Car Science #9: stupidly intelligent
Hello! You're reading Car Science; please consider subscribing, it's free and helps me.
Hey,
One thing I've largely steered clear of in Car Science, which is occupied with things that have factually been discovered, is autonomous vehicles. Because they haven't.
There's lots to get into, from the fact no one should be able to call something full-self driving when it isn't (and especially not sell it to the public to use on roads with other drivers) to the fact there's a lot of flat-out charlatanism and self-interest in promoting AI as the future of driving. And it's beginning to fall apart, after years of being the latest thing companies are throwing money at.
Here's a thing about selling cars: it doesn't make very much money. Just ask Stellantis CEO Carlos Tavares, who I'm legitimately obsessed with as a man who opened his tenure at the newly-formed car brand megazord by saying making EVs was stupid and annoying but he guessed they would have to do it because of How Things Are and has now moved on to the fact making petrol cars to Euro 7 standards is also stupid and annoying. Basically: making cars is stupid and annoying and he's right! The only thing that makes any dang money these days is luxury SUVs.
So how were car brands planning to survive? Well, aside from bringing out at least 5 luxury SUVs and crossovers every year until we all die from the effects of climate change, the overall plan for awhile now has been to become software companies. Just this week the new Chief Digital Officer at the beleaguered Jaguar Landrover, Anthony Battle, described their product as "laptops on wheels" that deliver "digital experiences." Which predictably upset a lot of car people but on the other hand the automotive press gets at least 2 articles a week about how VW's infotainment doesn't pair well with Apple Carplay.
Until I reviewed a car I had never heard of Apple Carplay because I own a 1993 Renault Twingo but one of the things I got wrong, when examining the BMW i4 and iX for flaws, was neither checking out the cup holder or whether it would pair well to Apple Carplay. People's entire experience of a car can be decided by whether or not it plays well with Apple Carplay. And, to be fair: I get it. Having something that's supposed to work to play music and podcasts and just doesn't and then affects the GPS and every time you get out to get another 9oz coffee for the cupholder you have to re-pair it will rapidly make even the best driving experience incredibly annoying; the truth is, people want their cars to do one thing, above all and that's just work.
In a digital world, that means just working with your phone. Which means you shouldn't even be aware that there's a process to connect to Apple Carplay. And so does all the other stuff like driver aide parking cameras and proximity warning sensors and everything else cars are packed full of that, honestly, is a good thing - they're safety features that avoid errors and can potentially save lives, which especially as cars get bigger and heavier every year is incredibly important.
Back to AI: selling AI solutions was how quite a lot of OEMs sold the future. While there was still quite a lot of dithering going on about EVs, AI was a sort of stop-gap, post-petrol concept. But it's turning out not to be so simple. We were supposed to have genuinely autonomous taxi services by now but we, well, don't and to even call the tech available currently Level 2 or Level 3 of autonomy (which still requires a human able to intervene at any time) is often flattering it. Truth is, making a cruise control made autonomous driving look within reach when it was still hundreds of miles off and the cost to get there is so vast it's not realistic, at a time when the industry is already facing enormous flux. Better to focus that programming resource on Carplay.
Last October the sinisterly-branded Argo AI commenced wind down, despite backing from Ford and VW, after it admitted that the project was just too far off achievable. Ford and VW's plan had been to make money off selling autonomous software but well, it was going to take an awful lot more money to get there. And a lot of computing power. And a lot of testing. And so far they'd really only made a slightly dorky version of the ID Buzz.
To be fair, people do love vans with hats.
Anyway, whew, long intro there but I promise there is a relevant bit of science to talk about here, which is that MIT researchers have been trying to work out what the cost of a predominantly autonomous fleet of cars, globally, would be. Not in monetary terms, although there is a huge one connected to this but how much energy all that computing power would take up.
The results are: a hell of a lot. You can read a full write-up of their research here - Computers that power self-driving cars could be a huge driver of global carbon emissions
The research modelled that the computing power alone needed for autonomous driving systems would be equivalent to 0.3% of global carbon emissions, similar to data centres currently. That's a drop in the ocean when you consider that transport is 72% of CO globally and passenger cars (yes, passenger cars) account for a really hefty chunk of transport emissions. 2020 was an odd year but 41% is still more than four times aviation's contribution, in terms of transport's overall tally.
So adding 0.3%, especially given that this is thinking about a theoretical future when hopefully vehicles will be zero emissions at tailpipe isn't really a big crime overall. Except when you think about how that would effect individual vehicle efficiency over a journey. Adding the equivalent of all the world's data centres to the energy needed for passenger cars is really quite a big demand and not just in terms of what resources the cars are using to run but all the ones involved in making the systems, from copper for wiring to chips and processors and cooling systems.
It's also just... a lot of data. From the research: "For example, if an autonomous vehicle has 10 deep neural networks processing images from 10 cameras, and that vehicle drives for one hour a day, it will make 21.6 million inferences each day. One billion vehicles would make 21.6 quadrillion inferences. To put that into perspective, all of Facebook’s data centers worldwide make a few trillion inferences each day. (1 quadrillion is 1,000 trillion)."
Logic says that's going to be stored somewhere - to review the AIs performance, to sort out liability in crashes, to process and improve systems, maybe just as a legal requirement for autonomous driving. So in fact it's gonna be creating massive data centres at the same time as consuming equivalent power to the ones we have.
Modelling scenarios in 2050 on the basis of hardware we have now is obviously like comparing an original Playstation to the latest iPhone but what the study says is that we're not just a long way from AI in terms of programming, we actually can't build it yet because we just don't have the parts to make it make any sense.
Years ago I interviewed Bryn Balcombe, who at the time was the chief technology officer at Roborace and ended up asking him how on earth they'd made an LMP3 car heavier while installing autonomous capabilities, compared to the safety systems for a human driver. The answer, boringly, was wiring for all the sensors - it was 2017 and of course things have probably since moved on but when even adding that to a car intended for no passengers is creating a weight and power demand it's hard to see how it could make sense in any kind of near future for, say, a family car.
Which isn't to say we're not somehow inches away from a breakthrough because nothing exists until someone does it. But for now it's hard to see how autonomy can be realised even in the practical and capacity terms, before the programming challenges.
I honestly see it as a positive that the car industry is, at least temporarily, giving up on this one. Driver aides are great but we need a lot more honesty about how far away cars are from being able to make decisions of their own, not follow patterns they've been taught, which is very different.
Anyway, something else next week (and I promise I will do one next week)
Hazel
x