Dear researchers: help me deal with incidents

Excellent document.

I will quibble (slightly) with only one piece. The author writes:

We aren’t going to get [SRE AI] from the vendors, nor from the proof-of-concept projects inside companies.

Actually, I am seeing some excellent results from "proof-of-concept" projects inside my company: things that are cutting down the response time for incidents by a substantial amount by collecting information immediately and surfacing a summary for operators. (Letting AI directly take action during an incident is something we will not permit until we have significantly more experience.) This is with an AI configured specifically for our system, not a generic tool.