tl;dr: A researcher compiled a working 32-bit RISC-V CPU (RV32IM) directly into the weights of Qwen 3.5 4B, using the model's Gated DeltaNet attention layers as exact, addressable RAM—no training, external tools, or sandboxes involved. The CPU runs during forward passes, executes C and MicroPython, passes official riscv-tests, and boosts arithmetic benchmarks to match or beat giving Qwen a Python tool (e.g., 48.6%→100% on NUPA long-integer math). The author also added an in-weight filesystem and suggests the technique as a testbed for AI safety research on RNN-state reasoning and tamper-proof monitoring.