Did the paper really prove its claim? It set aside more than half of Waymo's claims history and compared only the most recent portion of its own history to the all drivers benchmark. All drivers do things that Waymo doesn't like operate outside specific city centers, drive in snow or other severe weather, real people drive more during the day in traffic but Waymo is biased toward nighttime (less traffic), and real people drive freeways and high speed roads.
If you rigorously enforced those limits on regular drivers and additionally imposed an interlock device to prevent DUI and remote insurance company monitoring of driving habits-- that's the comparison population I want to see. Otherwise, this is a press release disguised as an academic paper.
It seems like you want them to explore a different question. Im confident that the best human drivers are better than Waymo. They are also better than all human drivers, by definition.
If you are worried about other drivers hitting you, the "all driver" comparison seems relevant. If you want to know if Waymo is a safer driver than you are, that will depend on where you are on the spectrum.
Maybe DUI's should be included, but a paper drawing the conclusion that "Waymo is safer because its computer drivers aren't drunk" suggests very different policy actions.
Unless we really can finally stop drunk driving overnight after all these years of trying I don't agree the comparison should result in significantly different policy. I want to know how Waymo statistically compares to all of all the actual drivers when setting policy, not how Waymo compares to what we wish real drivers would have been like. The only mistake would be to compare solely to drunk drivers, which is what you comment makes it sound like.
I would like to see the data sets filtered to match driving areas though. I.e. Waymo drives almost exclusively in urban environments so per mile data should be compared almost exclusively to urban driver data.
> ..not how Waymo compares to what we wish real drivers would have been like.
I want to see "all the things", but in specific reply to this point, i want to see this because i have no intention to commit a DUI myself. So the bar for me personally is how well it'll do compared to safe drivers.
Lots of stats are useful, but i don't think we should undersell the various "ai vs me" comparisons that individuals could use. Since those are the ones that will help people decide if they themselves can/should/shouldn't take a Waymo or w/e.
But in terms of policy decisions, whether they should be allowed on the road, these are the correct numbers. Convincing people like you to ride is not the goal of this paper or its claims.
If you rigorously enforced those limits on regular drivers and additionally imposed an interlock device to prevent DUI and remote insurance company monitoring of driving habits-- that's the comparison population I want to see. Otherwise, this is a press release disguised as an academic paper.