With FSD V9, Tesla Is Becoming An AI Robotics Company !!
Summary
Tesla is close to releasing Full Self-Driving Version 9 or FSD v9, the latest version of its AI-assisted city driving software.
FSD v9 is the result of significant progress by Tesla AI’s engineers and will itself accelerate progress by scaling up Tesla’s fleet data collection.
AI-assisted driving is not just a nice add-on for a fancy car; it’s a paradigm shift in the experience of driving and our idea of what a car is.
As this becomes more apparent to analysts, the consensus narrative on Tesla will shift and the stock price will increase something like two-fold within the next two years.
Tesla motors showroom with cars and illuminated logo branding at dusk London UK
AdrianHancu/iStock Editorial via Getty Images
Tesla (NASDAQ:TSLA) is becoming an AI robotics company. This is integral to any defensible valuation of the company. Yet so many equity analysts have either denied this fact or hand-waved it away by simply declaring it outside the scope of their analysis. This is an error of tectonic scale. I believe this error results in Tesla stock being priced at something on the order of 50% of what it ought to be.
What is an AI robotics company? It’s a company that uses software-controlled machines, i.e. robots, and trains those robots to perceive and act using the AI techniques of the contemporary decade: deep learning and deep reinforcement learning. Here is a brief overview of these techniques and their significance for the unacquainted. For those who already know, you can skip the next three paragraphs.
Contemporary AI, explained
In 2012, researchers at the University of Toronto published their design for a deep neural network called AlexNet. When AlexNet won a computer vision competition called the ImageNet Challenge, it ignited the field of computer science with interest in deep neural networks. Thus began the deep learning revolution.
The next year, the London AI research company DeepMind showed that deep neural networks could be used to store the implicit knowledge required to play Atari games skillfully. This implicit knowledge was acquired through an updated version of an old AI technique called reinforcement learning, which in its essence is the equivalent of trial and error. Deep reinforcement learning was born.
It can’t be overstated how different the field of AI is in 2021 compared to 2011 thanks to these discoveries. The possibilities for the creation of intelligent machines expanded from kludgey and fragile programs coded by hand to quasi-mysterious webs of mock neurons from which knowledge and intelligent behaviour emerge organically from a process of learning, often astonishing their creators. Before contemporary AI, when you wanted a robot to do something, you had to program it to do it. With deep neural networks, researchers and engineers can train AIs to do things that would be impossible to program — even things that would be impossible for humans to imagine how to do. Go players now study the moves of the AI AlphaGo to expand their notion of how the game can be played.
Close-up of a Go game board.
Photo by Alexandre Lecocq
Back to Tesla
While deep neural networks have had many dazzling successes living disembodied inside computers, they have yet to truly cross over into the physical realm of robotics. One of the very few AI robotics companies in existence is Covariant AI, which uses deep learning to train warehouse robots to pick up and sort the endless variety of physical objects that people buy. It is, in fact, Covariant AI from which I am borrowing the term “AI robotics”. Besides Covariant and Tesla, the only AI robotics companies I can think of with large-scale commercial operations are their competitors: Amazon (NASDAQ:AMZN) with its more opaque and mysterious AI robotics efforts and Mobileye (NASDAQ:INTC) with its decidedly slower and less ambitious effort to turn cars into AI robotics. Why are so few companies applying contemporary AI to robotics? Why is Tesla different from other car companies?
Tesla benefits from a culture and ethos of what I would call “mad scientist thinking”: from the C-level to the interns, researchers and engineers are encouraged to explore ideas that sound crazy if they can convincingly argue for them on pure scientific and engineering grounds. Kurt Kelty, Tesla’s former Senior Director of Battery Technology, described Tesla’s approach to R&D as “crazy” when describing the company’s partnership with Panasonic (OTCPK:PCRFY):
They’ve got a set way of working that really works well for them and then we come in and it’s just oddball, crazy ideas… the combination of the very conservative and just crazy works very well for the two companies.
Companies like GM (NYSE:GM), which owns the AI vehicle startup Cruise Automation, could dive head-first into AI robotics just as Tesla has but for lack of ambition and vision, both at the managerial level and among GM shareholders. It’s the company led by Elon Musk, the rocket man, who has drawn a legion of mad scientists to work for it and who has inspired a class of true believer shareholders who will put their money behind Tesla’s mad science.
How Tesla got here
Elon Musk deplores technological conservatism. He has an ideological conviction in science, engineering, and technology so fundamental to his worldview that, as a founder-CEO and a leader, he has a willingness to accept immense cost and risk when he believes it is necessary to cause radical change. Musk has also historically set up corporate governance to protect his vision from activist shareholders who might want to divert his companies onto a less radical course. He has at times been combative both privately, within the company, and publicly, with critics and competitors, who aren’t on board with his vision. Tesla’s uniqueness as a car company is in large part due to Musk’s influence from the company’s first prototypes to now, as well as the culture of the tech industry in Silicon Valley, where Tesla grew up. In Tesla we find a highly unusual cultural artifact when held up against the likes of GM, whose key minds spend more time thinking about dividends than deep learning.
Starting in 2015 and 2016, Musk began recruiting for Tesla’s AI division. Around the same time, Musk recruited an initially secret team to develop on-board computing hardware for Teslas that are specialized to run neural networks. Full Self-Driving Beta Version 9, or FSD v9, is the product of those seeds planted all those years ago.
Tesla (i.e. Musk and whatever other executives might have had a say) made two critical decisions around AI robotics that were publicly announced in 2016 and 2017. In October 2016, Tesla announced that it was upgrading the production design of its cars for full autonomy. Mainly, this meant the inclusion of eight HD cameras covering 360 degrees and a higher-powered, replaceable computer for running neural networks. The hardware change, along with Tesla’s ultimate ambitions, also necessitated that Tesla bring its AI software entirely in-house.
The implementation of the new hardware and software meant that, by mid-2017, the Tesla AI team had the ability to selectively query Tesla’s entire production fleet for sensor data and other data they thought would be useful for training their neural networks. AI director Andrej Karpathy said in July 2017:
The Tesla fleet is like a large, distributed, mobile data center. Each machine is attached to a big battery, a person, and moves very fast.
The decision to hire Karpathy as Director of AI in June 2017 was the second critical decision made at this time. Karpathy was a deep learning researcher from academia, corporate labs, and the (at the time) non-profit OpenAI. By putting an AI research scientist at the head of Tesla AI and, moreover, one who shared Musk’s conviction in the power of computer vision, Musk steered Tesla AI away from technological conservatism and towards technological radicalism. A more traditional corporate boss, even one from the software world, would not likely have been a practictioner and proponent of mad scientist thinking, as Karpathy is.
Silver Tesla Model 3 in fall.
Photo by Jteder
Why FSD v9 is a significant milestone
FSD v9 is the first (seemingly) reasonably mature production implementation of a collection of new ideas developed both within and outside of Tesla. One theme of the changes is the migration of more hand-coded elements of the software to neural networks. Another theme is the transition away from using two-dimensional still images as the atomic unit of data that human labellers annotate to train Tesla’s computer vision neural networks. Instead, annotators directly label 3D models of the driving scene. Moreover, the new computer vision neural networks make perceptual judgments by comparing sequential moments in time. With the additional dimension of time, this makes it a transition from 2D vision to 4D vision.
FSD v9 also incorporates the ongoing and increasing implementation of:
data curation: the art and science of devising large, diverse datasets that cause neural networks to tease out all the correct subtleties of the visual world, which Tesla’s large fleet facilitates uniquely well (e.g. Tesla can find many examples of cars without tires and tires unattached to cars to teach the neural net that the presence of one doesn’t always imply the presence of another)
multi-task learning, which turns perception into a global process wherein, for example, the probability that a patch of pixels is a vehicle up ahead is not computed separately from the probability that the same region of pixels is road surface
automatic labelling or weakly supervised learning, in which cues from human drivers create labelled data organically (e.g. manual human driving effectively labels the ideal trajectories for turning left through complex intersections)
self-supervised learning, in which data automatically labels other data (e.g. a neural network predicts what lane lines it will see once the car turns a corner and then the correct prediction appears in the form of the line lines that another neural network subsequently perceives)
using neural networks for action, not just perception, either by copying humans (imitation learning), using human input such as a deactivation of the AI to determine when an error has occurred (deep reinforcement learning), or both
Arguably even more important than the content of FSD v9, however, is the promise it holds to expand neural network learning within Tesla’s production fleet. The software for driving on city streets is becoming mature enough that, in combination with Tesla’s recently productionized visual driver monitoring, it seems like with v9 or maybe v9.1 it may finally be safe enough to roll out city street driving (still as a SAE Level 2 system) on a large scale. This is what Musk says Tesla intends to do and there is supporting evidence such as the activation of visual driver monitoring in production cars and the roll out of new neural networks to the whole fleet.
To take Tesla’s city driving software to the scale of tens of thousands or hundreds of thousands of vehicles would be a major accelerant of progress. While plenty of important data can be collected passively, the most rich source of information is to have the AI activated and actuating the vehicle. In particular, this accelerates the identification of errors, since a supervising human can quickly take over when the robot’s behaviour is incorrect.
FSD v9, then, seems to be both the culmination of significant R&D progress at Tesla AI and a precursor to the acceleration of that progress.
Elon Musk talking to astronauts.
Photo by NASA photographer Joel Kowsky
Brass tacks
When I make the sort of arguments I have above, I mainly receive two forms of critical response:
Reject the premise: dispute either the empirical performance of Tesla’s FSD software and/or my assertions about the general principles of deep learning (primarily that training data is a key differentiator for neural network performance) and the usefulness of Tesla’s fleet in light of those principles.
Reject the financial relevance: assert that Level 4 autonomy and robotaxis are a distant dream and conclude, therefore, that even if the premise is true, it has no impact on Tesla’s core business of selling cars for the foreseeable future.
In many past essays, I have cited academic and industry sources on deep learning and extrapolated them to Tesla in order to defend my premise against common objections. A recent example can be found here. In this essay, I will focus on the second form of counter-argument: even if true, what I’m saying doesn’t matter from a valuation perspective.
The flaw in this counter-argument is that it overlooks the potential for a Level 2 city driving system to provide value to customers. Stress and cognitive fatigue are major drawbacks of manual driving. A Level 2 city driving system can relieve that. AI-assisted driving, or human-assisted AI driving, is a fundamentally different experience from fully manual driving.
One improvement many folks such as myself have been hoping for is a visual driver monitoring system to help ensure human attentiveness. Tesla has delivered that.
Another area of software development that will be important for AI-assisted driving is the centre screen visualizations of what the AI is seeing and planning. The ideal for the visualizations is for the human to be able to assess, at a glance, whether they need to stop the AI and take over. Even better would be if Tesla can keep the driver in the loop via audio cues or even verbal narration without drawing their eyes away from the road.
The better the robot can monitor the human and the better the human can monitor the robot, the more viable AI-assisted driving will become. Eventually, with enough refinements of the AI, the driver monitoring, and the information about the AI conveyed to the driver, AI-assisted will most likely become much more attractive than unassisted driving. What Tesla is currently trialing in a limited private beta is not merely a software feature that customers can add onto a car, but software that introduces a new paradigm for vehicular transportation that supplants the idea of the car and the idea of driving.
People will still use the word “car” rather than “robot”, just like we still use the “phone” when we mean “hand-held computer”. But, in the context of a Tesla, the meaning of the word “car” will indeed be changed to “human-assisted robot”.
With just a few exceptions, equity analysts model Tesla out to 2025 or 2030 with the assumption that it will be an auto company like other auto companies, selling cars for a one-time price comparable to other brands, at margins comparable to other brands. Only a few analysts have been bold enough to imagine that Tesla will become primarily a software subscription business or that FSD sales will make up the majority of its automotive revenue.
Yet what is the justification for this conservatism? My best guess is that most analysts have either simply not connected the dots or are afraid of looking foolish by publishing a bold model that is an outlier among their peers.
Over the last 24 months, we saw Tesla’s stock price increase more than ten-fold as equity analysts herded around a new consensus with regards to its core auto business, in terms of projected sales and margins. I predict that within the next 24 months, analysts will again move abruptly together en masse toward a yet newer consensus that centres software and AI robotics in financial models of Tesla in 2027 or 2032.
I don’t think a new analyst consensus around AI-assisted driving would justify another ten-fold increase in the stock price. But it will probably justify something like a two-fold increase in the stock price. This is what analysts Alex Potter and Winnie Dong of Piper Sandler (NYSE:PIPR) argue in their deep dive report on Tesla. The authors write:
By the mid-2030s, we expect Tesla’s revenue from FSD software to have exceeded Tesla’s revenue from selling vehicles; the impact on profit is even greater … We assume 85% gross margin, with opt-in rates eventually rising to >50% (despite $39k pricing)
Their forecast on FSD software is integral to their $1,200 price target, based on discounted future cash flows out to 2040 and a roughly 10% discount rate. The authors write, “It would be difficult (probably impossible) to articulate a ‘Buy’ thesis without Tesla’s FSD software”.
Piper Sandler’s is an outlier among price targets, but I predict that by the end of Q2 2023, it will be closer to the average. Tesla is set for another grand shift in analyst consensus and a corresponding repricing.
Robot arm holding a cup.
Photo by David Levêque
My simple back-of-the-napkin math is as follows:
$25,000 FSD price x 45% EBITDA margin on FSD x 70% take rate = $7,900 in EBITDA
$7,900 average EBITDA x 172,000 cars sold (excluding leases) = $1.36B in total EBITDA
$10.4B actual Q121 revenue / $1.36B hypothetical Q121 EBITDA = 13% EBITDA margin
Tesla’s actual EBITDA margin averaged over the last four quarters: 13%
Actual current EBITDA margin of 13% + hypothetical modelled EBITDA margin from FSD of 13% = 26%
26% / 13% = 2
The 45% EBITDA margin on FSD software sales is borrowed from the EBITDA margins of mature software companies like Norton (NASDAQ:NLOK). The 70% take rate is a personal guess, as is the $25,000 price. The other figures come from Tesla’s public disclosures.
$25,000 might seem steep at first blush, but as a monthly subscription over the average ownership period for a new car, 7 years, it’s only about $300/month. I believe a not-too-distant version of FSD, maybe v10 or v11 or v12, will provide enough value to customers to justify the subscription fee.
The sustainable doubling in Tesla’s EBITDA, which will scale with automotive revenue, will justify a doubling in Tesla’s long-term valuation. Once this coming change in EBITDA is foreseen by a critical mass of analysts, it will begin to get priced into Tesla’s stock price.
AI
