Some thoughts on Tesla’s Autopilot & Robotaxi strategy:
1) I believe Tesla has a unique & superior strategy for solving Robotaxis which I categorise as the Intelligence/Data heavy approach, based on the assumption that Robotaxis are a very difficult intelligence problem to solve.
2) This is opposed to the approach of the rest of the industry which I would categorise as hardware heavy, Intelligence/data light which is based on the assumption that self driving is a relatively easy intelligence problem. I’ll explain this more later.
3) Currently I still put a small probability on Tesla solving Robotaxis in the near to medium term, but given the huge cashflow this business is almost certain to generate if solved, Tesla’s Robotaxi potential has very high present value even with low probability of success.
4) It's also worth noting that even without ever solving Robotaxis, Tesla’s current (and constantly improving) Autopilot driver assistance product alone is a hugely valuable and differentiating feature and is able to significantly reduce accidents and driving fatigue.
5) While my probability of success for Robotaxis is still low, It is clear to me that Tesla has the best strategy & I think this year Tesla’s move to video based “4D” neural nets (finally utilising its in-house neural net chip) may significantly increase my probability estimate.
6) So what exactly is Tesla’s Robotaxi strategy?

Central to Tesla’s strategy is the assumption: Self driving cars is a very difficult intelligence problem that needs a lot of data to solve.
7) Key here is to understand: When we teach a robot how to drive, in reality we are competing with a human brain structure that has already been developed by millions of years of data and reinforcement learning via natural selection...
8) .. and then 20+ years of training via Unsupervised Learning, Supervised Learning and Reinforcement Learning via interaction with the world and teachers.
9) All of this has trained humans to have an instinct for the laws of physics, the identification of objects and the predictions of their future actions and motion.
On top of this a human then gets ~ 1000 specific hours of training in actual driving lessons.
10) So, driving is very hard and requires a huge amount of training, learning and intelligence to understand the universe and predict objects paths and interactions. It can’t simply be learnt from a handful of hours of driving combined with expensive sensors.
11) Tesla's Robotaxi strategy is then built from the assumptions:
A: When solving Robotaxis, prediction is the hard problem – it requires a lot of intelligence and experience. Detection is easier.
B. Hence our strategy should be optimised to solving prediction, not detection.
12) C. To solve prediction of other objects paths we need billions of miles of real world driving experience to train the cars to understand the universe it is interacting with and to be able to solve enough edge cases to beat human driving accuracy by 2-10x.
13) D) We can't get billions of miles of driving experience without a hardware suite affordable to install in a consumer owned car.
E)Lidar won't be cheap enough to install in a high volume consumer car in a reasonable timeframe (& also has issues with rain & lidar interference)
14) F). Therefore we cannot use Lidar, even though it is a shortcut to solving the detection problem because it can easily measure distance and velocity of objects with hardware rather than software & intelligence.
15) G). Hence we have to solve distance and velocity estimates using machine learning with camera and radar data rather than lidar. If we can do this, lidar has no extra value anyway as its capabilities will only be a subset of what we can already do with computer vision.
16) H) If we have a large enough fleet it can experience enough different driving scenarios to collect enough data to train enough intelligence to solve both the detection and the prediction problems.
17) There is a false narrative spun by competitors that Elon's rational for choosing Tesla's camera first approach was just to save on manufacturing costs and/or an arbitrary rejection of lidar. This is completely false.
18) There are two mutually exclusive choices for self driving strategy: A) Data/Intelligence Heavy, Hardware Light or B) Hardware Heavy, Data/Intelligence Light.
19) It is impossible to choose a Hardware Heavy, Data Heavy approach because building a company owned R&D test car fleet of Tesla's size using this approach would cost the company towards $100bn capex and development costs.
20) You choose Tesla’s data heavy approach if you think driving is a very difficult intelligence problem and needs billions of miles of real world data to solve.
21) You choose Waymo's data light approach if you think driving is an easy intelligence problem - you only need tens of millions of miles of data (from a small test fleet) - most edge cases can be solved by developers thinking up & simulating driving edge cases from their desk.
22) So, I think Tesla has the best strategy for solving self driving cars.
I’ll follow with a thread on how Tesla may progress from its current Autopilot software along the march of 9s to full Robotaxis, & why the day of reckoning could be the video based architecture due in Q4.
You can follow @ReflexFunds.
Tip: mention @twtextapp on a Twitter thread with the keyword “unroll” to get a link to it.

Latest Threads Unrolled: