Autonomous Cars: The Level 5 Fallacy
Courtesy: Popular MechanicsA counterpoint to last week’s Alphabet/Google/Waymo SD Car winner-take-all thesis. Instead of a moonshot, we’ll see messy, helpful increments. It’s natural selection at work.
Last week, I “explained” how Google was set to become the Microsoft of Self Driving (SD) cars. Machine learning leadership, sophisticated simulations and a Central Valley test site, immense computational and financial resources…the Google/Waymo endeavor has secured an insuperable advantage. You’re a car manufacturer: What choice do have but to pay a premium for a Google SD license? You could go with cheaper, less advanced technology…and say goodbye to your company after the first ‘mistake’ on the road.
At $1,000 a car, the SD license will be Google’s Mother Lode 2.0, the dream of any tech company looking for a second growth wave.
I concluded the article by promising to temper my enthusiasm and to offer a counterpoint, starting with a Horace Dediu quote:
“Those who predict the future we call futurists.
Those who know when the future will happen we call billionaires.”
It’s one thing to predict, as Gordon Moore did in 1965, that semiconductors would double their computing power every 18 months (Moore’s Law, seemingly slower of late). It’s something else to know to invest in Apple’s 1980 IPO (+30,797%) or, even better, Microsoft’s 1986 stock market launch (+73,293%).
It feels good to predict the emergence of SD cars: It has to happen because it’d be cool if it did. But when? Betting your company on even an approximate estimate of “Level Five: Full Automation” is a risk that you don’t want to take.
As a refresher from last week, the Six Stages of Automation are…
Level 5 means going from point A to point B with a fraction, say 1/10th, of today’s accident rate. No ifs, no buts, no steering wheel.
It’s a great vision, but one that’s not likely to happen any time soon.
(Please, no ad hominem incivilities on Tweeter or elsewhere. I’m an unrepentant technophile, I’d love for true SD cars to happen, and soon. Born in 1944, I have to contemplate a future of lesser mobility where SD vehicles large and small would make a big difference for a large and growing segment of our population.)
In prior Monday Notes that discussed electric and autonomous cars, a subject of endless fascination, I evoked scenarios where SD cars can’t cope with circumstances that require human intervention. Today, I’ll offer the pedestrian crossing at the intersection of Hayes and Octavia in San Francisco:
Understandably, the Google Street View picture was taken in the early morning. Now, imagine the 1 pm Sunday scene with crowded sidewalks and sticky car traffic. In today’s world, pedestrians and drivers manage a peaceful if hiccuping coexistence. Through eye contact, nods, hand signals, and, yes, courteous restraint, pedestrians decide to sometimes forfeit their right-of-way and let a few cars come through. On the whole, drivers are equally patient and polite (an unceasing subject of amazement for Parisians walking the streets of San Francisco).
Can we “algorithmicize” eye contact and stuttering restraint? Can an SD car acknowledge a pedestrian’s nod, or negotiate “turning rights” with a conventional vehicle?
No, we can’t. And we don’t appear to have a path to overcome such “mundane” challenges.
But you don’t have to believe me, or think I’m not “with it”. We can listen to Chris Urmson, Google’s Director of Self-Driving Cars from 2013 to late 2016 (he had joined the team in 2009). In a SXSW talk in early 2016, Urmson gives a sobering yet helpful vision of the project’s future, summarized by Lee Gomes in an IEEE Spectrum article [as always, edits and emphasis mine]:
“Not only might it take much longer to arrive than the company has ever indicated — as long as 30 years, said Urmson — but the early commercial versions might well be limited to certain geographies and weather conditions. Self-driving cars are much easier to engineer for sunny weather and wide-open roads, and Urmson suggested the cars might be sold for those markets first.”
“[Recode’s Kara Swisher] So to finish up, give me the timeline. When is this going to be like done?
[Urmson] It’s never going to be done.
[Swisher] No I know, but like right now cars are done. There’s no more horses around.
[Urmson] Oh the transition? I think it’s going to take at least 30 years.
[Swisher] 30 years.
[Urmson] At least.”
Last week, we saw how the engineers who are in charge of Tesla’s self-driving technology keep leaving the company, quickly followed out the door by their replacements. Apparently, they disagree with Elon Musk’s overly enthusiastic representation of the future of Tesla’s SD technology. This is more telling than it might seem. Not about Musk’s enthusiasm — it has worked well for him so far — but about the engineers’ views of the SD timeline. A two-to-three year engineering timeline isn’t unusual; five years is considered longterm. Beyond the five-year horizon? No thanks, I’ll switch to a more spiritually and financially rewarding pursuit. We’ll leave the worthy but nebulous commitments to Carnegie Mellon and Stanford.
… we’re looking at the messy, lengthy, 30 year transition predicted by SD car expert Urmson.
(After leaving Google, Urmson started his own SD technology company, Aurora. Among the other domain experts that have joined him is Sterling Anderson, the former head of Tesla’s SD engineering.)
Instead of the pure, straight-to-Level 5 moonshot, we’ll see a progression of incremental improvements, percentages gained, more miles of roads successfully (and unsuccessfully) navigated. And we’ll be treated to vociferous arguments not unlike what we saw and keep seeing in PCs, smartphones, and other tech battlefields.
For example, commutes will become increasingly automated. Smooth, driverless sailing on highways and boulevards, a shrill warning when the human touch is necessary on gridlocked city streets or on the narrow, winding roads in the hills above Silicon Valley. If — and it’s a big if — these semi-SD cars can learn from their drivers’ actions, we might have a believable picture of a fully-SD future, and a justification for expanded investment in SD technology.
The messy “30-year transition”, the many uncertain steps in sensor and software engineering, the poorly understood problems of coexistence between conventional and SD cars leave much room for competitors large and small. SD cars are a much more complicated challenge than the PCs Microsoft helped standardize.
