Tesla’s Autonomy Day came and went, and investors and critics felt unimpressed by the ambitious presentation. We’ve all heard this story before; self-driving cars have been on the horizon for years. For many, the idea is pure fantasy, a misunderstanding of how difficult it is to search large datasets and engineer efficient intelligent machines. Driving isn’t easy, and it can take years to become reflexively familiar with all its challenges. Hydroplane, traffic cones, wild animals, road rage, drunks, and pedestrians all present horrific daily scenarios for victims and drivers, with more than a million traffic-related deaths globally each year.
As far back as 2013 is when Elon Musk first began toying with the notion of adding the mystic technology to Tesla Motors’s fleet. In particular, he mentions airplane autopilot systems as perhaps a good example of what is possible. Aeronautical use of pilotless steering has been around since 1912 when the Sperry Corporation created hydraulic connections to altitude monitors and heading indicators. A practical innovation considering flying through wide open air. Driving, however, is something altogether different.
Autopilot or otherwise, for more than half a decade Elon has stated his goal of making Tesla the safest of all self-driving vehicles. There was even an early 2015 tweet intimating autonomous cars were a foregone conclusion, potentially outlawing rebellious Sunday morning joyrides. This kind of language can understandably rile people up. Plausible explanations should accompany big claims. Given that Tesla’s present stock price hasn’t changed much since those words were said, the Silicon Valley CEO’s wild-eyed ideas are not selling investors.
Some of the confusion brewing between angsty traders and Tesla’s engineering team may lie in the rapid speed of progress building up in the machine learning field. Machine vision is one area of this new science experiencing breakthroughs. From Hubel and Wiesel’s work on the visual cortex of cats in 1959 to Fukushima’s Neurocognitron in 1980, academics have been trying to mimic the brain’s neurobiological methods for filtering light. But obstacles standing in the way of highly accurate visual recognition have always been data size and search efficiency. The smaller the discrepancy between any two similar objects, the larger the dataset needed to make those distinctions; and the larger the data, the faster your search efficiency must be to make reasonable requests. The white whale has always been how to achieve the best of both worlds: insanely accurate recognition paired with ridiculously fast results.
On stage with Musk during Autonomy Day was Tesla’s lead engineer Andrej Karpathy, tasked with solving the object recognition problem. Born in Slovakia, raised in Toronto, and a Stanford Ph.D. graduate, his fast-talking demeanor has made him a hero in the machine learning community; crafting deepnet theories and bare metal optimizations while setting the gold standard for vastly superior machine-learned image recognition, far surpassing any one person’s naturally gifted abilities. Additionally, his Stanford lecture series on the very techniques mentioned has been an invaluable resource for students and self-learners alike, providing nearly twenty hours of world-class instruction in the very innovations he discovered. What makes the course standout is the level of detail, the challenges, the solutions, from basic computational data structures to the most abstracted applications. Start to finish this training material has made Karpathy a legend amongst his peers, something most people only find at their local bar for the wrong reasons.
Karpathy’s path to Tesla no doubt began with his work at OpenAI, a research organization Elon founded with Sam Altman. In June of 2017, he joined the team and started to retool Tesla’s autonomous driving program from the ground up. His work over the last few years was on display at the April meeting, demoing the very strategies that earned him his field status. Path prediction, object detection, and long-tail accuracy flooded the slides, which assuredly glazed the eyes of many observers. Of all the insights, their detection confidence score was critical: 99.9999…%, the very number achieved by Karpathy’s previous work in improving the accuracy of machine-aided detection, open to the viewing public, setting the stage for Tesla’s future.
Before the presentation ended, Elon revealed his most forward-looking plan yet, to build a network of fully autonomous cars branded as RoboTaxi. Rated for mileage costs one-third that of gasoline car ownership, and one-tenth of ride-sharing options Uber and Lyft, the stakes are now clear. Aided with Karpathy’s machine vision object detection, every Tesla vehicle will now have the capability of identifying millions of distinct and similar objects and situations. Running deer, garbage cans, construction workers, cars swerving into lanes, and other innumerable variations in all their colors, shapes, sizes, and movements, whether during day or night, sleet or snow, rain or fog. RoboTaxi is a network of continuously adapting cars, reducing their object and path detection error rates for all the above conditions. As the fleet gets larger, so does its learning capacity, and thus its utility, and hence its accident rate begins to approach zero. The designs Tesla has for a world in which all driven miles happen across a self-learning network of machine-driven carriages is not a matter of calculating insurance; where Tesla is going, insurance is no longer needed.
Tesla was never about creating a fancy status symbol for wealthy customers, or even personal vehicles that afforded unique experiences. Tesla, from the beginning, has always been about creating a transportation monopoly using Palo Alto principles. One person who understands this culture more than anyone is Musk’s friend, investing partner, and Paypal co-mafioso, Peter Thiel. In 2014, Thiel published a book along with coauthor Blake Masters, titled Zero to One. The book was based off Thiel’s CS183 class at Stanford a few years prior. The tenets of the message reveal the mentality of Elon and his Silicon Valley cohort, a sensibility much different than anywhere else. For them, business, markets, are not just about products, recognition, giving-it-a-go, or having a day-job; fundamentally, it’s about escaping competition altogether—in other words, discovering monopolies.
In CS183 class notes made by Masters, and later recorded in the book, published online excerpts make plain this ethos. Much rather than putting yourself through the wringer of demoralizing perfect competition, where hundreds of thousands of industry colleagues fight over scraps, it’s best to figure out where to quickly achieve monopoly dynamics before anyone else. Sometimes the market already exists, other times you have to invent it, as Tesla is now doing with RoboTaxi. What is important is that you know your endgame, using scarce resources and scaling up to where the network effects become dominant, making it virtually impossible for competitors to break-in. These dynamics are now in play for Tesla, all the conditions put into place, the efficiency challenges with machine vision, the scaling of batteries, and also automated production. With the final piece yet to be set.
The media response to Tesla’s revealing Autonomy presentation was poor. Many who are not aware of the pursued winner-take-all strategy, or the technical efficiencies only recently discovered in machine vision, are disappointed in the two-year track of challenges and mishaps that have impacted the company, and indeed Musk himself. Most of the critiques are fair but rely wholly on impatience. Duke professor Mary Cummings and NYU’s Gary Marcus are often cited as experts — as indeed they are—but whose criticisms nearly always address the precision problem, mentioned before involving data and efficient search. Other flak is, of course, ankle biting from burned commentators, jealous rivals, and narrative-driven psychosis plaguing our social media-driven world — automated bots, attack pieces, clickbait, all clouding the frame. So far, Tesla has ridden out the storm, proudly announcing to investors production improvements, expanded product lines, and all the other elements of a vertically integrative expansion into every transportation ecosystem niche.
Over the next eighteen months, Tesla will have collected orders of magnitude more data, with an increasing curve of object detection accuracy, building millions of predictive models that will form their testable theory around safe autonomous driving. In the meantime, opponents are betting on all kinds of other possible solutions using much smaller data sets, none with the network effects required to reach escape velocity. One has to imagine the level of political and regulatory scrutiny Tesla will receive from the automotive and transportation industry once they painfully realize their capital stock is at risk, or at least subject to market-share deterioration. The intensity of competitive destruction will not be pretty. Automakers, dealerships, financiers, insurers, all seeing evaporating revenues as Tesla’s self-driving borg absorbs their customer base. And it will happen rapidly, as network dynamics often show. Once the reaction takes place, Tesla will have reached the necessary speed carrying it to its final destination; summoning not a ride, but a monopoly.