Well, first of all the majority of the discussion was around assets which are not what most people have in mind when they talk about software development. But also LLMs work great for their intended purpose of auto completion, even a local model running on your own machine is miles ahead of the snippet libraries we used to use until recently. Also they can spot errors in your code much better than you yourself can, as they’re essentially impartial and looking for patterns that have previously been reported as causing issues statistically, while you are biased towards the mistakes you usually make.

I wouldn’t trust anything it outputs any more than what I would trust something written by an 8 year old on heavy doses of stimulant drugs, but in reduced scope and limited capacity they’re quite useful, and can help in some tricky situations like memory corruption or threading issues.

That being said I strongly support that vibe code is not trustworthy and anyone who does that is essentially gambling with that code.