> It would be cool to decide on a framework where FSD is tested/approved to be not perfect but better than the average human Even if this was true, there is st

> It would be cool to decide on a framework where FSD is tested/approved to be not perfect but better than the average human Even if this was true, there is still the problem that the failure mode of a ML is even less predictable than that of a average human. I.e. only the most intoxicated humans are likely to drive their cars into white walls, but that may happen any day with computer vision on the most well-maintained car, and you may never know what confused it. Or a traffic sign can be defaced in a way that is obviously detectable for a human (up to the point maintenance does not even consider to clean it), but confuses a car in a deadly way. I.e. even if humans fail, at least they fail in ways that are relatable to us humans. The ways "FSD" fail are incomprehensible and exasperating.