> If an AI can react to terrible human drivers

That's not sufficient - nowhere near sufficient. If an AI can react properly when kids are playing in a 5mph zone with cars parked on either side of the street, before their ball bounces into the street. If it can react appropriately to seeing a young kid clearly struggling to ride a bike. If it can understand that the tennisball bouncing across the road will be followed shortly by an enthusiastic dog... et cetera, et cetera.

I don't need an AI to compensate for shitty driving (taking a roundabout the wrong way). I need it to drive so as to protect non-vehicles that it encounters.

A) also doesn't work unless all other road users (bikes, scooters, skateboards, pedestrians, etc.) are also banned from the roads. The other option is limiting the autonomous cars to tunnels or dedicated tracks, but then we're just reinventing trains again.

L4 is basically reinventing trains, with invisible tracks in the form of high-def maps. Even that is just so hard ;_;

Given how many people you see driving down the road staring at their phones, reinvention of trains is probably still a good thing for society.

Good point. On the other hand, if we could encourage people to use actual trains, that would be much better for the environment.

Why do you need A and B? Waymo is operating driverless taxis at Level 4 in Phoenix; what do you see as fundamentally missing (aside from expanding range) for Carmack to be able to win?

The gap between level 4 and level 5 is huge.

Basically, if you can find videos online of Waymo giving up and requesting backup for a situation not severe enough for humans to do the same, then you are not really at level 5 yet. If you cannot rely on the car to handle driving though any circumstance you could reasonably expect your hired chauffeur to handle, well then you are not level 5 yet. Examples of situations where humans give up are hurricanes, and blizzards.

And there are plenty of videos of it stopping for situations where no human would ever be equally confused, plus a good handful of videos of the cars doing absolutely unacceptable things resulting in near misses.

Are you really that confident that by 2030 Waymo's roadside assistance program will be disbanded, or at least reduced to a killswitch, summoning a replacement car and a tow truck (because obviously the car must be broken if assistance was called)? Because if they need more than that, it sure isn't level 5 yet.

Those taxis can’t even reliably drive in the rain yet, hence Arizona for the staging grounds. So they’re a far cry from meeting the standard, unless rain is now considered a natural disaster.

From what I'm reading, L5 would need to handle the annual meeting of the summer tires club... Assuming that's a problem in NY or Chicago.

I’ve driven in one. Felt like I was driving with a 16 year old driver. I’d never drive in the rain with it. Through intersection with faulty traffic lights, and police directing traffic? Nope. It never got onto the freeway, and it avoided complicated driving/merging decisions by driving through neighborhoods.

The claim is that you need A or B, not A and B (“Fully self-driving cars need one of two things:”)

You're assuming near perfect driving.

IMO, minimum requirement for level 5 driving is merely to match human driver accident rates.

> IMO, minimum requirement for level 5 driving is merely to match human driver accident rates.

Minimum requirement for nerds, maybe. Minimum requirements for ordinary folks will probably be more along the lines of "will be noticeably better than me on average" and "will not make any given mistake more frequently (or more severely) than I would".

i.e. I expect people a lot of people would want "strictly better than me on practically every axis" as their minimum requirement, not just "on par with me based on the average accident rate". People just aren't going to put up with seeing cars make dumb mistakes that they would never make, even if the accident rate ends up being good on average.

Not in the courts, no. If switching to self-driving reduces accident rates in 99% of situations, but in that last 1% someone gets killed, then there's a lawsuit waiting to happen.

Lawsuits happen with existing cars. Life goes on.

I thought an all self driving fleet today is statistically safer than humans

I worry about the ethical implication of a self-driving car (a slave AI smarter than it's occupant) being forced to go to McDonalds.