Tesla’s second annual AI Day is coming up on September 30th. While the event is primarily geared toward recruiting talent for AI, software and chip development (confirmed by Elon on Twitter), it serves a second purpose of outlining the company’s progress and long-term vision on the topic.
AI Day II is already a win for Tesla. As investors, we believe that AI Day will be combination of hard-to-understand technical jargon along with optimistic comments from Elon regarding the timeline of FSD. In the end, many—potentially including ourselves—may leave the event thinking there was little new added to the story. That takeaway misses the point: Autonomy is coming, and no other automaker has made the progress that Tesla has in advancing its potential. Eventually, Tesla will get it right and FSD will ship. And, once again, the company will be years ahead of its competition.
Where is FSD today?
The promise of Level 5 FSD has been on the table since 2018, when Musk said that FSD would be “safer than humans,” enable drivers “to fall asleep” at the wheel and deliver passengers “without intervention.” Musk also claims that, eventually, the cost of FSD will be “less than the subsidized value of a bus ticket.” While we believe this will be true in the future, today, the software is still at the L2 conditional automation, a light year gap to L5 and costs $15k.
We estimate that there are about 100k FSD betas on the road since its launch in October 2020. These beta testers are part of our estimated 400k FSD owners total, with 75% of those based in the US. This implies about a 20% adoption rate life to date.
While there is significant work to be done in order to ramp the current FSD beta to Full Self-Driving, Tesla just released an update (10.69.1) that includes improved speed while entering highways, improved driving smoothness, better prediction of the trajectory and speed of objects, better navigation around forking lanes and reduced false slowdowns near crosswalks.
The March of Nines
We believe that the latest updates underscore how even some of FSD’s basic driving behaviors still need refinement. This begs the question: How reliable does the technology need to be before lawmakers fully support it? We believe it comes down to the “March of Nines” theory, suggesting that an autonomous vehicle needs to be right 99.9999% of the time—or, involved in a single fatal crash for every 85 million miles driven—for it to be as safe as a human driver.
This theory is weighted by the importance of the corner cases, which are extremely infrequent events on the road. An example of a corner case is a highway sign falling down in front of a moving car.
The belief in the March of Nines tempers the willingness of lawmakers to pave the way for FSD. But, AI Day is a chance for Tesla to present why such cases are less relevant and that improving the software will save lives.
Getting geeky, the workings of FSD
For starters, it’s about the data. Since 2015, the company has been collecting actual miles driven data and, as of December 2018, collected 1B miles. By 2020, we estimate that Tesla hit 2B miles of data collected compared to Waymo’s 6.1m miles reported the same year.
When Tesla says its “collects” data, it means it’s capturing gateway log files (seatbelt, AP, cruise-control, speed settings, brake usage, steering input), stored in the vehicle’s event data recorder (EDR), and data records of FSD/AP enabled vehicles. Video is submitted to Tesla when a corner case is observed. The data is stored on a small 8GB chip, according to Spectrum IEEE, and sent back to Tesla over mobile data.
One reason why Tesla’s approach to solving FSD is different than other companies (like Waymo that uses Lidar for its L4 software) is that Tesla’s 7nm D1 chip, Dojo, learns primarily by camera images. Dojo takes in various inputs including singular snapshots of a scene, a bird’s eye view and vector space (the conversion of 2D images into a digitized 3D space) to conduct its AI training. Dojo then makes predictions based upon the movement of objects in the 3D image and self-checks its predictions. This process is repeated, again and again, for the purpose of learning pattern recognition and improving prediction accuracy. Then, Dojo can implement actions such as braking at a crosswalk for pedestrians.
In order for FSD to go mainstream, additional step functions in Dojo’s efficiency will be required. While the company has been clear that all Teslas made today will not require a hardware update for full autonomy, we are betting that additional hardware upgrades will be needed in the future.
In terms of privacy
While the data collected by Tesla is said to replace VIN numbers with temporary, generic IDs, it is possible to temporarily disable collection of personal data (Software > Data > Sharing) on the Tesla dashboard. The only way to fully deactivate Connectivity is to contact Tesla.
A chance to address its critics
AI Day is also a chance for the company to address FSD critics, including:
- NHTSA: Currently assessing the performance of Autopilot in nearly 830K Tesla vehicles to better understand “the role that the Cabin Camera plays in the enforcement of driver engagement/attentiveness.” Last month, the organization issued a statement that the technology is far from reliable and that “no one should risk their life, or the life of anyone else, to test the performance of vehicle technology.”
- Ralph Nader: Called the company’s release of FSD “one of the most dangerous and irresponsible actions by a car company in decades.”
- The Dawn Project: Implied that FSD makes 1000X more critical errors than human drivers (which is a a difficult claim to prove, in our opinion.) Tesla has since filed a cease and desist letter to founder Dan O’Dowd for “disseminating defamatory information to the public regarding the capabilities of Tesla’s FSD technology.”
The Toy Department: Optimus
It’s worth noting that Musk hinted we may have a working Optimus prototype by September 30. While the odds of seeing a working prototype at the event are low, we view Tesla outlining its long-term obstacles related to the humanoid opportunity as time well spent given the size of its addressable market. In a nutshell, that market would be defined as physical labor. The initial industry that Optimus would likely disrupt is US manufacturing jobs that account for about 10% of labor, or about $500B in annual wages. The global market for physical labor is many times larger than US manufacturing labor. It’s a massive market, even bigger than the $2.5T in annual global car sales. We will continue with more about Optimus after AI day.