Duration 15 years (02022-02037)
Michael P Dierickx
will go to GiveWell Maximum Impact Fund if Bouchard wins,
or Médecins Sans Frontières if Dierickx wins.
As I write this, Tesla has around 60,000 people testing the Beta 10.9 version of its "Full Self-Driving" software. Even adjusting for a roughly 1:10 ratio between the hours driven by a volunteer test driver versus a professional, this far exceeds not only Tesla's largest competitor, Waymo, but also significantly exceeds all of its U.S.-based competitors combined. A truism in deep learning is that model size and quantity, quality, and diversity of training data are two of the fundamental inputs to a system's performance. Tesla is the unparalleled leader with regard to training data and it is expending significant effort and money on onboard compute to keep its model sizes increasing. From the first principles of deep learning, it stands to reason that if anyone truly achieves SAE Level 4 autonomy within 15 years (barring any loopholes that match the letter but not the spirit of the definition of L4 autonomy, which I have tried to preclude in the wording of my prediction), it will be Tesla.
(1) Data suggest most FSD owners do not turn the system very often, and what little time they do turn it on is on the freeway.
(2) Training data are not magic, and they cannot make things appear magically. Just as stone bridge over a creek cannot be "machine-learned" into the Golden Gate Bridge, an ADAS system cannot be ML'ed into a true L4 system
(3) Tesla is not currently geo-fenced and its approach is not based on local road knowledge, by way of my bridge analogy. This is a fundamentally ADAS approach. As such, the assertion that 90% of public roads (by length, let's say) will be autonomously-driveable by Tesla is extremely aggressive as Tesla has not even begun to consider which roads can and cannot be driven autonomously.
This is a hard problem and I believe no company or entity will reach the milestone of "SAE Level 4 autonomy on over 90% of public roads in the contiguous United States" within fifteen years, to say nothing of whether or not it's Tesla or whether or not there are 1 million cars.
• Tesla would need at least 1 million cars on the road.
• No other no other entity will have reached this milestone before Tesla
• These cars would be SAE Level 4 Capable (as defined by: https://www.sae.org/standards/content/j3016_202104/) for the full US cross country journey between 1652 Maiden Hair Ct, South Lake Tahoe, CA 96150, USA and 46 Klein Rd, Fort Kent, ME 04743 during all normal weather conditions that leave the road open and passable in a standard car.
• SAE Level 4 will be considered safer than human drivers
• This capability will be usable by the commercial public (i.e. at least tens of millions of consumers will have the option to pay for and use SAE Level 4).