Karpathy's AutoResearch with Memory-in-the-Loop States by The Adimension#302

Open

habanwer wants to merge 11 commits intokarpathy:masterfrom

habanwer:autoresearch/memory-in-the-loop

Commits on Mar 16, 2026

ground.json [NEW] — externalize platform config from karpathy/autoresearch prepare.py
habanwer
committed
model.json [NEW] — externalize hyperparameters from karpathy/autoresearch train.py
habanwer
committed
prepare.py [MODIFIED +125/-17] — add GPU platform detection, read config from ground.json
habanwer
committed
train.py [MODIFIED +194/-55] — load from model.json, import PLATFORM, GPU-safe numerics
habanwer
committed
program.md [MODIFIED +96/-114] — structured agent protocol replacing free-form narrative
habanwer
committed
.gitignore [MODIFIED +5/-2] — update ignore rules for multi-GPU workflow
habanwer
committed
analysis.ipynb [MODIFIED metadata] — update kernel to Python 3.12.10
habanwer
committed
README.md [MODIFIED] — add Memory-in-the-Loop fork introduction with DEITY Principles mapping
habanwer
committed

Commits on Mar 17, 2026

Address review feedback: fix eval_steps guard, ground.json validation, PLATFORM[compile] honoring, rotary dtype cast, unused imports, TSV header format, README grammar, notebook version alignment
habanwer
committed

Commits on Mar 18, 2026