Waymo needs to get this fixed and I support the NTSB investigation.
I’ll push back on how you are framing the safety debate overall.
If AVs are much safer overall but less safe in a set of rare edge cases, and the net result is many fewer deaths and injuries, is that a tradeoff worth making?
Almost everyone reading this post commits 10+ moving violations a day. Speeding, rolling through stop signs, failure to signal turns, running stale yellow lights, etc.
That is the baseline that AVs are trying to improve on. I don’t what them to have a blank check but I do want the approach that saves lives overall.
Thanks, I did watch it after seeing it on a previous post. I agree that the comparison can be used to wave away AV mistakes, and that's not a good faith approach to the discussion.
You call it a trope that humans are bad drivers - in your experience do you not see people on their phones, speeding, etc. on a continual basis as you drive currently (not sure where you live, I'm in Austin)?
I like the idea that we can improve safety to European levels but that also not the reality we live in. I don't see the movement to implement European car sizes and driving standards gaining momentum in the US the same way AVs are. The promise (which I agree will need lots more data to continually reconfirm) is that AVs can work in practice at levels that are safer than even the European comparisons looks very real.
I'll leave it here. Appreciate you being willing to engage in real discussion. Keep up the good work :)
Kevin, thanks for your thoughtful feedback. This is a central tension point right now: potential net statistical safety gains vs. specific unsafe behaviors, especially with poor optics (e.g., putting school kids in danger).
It will be years before we really know how net statistical safety will turn out at scale. (Waymo claims are overstated, and even if true do not apply to other companies.) In the meantime I think we will end up comparing robotaxis to humans on an incident-by-incident basis. Not because this is optimal on a purely utilitarian basis, but rather because this is what society will demand in practice. The comparison to humans is valid, but the net statistical approach is not.
Thanks for the article, enjoyed it.
Waymo needs to get this fixed and I support the NTSB investigation.
I’ll push back on how you are framing the safety debate overall.
If AVs are much safer overall but less safe in a set of rare edge cases, and the net result is many fewer deaths and injuries, is that a tradeoff worth making?
Almost everyone reading this post commits 10+ moving violations a day. Speeding, rolling through stop signs, failure to signal turns, running stale yellow lights, etc.
That is the baseline that AVs are trying to improve on. I don’t what them to have a blank check but I do want the approach that saves lives overall.
The trope that humans are terrible drivers as a motivation for autonomous vehicles is IMO misguided.
8-minute explainer here: https://youtu.be/pYb4X5aJhgU
Thanks, I did watch it after seeing it on a previous post. I agree that the comparison can be used to wave away AV mistakes, and that's not a good faith approach to the discussion.
You call it a trope that humans are bad drivers - in your experience do you not see people on their phones, speeding, etc. on a continual basis as you drive currently (not sure where you live, I'm in Austin)?
I like the idea that we can improve safety to European levels but that also not the reality we live in. I don't see the movement to implement European car sizes and driving standards gaining momentum in the US the same way AVs are. The promise (which I agree will need lots more data to continually reconfirm) is that AVs can work in practice at levels that are safer than even the European comparisons looks very real.
I'll leave it here. Appreciate you being willing to engage in real discussion. Keep up the good work :)
Kevin, thanks for your thoughtful feedback. This is a central tension point right now: potential net statistical safety gains vs. specific unsafe behaviors, especially with poor optics (e.g., putting school kids in danger).
It will be years before we really know how net statistical safety will turn out at scale. (Waymo claims are overstated, and even if true do not apply to other companies.) In the meantime I think we will end up comparing robotaxis to humans on an incident-by-incident basis. Not because this is optimal on a purely utilitarian basis, but rather because this is what society will demand in practice. The comparison to humans is valid, but the net statistical approach is not.
I have a talk on this very subject here (starting on slide 14): https://philkoopman.substack.com/p/keynote-talk-embodied-ai-safety?r=1tnaed&triedRedirect=true