olmoearth-mlx

OlmoEarth MLX: Spatio-Temporal Earth Intelligence

Trit: +1 (PLUS - world creation via planetary observation)
Foundation: AI2 OlmoEarth + Apple MLX + GeoACSet + Dune.xyz Geographic WEV

Overview

OlmoEarth is AI2's open spatio-temporal foundation model for planetary intelligence, trained on:

• Sentinel-2 L2A (12 bands, 10-60m resolution)
• Sentinel-1 SAR (VV, VH polarization)
• Landsat (historical continuity)
• WorldCover (11 land cover classes)
• OpenStreetMap raster
• SRTM elevation
• WRI Canopy Height Map

This skill enables:

1. Geographic embedding of crypto wallet activity from Dune.xyz
2. Impact area identification for Protocol Labs infrastructure
3. GeoACSet materialization for categorical spatial reasoning

Architecture

┌─────────────────────────────────────────────────────────────────────────┐
│                     OLMOEARTH-MLX PIPELINE                              │
├─────────────────────────────────────────────────────────────────────────┤
│                                                                         │
│  Dune.xyz Query ─┬─► IP Geolocation ─┬─► OlmoEarth Embedding            │
│                  │                   │                                  │
│  Wallet Activity ┘   Region Bounds  ─┘   FlexiVit Encoder               │
│                                              │                          │
│                                              ▼                          │
│                              ┌───────────────────────────────────┐      │
│                              │     GeoACSet Materialization      │      │
│                              │   Regions → Districts → Parcels   │      │
│                              │        with GF(3) trits           │      │
│                              └───────────────────────────────────┘      │
│                                              │                          │
│                                              ▼                          │
│                              ┌───────────────────────────────────┐      │
│                              │     Tenderloin WEV Integration    │      │
│                              │   Geographic Value Extraction     │      │
│                              └───────────────────────────────────┘      │
└─────────────────────────────────────────────────────────────────────────┘

OlmoEarth Model Specs

Model	Encoder Params	Decoder Params	Embed Dim	Depth
Nano	5.8M	3.7M	192	12
Tiny	22M	14M	384	12
Base	86M	55M	768	12
Large	307M	197M	1024	24

Supported Modalities

MODALITIES = {
    "SENTINEL2_L2A": {
        "bands": ["B02", "B03", "B04", "B05", "B06", "B07", 
                  "B08", "B8A", "B11", "B12", "SCL", "CLD"],
        "gsd": [10, 10, 10, 20, 20, 20, 10, 20, 20, 20, 20, 20],
        "temporal": True
    },
    "SENTINEL1": {
        "bands": ["VV", "VH"],
        "gsd": [10, 10],
        "temporal": True
    },
    "WORLDCOVER": {
        "bands": ["LC"],
        "classes": 11,  # Tree, Shrub, Grass, Crop, Built, Bare, Snow, Water, Wetland, Mangrove, Moss
        "gsd": [10],
        "temporal": False
    },
    "SRTM": {
        "bands": ["elevation"],
        "gsd": [30],
        "temporal": False
    }
}

MLX Inference

# olmoearth_mlx.py - Apple Silicon optimized inference

import mlx.core as mx
import mlx.nn as nn
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple

@dataclass
class OlmoEarthConfig:
    """OlmoEarth model configuration for MLX."""
    embed_dim: int = 768
    depth: int = 12
    num_heads: int = 12
    patch_size: int = 4
    num_modalities: int = 8
    temporal_tokens: int = 4

class FlexiVitEncoder(nn.Module):
    """FlexiVit encoder for OlmoEarth on MLX."""
    
    def __init__(self, config: OlmoEarthConfig):
        super().__init__()
        self.config = config
        
        # Patch embeddings per modality
        self.patch_embed = nn.Linear(
            config.patch_size * config.patch_size * 12,  # Max bands
            config.embed_dim
        )
        
        # Positional encoding
        self.pos_embed = mx.zeros((1, 256, config.embed_dim))
        
        # Transformer blocks
        self.blocks = [
            TransformerBlock(config.embed_dim, config.num_heads)
            for _ in range(config.depth)
        ]
        
        self.norm = nn.LayerNorm(config.embed_dim)
    
    def __call__(self, x: mx.array, timestamps: mx.array) -> mx.array:
        # Patchify and embed
        x = self.patch_embed(x)
        x = x + self.pos_embed[:, :x.shape[1], :]
        
        # Transform
        for block in self.blocks:
            x = block(x)
        
        return self.norm(x)

class TransformerBlock(nn.Module):
    def __init__(self, dim: int, num_heads: int):
        super().__init__()
        self.attn = nn.MultiHeadAttention(dim, num_heads)
        self.mlp = nn.Sequential(
            nn.Linear(dim, dim * 4),
            nn.GELU(),
            nn.Linear(dim * 4, dim)
        )
        self.norm1 = nn.LayerNorm(dim)
        self.norm2 = nn.LayerNorm(dim)
    
    def __call__(self, x: mx.array) -> mx.array:
        x = x + self.attn(self.norm1(x), self.norm1(x), self.norm1(x))
        x = x + self.mlp(self.norm2(x))
        return x

Dune.xyz Geographic Mapping

Trackable Interactions

Query	Geographic Signal	GeoACSet Mapping
Filecoin Storage Deals	Provider location	Region → District
FVM Contract Deploys	Creator wallet region	District → Parcel
Bridge Transactions	Source/dest chains	Region ↔ Region
Grant Disbursements	Recipient geography	Parcel → Building

SQL Queries (Dune)

-- Filecoin storage provider geographic distribution
SELECT 
    provider_id,
    COALESCE(ip_geolocation.country, 'Unknown') as country,
    COALESCE(ip_geolocation.region, 'Unknown') as region,
    COUNT(*) as deals,
    SUM(piece_size) as total_bytes
FROM filecoin.storage_deals sd
LEFT JOIN ip_geolocation ON sd.provider_ip = ip_geolocation.ip
WHERE deal_timestamp > NOW() - INTERVAL '30 days'
GROUP BY 1, 2, 3
ORDER BY deals DESC;

-- FVM contract geographic heat map
SELECT 
    DATE_TRUNC('day', block_timestamp) as day,
    creator_country,
    creator_region,
    COUNT(DISTINCT contract_address) as contracts,
    SUM(gas_used) as total_gas
FROM filecoin.fvm_contracts
WHERE block_timestamp > NOW() - INTERVAL '90 days'
GROUP BY 1, 2, 3;

-- Cross-chain bridge flows by geography
SELECT
    source_chain,
    dest_chain,
    sender_region,
    receiver_region,
    SUM(amount_usd) as volume_usd,
    COUNT(*) as tx_count
FROM bridge_transactions
WHERE timestamp > NOW() - INTERVAL '7 days'
  AND (source_chain = 'filecoin' OR dest_chain = 'filecoin')
GROUP BY 1, 2, 3, 4;

GeoACSet Materialization

# GeoACSet schema for crypto-geographic data
@present SchCryptoGeo(FreeSchema) begin
    # Objects
    Region::Ob
    District::Ob
    Wallet::Ob
    Transaction::Ob
    EarthTile::Ob
    
    # Morphisms
    region_of::Hom(District, Region)
    district_of::Hom(Wallet, District)
    sender::Hom(Transaction, Wallet)
    receiver::Hom(Transaction, Wallet)
    tile_of::Hom(Wallet, EarthTile)
    
    # Attributes
    lat::Attr(Region, Float64)
    lon::Attr(Region, Float64)
    embedding::Attr(EarthTile, Vector{Float32})  # OlmoEarth embedding
    trit::Attr(Transaction, Int)  # GF(3) classification
end

@acset_type CryptoGeoACSet(SchCryptoGeo)

# Materialization from Dune query results
function materialize_crypto_geo(dune_results::DataFrame)::CryptoGeoACSet
    acset = CryptoGeoACSet()
    
    # Add regions
    for row in eachrow(unique(dune_results, :region))
        add_part!(acset, :Region; 
            lat=row.lat, 
            lon=row.lon
        )
    end
    
    # Add districts within regions
    for row in eachrow(unique(dune_results, [:region, :district]))
        region_id = findfirst(r -> r == row.region, acset[:Region])
        add_part!(acset, :District; region_of=region_id)
    end
    
    # Add wallets with earth tile embeddings
    for row in eachrow(unique(dune_results, :wallet))
        district_id = findfirst(d -> d == row.district, acset[:District])
        tile_embedding = get_olmoearth_embedding(row.lat, row.lon)
        tile_id = add_part!(acset, :EarthTile; embedding=tile_embedding)
        add_part!(acset, :Wallet; district_of=district_id, tile_of=tile_id)
    end
    
    return acset
end

Impact Area Identification

Protocol Labs Infrastructure Zones

PL_IMPACT_ZONES = {
    "high_density": [
        {"name": "SF Bay Area", "lat": 37.7749, "lon": -122.4194, "radius_km": 100},
        {"name": "Greater NYC", "lat": 40.7128, "lon": -74.0060, "radius_km": 80},
        {"name": "Singapore", "lat": 1.3521, "lon": 103.8198, "radius_km": 50},
    ],
    "emerging": [
        {"name": "Shenzhen", "lat": 22.5431, "lon": 114.0579, "radius_km": 60},
        {"name": "Berlin", "lat": 52.5200, "lon": 13.4050, "radius_km": 40},
        {"name": "Bangalore", "lat": 12.9716, "lon": 77.5946, "radius_km": 50},
    ],
    "frontier": [
        {"name": "Lagos", "lat": 6.5244, "lon": 3.3792, "radius_km": 30},
        {"name": "São Paulo", "lat": -23.5505, "lon": -46.6333, "radius_km": 60},
        {"name": "Jakarta", "lat": -6.2088, "lon": 106.8456, "radius_km": 40},
    ]
}

def classify_impact_zone(lat: float, lon: float) -> Tuple[str, str, float]:
    """Classify a coordinate into PL impact zone."""
    from math import radians, sin, cos, sqrt, atan2
    
    def haversine(lat1, lon1, lat2, lon2):
        R = 6371  # Earth radius in km
        dlat = radians(lat2 - lat1)
        dlon = radians(lon2 - lon1)
        a = sin(dlat/2)**2 + cos(radians(lat1)) * cos(radians(lat2)) * sin(dlon/2)**2
        return 2 * R * atan2(sqrt(a), sqrt(1-a))
    
    for zone_type, zones in PL_IMPACT_ZONES.items():
        for zone in zones:
            dist = haversine(lat, lon, zone["lat"], zone["lon"])
            if dist <= zone["radius_km"]:
                return zone_type, zone["name"], dist
    
    return "global", "Other", float("inf")

GF(3) Geographic Classification

Trit	Zone Type	WEV Signal	Strategy
+1	high_density	Infrastructure concentration	Long storage deals
0	emerging	Growth potential	Bridge liquidity
-1	frontier	Adoption catalyst	Grant allocation

Tenderloin Integration

# Connect OlmoEarth embeddings to Tenderloin WEV
def geo_wev_extraction(
    dune_query: str,
    olmoearth_model: FlexiVitEncoder,
    geoacset: CryptoGeoACSet
) -> List[Dict]:
    """Extract geographic WEV from Dune data with OlmoEarth embeddings."""
    
    # Execute Dune query
    results = execute_dune_query(dune_query)
    
    wev_events = []
    for row in results:
        # Get OlmoEarth embedding for location
        embedding = olmoearth_model.embed(row.lat, row.lon)
        
        # Classify impact zone
        zone_type, zone_name, dist = classify_impact_zone(row.lat, row.lon)
        
        # Map to GF(3) trit
        trit = {"high_density": 1, "emerging": 0, "frontier": -1}.get(zone_type, 0)
        
        wev_events.append({
            "source": "dune_geo",
            "lat": row.lat,
            "lon": row.lon,
            "zone": zone_name,
            "trit": trit,
            "embedding": embedding.tolist(),
            "value": row.volume_usd,
            "category": "geographic"
        })
    
    return wev_events

Usage

# Install OlmoEarth
pip install olmoearth-pretrain

# Run geographic WEV extraction
python -c "
from olmoearth_mlx import OlmoEarthMLX
from tenderloin.fund_wallets import WalletFundingEngine

# Load model
model = OlmoEarthMLX.load('olmoearth-base')

# Query Dune for geographic data
geo_wev = model.extract_geographic_wev(
    query='filecoin_storage_providers_by_region',
    timeframe='30d'
)

# Fund wallets based on geographic WEV
engine = WalletFundingEngine()
for event in geo_wev:
    engine.ingest_wev(event['category'], event['value'], trit=event['trit'])
"

Canonical Triads

olmoearth-mlx (+1) ⊗ geoacset (0) ⊗ dune-geographic (-1) = 0 ✓
tenderloin (+1) ⊗ prediction-markets (0) ⊗ olmoearth-mlx (-1) = 0 ✓

References

• OlmoEarth arXiv Paper
• allenai/olmoearth_pretrain
• bmorphism/GeoACSets.jl
• Dune Analytics
• Filecoin Data Portal

Scientific Skill Interleaving

This skill connects to the K-Dense-AI/claude-scientific-skills ecosystem:

Autodiff

• jax [○] via bicomodule
  • Hub for autodiff/ML

Bibliography References

• general: 734 citations in bib.duckdb

Cat# Integration

This skill maps to Cat# = Comod(P) as a bicomodule in the equipment structure:

Trit: 0 (ERGODIC)
Home: Prof
Poly Op: ⊗
Kan Role: Adj
Color: #26D826

GF(3) Naturality

The skill participates in triads satisfying:

(-1) + (0) + (+1) ≡ 0 (mod 3)

This ensures compositional coherence in the Cat# equipment structure.