Tesla AI FSD Will Vastly Improve in November

by

Tesla AI has updated its FSD (full self driving) achievements. There will be huge progress towards nearing human level driving. Tesla expects to surpass human level driving in mid-2025 (6-9 months).

Solving robotaxi and full self driving is key to Tesla’s future. Version 13.X should get them from 12% FSD adoption (% of people who buy a Tesla who get FSD) in the US to 25% adoption worldwide (China, Europe etc…) it will be smooth and comfortable. Everyone else is at 12.1 or v 11 or worse. Everyone other company self driving system is 600X worse than Version 13.

Tesla is transition to superhuman driving. This is insanely hard for Tesla. Tesla FSD barely works for version 3 hardware. It really needs hardware 4. Every other car maker needs version 4 hardware or better installed on many cars to replicate and they need an AI training cluster.

Going from V 12.1 to version 13 needed 2 billion miles of training data. Waymo and Apollo go and Xpeng have started to follow. They are at best at 12.1 level and most experts have said the Chinese self driving systems are at Tesla FSD version 11 level. Tesla used 500,000 cars driving an average 4000 miles using FSD in 2024 to improve the AI. If another company had 5000 cars training then they would need 400,000 miles driven. This would take 4 years of cars permanently on the road with human taxi driver. 100,000 cars driving 20,000 miles or 50,000 cars driving 40,000 miles. This means that Tesla is a minimum of 4-6 years ahead of any possible competitor to get to where Tesla will be in November. Tesla will improve 1000 times in miles driven per intervention in 2025. Catching up to Tesla’s constant improvement will be impossible if Tesla licenses to car makers building over half of the cars in the world.

Tesla FSD should go global in Q1 2025 with a high adoption rate. Tesla could reach 5 million cars with real life usage of FSD by mid 2025. This would be for highway and city driving (aka all driving). This means about 1000 miles per month per car. This would mean 5 billion miles per month of training data. This is going from 1 billion miles per month by December 2024 to about 2-3 billion miles per month in April 2025 and then 5 billion miles per month in July 2025. This could be 40 billion FSD driving miles in 2025.

A 5-6X improvement in interventions would be greater than a 3X leap from a new teenage human driver and an experienced human driver.

Tesla has integrated several of these improvements and are already seeing a 4x increase in miles between necessary interventions compared to v12.5.4. This is likely a 12.5.6 or version 13.0 is already showing a 4X increase in miles per intervention.

50,000 FSD customers have end to end highway for FSD. This is FSD that replaces the Autopilot that was already ten times safer than human driving.

FSD version 13 is releasing to Tesla employees and a FSD Version 13.3 will go wide to AI4( Hardware 4 users around the end of November. Tesla FSD Version 13 appars be a day later than the end of October (today) with a November 1st or 2nd release.

There is actually smart summon that will to to European and Asian customers. This means driverless operation within line of sight to a parked Tesla.

Tesla FSD Improvement List
– Full rollout of end-to-end highway driving to all AI4 users, targeted for early next week, including enhancements in stop smoothness, less annoying bad weather notifications, and other safety improvements
– Improved v12.5.x models for AI3 city driving
– Actually Smart Summon release to Europe, China and other regions of the world
– v13 is a package of following major technology upgrades:
– 36 Hz, full-resolution AI4 video inputs
– Native AI4 inputs and neural network architectures
– 3x model size scaling
– 3x model context length scaling
– 4.2x data scaling
– 5x training compute scaling (enabled by the Cortex training cluster)
– Much improved reward predictions for collision avoidance, following traffic controls, navigation, etc.
– Efficient representation of maps and navigation inputs
– Audio inputs for better handling of emergency vehicles
– Redesigned controller for smoother, more accurate tracking
– Integrated unpark, reverse, and park capabilities
– Support for destination options including pulling over, parking in a spot, driveway, or garage
– Improved camera cleaning and handling of camera occlusions

9 thoughts on “Tesla AI FSD Will Vastly Improve in November”

  1. Touchwood, it looks like there may have been a significant improvement in latest 12.4.2 12.5.2 and 12.6.1 updates, with only a single critical disengagement in 945miles of testing recorded on open source FSD community tracker:
    https://teslafsdtracker.com/

    That would be a significant improvement after performance plateau of last 1-2 years, and might mean only 2-3 orders of magnitude improvement still needed

  2. I don’t understand the chart. It says that going from 12.3.6 to 12.5 is a 30x improvement in miles between disengagements. But teslafsdtracker.com says it didn’t improve at all over that period, and actually got slightly worse.

    I wonder which is correct. If it got 30x better, then version 13 ought to be amazing. But if it got slightly worse, then version 13 might be roughly the same.

    The two charts are showing slightly different things (all disengagements vs only critical disengagements). But I would still expect about the same improvement either way. Not 30x vs less than 1x.

    So I really would like to understand how 12.3.6 and 12.5 actually compare.

  3. “at or above human driving levels” is a pretty low bar. From watching Florida drivers, the typical driver makes 4-5 law violations per trip. Illegal turns, weaving in traffic, speeding, no turn signals, etc.

  4. In EU, there will be no “x years ahead of competitors”. The car industry lobby is powerful enough to stop Tesla FSD by regulation until the European car makers have useful systems to deploy.

  5. Seems like the next release is always the big FSD moment, but then it’s not. City and Highway driving are not aka “all driving”. How about winding dirt roads with trees and culverts. How about towing a trailer? As long as one can turn all it all off, at least it won’t do harm.

  6. One of the hardest challenges for FSD seem to be single bar gates as often used in Corporate and commercial car parks. I can imagine many robotaxi journeys start after dropping the owner at a work or a city center car park. If the car then crashes through the security gate that is a big issue. and if through a level crossing potentially fatal The bar is easily confused with road markings and cannot be detected by parallax easily. This would be easier to detect with radar, but Tesla has no radar. It will be interesting to follow Tesla’s roadmap to fix this.
    https://www.youtube.com/watch?v=Q3V82HZbVX4&pp=ygUhZnVsbCBzbGVmIGRyaXZpbmcgc2luZ2xlIGJhciBnYXRl

      • This is the problem. FSD works MOST of the time, but it can fail spectacularly, with not even an attempt to brake to avoid the worst end result of a crash. And it can do so in bright daylight conditions too, for reasons that no human driver would ever anticipate, which means the human driver will be too complacent to retake control in time.
        I’d also like to see more actual crash stats to compare, on a similar mile to mile basis. Long empty stretches on highways are unimpressive, especially if you live in NYC like I do. Using just miles alone, the Apollo missions would be the safest method of travel ever. After all, no one died (aside from the test firing in Apollo 1. Oops….). But we know intuitively that’s not true.
        Let us know when insurance companies start dropping rates for FSD cars compared to human drivers. That’ll be the best real-world case for FSD.

Comments are closed.