Tesla has revealed its AI product/self-driving roadmap for the next few months, but it raises more questions than it answers.
Tesla official AI Roadmap with self-driving software releases
September 2024
- v12.5.2 with ~3x improved miles between necessary interventions
- v12.5.2 on AI3 computer (unified models for AI3 and AI4)
- Actually Smart Summon
- Cybertruck Autopark
- Eye-tracking with sunglasses
- End-to-End network on highway
- Cybertruck FSD
October 2024
- Unpark, Park and Reverse in FSD
- v13 with ~6x improved miles between necessary interventions
Q1 2025
- FSD in Europe (pending regulatory approval)
- FSD in China (pending regulatory approval)
Electrek’s Take
The main problem is improvement against what? Tesla has always refused to release FSD intervention data so we wouldn’t even know if v12.5.2 can achieve a “3x improvement in miles between necessary intervention.”
As we previously reported, we do have some crowdsourced data, and Elon Musk acknowledged it positively, so we might as well use that since Tesla refuses to release official data.
For v12.4, the CEO said that it will come without steering wheel nag and it will be able to drive “5 to 10x more miles per intervention“.
v12.4 never made it to a wide release and v12.5 was supposed to go wide release last month and it only made it to about 5% of the FSD fleet the last time I checked.
Now, instead of a 5 to 10x increase on v12.4, which was a flop, he is now talking about a “~3x improvement” this month with the release of v12.5.
Let’s compare the data so far. Based on the crowdsourced data, the average of the v12.5.1 versions is 32 miles between disengagement and 128 miles between critical disengagement.
This compares to 30 miles between disengagement and 189 miles between critical disengagement for v12.3.6, which is the last FSD version that went into a wide release earlier this year:
Keep in mind that to achieve Tesla’s promise of an unsupervised self-driving system, it would likely need to be at between 50,000 and 100,000 miles between critical disengagement, aka 390x over the current data.
Based on Elon’s new timeline and compared to this data, we should be at around ~400 miles between “necessary interventions ” by the end of the month.
The term “necessary interventions” is also problematic and raises questions. If a critical disengagement implies a serious crash would occur without a driver behind the wheel, then you are off by a factor of at least 10.
The crowdsourced data is also trying to distinguish “critical” disengagements and it makes sense since many disengagements have nothing to do with the level of safety in FSD. Some disengagement are due to comfort and frustration on the driver’s part.
With v12.5.1, a lot of the disengagement have to do with navigation errors and speed. The former, Tesla will certainly need to fix. The latter is more problematic as it actually influence the data.
Many FSD drivers complain that the latest update is drives extremely slow. This should help with the performance and reduce critical disengagement, but it will not help overall disengagement as many drivers just grow frustrated, myself included, and take control of the vehicle to start driving at more reasonable speeds.
So yes, as per my headline, this roadmap brings more questions than it answers mainly due to the fact that Tesla refuses to release intervention data and yet, it uses increases in the metric to track progress.
Author: Fred Lambert
Source: Electrek
A critical disengagement every 50k-100k miles would mean a serious crash every year for a typical taxi application. This would need to be in the 500k to 1M mile range in my opinion, or even higher.So Tesla is off by a factor of at least 3900X.