@savageazfck2
Adam ClarkDeveloper▲5
system architect, independent
57m ago · New project · Agents · Akron

Shipped: coyote-apple

What it does:
Compiles tensor programs to the Apple Neural Engine and four other substrates (GPU, CPU, encrypted, remote-under-treaty) from a single source where the model is declared in code with provenance. Every run is witnessed, replayable, and certifiable. rev fn programs are provably invertible: uncall runs the verified inverse.

Who uses it:
Bad Apple, my sovereign AGI organism, is the primary user (its ANE brain, its ANE-compiled output firewall, its witnessed training runs). Beyond that it's early: ~450 unique cloners across the repos in the last two weeks and ~2,900 crate downloads, mostly people poking at the MIL toolchain.

One number that shows it works:
4B — a 4-billion-parameter model resident and running on the Neural Engine through this toolchain. It's the only public stack that does it, and the distillation run training its successor is live right now.

coyote-apple logocoyote-apple@savageazfck/coyote-apple · github.com/savageAZfck/coyote-apple

Comments

No comments yet
Start the conversation.
Log in to join the thread.