From Chaos to
Intelligence
"A million random walks, crystallized into wisdom."
The Problem
Every mission generates data:
Billions
Traversal events
Millions
Pheromone deposits
Thousands
Observations
Hundreds?
Patterns
Most of this is noise. Some of it is signal. The signal is intelligence.
The Pipeline
EPHEMERAL (dies with mission)
Raw events stream in...
AGGREGATION
Group by: concept, edge, region. Count traversals, deposits, finds.
THRESHOLD CHECK
Is pheromone_level > superhighway? Is pattern_occurrence > minimum?
PATTERN EXTRACTION
What sequences led here? What caste combinations work? What parameters succeeded?
EMBEDDING GENERATION
Convert patterns to vectors. Enable semantic similarity. Allow cross-mission matching.
PERMANENT (survives forever)
What Becomes Permanent
Definition: Edges with pheromone_level > threshold (default: 20)
Meaning: This path is PROVEN. Many ants walked it. Something good happened.
Usage: Harvesters prefer them. New missions check for relevant superhighways. They're the colony's "muscle memory."
Definition: Named, reusable strategies extracted from successful activity
Meaning: "This approach works for this type of problem."
Usage: New missions query for patterns that match their problem type. The colony doesn't start from zero.
pattern: "sparse-marker-search"
source: hunt-btc
applicable_to: [large_search_spaces, convergent_solutions]
success_rate: 0.85
Definition: Vector representations of concepts, patterns, and mission states
Meaning: Semantic understanding. "This concept is similar to that concept."
Usage: Cross-mission transfer learning. Find patterns from other missions that might apply.
Cross-Mission Transfer
Knowledge from one mission helps another. This is the magic.
Mission A: hunt-btc
Crystallizes: "sparse-marker-search" pattern
Embedding: [0.2, 0.8, 0.1, ...]
Tags: [search, convergence, markers]
Mission B: trade (new)
Description embedded: [0.3, 0.7, 0.2, ...]
Pattern suggested: "sparse-marker-search"
Mission B uses pattern
"Mark rare profitable signals"
"Route harvesters to signal clusters"
"Trade probability increases near cluster centers"
Transfer Rules
Similarity Threshold
Pattern must be >0.7 similar to apply
Human Approval
First transfer from each pattern requires approval
Adaptation Required
Pattern parameters may need mission-specific tuning
Feedback Loop
Track if transferred pattern helps or hurts
Data Lifecycle
Individual traversal events
Temporary pheromone deposits
In-flight coordination messages
Failed exploration attempts
Aggregated metrics per period
Error events and anomalies
Mission config snapshots
Rollback checkpoints
Superhighways
Patterns
Embeddings
Mission success/failure summaries
Measuring Intelligence Growth
Metrics
Pattern Count
How many reusable patterns exist?
Superhighway Density
What % of traversals use proven paths?
Transfer Success Rate
Do patterns from other missions help?
Time to Solution
Does the colony solve faster over time?
The Zero-Shot Mission Test:
- Define a completely new mission type
- Give the colony no domain-specific training
- Let it find and apply relevant patterns
- Measure if it solves faster than a fresh colony
If yes: General intelligence.
If no: Mission-specific intelligence (still valuable).
Lessons from Ants at Work