autoresearch-ane brings Karpathy loop to ANE

This is the most interesting kind of AI tinkering: not a new model, but a new way to squeeze useful research cycles out of hardware that was never meant to be open for training. If it holds up, the real story is throughput-per-watt and overnight experimentation on commodity Macs, not raw benchmark glory.

• –The fork stands on two timely trends at once: Karpathy’s `autoresearch` loop and the recent reverse engineering work that exposed ANE training paths.
• –Its biggest reported gain is architectural, not algorithmic: eliminating per-batch recompilation reportedly boosts throughput from roughly 120 to 1340 steps in the same 5-minute budget.
• –The repo is honest that this is a separate training stack with different data, metrics, and implementation details, so its numbers are not directly comparable to the original CUDA version.
• –Using private Apple APIs makes this exciting for researchers and hackers, but fragile for anyone hoping for a stable production path.
• –As a prototype, it is still early, but it points toward a broader idea: autonomous model experimentation could move down from datacenter GPUs to personal silicon.