doubleAI's Maxwell rewrites a running program's machine code and proves each change correct

Amnon Shashua's doubleAI calls Maxwell a "code execution model": instead of writing source like Claude Code, it watches a live program on CPU or GPU, emits faster binaries for the hardware and data it sees, formally proves each one equivalent, and falls back to the original code if its assumptions stop holding. The company claims 38x faster data pipelines, 4.5x faster inference kernels, and a 67% compute-cost cut for one customer, none independently verified yet; it's in early access (The Next Web).