Initial commit: GKERN glyph kernel

2026-07-09 13:28:07 -04:00
commit 0807c58eae
43 changed files with 6006 additions and 0 deletions
@@ -0,0 +1,47 @@
+#!/usr/bin/env python3
+"""COMPARISON REPORT"""
+print("""
+\033[1;36m============================================================\033[0m
+\033[1;36m        GLYPHOS vs TRANSFORMER - COMPARATIVE ANALYSIS         \033[0m
+\033[1;36m============================================================\033[0m
+
+\033[1;33m1. WHAT EACH BENCHMARK MEASURES\033[0m
+------------------------------------------------------------
+\033[1mTRANSFORMER:\033[0m
+  ✓ Full vocabulary embedding (10,000+ classes)
+  ✓ Multi-head attention O(N²) for ALL token pairs
+  ✓ Softmax normalization (exponential operations)
+  ✓ Residual connections + Layer Normalization
+  ✓ Language model output head
+  
+\033[1mGLYPHOS:\033[0m
+  ✓ Sparse graph with 4 edges per node
+  ✓ Simple weighted averaging of neighbors  
+  ✓ Binary similarity check (fixed threshold)
+  ✓ Sigmoid activation (cheap approximation)
+  ✓ No vocabulary, no language modeling
+
+\033[91mKEY POINT: These solve fundamentally DIFFERENT problems!\033[0m
+
+\033[1;33m2. OPERATIONAL COST COMPARISON\033[0m
+------------------------------------------------------------
+| Component          | Transformer      | GlyphOS          |
+|--------------------|------------------|------------------|
+| Attention Ops      | \033[91m~500M/token\033[0m     | \033[92m~16K/node\033[0m     |
+| Memory Pattern     | \033[91mRandom/Cache miss\033[0m|\033[92m Sequential/Clean\033[0m|
+| Scaling Behavior   | \033[91mO(N²)\033[0m          | \033[92mO(edges) ≈ O(N)\033[0m |
+| Training Required  | \033[91mYes (weeks)\033[0m     | \033[92mNo (static)\033[0m    |
+| Capability         | \033[93mText generation\033[0m  | \033[93mGraph relaxation\033[0m |
+
+\033[1;33m3. APPLES-TO-APPLES COMPARISON NEEDS\033[0m
+------------------------------------------------------------
+□ Same task (e.g., text completion)
+✓ Same sequence length
+✓ Same hardware  
+□ Same evaluation metric (perplexity, BLEU, etc.)
+□ Same parameter budget
+
+\033[91mWithout these, performance claims are misleading.\033[0m
+
+\033[36mRun actual benchmarks to see real timings.\033[0m
+""")