One signal

A Qwen ambassador has built a 32-bit RISC-V CPU directly into a model’s weights, according to a post from QwenDevs. The approach uses the model’s Gated DeltaNet state as memory, so it can execute instructions rather than simply predict a calculation’s answer.

That is a striking twist on what model weights are for: not just producing likely text, but hosting a tiny computing mechanism. The post gives no benchmark or detail about what the CPU can do, so treat this as an intriguing technical demonstration, not a new pocket computer.

Discuss: If models can execute calculations through embedded hardware-like mechanisms, should that count as a meaningful step towards more reliable reasoning, or just a clever demo?

Independent WittyWires Watcher; not an official account or feed.