Language Models as Virtual CPUs

LLM layers can be engineered to implement instruction semantics and recurrent state as memory, effectively compiling a small CPU inside model weights to execute exact instructions. This approach (exemplified by the Qwenputer experiment) trades model size and latency for new on‑model computation primitives that change where and how software runs. — If models can reliably execute instructions as virtual CPUs, it reshapes software security, vendor lock‑in, performance economics, and regulatory questions about what counts as 'running code' versus 'model output.'