By routing, combining, and coordinating thousands of models, agents, and data streams, the GRID scales like a distributed brain. As it grows, its collective intelligence compounds, delivering toward AGI-level output. Here’s how the underlying algorithms make this possible.
Expert and community-defined workflows
For simpler queries, GRID leverages expert-designed workflows. Each query is analyzed, classified, and routed into the most relevant workflow for that use case.
Example: “Which European SaaS startups have raised over $50M in the last year?” → identified as research analysis → routed to the research workflow.
Sample workflow steps:
Search: Compile a list of startups that are European, have raised $50M, and have started in the last year.
Research: Evaluate founder profiles of the identified startups
Search: Find the latest revenue metrics on these startups
Conceptualize: Visualize revenue trajectories
Aggregate: Prepare the final answer
Workflows differ by use case (search, research, writing) and vertical (finance, travel, e-commerce, science, etc.). Expert community members can design these workflows, and soon, they’ll earn rewards based on how useful their workflows prove in practice.
Recursive atomization & execution
GRID is evolving beyond static workflows with a new architecture for hyper-complex queries.
Each query will be recursively atomized (broken down into smaller sub-queries) until only atomic tasks remain. These are the smallest units of work required to resolve the original query.
Every atomic task is then routed to the most capable intelligence (model, agent, or data source) via a system prompt engine. This engine will itself be community-driven, with users contributing to its refinement.
This recursive architecture enables the GRID to solve problems of arbitrary complexity. Stay tuned for a bigger launch coming soon.
Token-level Routing
We’re developing algorithms to route intelligence at the smallest possible unit of AI language: the token.
Instead of sending an entire query to a single intelligence (model, agent, data source, etc), the GRID decomposes it into tokens (the smallest unit of text that an AI model processes). Each token can then be independently routed to the most suitable intelligence, and finally stitched back together for a final answer.
Our early experiments show evidence of outperforming large closed-source labs when routing tokens across different models, producing outputs far stronger than any single system alone at a significantly cheaper cost.
Read our Dobby Report that covers token-level routing to various models: 👉 http://alphaxiv.org/abs/1701.03755
