Day 11: Eleven Pipers Piping | FSD Or Bust
Part of
12 Days of Christmas - Tesla Edition a series (c) by the Artful Dodger, Dec 2023
Over this Yuletide season, I will post a daily installment focusing on Tesla products, past, present, and future (please note that I will express
major themes as short-hand bullet points as I
roll tide. Here's the series so far:
Day 01: A Partridge in a Pear Tree | Roadster Proof of Concept
Day 02: 2 Turtle Doves | S/X Fraternal Twins go Mainstream
Day 03: 3 French Hens | Model 3 Bets the Company
Day 04: 4 Calling Birds | Model Y Built at Four Factories
Day 05: Five Golden Rings | Semi Breaks Physiks
Day 06: Six geese a-laying | Megapack To Excel
Day 07: Seven Swans a-Swimming | Cybertourdeforce
Day 08: Eight Maids A-Milking | Model 2 World Car
Day 09: Nine Ladies Dancing | Dojo as a Service[/HEADING]
Day 10: Ten Lords A-Leaping | Teslabot FTW
Intro to Part 11: Eleven Pipers Piping | FSD Or Bust
Back in Oct 2016, Tesla parted ways with its 'driver-assist' vendor, MobilEye. Some of the reason was their unwillingnes to give Tesla access to their code, but also because Elon was convinced Tesla could do
ADAS better and faster in-house. The
FSD computer project was started around this time, with Tesla designing its own custom silicon optimized as an AI inference computer (with low-power draw and over 20x the compute provided by the MobilEye chip). Would this be enough? Well, only one way to find out: enter FSD beta.
1. Creeping Scope, Hidden Dragons
- The core problem with writing software to drive a car is that people don't think about how they actually drive, they just do it intuitively. That's also why dads are lousy driver instructors because they can't articulate verbally (forebrain) what is esentially a hindbrain function (how do you catch a fly ball, dad? Well, lets go practice, little Timmy...)
- Meet Elon Musk, the boy who never met a problem he couldn't solve. Elon is naturally optimistic, so he approached this problem in his usual way, that is to specify the minimum viable system that might work, build that, and see how well it performs
- Turns out, early progress was quick (aided by some 1st-principles avoidance of dead-ends like LIDAR), but soon enough a series of 'local maxima' appeared along the road to autonomy, and lengthy rewrites were the only way to continue making measurable progress in self-driving skill
- This methodology lead us to FSD beta v11.x the current wide-spread branch of this effort, which is statistically safer than average U.S. drivers, but lacks certain human-like traits in terms of smoothness and planning ahead
- the current architecture is somewhat like having a troop of Jr. Woodchucks operating your car; one working the pedals, another steering, a 3rd one watching for stuff crossing in front of the car (No. 4 randomly triggers those wipers...) But all these kids don't neccessarily talk to each other to reach a group concensus before yanking on that wheel or jamming on those brakes: Adult supervision is still required!
- FSD beta v12 holds the potential to finally get past all these hand-written heuristics and compound Neural Nets (NN). The new architectue is to create an end-to-end NN which in takes photons from sensors and outputs control commands to the car. So, instead of woriking like a virtual knot-hole gang of raucous 11 yr olds, I hope that v12 debuts more like a 16-yr old only-child
Lesson 1: Make it So, Number 1
2. Slaying the Beast
- Tesla is following the modern trend in computing science as expressed in Moore's Law: CPU cycles are cheap!
- For example, Tesla has replaced 2,000 human labelers with a single program called "Auto-labeler", which annotates video data (3D object recognition) coming from the fleet, and prepares it for use in the NN training stage
- Eventually, most human-written code will be in the data centre, primarily focused on managing workflow and automating the creation of NN deliverables. Tesla already has a small, highly productive Autopilot team, now they are adding massive amounts of more training compute
- Most importantly, Management will not give up on the dream of autonomy. Its central to the future value of Tesla, and will fundamentally change the way people purchase transportation in the future. The implication for 'dumb' ICE cars is stark, even the total fleet size will change as a result of autonomy.
Lesson 2: It's always harder than you thought (until it becomes automatic)
3. How Good is Good Enough?
- Eventually, even v12 and its siblings will reach a point of dimishing returns, where large amounts of extra work produce diminishing improvements
- ideally, this point is reached long after FSD is already multiple times safer than human drivers, though there will always be finger pointing at individual incidents even though safety is increased as a whole
- finally, it comes down to the cognitive style: Inductive reasoning ie: learn from example vs Deductive reasoning - applying a set of general principles
- Governments will always want to write laws and have the police interpret and enforce them. This is a classic application of deductive reasoning abilities, which may not be achievable before the advent of AGI
- But Tesla FSD doesn't have to enforce the laws, only abide by them. An experienced Driving Instructor would be a good hire for the AP team
- Tesla FSD will never be accident free (you can't control other cars), but this style of NN learning-by-example from copious amounts of data will make FSD the best that's possible with this approach
- BTW this is how most humans learned to drive. People hold valid drivers licences but didn't pass the Turing Test (or a touring test). FSD in its current form has the potential to become much better than the average driver, w/o needing to be Einstein
Lesson 3: Run What Ya Brung
Conclusion:
The FSD development project has been a long and winding road, full of washed-out bridges and dangerous curves. Still, it the skill and determination of the Tesla AP team, coupled with steadfast backing from Tesla Management / Governance, which will see them though to ultimate success. To paraphase JFK, '
We do these things not because they are easy, but because we thought they'd be easy...' Well we know better now, but we're not bitter.
Forward! Excelsior!
Next up: Coming Soon to Your Town: Robotaxi -
Void where Prohibited
Tomorrow's Topic:
Day 12: Twelve Drummers Drumming | Robotaxi at Last