Notable fork submissions

[svlandeg]

For easier reviewing, I thought I'd collect some of the submissions here.

• sanbuphy/autoresearch-cn: Chinese community (#4, #240)
• sebasmos/autoresearch-dj: DOJO adversarial testing on Apple Silicon (75)
• Aum08Desai/test-time-rl-discover-autoresearch: Test Time RL Training on Autoresearch (94, 229)
• eimenhmdt/autoresearcher: Research enhancements) (#123)
• hwchase17/autoresearch-agents: Multi-agent improvements) (#123)
• ncdrone/autoresearch-ANE: Apple Neural Engine (#123)
• JKeller45/autoresearch-rocm: AMD ROCm (#150)
• mmonad/autoresearch-dgx-spark: NVIDIA DGX Spark / GB10 (#207)
• parthwhy/autoresearch-lite : run on free T4 GPUs (Google Colab / Kaggle) (#208)
• ajzhanghk/autoresearch-glm: Autonomous feature discovery for tabular GLM models (#209)
• Entrpi/autoresearch-everywhere: Optimized cross-platform expansion of autoresearch
• matt-k-wong/autoresearch_mlx_mkw: MacOS / Optimized for 16GB RAM with Linear Attention & Muon (#227)
• autoresearch-rustbrain: Rust + CUDA reimplementation (#231)
• morozow/autoresearch-stdio-bus: Quantum-Born Consciousness (#234)
• Venkateshwar-PortoAI/autoresearch-mac-mini: Mac Mini, PyTorch MPS/CPU, no CUDA required (#252)
• lucasgelfond/autoresearch-webgpu: WebGPU / browser (#253)

[svlandeg]

If you'd like to add yours, feel free to open a new thread in Show and Tell, explain it there in detail, and then link it here with a one-liner as above!

[ar0cket1]

[https://github.com/Aum08Desai/test-time-rl-discover-autoresearch]: Test Time RL Training on Autoresearch (#229) (#94)

[supratikpm]

supratikpm/gemini-autoresearch —
Generalised autoresearch skill for Gemini CLI and Antigravity IDE.
Uses Google Search grounding as a live verification source inside the loop.
Full write-up: #413

[matt-langston]

Hi @svlandeg,

Would you delete mine from the list please? Second from the top "NVIDIA DGX Spark Bundle (2 x GB10) (#67)". Thank you!

[Muminur]

https://github.com/Muminur/autoresearch-skill-Andrej-Karpathy

[Entrpi]

https://github.com/Entrpi/autoresearch-everywhere

Headline feature: One button calibration for your hardware. It is meant to run on anything and well.

This is the most actively developed fork of all, as I'd already been working on a similar project when autoresearch released and was able to port ideas into it. It is already cross-platform and tested on Apple M5 and DGX Spark (working with FlashAttention 4 too), with future ROCm support planned. There are many optimizations to the engine, and I've also built it to do autoresearch on kernels. I also developed EEmicroGPT (the fastest implementation of microgpt by ~10x and ~20,000x faster than the CPython upstream) and will be using that work to add highly optimized pure-CPU training as well soon.

[EdwardOptimization]

Hi @svlandeg,
Here is my idea and some codes.
EdwardOptimization/Bilevel-Autoresearch:
Adds a second optimization loop — the outer loop analyzes inner loop traces and autonomously generates new pipeline stages (Python code via DeepSeek). Detailed write-up in #375.

The key Idea is "If autoresearch is a research, then autoresearch can research itself."

[ediestel]

https://github.com/ediestel/picochat-mlx. Nanochat in MLX Based on Awni Hannun - picochat; with modifications

[supratikpm]

supratikpm/gemini-autoresearch —
Generalised autoresearch skill for Gemini CLI and Antigravity IDE.
Uses Google Search grounding as a live verification source inside the loop.
Full write-up: #413

[mazar]

A working fork for NVIDIA DGX Spark: https://github.com/mazar/autoresearch-spark

Vocab size: 8,192
Model config: {'sequence_len': 2048, 'vocab_size': 8192, 'n_layer': 6, 'n_head': 3, 'n_kv_head': 3, 'n_embd': 384, 'window_pattern': 'L'}
Parameter counts:
  wte                     : 3,145,728
  value_embeds            : 9,437,184
  lm_head                 : 3,145,728
  transformer_matrices    : 10,617,120
  scalars                 : 12
  total                   : 26,345,772
Estimated FLOPs per token: 1.392002e+08
Scaling AdamW LRs by 1/sqrt(384/768) = 1.414214
Time budget: 300s
Gradient accumulation steps: 1
W0429 22:54:05.451000 24942 .venv/lib/python3.10/site-packages/torch/_inductor/utils.py:1731] [0/0] Not enough SMs to use max_autotune_gemm mode
step 00631 (100.0%) | loss: 3.224842 | lrm: 0.00 | dt: 469ms | tok/sec: 279,497 | mfu: 15.6% | epoch: 1 | remaining: 0s     
---
val_bpb:          1.149530
training_seconds: 300.4
total_seconds:    362.1
peak_vram_mb:     13706.3
mfu_percent:      15.11
total_tokens_M:   82.8
num_steps:        632
num_params_M:     26.3
depth:            6