How much should we understand LLMs before deploying them at scale?

One post jokes about demanding an exact account of human brains before inventing college and letting people get smarter. A poster quoting it says a complete understanding of LLMs is not needed for safe deployment, but argues that studying how models work internally adds valuable evidence as other monitoring becomes less useful. They say the human comparison misses the scale of potential AI deployment, imagining millions of fast, capable copies with access to private files and networks.