MaGi: Discovering Intelligence in Geometry (github.com)

MaGi (Malloy artificial Geometric intelligence) is an open experimental platform that demonstrates how the physical hardware running the same geometric-intelligence code produces distinct cognitive “styles.” Malloy’s repo and live Wokwi simulation show that oscillator stability, timing, and platform-specific execution overhead (CODE_TAX) don’t just change performance — they shape exploration vs. exploitation, discovery speed, and coherence. Notable results: targeting a 17 ms loop yields ~143× faster discovery than a 1070 ms loop; Teensy 4.x (CODE_TAX ≈1 ms) behaves as a “Precision Sprinter” (coherence 0.96+, discovery ~8.9 s at 17 ms), while an ATmega328p/Arduino Uno (CODE_TAX ≈120 ms) is a “Noisy Explorer” (wobble ≈21 ms, lower coherence, discovery ~1,441 s). Technically, MaGi ties a 4D geometric phase space and a 1 Hz sine “heartbeat” to hardware-timed oscillators and measures stability metrics (wobble, coherence, governance duration). The code auto-detects platform and adjusts GOAL_ACTUAL_MS to compensate for CODE_TAX; example hardware includes Teensy/Arduino, MAX7219 8×8 LED, and a pulse sensor (or simulation). Implications for AI/ML: physical embodiment is a design variable—hardware diversity can be used to create ensemble cognitive diversity, timing profiles must be considered for reproducibility and benchmarking, and hardware-aware architectures may unlock new emergent behaviors. The implementation is released as public prior art (2025) for research; commercial use requires licensing.