It’s pretty clear that autonomously driving robot cars will be everywhere sooner or later – the question is when.
I had my first ride in a robot cab! I’m in San Francisco this week. And then another experience yesterday, in a different way. Let me tell about them.
Day 1.
My friend and social technology expert Tom Coates took me in a Waymo robot cab up to Twin Peaks to see the city.
I mean - beautiful to an absurd degree. Downtown towers lit in the evening sun; a blimp floating over the Golden Gate Bridge; a rainbow joining the peak of the hill parks to the low bruised clouds.
My photos on Insta barely capture it. What you’ll also see there is a video of the steering wheel of the robot car moving on its own, as it drove us down the hill.
It wasn’t a great first ride, on the way up.
We approached a school bus and blocked the road. The Waymo didn’t leave room. While we were waiting, the school bus drive waved at me to get out of the way. I sat in the front passenger seat. I gesticulated at the empty driver’s seat –
It’s empty! (I tried to say by waving my arms.) I’m in a haunted car! I can’t tell the ghost what to do!
It reversed eventually.
Only to get frustrated at a slow cyclist, avoid them by turning left down a street which turned out to be closed to traffic, and then it panicked: the Waymo came to a halt in the middle of the lane, turned on its hazards, and then… nothing.
What happened next was reassuring. Tom called customer service and someone took over, turned us round, and we set off again. There were a couple of other erratic moments.
The ride back down from Twin Peaks had no sweaty palm moments at all, so I got to observe.
The dash of the car has a screen with a graphic that shows what the car can see: it’s a cartoon of the room, coloured blocks being other vehicles, and dots being people.
The way Waymo spots people is not magical. It sees the same people that I do. (So if a person vanished behind something, the dot disappears; no object permanence).
Only it sees much quicker, and all at once. On one particular intersection the car was aware of maybe 20 people. I had to look around to double-check I had seen all the same people.
Now, my mental model of attention is a pyramid. The peak is focus; I can focus on one object at once. Below that are things I’m attending to, maybe a half dozen. Below that I have awareness. Objects move up and down the pyramid, deliberately and automatically.
The Waymo has a broad base of awareness that is bigger than mine. It has more a great attentional bandwidth than humans. This is unusual.
So, I had time to have these observations. It felt very quickly very normal to be in this robot car with the empty driver’s seat.
(btw I suspect robot cars will have empty driver’s seats mandated by legislation for a long time; it’s a great visible symbol of the “mind” of the car.)
Day 2.
On the following day, a very different interaction with a robot car.
I was driving down the I-280 and had just gotten into enjoying the sun, the scenery and the music. Cruising at 60mph, a (human-driven) car crosses directly in-front of me, swerving across 2 lanes having almost missed their exit, simultaneously slowing to about 20mph.
I hit the brakes hard and luckily there was some distance to the next car behind me: I didn’t hit the erratic driver, and nobody hit me. Phew.
Later I saw a Waymo in one of the middle lanes at 60mph.
I don’t know if it was in autonomous mode, maybe it had a driver taking it back to base.
Which car would I prefer to be driving near to on the interstate, the human or the robot?
Well.
I can mentally model the robot car way better than other human drivers. I can “contain” my theory of mind of the Waymo.
The Waymo is consistent, safe and dumb. Faced with almost missing its exit, the Waymo would re-route. It wouldn’t swerve.
It’s wild how quickly I went from: oh this will be a white knuckle ride. To: yeah ok, this fits right in, actually I’d prefer to drive near these.
(Now the empty driver’s seat is also the removal of somebody’s job. I’m not thinking about that particular social impact here.)
What I’ve learnt this week is: it’s clear that this works.
At some point robot cars will be coming onto our roads as fast as they can manufacture them.
This works, whatever “this” is…
…because what “this” is is not clear. It’s not merely robot cars.
Like many things in Silicon Valley, the overall “this” is broad and obscure with scaffolding papered over with money and extraordinary effort. What is real?
- “This” could be robot cars PLUS many, many drivers in offshore call centres with Xbox controllers remotely piloting the car at any semi-tricky point. A wage arbitrage at best; economically unsustainable at worst.
- “This” could be robot cars PLUS a required millimetre-precision 3D model of the entire operating environment with wildly detailed statistics about expected environmental behaviour.
Whether or not our streets are full of robot cars in, say, 2027, with millions of the things rolling off the production lines depends on which “this” is the reality.
Looking at each…
Perhaps each robot car rests on the frequent, fractional effort of many, many remote human drivers?
If there are many drivers, the economics won’t work.
So the assumption is that the artificial, artificial intelligence can and will be replaced by actual AI. But can it? Is that moment 2 years away, 5 years away, 10 years away?
Is that moment dependent on collecting sufficient training data, or is it a matter of a fundamental breakthrough in the way AI works?
So, in that scenario, we’ll be waiting an unknown amount of time before rollout.
Perhaps the blocker is the ultra high resolution dynamic 3D model of the city?
Every road, every freeway, every place it might expect to see a person: a map like this is expensive to produce.
So ultimately this requires a financial structuring innovation.
A map is not low-cost, high margin software. Think of it as a large capital investment that requires maintenance, like a developer building a new city quarter. Then it produces a yield as people pay for it over many years.
This is not a venture-shaped investment. It’s more like private equity, or infrastructure. Someone will need to product the map and maintain it (like, half a billion dollars globally, say?) then rent it out.
Then the yield will need to be repackaged and sold to pension funds who require the rock-steady reliability.
I understand that financial structuring is how EV charging networks rolled out in the Nordics in Europe. This is half-remembered, so apologies, but I believe the breakthrough was to take the company building the EV charging network and de-merge it into two entities.
One, an entity that rents the land and owns the physical infra of the charging points, and buys and sells the electricity. This is a private equity shaped, capital intensive, yield operation.
The second entity owns the customer relationship. It’s mainly software. It innovates on brand, reaches customers with electric vehicles, innovates on pricing and bundles. It’s venture-shaped.
Separated, these two business types can flourish.
In the case of building the 3D operating environment model, it’s like building a chip fab or any other long-lived high-cost asset. The blocker for PE investment is sustainable demand for the map.
So demand can’t just come from Waymo running a taxi operation. There need to be many companies offering differentiated use cases: yes cabs and ride sharing, but also commercial fleets; autonomous driving as a feature in individually owned vehicles; last-mile delivery operations; and so on, all paying per-use.
So that will take a period of time to develop too.
Downstream questions: local legislation, and route to market and supply chain (like, will existing car companies license Waymo’s technology or assemble their own). I feel like the determining points are the ones above.
My takeaway.
“This” works.
Robot cars work, feel normal, and are even preferable to be in and to be near! There will be very quick cultural acceptance and there’s already cultural readiness. It’s a matter of time. I didn’t take that as a given; that was a surprise to me.
I would love to know what the actual blocker to immediate global rollout is. Those were my guesses.
It’s pretty clear that autonomously driving robot cars will be everywhere sooner or later – the question is when.
I had my first ride in a robot cab! I’m in San Francisco this week. And then another experience yesterday, in a different way. Let me tell about them.
Day 1.
My friend and social technology expert Tom Coates took me in a Waymo robot cab up to Twin Peaks to see the city.
I mean - beautiful to an absurd degree. Downtown towers lit in the evening sun; a blimp floating over the Golden Gate Bridge; a rainbow joining the peak of the hill parks to the low bruised clouds.
My photos on Insta barely capture it. What you’ll also see there is a video of the steering wheel of the robot car moving on its own, as it drove us down the hill.
It wasn’t a great first ride, on the way up.
We approached a school bus and blocked the road. The Waymo didn’t leave room. While we were waiting, the school bus drive waved at me to get out of the way. I sat in the front passenger seat. I gesticulated at the empty driver’s seat –
It’s empty! (I tried to say by waving my arms.) I’m in a haunted car! I can’t tell the ghost what to do!
It reversed eventually.
Only to get frustrated at a slow cyclist, avoid them by turning left down a street which turned out to be closed to traffic, and then it panicked: the Waymo came to a halt in the middle of the lane, turned on its hazards, and then… nothing.
What happened next was reassuring. Tom called customer service and someone took over, turned us round, and we set off again. There were a couple of other erratic moments.
The ride back down from Twin Peaks had no sweaty palm moments at all, so I got to observe.
The dash of the car has a screen with a graphic that shows what the car can see: it’s a cartoon of the room, coloured blocks being other vehicles, and dots being people.
The way Waymo spots people is not magical. It sees the same people that I do. (So if a person vanished behind something, the dot disappears; no object permanence).
Only it sees much quicker, and all at once. On one particular intersection the car was aware of maybe 20 people. I had to look around to double-check I had seen all the same people.
Now, my mental model of attention is a pyramid. The peak is focus; I can focus on one object at once. Below that are things I’m attending to, maybe a half dozen. Below that I have awareness. Objects move up and down the pyramid, deliberately and automatically.
The Waymo has a broad base of awareness that is bigger than mine. It has more a great attentional bandwidth than humans. This is unusual.
So, I had time to have these observations. It felt very quickly very normal to be in this robot car with the empty driver’s seat.
(btw I suspect robot cars will have empty driver’s seats mandated by legislation for a long time; it’s a great visible symbol of the “mind” of the car.)
Day 2.
On the following day, a very different interaction with a robot car.
I was driving down the I-280 and had just gotten into enjoying the sun, the scenery and the music. Cruising at 60mph, a (human-driven) car crosses directly in-front of me, swerving across 2 lanes having almost missed their exit, simultaneously slowing to about 20mph.
I hit the brakes hard and luckily there was some distance to the next car behind me: I didn’t hit the erratic driver, and nobody hit me. Phew.
Later I saw a Waymo in one of the middle lanes at 60mph.
I don’t know if it was in autonomous mode, maybe it had a driver taking it back to base.
Which car would I prefer to be driving near to on the interstate, the human or the robot?
Well.
I can mentally model the robot car way better than other human drivers. I can “contain” my theory of mind of the Waymo.
The Waymo is consistent, safe and dumb. Faced with almost missing its exit, the Waymo would re-route. It wouldn’t swerve.
It’s wild how quickly I went from: oh this will be a white knuckle ride. To: yeah ok, this fits right in, actually I’d prefer to drive near these.
(Now the empty driver’s seat is also the removal of somebody’s job. I’m not thinking about that particular social impact here.)
What I’ve learnt this week is: it’s clear that this works.
At some point robot cars will be coming onto our roads as fast as they can manufacture them.
This works, whatever “this” is…
…because what “this” is is not clear. It’s not merely robot cars.
Like many things in Silicon Valley, the overall “this” is broad and obscure with scaffolding papered over with money and extraordinary effort. What is real?
Whether or not our streets are full of robot cars in, say, 2027, with millions of the things rolling off the production lines depends on which “this” is the reality.
Looking at each…
Perhaps each robot car rests on the frequent, fractional effort of many, many remote human drivers?
If there are many drivers, the economics won’t work.
So the assumption is that the artificial, artificial intelligence can and will be replaced by actual AI. But can it? Is that moment 2 years away, 5 years away, 10 years away?
Is that moment dependent on collecting sufficient training data, or is it a matter of a fundamental breakthrough in the way AI works?
So, in that scenario, we’ll be waiting an unknown amount of time before rollout.
Perhaps the blocker is the ultra high resolution dynamic 3D model of the city?
Every road, every freeway, every place it might expect to see a person: a map like this is expensive to produce.
So ultimately this requires a financial structuring innovation.
A map is not low-cost, high margin software. Think of it as a large capital investment that requires maintenance, like a developer building a new city quarter. Then it produces a yield as people pay for it over many years.
This is not a venture-shaped investment. It’s more like private equity, or infrastructure. Someone will need to product the map and maintain it (like, half a billion dollars globally, say?) then rent it out.
Then the yield will need to be repackaged and sold to pension funds who require the rock-steady reliability.
I understand that financial structuring is how EV charging networks rolled out in the Nordics in Europe. This is half-remembered, so apologies, but I believe the breakthrough was to take the company building the EV charging network and de-merge it into two entities.
One, an entity that rents the land and owns the physical infra of the charging points, and buys and sells the electricity. This is a private equity shaped, capital intensive, yield operation.
The second entity owns the customer relationship. It’s mainly software. It innovates on brand, reaches customers with electric vehicles, innovates on pricing and bundles. It’s venture-shaped.
Separated, these two business types can flourish.
In the case of building the 3D operating environment model, it’s like building a chip fab or any other long-lived high-cost asset. The blocker for PE investment is sustainable demand for the map.
So demand can’t just come from Waymo running a taxi operation. There need to be many companies offering differentiated use cases: yes cabs and ride sharing, but also commercial fleets; autonomous driving as a feature in individually owned vehicles; last-mile delivery operations; and so on, all paying per-use.
So that will take a period of time to develop too.
Downstream questions: local legislation, and route to market and supply chain (like, will existing car companies license Waymo’s technology or assemble their own). I feel like the determining points are the ones above.
My takeaway.
“This” works.
Robot cars work, feel normal, and are even preferable to be in and to be near! There will be very quick cultural acceptance and there’s already cultural readiness. It’s a matter of time. I didn’t take that as a given; that was a surprise to me.
I would love to know what the actual blocker to immediate global rollout is. Those were my guesses.