4d ago

Tesla Reports Drop in Self-Driving Safety After Introducing “End-to-End Neural Networks”

fuelarc.com

Tesla Reports Drop in Self-Driving Safety After Introducing “End-to-End Neural Networks” - FuelArc News

TL;DR: Tesla self-driving tech is becoming less safe per mile, according to Tesla’s own data.

Q1 2025 was 2.5% worse than Q1 2024.

Q2 2025 was 2.8% worse than Q2 2024.

Not a great look.

Fuck AI @lemmy.world

KayLeadfoot @fedia.io

2d ago

Tesla Reports Drop in Self-Driving Safety After Introducing “End-to-End Neural Networks”

fuelarc.com /news-and-features/teslas-self-driving-data-reports-safety-decline-after-introducing-end-to-end-neural-networks/

BoycottUnitedStates @europe.pub

tfm @europe.pub

4d ago

Tesla Reports Drop in Self-Driving Safety After Introducing “End-to-End Neural Networks”

fuelarc.com /news-and-features/teslas-self-driving-data-reports-safety-decline-after-introducing-end-to-end-neural-networks/

28 comments

Move fast and break things - literal edition
- The New Self-Driving Package presented by Fred Durst
  
  Apt since Durst is such a shitbag
This is taking “testing in production” to a whole new level. How did this get past the regulations?
On second thoughts, does any country have concrete regulations for self driving vehicles? I am curious what they would be, and how they would quantify the thresholds since no self driving solution would be 100% accident-free.
- How did this get past the regulations?
  It's almost like DOGE specifically dismantled the parts of the government that were investigating and attempting to regulate Musk's companies.
  
  Wow, that's crazy -- I thought they impartially reviewed the federal government and determined that all the waste was coming from the 10-person autonomous safety office at the NHTSA, and whoever does the science to make sure rockets don't poison us all at the FAA.
- Well, we know how he got passed regulations
- Teslas "FSD" is not legal in the EU AFAIK.
  
  The Cybertruck isn't road legal in the UK. And if it was it'd need a special license due to it's weight.
- Switzerland:
  https://www.saam.swiss/switzerlands-regulatory-framework-for-automated-and-driverless-vehicles-implementation-on-1st-march-2025/
Teslas are already the least safe cars sold in america!
- If you ask Tesla drivers, they often purchased the car because someone told them Teslas are safe, despite them being the statistically most lethal car you could drive in America.
  We live in a dystopian information environment.
  
  I remember back when elon bragged how save they are and how they smashed the safety tests (american, not real ones) and i though wow interesting and looked at the tests. And literally every test they smashed was because the car is super bottom heavy because that's where the battery is. You could weld some railroads under every car and they would do just as well. I even remember thinking: oh he's gonna get roasted once people realise that. Still waiting
End-to-end ML can be much better than hybrid (or fully rules-based) systems. But there's no guarantee and you have to actually measure the difference to be sure.
For safety-critical systems, I would also not want to commit fully to an e2e system because the worse explainability means it's much harder to be confident that there is no strange failure mode that you haven't spotted but may be, or may become, unacceptable common. In that case, you would want to be able to revert to a rules-based fallaback that may once have looked worse-performing but which has turned out to be better. That means that you can't just delete and stop maintaining that rules-based code if you have any type of long-term thinking. Hmm.
- I've thought about it in the past... what if there was a bug in an update and under some specific conditions the car will just vere to the side and crash. There's a possibility that every self-driving Tesla travelling west into a sunset suddenly slams on the brakes causing a pile up. Who knows what kind of edge cases could exist?
  Even worse, what if someone hacks the wireless update and does something like this intentionally?
- yeah i wanna see what the fuck metrics made them think this was a good idea. what is their mean average precision. did they recall@1 for humans on the road
Ha funny. I highly suspected this could happen when I heard they have quotas for how many changes the people training the AI make to the AI behavior. That's a recipe for flooding the system with bad data.
No AI can be better than the info it is given, and if X is an indicator on that, it's just about a certainty that Tesla AI will rot in misinformation.
Could this be attributed to the driver mix changing?
It's quite possible tesla drivers are worse in 2025 than 2024
- The politically motivated sell off should have something to do with it 😂
  
  Yeah exactly!
Replacing code with networks has the potential to be much faster with similar quality. The idea is good.
- At least when it's Teslas code getting replaced
  
  No, in general. You want to do as little pre and post processing as possible for neural networks.
It will be fine they said, it will get better they said.
Somehow it gets even worse.

28 comments