🌀 Recursive Self-Improvement (RSI) Cheat Sheet

An ultra-condensed, emojified quick-start guide for the fractalsearch Mandelbrot set training loop. Inspired by Andrej Karpathy’s autoresearch project and showcased in video t7_ZXgfJVG8 (“Recursive Self-Improvement - Computer, load up celery man. Can AI build AI?”).

🧠 The Concept: Can AI build AI?

RSI (Strong): AI agents autonomously writing, training, evaluating, and committing code modifications to improve their own architectures, algorithms, or ML pipelines in a closed feedback loop.
The Subject: Neural networks (MLPs, Fourier features, Hash Grids) forced to learn the infinitely complex boundaries of the Mandelbrot Set ((real, imag) -> target value in [0, 1]).

📊 The Immutable Eval Harness (`harness/`)

DO NOT TOUCH files under harness/. They represent the fixed ground truth.

File	Role	Key Metric
📍 `groundtruth.py`	Defines the math of the Mandelbrot set (smooth normalized escape time).	Immutable target function.
🧬 `interface.py`	Declares the `Solution` contract (`fit(ctx)` and `predict(coords)`).	API template for solutions.
🎛️ `evaluate.py`	Runs training and evaluates MSE against ground truth.	MSE (Mean Squared Error) 👈 lower is better!

⏱️ Time & Grid Constants to Memorize

⏱️ Training Budget: 300 seconds (5 minutes) target training window inside fit().
🚨 Hard Kill: 600 seconds (10 minutes). Exceed this and SIGALRM force-kills the run (logs as timeout).
🖥️ Eval Grid: Dense 4K UHD aspect-correct grid ( $3840 \times 2414 \approx 9.3 M$ coordinates). Aspect ratio is anchored at $1.59$ (not square).

🔄 The Infinite Research Loop (Step-by-Step)

[ Git master ]
     │
     ▼
1. 💡 Formulate hypothesis (e.g. Add Fourier Features, change activation)
     │
     ▼
2. 📄 Copy prior winning solution file under `solutions/` (e.g. `fourier_features.py`)
     │
     ▼
3. 💾 Run Git commit: `git add -A && git commit -m "feat: <hypothesis>"`
     │
     ▼
4. 🚀 Launch evaluation:
        uv run python -m harness.evaluate solutions/my_new_solution.py > run.log 2>&1
     │
     ▼
5. 🔍 Read metrics: `grep "^mse:\|^status:" run.log`
     │
     ▼
6. 📈 Inspect runs.jsonl or monitor local dashboard (localhost:8000)
     │
     ▼
7. 🔁 Loop forever (refine hypothesis and start again!)

🚫 Non-Negotiable Laws

✅ DO

Copy & mutate: Duplicate the closest prior winner under a new filename inside solutions/. Keep each script self-contained.
Explore wild ideas: Try custom activations, boundary-weighted sampling schedules, multi-resolution curriculum training, learning rate warm-ups, or grid-based embeddings (e.g., instant-ngp hash grids).
Write notes: Document failure modes and wins inside solutions/notebook.md so future agent runs have context.

❌ DON’T

No hardcoding: Do not hardcode Mandelbrot math inside your candidate solution. The model must remain a universal function approximator.
No extra deps: Do not add dependencies beyond PyTorch, NumPy, and Pillow.
Never pause: The loop must run continuously without asking the human for permission to start the next experiment.

Recursive Self-Improvement (RSI)