Working with a team of SREs using LLMs to troubleshoot production issues and holy shit - the rate at which it uses that exact language and comes to completely fabricated or absurd conclusions is close to 80-90%