Mosaic Labs // composite-intelligence research
Intelligence through
composite evolution
Publicly funded, openly organized AGI research — and the Mosaic Protocol, a launchpad for stocks not yet onchain. The lab builds the engine; the protocol funds the lab.
One intelligence, three layers
Research
Five converging routes to AGI: semiodynamic compression, spatial reasoning, context-state diffusion, and reworked RL. Micro-models, not Google-scale compute.
Layer 02Protocol
The same composite intelligence, pointed at markets. Verify a private-equity position, price it, give it a 24/7 onchain market.
Layer 03$MOSAIC
Access, integrity stake, governance, fee sink. Listing flow funds compute; the reference rises as the mosaic fills.
RESEARCH → ENGINE → PROTOCOL → $MOSAIC
Layer 01 — the engine room
Five routes into the same door
Compression as cognition. Space as reasoning. Diffusion as agency. Each track is a small model built to break a specific ceiling.
Read the research →Five development tracks
1. Errloom
The post-training layer. Musical dynamics, a .mos training DSL, physically-recast RL abstractions, and procedural learning.
current priority2. Thauten
Discrete auto-encoding of reasoning into a compressed intermediate representation. The model reasons in zip-space, never expanding to tokens.
spatial3. SAGE
Semantic Automaton in Geometric Embeddings. An HRM on a 2D grid of LLM embeddings — an externalization surface for the model’s world model.
agency4. Diffusion ASI
Context-state direct editing. Every token in context can change at every step — edit, apply and validate collapse into one forward pass.
live5. Market Intelligence
The first output that leaves the lab. CLI-native, cross-chain, real-time formation detection — the module the Mosaic Protocol runs on.
Layer 02 — the application
A market for what has no picture
Pre-IPO shares, private equity, gated tickers, NDA'd secondaries. The protocol tessellates them into one onchain surface — one estimator at a time.
The protocol →Tiles without a picture
Fragmented, illiquid, face-down assets. The engine gives each one a composite fair value and a confidence band; the protocol gives it a line. The mosaic fills as more tiles are verified and priced.
The vision
“Deploy an intelligence fractal decompression phenomenon, wherein a model infinitely decompresses and recompresses information until it tiles consciousness across the universe.”