flower SuperGrid is the foundational layer designed to simplify the deployment of decentralized AI systems. Traditionally, building such systems required hundreds of manual steps for onboarding and configuration. SuperGrid reduces this complexity to a few clicks via a user-friendly interface at flower.ai.

Key Technical Capabilities

  • Supernodes: Individual nodes running on data silos that execute distributed workloads.
  • Superlink: A coordinator that manages the network without holding any actual data.
  • Heterogeneous Confidential Compute: The first integration allowing the use of confidential compute from different vendors within a single system, a critical requirement for scaling workloads.
  • Flower Hub: A community-driven repository for decentralized apps with a trust and review system.
  • Isolation and Auditability: A strong isolation model with auditable communication between components and support for streaming large model weights for LLM training.

Project Kaya: Collaborative AI Agents

Most current AI agents are limited to public web data or a single organization's private data. Project Kaya is a collaborative AI agent built on SuperGrid that allows agents to communicate across different organizations or data silos to solve tasks that no single agent could solve alone.

How Collaborative Agents Work

  1. Task Decomposition: A coordinating agent breaks down a user task and sends messages to individual supernodes.
  2. Autonomous Participation: Supernodes are fully autonomous; they can decide whether to accept or refuse a request based on their own governance principles.
  3. Local Processing: Agents on supernodes process the request using local data without the data ever leaving the premises.
  4. Controlled Response: Supernode operators can review and reject sensitive information before a response is sent back to the coordinator.
  5. Aggregation: The central agent aggregates the results from participating supernodes to provide a final answer to the user.

SuperGrid Frontier: Decentralized Training Pipelines

To improve agents using local data, Flower Labs provides SuperGrid Frontier, a decentralized training pipeline. This allows for the training of large models without moving the underlying data, adhering to the principle that "data never moves, only the learnings move."

Training Milestones and Research

  • Communication Efficiency: SuperGrid Frontier has demonstrated up to a 1,000x reduction in communication costs during decentralized training runs.
  • Llama 7B (UK Edition): Flower Labs released an open-weight Llama 7B model specifically optimized for the UK, trained using the SuperGrid Frontier pipeline.
  • Large-Scale Collaboration: In partnership with the US Department of Energy and Sandia National Labs, Flower Labs is currently training a 70 billion parameter LLM decentralized across three different sites.
  • Academic Contributions: Research published at ICLR 2025 includes work on decoupled embeddings for pre-training language models and "Photon" for federated pre-training of LLMs.

Summary of the Flower Ecosystem

Flower Labs is advancing "collaborative superintelligence" through three integrated building blocks:

Component Purpose
Flower SuperGrid The decentralized AI platform for running workloads on distributed data.
Project Kaya Collaborative agents that can communicate and reason across networks.
SuperGrid Frontier The decentralized training pipeline for building models on private data.

Sources