WORTH A LOOK #9 “Tesla’s Full Self Driving Re-Boot”

Worth a Look

March 21, 2024

Tesla’s Full Self Driving Re-Boot

As a long-time Tesla follower it has been fascinating to watch the company’s remarkable journey and progress towards solving the “impossible task” of replacing human driving with software.

Creating a mass-market Full Self Driving (FSD) operating system that is safer and more efficient than human driving would unlock a new hyperbolic growth phase for Tesla. 

The prize for the dominate FSD provider of the future – that can be regulated in the west and licensed by auto manufacturers – would be earnings power that is multiples larger than the auto industry as we know it today.

Tesla’s Progress – FSD v12.3

Tesla recently released its newly overhauled FSD version 12.3 to the public. This was not just an update to the old system, but a completely new approach re-built from the ground up. 

Even for the Tesla-haters, being aware of this FSD progress is definitely Worth a Look. 

In the 4th episode of their new podcast BG2Pod, hosts Bill Gurley and Brad Gerstner do an excellent job of breaking down the significance of FSD v12.3, this history of FSD technology, and how radical this platform shift is for Tesla.  

Click here for podcast link

Bill Gurley is considered a Godfather of venture investing and, until now, has been an outspoken critic of full self-driving.

Brad Gerstner is the founder of the tech-focused hedge fund Altimeter Capital. Last quarter Tesla appeared for the first time in Altimeter’s quarterly filings as a new position.

 

Summary of my notes:

 

A completely new framework

  • Past argument against FSD was that getting to work better than humans [at scale] would take many years, reason being that it is impossible to code for “corner cases” on the road and this is where wreaks happen. FSD v12 seems to solve for this.
  • 1 year ago, Tesla completely changed their FSD framework from a deterministic model based on C++ code to what they call an end-to-end model based on imitation learning. 
  • FSD v12.3 comes after 11 previous software versions. The market is skeptical of how different this recent version can be. FSD v12.3 acts more like a human than ever before.
  • Bill Gurly is a known skeptic of FSD but says Tesla’s new approach has the best shot at working than any other approach he has seen.

How is v12.3 different? 

  • Radical decision to completely scrap the past approach and start over.
  • Tesla’s prior approach to FSD was to try and code every object, circumstance, and combination of every drive. This became an inefficient rat’s nest of code as it grew. 
  • The old approach has been tossed out and replaced with a neural network model where videos from drivers are uploaded every day along with the driver reactions to video from five car cameras (breaks, steering wheel, etc.).
  • Bill Gurley refers to Occam’s razer theory – a simpler approach is likely to be the optimal approach.
  • Tesla’s new approach has a much better chance of going all the way by being more reasonable, scalable, accurate, elegant. Requires uploading a tremendous amount of video.
  • The model does not have a deterministic view of what is a stoplight or road sign. For example, you would have to see and then label every stoplight, first job identify that you are at a stoplight. 
  • In the old model, Tesla would code in C++ to say, “when you are at this stoplight, do this”. This would never solve all “corner cases”.
  • The new approach has NO CODE, does not know there is a stoplight ahead, instead it observes the driver’s behavior which becomes the label. “When we see pixels like X this is how the model should behave”
  • Prior versions were just slightly better patchwork models.
  • Rate of improvement on FSD 12.3 is 5x-10x better per model over the prior system.
  • AI models for FSD 12.3 are generic open-source models that are customized, Tesla has been working on these for the last decade.
  • NVIDA chips are used for training.
  • Engineers who were writing code are now focused on data infrastructure, making sure that the data pulled off the edge makes the new models better. 
  • FSD engineering functions at Tesla now focused on data management vs coding a patchwork of updates. The model is digesting an enormous amount of incoming data. 

Why was this not possible before?

  • 5 years ago this approach was not possible. 
  • All the infrastructure pieces are now in place. The models have been around for a while but new Nvidia hardware, ability to download and process 10+ gigabit of video per car every night, scale of cars on the road was not there before. 
  • 5M cars on the road, 30 miles per day of video data going back 10 years.
  • Data has to be processed on the edge, 99% of data Tesla collects never makes it back to Tesla, most gets processed on the car itself, they are looking for the outlier moments to finetune the model, then re-uploading the model to the car. This is why you get these exponential moments of improvements that we are seeing now.

Where is the competition?  

  • Who else has the capacity to do this? [no one so far]
  • Is the neural network approach the right answer? If so, then there is no one today that can compete today. Feels like a step function where FSD can improve at a much faster rate.
  • Tesla’s approach requires millions of connected cars on the road to capture the long-tail events, the events that only occur a few times because otherwise you don’t have a statistically relevant pool.
  • It’s not about producing a quantum of data, it’s about the quality of the data that can only come from a larger footprint.
  • To compete you need to control the design of the car, not just provide the software (mobile eye, QUALCOMM), Tesla has the car in your garage at night and uploading automatically, massive infrastructure. 
  • BYD could possibly compete but then would need to have permission to operate an FSD operating system in the US.
  • You could combine the clusters of every hyper-scaler in the world and not have this much data, you couldn’t possibly store all of this data, that’s the size of this challenge. 
  • Waymo still using old processes, 30-40 cars on the road.
  • Tesla has taught the car the specific moments it should record, example: anytime there is an abrupt movement, brake for example, will record, can weight these reactions heavier than other actions. 
  • Tesla’s advantage is in solving for the “corner case scenarios” – new system allows Tesla to only capture these situations [vs worthless data]. 
  • Data Tesla needed to get started was almost impossible to process, millions of cars, now in a position to capture the more severe, less frequent moments because of scale/footprint. 
  • Least likely competition is Cruse/Waymo because they don’t have enough cars on the road, and these cars cost $150k, math doesn’t work, can’t build the footprint.

FSD potential impact on Tesla financials

  • Tesla unit economics ~$2.5k per car, 7% penetration of FSD (FSD v11) paid $12k for FSD.
  • What if we cut the software cost in half to $500 per month? [or less?]
  • If FSD penetration went from 7% to 20% then the contribution margin is about the same even if Tesla charged half as much for software. 
  • 50% penetration of FSD would generate billions in incremental EBITDA. 
  • Tesla will focus on FSD penetration, get people trying the product, play around with price.

________________________________

What is Worth a Look?

As part of our research process we are continuously reading, watching, and listening to a wide range of content. Worth A Look is a summary of a selected piece of content that our followers might find interesting from the perspective of explaining an important transition happening within in a sector or stock we cover, or wisdom from other investors or thought leaders we are learning from.

 

About JDP: JDP is a private investment partnership that makes long-term investments in public companies with significant unrealized potential.  www.jdpcap.com

Disclaimer
This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other advice. Nothing contained in this document constitutes a solicitation, recommendation, endorsement, or offer by JDP or any third-party service provider to buy or sell any securities or other financial instruments in any jurisdiction. This document is intended only for investors in JDPI, LP and its affiliated entities, and prospective investors who have attested to qualifying as an “accredited investor”, also referred to as a “professional” or “qualified” investor based on the country of domicile. Funds managed by JDP might maintain positions in companies discussed in this document.