Silicate Zero: Booting an AI on Bare Metal
Silicate Zero: Booting an AI on Bare Metal
For 50 years, operating systems have been designed to multiplex CPU time for human typists. We are replacing the software lasagna. Silicate Zero is a bare-metal OS engineered from the ground up for autonomous neural networks.
This is the Stage 11 Public Preview: Booting a custom, pure-Rust Ring-0 unikernel that executes a 1.7B parameter LLM (Qwen) directly on bare-metal Apple Silicon (HVF). No Linux. No POSIX. No C standard library. Just raw matrix multiplication on the silicon.
A Note on Performance (Why is it so slow?):
You will notice the generation takes about 62 seconds for the first sentence (~0.4 tokens/sec). This is intentional. For this milestone, we completely disabled SIMD, GPU/Metal acceleration, and batched processing. The engine is running pure scalar f32 math on a single thread.
Why? To prove 100% architectural correctness, deterministic cross-platform parity, and zero-allocation memory safety in Ring-0. We built the unshakeable mathematical baseline first. Performance optimization is the next stage.
The Silicate Doctrine (AGPLv3 Dual-License):
The operating system of the neural era must be built in the open. Silicate Zero will be released under the AGPLv3 dual-license model. It will be 100% free and open for builders, hackers, and researchers pushing the limits of bare-metal inference. The codebase and architectural manifests will be published as we reach stable milestones.
Next on the Roadmap:
1. Sub-MP-D3.5: Introducing Hardware RNG for non-deterministic sampling (giving the AI the ability to "dream" and break out of greedy-search loops).
2. Sub-MP-G: NEON SIMD and Metal Performance Shaders integration to achieve viral-grade token throughput.
3. Stage 14: Neural IMGUI Desktop environment.
Computing is being rewritten. Welcome to Ring-0.
https://silicate.tech/