[slowmovintarget]

It's unclear that what we currently have in LLMs can devise tests, perform the measurement, and act on the results (because it saw to the results being in a form it could ingest).

We've given GPT models access to REPLs to see if they'd become Skynet all of a sudden, but they still need motivation. That motivation loop is currently run by humans... which perhaps are the least trustworthy part of the whole technology stack.

Hook up sensors to induce prompts... yikes.