It's my honor to announce that John Carmack and I have initiated a friendly bet of $10,000* to the 501(c)(3) charity of the winner’s choice:
(Disclosure, I work at Google but not on Waymo, and I don’t know anything that isn’t public)
Firstly, this is is a really fun bet. Kudos to both of you for participating. Secondly, I would take the Carmack side of the bet, if only because Waymo is already letting anyone book driverless cars in Tempe, AZ. Phoenix is a top 10 city in the US. That doesn’t seem like an insurmountable gap as they continue to expand the service area… Frankly I’m more concerned about how quickly this can become a sustainable business model.
Random youtube video from a year ago: Waymo FSD - Chandler to Tempe with Zero Interventions | JJRicks Rides With Waymo #23 - YouTube
Given that the statement of the bet says “…for passenger use in major cities” - plural - I doubt the intent here is that its terms would be satisfied by achieving level 5 in just one of these ten cities.
So there’s a question, if it’s possible in 9 out of the 10 top cities is that a win or a loss?
Referring to major cities in the context of SAE L5 automation seems redundant, at least my read of J3016 is that it defines L5 systems to be capable of driving anywhere a human would reasonably be expected to drive a comparable vehicle. Any kind of geofencing to a specific city or set of cities makes it automatically L4.
With that in mind, I am 100% with you. While I expect that L4 local driverless taxi and delivery services will continue to expand and we might even get consumer vehicles that offer full L4 capabilities on interstate highways, actual “any road, any time” L5 is a problem that I believe requires strong AI to solve. I’m not convinced we’ll ever get there, and if we do the idea of trapping such an AI in a car brings its own concerns both moral and practical.
It is up to John and Jeff, of course, but I would take a failure in any of the ten to show that we would not be at level 5 yet.
Is Level 5 the max? That’s aiming pretty low. There should be at least 5 levels above that! Level 5 is “driver’s permit” level. There should be “newly licensed 16 year old”, “can rent a car (abilities of human at age 25)”, “seasoned driver (20 years of experience)”, “stunt car driver vs demolition derby adversaries”, “that dude Bishop from Alien with the knife trick driving at least 100 mph through Manhattan rush hour without hitting anything”. Each of these has additional abilities.
But isn’t the goal of having a car drive like a human a red herring?
Computers can do stuff that humans find very difficult, such as staying between two white lines for hours on end, a pretty easy task. On the other hand, humans excel at more fuzzy logic (well, after gaining some experience).
My expectation is that roads will have to be adapted to suit autonomous cars, where cars can detect information and position themselves through electronics rather than looking at signs and lines on the road. And then some cities will go all in adapting their infrastructure to allow for driverless cars, and others won’t.
BTW, at least currently the number of vehicles on the roads is limited to the number of driver’s licenses. With fully driverless cars, big families may buy a car for every kid in the family. And have those cars drive back home empty after dropping of a passenger. Traffic congestion will be unbelievable.
Isn’t Waymo already doing this in my hometown (Chandler, AZ)? I often see driverless Waymos as I go about town…
Maybe you meant to say “included as a standard ability in a typical new family car purchase” instead of “available for passenger use”? I suspect those Waymo cars are very expensive to build: Lots of sensors, lots of power, and lots of trunk space needed for that on board supercomputer (I would suspect).
In the US, my feel is that Driverless trucks might come into wide use before driverless cars. Highways tend to have less unpredictable pedestrians, bicyclists, etc.
I also feel that China might be able to get driverless cars in wide use before we do (since they are still initially building some of their infrastructure). Do Chinese cities count as any part of this bet?
Good luck! I would take the same side of the bet as you.
Does either of your definitions of Level 5 include remote intervention? The formal definition doesn’t seem to make this clear. For example, the car may not have a wheel, pedals, or driver/assistant, but in case of a deadlock of some kind a remote operator can step in and make the necessary decisions to get the car moving again.
My understanding is that this is a part of the solution for existing autonomous taxi services, and it’s one reason why self-driving (if it’s ever available for private sale) is more likely to be a service with ongoing maintenance fees rather than an all-in sale.
Having to have dealt with an ice storm with 7"/~18 cm of snow on top, I wonder how driver-less vehicles would cope with ice-covered roads, black ice on bridge ramps and even just on curves. Road conditions change rapidly traveling down any highway. Is the AI in the vehicles going to check the morning weather on TV or radio? 
Driver-less vehicles may be well suited for many weather conditions, but I think winter driving sometimes includes a little ingenuity - especially when your vehicle can’t make a hill and begins to slid backwards and sideways, or takes off like a rocket down an ice covered road on a steep hill. A human driver may stop and decide it’s unsafe to attempt driving down that hill. What would an autonomous car do… “go for it”? 
This is kind of pedantic, but I think what you mean is a level 4 vehicle that can drive anywhere within one of the top 10 city in the US. Level 5 means that the vehicle can drive everywhere, so adding a condition to where the vehicle operate is kind of redundant. Like for instance, when Waymo launched in Chandler, they didn’t claim that it was a level 5 in Chandler.
There is also the issue what constitute driving within a city, like usually self-driving taxis operate within a subset of the streets (a geonet). So at which point is the subset large enough to constitute it “drive anywhere”? 95%? 99%?
The issue with inclement weather is interesting. Mercedes is in the process of launching the first privately available Level 3 system (“Drive Pilot”) but last I saw it was only cleared to operate at that level in basically perfect weather. If there’s rain, it will notify the driver to take over. To count as fully Level 5, does it have to be certified to operate in all conditions?
I didn’t read this as “the cars can only drive in mayor cities” but as the “service is available at least in mayor cities”.
So the way I imagine it is as a an Uber/Lyft like ride service, or maybe a Zip-Car like carshare, which use fully autonomous self-driving vehicles that work at SAE level 5. Now if such enterprise came to start in 2028, I could easily see that they wouldn’t be available in all the country, but they should be in the mayor cities by then (given that they only need to have a fleet out there, and even parking challenges are a lot easier with self-driving vehicles).
Now just as when Uber/Lyft started and they were only available in mayor cities (or how ZipCar is only available in mayor cities) the limit was, in theory at least, where I could start, not where I could go. This would be doubly so here in the scenario where Carmack wins the bet. That is I may not be able to call and get an L5 self-driving car to pick me up in Springfield Illinois and use it to drive around; but I could get a self-driving L5 car in Chicago, have it drive me to Springfield Illinois, and then drive around anywhere I want to, because the cars have no limit on where they can be.
In theory again, in practice, fuel and other considerations probably add an extra layer of challenge to defining what is a pass or a fail, but again that’s one of the reasons one would think this will take more than 8 years. From my understanding the SAE levels here do not include refueling or other behavior other than “knowing how to move around”, the cars may simply refuse to go beyond a range not due to not being able to drive, but due to needing to be able to go back for refueling at a specific fuel center (belonging to the company I guess). The core goal is to make people look and notice, and in 8 years celebrate that we did it, or see how close we are now.
Sadly another crash of a Tesla while on autopilot. It seems there’s a long way to go with getting the AI to recognize conditions ahead.
Jeff may just win his bet. 
That is the intent, just one is enough, but it has to be a “major” metro area, top 10 cc @MarkRansom @araybold
That would be an interesting outcome! Difficult to achieve in practice though? Wouldn’t this be like glorified cable cars?
Absolutely, I expect to see driverless highway trucks in our lifetimes without question.
If it’s human intervention, then it is human intervention, whether remote or not.
Everything except emergency conditions, yes. Not during a hurricane for example. That’s an unreasonable expectation. Nobody should be out in emergencies, or at least only emergency vehicles 
.
Yes, that’s my understanding as well. It needs to be available in at least one of the top 10 US metro areas in some form that you can pay for.
I am pretty sure this is going to happen by 2030. For major cities both Waymo and Tesla will be able to do it. For most of the US Tesla is going to take the lead and their FSD tech will be miles ahead of anybody else.
@codinghorror
What do you think would contribute to the difficulty of getting to that level? I think the obvious ones is the problem where (in example) the car may have to choose between killing two different human beings (in theory). AI bias is certainly going to be a contributing factor such as tesla cars thinking an orange moon is a traffic stop. I wonder if by 2030 if level 5 will still be possible if government bodies implement laws that prevent non-human intervention driving on the road (or maybe without some enormous tax or specialised insurance) which would make it impossible for all self driving cars. I’m sort of ranting here but do you have any particular reason you think it would be impossible?
Also do you think autonomous public transport become mainstreamed before or after cars? “mainstreamed” in the terms that it is far more popular compared to the alternative. While cars are already being developed for, it would be easier for companies such as uber to buy and implement it faster then it would to force citizens to buy autonomous cars but then again, public transports owned by governments often aren’t profitable as companies and the AI might have to be recreated for transport such as trains.
@meerfall
Have you ever witnessed a software development process? Jeff Atwood has; Carmack I don’t know about. But that is why I take Jeff’s side of this one. The future of AI is THERAC-25, full stop.
I’m a software developer so I should know what a software development process is! 
I think you misunderstood my comment as I am on Jeff’s side. We will never be 100% sure if the AI driving the cars is on any sae level, unless the code is independently reviewed.