larql-lql-demo

Replication of Chris Hay's "I Decoupled Attention from Weights — Gemma 4 26B" YouTube demo, using the cross-platform cronos3k/larql fork and the chrishayuk/gemma-3-4b-it-vindex.

What this is: mechanistic interpretability via LQL (Lazarus Query Language). You are NOT running the LLM for text generation. You are querying the model's FFN weight structure as a graph — inspecting which features activate, comparing concept representations across layers, and optionally editing knowledge directly in the vindex.

---

Quickstart

# 1. Build the LARQL binary (cross-platform fork, Linux/macOS/Windows)
bash setup_env.sh

# 2. Download the Gemma 3 4B vindex (~4-5 GB)
bash fetch_vindex.sh

# 3. Launch the interactive LQL REPL
python lql_repl.py

# 4. Or run the YouTube demo queries non-interactively
python lql_repl.py --script demo_queries.lql

---

What LQL Queries Look Like

-- Which features fire most strongly at layers 20-23 for this prompt?
WALK "The capital of France is" TOP 10;

-- Compare concept representations side-by-side at the same layer
PROBE "France is" vs "Germany is" vs "Japan is" AT LAYER 22;

-- Predict what the model thinks comes next
INFER "The capital of France is" TOP 5;

-- Edit the model's knowledge directly (not a fine-tune — writes to the graph)
INSERT INTO EDGES (entity, relation, target)
VALUES ("Atlantis", "capital-of", "Poseidon");

---

Repository Layout

.
├── setup_env.sh        # Rust toolchain + clone cronos3k/larql + cargo build --release
├── fetch_vindex.sh     # HuggingFace CLI download of gemma-3-4b-it-vindex
├── lql_repl.py         # Interactive REPL + --script mode for batch query files
├── demo_queries.lql    # The exact query sequence from Chris Hay's YouTube demo
└── data/
    └── gemma3-4b.vindex/   # Populated by fetch_vindex.sh

---

VM Sizing

LQL queries do sparse KNN lookups, not full autoregressive generation. A DigitalOcean Basic 8 GB / 4 vCPU droplet (~$0.07/hr on-demand) is sufficient. SSH in from Codespaces, run bash setup_env.sh, done.

---

Credits

• Original LARQL + LQL language: chrishayuk/larql
• Cross-platform fork (Linux/CUDA builds): cronos3k/larql
• HuggingFace Space (zero-install browser demo): https://huggingface.co/spaces/cronos3k/LARQL-Explorer