Doesn't this foolish monk not know … that vision alone in self driving vehicles makes for self-crashing vehicles? Susceptible to the literal "Wile E. Coyote" tactic of a road painted on a wall, for example TESLA. 🤡

@AlanTennant You could probably get a human to drive into a painted wall if you painted it just right. And the AI just has to be better than humans. So this is perfectly possible in principle. But lidar is getting cheap so I agree it's dumb not to use it at this point. And it's not like we stop caring about additional safety once we hit human-level. (See also my Musk vs McGurk post.)

And do @traders think we need a stipulation that it be safely deployed? I'm thinking it's hard for it to reach our threshold for "widely" if it's killing people more often than human drivers do. So maybe the resolution criteria can ignore the safety question.

What do @traders think of using 100M autonomous miles as the threshold for "widely deployed". Waymo passed that recently and it's roughly the number of miles between fatalities for human drivers. So if a self-driving system has a perfect safety record but much fewer than that many miles, we only have weak evidence that it has human-level driving ability, safety-wise.

Conclusion: we'll go with 100M autonomous miles as our threshold, with no fatalities that are the fault of the AI.

@dreev If Tesla robotaxis go 11M miles before they have first AI fault fatality and then to a cumulative 111M miles without a second AI fault fatality does this count?

If not, would this mean Tesla is forever banned from qualifying or is it specific to major version eg if v15 does first 20M cumulative miles then could v16 qualify but only when v16 reaches 100M miles (i.e. 120M cumulative miles)?

Do miles driven by robotaxis and FSD (unsupervised) if ever release to tesla owners get aggregated for the 100M miles test? I sort of expect there is some different software which for want of a better name I call a 'command and control shell' but these different front ends may well load and use the same driving software. How do you distinguish what counts as the same?

@ChristopherRandles The first thing you said sounds right to me. Clock resets on fatality and just wait for 100M miles, without worrying about version. We don't even necessarily have to worry about company. Like say Zoox drops their lidar and Tesla and Zoox combined hit 100M vision-only autonomous miles, it might be in the spirit of the question to count that. What do you think? (Not that it's likely to matter -- everyone but Tesla's all in on lidar, not to mention radar.)

@dreev Not sure about combining Tesla and someone else outside that group. Suppose Tesla did 50m miles the Zoox drops to vision only. If Zoox had a fatality after 11 million miles when Tesla had done 80 million miles and then Tesla does their remaining 20M miles without fatality.

Do you then say Tesla has done 100 million miles and that should count, or
do you have to wait for a combined 50+11+100=161M miles?

Is it combined miles only, or
Individual company/group only, or
either one of above counting?

Individual company/group only makes more sense to me (assuming they remain separate products) but your call.

If one group merged or bought out another and they produce a later combined group product then treating as one group all along may make sense? Hence my company/group wording.

@ChristopherRandles Good point. How about if we need 100M miles from one company, but we don't worry about software versions? The question about mergers is confusing me, but that one seems especially unlikely that this market would hinge on it. I'm fine defaulting to your suggestion that we declare success when one company can claim 100M miles, without worrying about whether all of those miles were under one company at the time.