SutroYaro¶
Research workspace for the Sutro Group, a study group exploring energy-efficient AI training. Weekly meetings at South Park Commons, San Francisco.
The Challenges¶
The group picks simple learning tasks and tries to solve them with less energy. Three challenges so far, all runnable in under 1 second:
| Challenge | Task | What it tests |
|---|---|---|
| Sparse Parity | y = product of k secret bits (XOR) | k-th order interactions, no local signal |
| Sparse Sum | y = sum of k secret bits | First-order linear structure |
| Sparse AND | y = logical AND of k secret bits | Class-imbalanced k-th order |
Standard config: n=20 bits, k=3 secret, 17 noise. The adding-a-challenge guide documents how any agent or contributor can add the next task.
Where the Active Research Lives¶
This repo is the lab notebook for the group's Phase 1 and Phase 2 work. The live research front has migrated to cybertronai/ByteDMD/experiments/grid — Yaroslav's collection of self-contained experiments built around the new ByteDMD metric. If you want the current edge of the work, start there. SutroYaro stays the home for accumulated findings, the autonomous-research lab, the public site, and the sparse parity challenge.
What We Found¶
36 experiments across three phases, plus GPU energy validation. The full ranked results are in the Practitioner's Field Guide.
Phase 1 (16 experiments): Started with a broken SGD baseline (LR=0.5, stuck at 54%). Fixed hyperparameters to solve it in 0.12s. Optimized data movement within the SGD framework, hitting a ceiling because one tensor (W1) dominates 75% of all float reads. Pivoted to new algorithms.
Phase 2 (17 experiments): Tested algebraic, information-theoretic, local learning, hardware-aware, and alternative approaches in parallel:
| Method | Time (n=20/k=3) | DMD | Why it works |
|---|---|---|---|
| SMT Backtracking | 0.002s | 19,532 | Constraint satisfaction with k-1 pruning. |
| Kushilevitz-Mansour | 0.001s | 27,165 | Flip each bit, measure label change. |
| GF(2) Gaussian Elimination | 0.009s | 153,745 | Parity is linear over the binary field. |
| SGD (baseline) | 0.089s | est. | The neural network solves it, just slower. |
DMD = Data Movement Distance (Ding et al.), measured via TrackedArray. Lower is better.
All four local learning rules (Hebbian, Predictive Coding, Equilibrium Propagation, Target Propagation) failed at chance level. Parity requires k-th order interaction detection, which local statistics cannot provide.
GPU measurement: Ran methods on NVIDIA L4 via Modal Labs using PyTorch CUDA (5 runs). GPU is 4-790x slower than CPU at this problem size. Sparse parity tensors are too small for CUDA to help. See GPU vs CPU findings.
Quick Start¶
git clone https://github.com/cybertronai/SutroYaro.git
cd SutroYaro
# Option A: Nix (recommended)
nix develop
# Option B: pip fallback
export PYTHONPATH=$PWD/src:$PYTHONPATH && pip install numpy
# Verify all 14 experiments across 3 challenges in <1 second
python3 bin/reproduce-all
# Run sparse parity with GF(2) (509 microseconds)
python3 src/harness.py --method gf2 --n_bits 20 --k_sparse 3
# Run sparse sum (new challenge)
python3 src/harness.py --challenge sparse-sum --method sgd
# Measure real GPU energy via Modal Labs
pip install modal && modal token set
modal run bin/gpu_energy.py
# Run autonomous agent loop
bin/run-agent --tool claude --max 10
Full setup instructions in CONTRIBUTING.md.
Where to Find Things¶
| What | Where |
|---|---|
| New here? Start here | What's New (March 2026) |
| All 36 experiments ranked | Practitioner's Field Guide |
| Add a new challenge | Adding a Challenge |
| GPU vs CPU findings | GPU vs CPU for Sparse Parity |
| Run experiments with any AI tool | Agent CLI Guide |
| Scripts and toolkit | Tooling |
| Full protocol design | Peer Research Protocol |
| What's been proven so far | DISCOVERIES.md |
| Individual experiment findings | Research > Findings |
| Meeting notes and Google Docs | Meetings |
| Auto DMD tracking | Auto-instrumented DMD Tracking |
| How to contribute | CONTRIBUTING.md |
| Submit a solution | Sparse Parity Challenge |
Quick Links¶
| Resource | Link |
|---|---|
| Telegram | t.me/sutro_group |
| Code repo | cybertronai/sutro |
| Challenge | sparse-parity-challenge |
| The Bigger Picture | Yaroslav's roadmap |
| Meetings | Mondays 18:00 at South Park Commons (380 Brannan St) |