| # Fuga 2.0 — VSA Hierarchical Predictive Memory & Code Generation |
|
|
| **Non-transformer AGI core** built on Vector Symbolic Architectures (VSA), Hierarchical JEPA, Temporal Memory, and locality-sensitive binding. Replaces backpropagation with local Delta Rule updates and generates Rust code via pure VSA/TM autoregression — no LLM dependency. |
|
|
| --- |
|
|
| ## Training Pipeline |
|
|
| ### 1. Index source files into phase graph |
|
|
| ```bash |
| # Index a directory — reads .rs files, creates PhaseNodes with SDR encoding |
| cargo run --release -- mirror-index src/ai |
| |
| # Load existing mirror and index another directory |
| cargo run --release -- mirror-index src/core |
| ``` |
|
|
| Creates `fuga_mirror_nodes.bin` (phase nodes), `fuga_mirror_tm.bin` (TM), `fuga_mirror_jepa.bin` (HJEPA). |
|
|
| ### 2. Train the predictor (HJEPA + TM) |
|
|
| ```bash |
| # Train on existing mirror nodes (5 epochs, chunk=1) |
| cargo run --release -- train-predictor 5 |
| |
| # With larger chunks for sequence patterns |
| cargo run --release -- train-predictor 10 --chunk 3 |
| ``` |
|
|
| ### 3. Train token vocabulary (embedded in generation) |
|
|
| ```bash |
| # Builds top-4000 char-level token vocab from indexed .rs files |
| # then trains TM on 20000+ token bigram steps |
| # then generates tokens |
| cargo run --release -- generate-code "fn new" --tokens |
| ``` |
|
|
| The token trainer: |
| - Char-level tokenizer: splits identifiers from operators, recognizes `->` `::` `=>` `!=` `==` `>=` `<=` `+=` `-=` `&&` `||` |
| - Syntactic pattern injection: 14 hardcoded Rust patterns × 5 repeats |
| - WTA (Winner-Take-All) prediction with Inhibition of Return |
| - Anti-repetition window (16 tokens) |
|
|
| --- |
|
|
| ## Generation |
|
|
| ### Token-level (syntactic) |
|
|
| ```bash |
| cargo run --release -- generate-code "fn new" --tokens |
| ``` |
|
|
| Outputs real Rust tokens: `( ) { } [ ] , :: . ' -> \` + identifiers, numbers |
|
|
| ### PhaseNode-level (semantic) |
|
|
| ```bash |
| # Beam search over PhaseNode graph |
| cargo run --release -- generate-code "struct Foo" |
| |
| # Autoregressive mode (generates full snippets) |
| cargo run --release -- generate-code "fn new" --gen |
| |
| # With beam width and temperature |
| cargo run --release -- generate-code "async fn" --beam 3 --temp 1.2 |
| ``` |
|
|
| --- |
|
|
| ## Query & Evaluation |
|
|
| ```bash |
| # Self-query — find matching phase nodes |
| cargo run --release -- self-query "async fn handle" |
| |
| # Evaluate mirror quality |
| cargo run --release -- eval |
| |
| # Inspect text or file |
| cargo run --release -- inspect "fn new() -> Self" |
| cargo run --release -- inspect src/main.rs |
| ``` |
|
|
| --- |
|
|
| ## Tests |
|
|
| ```bash |
| # All library tests |
| cargo test --lib |
| |
| # Anomaly detection (Inhibition of Return, overshoot) |
| cargo test --test test_anomaly_detection -- --nocapture |
| |
| # JEPA / TM / MoE tests |
| cargo test --test jepa_test |
| cargo test --test hierarchical_jepa_test |
| cargo test --test moe_routing_test |
| ``` |
|
|
| --- |
|
|
| ## Architecture |
|
|
| | Component | Description | |
| |---|---| |
| | Hypervector | 8192-bit, ~2% density (~164 active bits), XOR bind / sum bundle / permute | |
| | Hierarchical JEPA | L0 (static), L1 (macro), L2 (metacognition) with ls_bind phase-shift | |
| | Temporal Memory | Cells with DendriteSegments, learn_segment / reinforce / prune / predict_next | |
| | SDR (Sparse Distributed Representation) | `encode_text()` → deterministic hash-based sparse binary vector | |
| | Tokenizer | Char-level: splits identifiers from operators, multi-char operator recognition | |
| | WTA | Winner-Take-All with Inhibition of Return (fatigue = wins × 10, decay every 10 steps) | |
| | AnomalyEvent | Detects phase overload — `pred_count > 100` or `power_mw > 500` triggers overshoot | |
|
|
| ## License |
|
|
| Apache-2.0 |
|
|
