Related repos in the cybertronai GitHub org¶
A single source of truth for the repositories adjacent to SutroYaro. Updated 2026-05-09.
Active research front (May 2026)¶
These repos are where current work lives. If you're picking up the project, start here.
| Repo | What it is | Latest activity |
|---|---|---|
SutroYaro |
This repo. Lab notebook for Phase 1+2 work, autonomous-research infrastructure, public docs site, sparse-parity scoreboard. | Active |
ByteDMD |
The primary cost metric (data-movement complexity, byte-granularity LRU stack). Yaroslav's active research front lives at experiments/grid — self-contained experiments. |
Active |
simplified-dally-model |
Yaroslav's 2D Manhattan-grid cost model. Single-processor, explicit communication cost. Goal: minimum instruction set that captures the relevant cost dynamics. Will eventually replace 1D Dally pJ numbers in the SutroYaro eval harness. | Active (2026-05-08) |
sutro |
Main code repo. Owns sparse_parity_benchmark.py and the original solver implementations. |
Reference |
sutro-problems |
Small reproducible problems collection. Has the matmul energy-metric work. Andy Zhang is now an owner (Telegram, 2026-05-06) and active on this. | Active |
sparse-parity-challenge |
Submission pipeline: submitters open a GitHub Issue with a solve() function; CI auto-evaluates under ByteDMD and posts the score. |
Active |
Companion baseline catalogs (shipped May 2026)¶
Reproducible-baseline implementations of the synthetic learning problems from two distinct paper lineages. Both shipped via the agent-team build pattern; see each repo's BUILD_NOTES.md for the session-level details.
| Repo | Paper lineage | Stubs | Site |
|---|---|---|---|
hinton-problems |
Hinton 1981–2022 — representational toy tasks (4-2-4 encoder, family trees, shifter, capsules, Forward-Forward) | 53 v1 + 2 add-ons | site |
schmidhuber-problems |
Schmidhuber 1989–2025 — algorithmic capability (long-time-lag indexing, key-value binding, Levin/OOPS search, controller+model+curiosity, World Models) | 50 v1 + 8 v1.5 | site |
Together: the representational + algorithmic baseline pair. Both are pure numpy + matplotlib, laptop-runnable, with paper-comparison metrics per stub. The follow-up v2 work is to instrument these baselines with ByteDMD and compare data-movement cost across algorithm families.
Tracking issues: - hinton-problems #45 (v2 ByteDMD) and #46 (v1.5 paper-scale) - schmidhuber-problems #17 (v2 ByteDMD) and #18 (v1.5 paper-scale + original-simulator)
Adjacent / external¶
| Repo | Connection |
|---|---|
adotzh/SutroAna (not in cybertronai org) |
Anastasia Zhiboedova's auto-research-loop framework. Presented at meeting #16 (04 May 26). Independent of SutroYaro but solving overlapping problems. |
Older cybertronai repos (reference, not active for this work)¶
The org has a long history. These are not currently part of the energy-efficient-training thread but are useful background:
scaling-laws,autograd-hacks,autograd-lib— Yaroslav's instrumentation toolspytorch-sso,pytorch-lamb,pytorch-fd— second-order methods, LAMB, fluctuation-dissipationgradient-checkpointing— memory-efficient trainingtransformer-xl,imagenet18,Megatron-LM,bflm— training-at-scale workncluster,pytorch-aws,aws-network-benchmarks,autotune— AWS / training-infra tools
How they connect (conceptual map)¶
graph LR
%% Cost / metric layer
ByteDMD["ByteDMD
byte-granularity LRU
stack-distance tracer"]
SimpleDally["simplified-dally-model
2D Manhattan-grid
communication cost"]
%% Problem layer
SPChall["sparse-parity-challenge
submission pipeline,
auto-evaluated"]
Hinton["hinton-problems
53 representational stubs
1981–2022, pure numpy"]
Schmid["schmidhuber-problems
58 algorithmic stubs
1989–2025, pure numpy"]
SutroP["sutro-problems
matmul + small
reproducible problems"]
Sutro["sutro
sparse_parity_benchmark.py
+ original solvers"]
%% Lab
SY["SutroYaro (this repo)
lab notebook · scoreboard ·
autonomous research · public site"]
%% Adjacent
Ana["adotzh/SutroAna
auto-research loop
(meeting #16)"]
%% Edges
ByteDMD -->|"cost metric"| SY
SimpleDally -->|"cost metric (next)"| SY
SPChall -->|"problem set"| SY
Hinton -->|"problem set (v2 target)"| SY
Schmid -->|"problem set (v2 target)"| SY
SutroP -->|"problem set"| SY
Sutro -->|"reference solvers"| SY
Sutro -.benchmark spec.-> SPChall
Ana -.adjacent.-> SY
%% Styling — three families: cost (purple), problems (blue), lab (green), adjacent (gray)
classDef cost fill:#e8d5ff,stroke:#7e3ff2,stroke-width:2px,color:#000
classDef problem fill:#cfe5ff,stroke:#1f6feb,stroke-width:2px,color:#000
classDef lab fill:#d1f4d1,stroke:#1a7f37,stroke-width:3px,color:#000
classDef adjacent fill:#eee,stroke:#999,stroke-dasharray:5 3,color:#000
class ByteDMD,SimpleDally cost
class SPChall,Hinton,Schmid,SutroP,Sutro problem
class SY lab
class Ana adjacent
Reading the diagram:
- Purple (left) — cost-metric repos.
ByteDMDis current;simplified-dally-modelis the 2D grid successor Yaroslav is iterating on. - Blue (middle) — problem repos. The submission pipeline (
sparse-parity-challenge), the two companion baseline catalogs (hinton-problems,schmidhuber-problems), andsutro-problemsfor everything else (matmul energy-metric work, etc.).sutroholdssparse_parity_benchmark.pywhich feeds the challenge spec. - Green (right) — the lab.
SutroYaroconsumes the cost metric + the problems and produces: scoreboard, accumulated findings (DISCOVERIES.md), autonomous-research infrastructure, public docs site. - Gray (dashed) — adjacent.
adotzh/SutroAnais solving overlapping problems independently (presented at meeting #16, 04 May 26).
Updating this doc¶
Maintained by hand for now. Bump alongside the active-threads catch-up when:
- A new repo is added to the
cybertronaiorg - Yaroslav signals "this older repo is now active again"
- The conceptual map changes (e.g.,
simplified-dally-modelbecomes the primary metric and 1D ByteDMD goes to reference)
The gh repo list cybertronai --limit 50 command is the authoritative source of what exists; this doc is the curated narrative.