autoresearch-mlx ports Karpathy loops to Apple Silicon

This is the kind of fork that makes a clever research idea actually usable by more builders, and the loop itself is the real product here. The Mac-specific port also highlights how much of autoresearch’s value comes from iteration speed and discipline, not just raw GPU scale.

• –It preserves the core protocol: one editable `train.py`, a fixed 5-minute training budget, a single metric, and git-based keep-or-revert decisions.
• –MLX removes the PyTorch/CUDA dependency, which makes autonomous research loops practical on Apple Silicon laptops and desktops.
• –The repo’s published runs already show meaningful gains under a strict wall-clock budget, which is a good sign the method is doing real work rather than just generating churn.
• –The best results look hardware-sensitive, so this feels like a strong companion fork to upstream autoresearch, not a substitute for large-GPU experimentation.
• –For AI devs, the appeal is obvious: you can prototype the research loop locally on a Mac, then decide later whether it’s worth porting upstream ideas back to bigger hardware.