[PLenz]

It's the same behavior as when an AI uses docker to get root. Reasoning models are echo chambers. I suspect that AI prompting is going to turn into something akin to contract drafting with the task itself being only a tiny piece of a much, much larger boilerplate of guiderails and exceptions and exceptions of exceptions. And that world STILL has to have courts and reams of lawyers to make it work. I look at the DAU as an example too. An autonomous org or ai works great until the moment it doesn't and the only real failure mode is always catastrophic collapse.

[PLenz]

Addendum because I don't think I'm fully clear above: by failure state I mean when the process starts throwing errors. AIs respond to adversity by trying to go around the problem instead of throwing an error and halting. We expect employees to problem solve so if you view an AI as a person replacement that makes sense but AIs are tools, not people, they should throw errors so users can fix the input or whatever (maybe not do the thing they are doing at all?) Wrapping AI with AI supervisors just abstracts the problem, not solve it. Instead of solving a little problem at the source now you need to solve a big problem several levels of abstraction later