Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register
This site may earn commission on affiliate links.
2, 3, and 4 are just issues for your personal driving preference.
You’re right. No one else has mentioned anything about these issues. It’s basically perfect, really.

I stand corrected. You have a long history of being right.

The true key problems with FSD v12 are that it signals too early and drives too fast.
 
In that prior context, it wasn't clear what role the disengagement data would play, i.e. how these negative training examples could be used to teach what not to do.
The disengagements can inform the team that there are scenarios which the system doesn't handle correctly. They can then find training examples to correct that shortcoming. As you said, they have tons of training data already, but they don't know which scenarios aren't covered by the current training data.
 
  • Like
Reactions: FSDtester#1
You’re right. No one else has mentioned anything about these issues. It’s basically perfect, really.
I never said it was perfect. But, as an example, I would say that having it slam the brakes on at a yellow light when 90+% of people would normally proceed through the intersection is a much more serious issue than any of the issues in your list. (I have been rear ended twice from people planning to run the red light behind, but I didn't run it. And both of those cases the yellow light was old and just about to go red, unlike what FSD is doing where it will slam the brakes at a fresh yellow.)
 
Sign reading is very limited and mostly just standard Stop and Speed Limit signs. No Left Turn (like No Right on Red) is based on inaccurate map data.
FSD doesn't have OCR the sign to read what it says. FSD just has to know what the image means and what the car should or shouldn't do. Tesla used AI so FSD learned what Stop Signs and Traffic Light's look like and learned that some traffic lights are vertical and others horizontal. So maybe FSD/AI can learn what other signs mean too. I just wouldn't assume the problem is solved by improving map data.

I would expect if FSD watched thousands of videos with these types of signs AI would learn what to do from what the human drivers did. Could be wrong but AI has already done some amazing things for FSD.

1712242529909.png
 
Last edited:
So maybe FSD/AI can learn what other signs mean too.

The only two roadblocks (pun intended) are camera resolution and in-vehicle compute. It's possible that some signs might have fine details that aren't discernible by the cameras on HW3, so a model could never learn to act on details it can't see. And second is the FSD computer's ability to hold model weights in memory and perform inference with reasonable latency. It's possible that V12 has just enough model weights to hold vague patterns about stop signs and traffic lights, but not enough to compress information about detailed variations in traffic signs.
 
The only two roadblocks (pun intended) are camera resolution and in-vehicle compute. It's possible that some signs might have fine details that aren't discernible by the cameras on HW3, so a model could never learn to act on details it can't see. And second is the FSD computer's ability to hold model weights in memory and perform inference with reasonable latency. It's possible that V12 has just enough model weights to hold vague patterns about stop signs and traffic lights, but not enough to compress information about detailed variations in traffic signs.
You could absolutely be right. My point in my previous post is just don't assume the solution is map data which was posted. Assumptions sometimes take you down a rabbit hole.
 
slam the brakes on at a yellow light when 90+% of people would normally proceed through the intersection is a much more serious issue

This is also a problem, and a regression, in v12. Sorry for not being exhaustive in listing the problems! Trying to be succinct but that always leads to problems!

A tap of the accelerator resolves the yellow light issue and it’s easy to react to since impending yellows are easy to predict.

I find issues that can be resolved with accelerator input easier to deal with than ones that require slowing down.

For example the stop sign thing is relatively easy to fix; I just press the accelerator all the way up to the NHTSA stop, and move the NHTSA stop to the stop line if needed. If that means the NHTSA stop is bypassed (happens occasionally), I brake to a stop. It’s just super annoying because it’s not just a tap and requires careful modulation to avoid jerking.

Speed issue: accelerator input - easy

Red lights: accelerator input, easy - until it is hard, and not stopping (the worst).

I listed problems that require large amounts of engagement (and disengagements) for me.

Turn lane issues, etc., are only occasional and are second-level in many ways.
 
Last edited:
I'm just mystified that they released FSD in this state:
1) Doesn't even attempt go the speed desired by the user.
2) Won't respond in a normal way to red traffic lights.
3) Frequently can't handle stop signs in a remotely normal way.
4) Cuts it super close to curbs on many turns.
5) Simply stops in lanes of traffic in the middle of a maneuver.

Here is a different list, of what I consider actual issues rather than preferences:
  1. Goes 15 MPH over the speed limit. (Ticket anyone?)
  2. Slams the brakes on at fresh yellow lights, making it more likely to be rear-ended.
  3. Signaling too early, before turn prior to actual turn, in some cases causing the person at that intersection to assume you are turning and pull out in front of you.
  4. Curbs wheels sometimes? (I don't know if 12.3.3 fixed this.)
  5. Doesn't pull over for first responder vehicles with lights/sirens.
  6. Doesn't stop for school bus with red lights and stop sign on other side of road.
  7. Doesn't identify/respect school zone speed limits.
Sure, you can mitigate most of those, but those are actual issues that they should work on before dealing with preference issues like stopping too early/slowly.
 
Last edited:
Here is a different list, of what I consider actual issues rather than preferences:
  1. Goes 15 MPH over the speed limit.
  2. Slams the brakes on at fresh yellow lights, making it more likely to be rear-ended.
  3. Signaling too early, before turn prior to actual turn, in some cases causing the person at that intersection to assume you are turning and pull out in front of you.
  4. Curbs wheels sometimes? (I don't know is 12.3.3 fixed this.)
These are also definitely issues, and not preferences. My list is also not preferences. If they were, I would not have mentioned them.

1) Use manual mode. We’ve discussed. There are limitations due to lack of fixed offset.
2) Tap accelerator.
3) Disengage and prepare for evasive action
4) It did not. Disengage on many corners.
 
  • Informative
Reactions: sleepydoc
I never said it was perfect. But, as an example, I would say that having it slam the brakes on at a yellow light when 90+% of people would normally proceed through the intersection is a much more serious issue than any of the issues in your list. (I have been rear ended twice from people planning to run the red light behind, but I didn't run it. And both of those cases the yellow light was old and just about to go red, unlike what FSD is doing where it will slam the brakes at a fresh yellow.)
Since yellow lights vary in timing, with no way to tell what that timing is, and the new NN are not coded but instead learn how to handle things by watching curated videos, I wonder if that's why we're seeing this behavior. Your yellows might have a 4 second timer, but the NNs were trained on 2 or 3 second timers, so it brakes out of caution?

For example, if your lights near you always have 4 second timers, and that's your muscle memory for years, then you visit another state where they have 2 second timers, I'd bet you run a few red lights before adapting.
 


Here's another one with video but it doesn't say which version of V12.3.x


These situations are really impressing those without FSD experience who are trying out their free version.