Comment by addaon

Yep. This is why the estimates of compute needed for AI (if it turns out to be useful) are many orders of magnitude too low — the technology isn’t mature until it actually succeeds at tasks, with fully homomorphic encryption from my prompt through the response.