Project Prism is an experimental AI architecture research project by SmilyAI Labs.
We are investigating a simple question:
Does one neural network really need to learn knowledge, computation, and working memory in the same way?
Our current research model, Orion Flagship, explores an architecture we call Transformer 2 (T2).
Traditional language models largely rely on a homogeneous stack of Transformer layers to learn many different functions at once.
T2 explores separating several of these roles:
KNOW โ sparse learned memory for persistent knowledge
DO โ reusable neural procedures for computation and transformation
THINK โ persistent working state for intermediate reasoning
VERIFY โ additional computation for checking and refining results
The goal is to investigate whether this separation can improve parameter and compute efficiency while retaining strong general capabilities across language, code, mathematics, and reasoning.
Orion Flagship is the primary experimental model being developed under Project Prism.
The planned flagship architecture targets approximately:
Before scaling to the full model, we are testing the architecture with smaller prototypes and controlled ablations.
Our current Mini-T2 experiments are investigating:
Results are experimental. Claims about architectural improvements will be based on controlled evaluations and ablations rather than training loss alone.
KNOW ยท DO ยท THINK ยท VERIFY
Exploring what comes after the conventional Transformer.