We're debugging LLMs in production by reading chat logs

"We don't usually test, but when we do we test in production"