The stats don't matter much when it's clear they can't handle collision avoidance with anything outside their limited training set like airplanes. You have to ask what else can't they detect in front of the car if they're 100% dependent on ML to decide there's an obstruction. That is an irresponsible threat to public safety no matter how lucky they are scraping by on their current architecture.

Except the stats *do* matter. It’s how we measure things that operate in the real world. The blind baby stroller benchmark might be academically interesting, but if distracted drivers are smashing kids on public roads for another decade while Waymo perfects it’s craft the net result is just more flat people.

Talking of FSD beta here, I don't think there has been any death or even major accident to date. But that's mostly because (especially until a few months ago) it was bad enough that nobody would trust it, so people were always alert to intervene.

If you were to let FSD beta just drive by itself enough time, and intentionally never intervene, it would eventually crash, no doubt about it. Before v10.69 it was hard to get a 20 minute drive with no interventions (unless it was mostly straight roads).

Relevant HN: https://news.ycombinator.com/item?id=33984922

Comparing disengagement and driver intervention data might be useful to compare Tesla FSD vs others.