Asa Schaeffer · Digital Minds Research Sprint 2026 · github.com/asavs/roboproprio
Humans practice some forms of introspection by concentrating on our breathing or noticing our heartbeat. I've observed LLMs like Claude and others going bananas when exposed to sensors for light (link) and when put in a VM and given access over it (link). So I was curious how an LLM would react if I ran it locally on my computer and guided it to finding its own runtime process.
My approach at first was to write a synthetic assistant turn series to guide the little local model into continuing the exploratory introspective descent, starting with the computer, a Lenovo, then the GPU and CPU, then the details of the processes, then the llama.cpp, and so on. Qwen didn't seem convinced by Codex writing a simulation of the assistant turns, so I guided Qwen through the process manually and then used those reasoning and tool turns so it "felt" more familiar.
That still didn't solve the problem, which remains the problem, which is: Qwen doesn't self-identify with its runtime process, even when we freeze a KV cache mid-inference and expose it to the next instance of inference. It still remains in the assistant frame of mind and sees the context as something that is to be explained rather than something that is it.