autoresearch gets Apple Silicon with M3 Ultra port

Two competing PRs bring MPS support to single-GPU training, opening the door for researchers on Mac hardware.

Contributors steefvw and jeong-sik independently submitted MPS (Metal Performance Shaders) ports this week, each targeting Apple Silicon acceleration. The Bach inventions PR and the M3 Max support PR represent parallel efforts to eliminate GPU dependency for the research automation pipeline. @karpathy now faces a pleasant triage problem: reconciling two paths to the same goal. Community is clearly signaling demand for Mac-native training.

nanochat tokenizer switches from pickle to JSON, fixing Docker on ARM

A serialization swap solves crashes on ARM-based container platforms without code bloat.

Contributor abcdedf identified that pickle-based tokenizer serialization was crashing nanochat inside ARM Docker environments. The fix trades pickle for JSON serialization, eliminating architecture-specific binary incompatibilities. Separately, vivian-my posted a 4.3% speedup through architecture and optimizer tuning (94.6 minutes baseline), showing the repo remains under active performance pressure.

autoresearch security PR demands weights_only torch.load

A hardening measure addresses arbitrary code execution risk in model loading.

Contributor robotlearning123 submitted a security fix requiring weights_only mode during torch.load operations, blocking deserialization of untrusted pickled objects. This is a high-signal defensive change for a tool that automates research pipelines—malicious checkpoint files could compromise experiments at scale. The PR also specifies safe exit protocols.

autoresearch experiments now print hyperparameters in final summary

A UX improvement logs training configuration for reproducibility.

Contributor mvanhorn added hyperparameter printing to the experiment summary, surfacing learning rates, batch sizes, and other config directly in final output. Small change; large impact for researchers reviewing logs weeks later.