LLMs don't learn anything after being deployed. They're probabilistic and cost you more the larger and more complex your system becomes.
Are we thinking of tests like we've never done before?
Are we prepared to build platforms again from scratch when they become untenable?