A researcher finds an accelerator 80x more efficient than an A100 already shipped

Claim

A solo researcher using Claude Code ran Karpathy's llama2.c on Apple's M4 Neural Engine at under a watt by reverse-engineering undocumented APIs, uncovering an accelerator eighty times more efficient than an A100 already present in hundreds of millions of devices.

Reported metric: 80x / <1 W

Reporting date: 2026-03-03 (not necessarily the event date)

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

• Welcome to March 3, 2026 (theinnermostloop.substack.com)