Qwenputer runs deterministic RISC-V CPU inside Qwen models

Qwenputer's demonstration that an open-weight model can hold a deterministic register state challenges the common assumption that generative AI is inherently probabilistic and unreliable for exact computation. By enforcing structured outputs and a strict feedback loop, the project achieved stability over hundreds of thousands of steps — a result that suggests LLMs can be repurposed as verifiable state machines in narrow domains. This pairs with a second incident: a heavily obfuscated fingerprinting script was mistakenly flagged as containing bioweapon design data because its byte-array shuffling and encoding happened to match DNA signature patterns. Together, these cases point to a shared engineering concern — the need for verifiability in black-box processing — and combining static AST analysis with dynamic sandbox evaluation is recommended as a practical audit approach. It also cautions that running such emulation locally requires FP16 or BF16 precision to avoid quantization noise breaking determinism, and that over-obfuscated third-party scripts risk false positives in security scanners.