LLM.Genesis packs LLM inference into 64KB SRAM

This feels less like a faster llama.cpp and more like a manifesto for turning LLM execution into a tiny virtual machine. The 64KB SRAM target is the real differentiator: it makes the project genuinely interesting for embedded setups, but it also means I/O, tooling, and format friction will matter more than raw throughput.

• –The `STREAM` opcode and paged weight loading push the bottleneck toward storage, so latency will be dominated by flash or SD-card performance.
• –GCS DNA is a clever separation of model logic from the runner, but it creates a new format the ecosystem has to adopt and debug.
• –The runtime claims zero external dependencies, yet the repo still leans on Python for compilation tooling, so the build story is light, not pure C++.
• –There are no releases yet, which makes this read more like an architectural prototype than production-ready inference infrastructure.