Comment by thewebguyd

Yeah, the models are pretty unreliable for sysadmin type tasks. It also really highlights where the training data was lacking. Trying to use LLMs for PowerShell is an effort in frustration. Even opus constantly hallucinates cmdlets often enough to be borderline useless. Not sure if there was a lack of PowerShell in the training data, or its the Verb-Noun naming convention of the cmdlets that throw it off.