Benchmarks
Performance targets and current results. Updated each session.
Current Results
| Metric | Target | Result | Notes |
|---|---|---|---|
| Sparsity | 1–4% | 2.43% ✅ | Fixed in Session 2 via clusters |
| Search efficiency | > 95% | 99.1% ✅ | Only 1% of nodes searched per query |
| Compression ratio | > 1.0x | 1.78x ✅ | 44% less memory than dense approach |
| Build time (100k nodes) | < 5 sec | 0.87s ✅ | 257x speedup via vectorized NumPy |
| Search speed (bucket) | < 10ms | 3.5ms ✅ | Fast enough for real-time inference |
| Clustering (100k → 32) | < 30 sec | 4.4s ✅ | K-means with vectorized cosine similarity |
| Search speed (cluster) | < 50ms | 30.9ms ✅ | Searches whole cluster (~3,125 nodes) |
| Context boost | boosted > unboosted | 3431→3506 ✅ | 2% more nodes activated |
| .nci save (100k) | < 5 sec | 0.4s ✅ | Binary format, ~34 MB |
| .nci load (100k) | < 5 sec | 1.6s ✅ | Full round-trip verified identical |
| Consolidation | reinforced > fresh | 0.50→0.70 ✅ | 10 reinforcements tested |
| Inter-cluster connections | > 0 per cluster | 3.0 avg ✅ | 96 total across 32 clusters |
| Long-term compression | 10x–100x | Future ⚪ | Requires semantic signatures |
| RPi inference speed | 2–5 tok/sec | Future ⚪ | End goal |
Brain vs NCI vs Traditional AI
| Concept | Human Brain | NCI | Traditional AI |
|---|---|---|---|
| Knowledge unit | Neuron — fires on right input | Resonance Node — activates on signature match | Weight in a matrix — always participates |
| Sparsity | 1–4% active at any moment | Targeting 1–4% — built in | 100% of weights used every inference |
| Compression | Native — concepts as relationships | Native — compression is the model | Optional afterthought |
| Consolidation | Sleep consolidates memory | Consolidation score on every node | No equivalent |
| Autonomous learning | Continuous | Designed for it | Impossible without full retraining |
| Power required | ~20 watts | Target: modest CPU, no GPU | Thousands of watts |