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.
The problem with OpenSteet Maps is shown below. See if you can find it.

View attachment 1048809
Whether that actually existed online for more than a few minutes is the real question. My experience with making innocuous changes in OSM to things like speed limits is that the moderators take days and sometimes weeks to accept a change. And even valid changes are often rejected. I assume your post was designed to provide comic relief, as it does nothing to undercut the credibility of OSM. People are constantly messing with specific entries on Wikipedia, but every AI chatbot has vacuumed Wikipedia clean. And probably OSM also. Is that why chatbots hallucinate? I doubt it.
 
Anyone received updated maps for NA?

1716254072653.png
 
Last edited:
Right, that's great theoretically, but if you recall, I kinda brought this all up because Elon chose to use the term emulation. It could be absolutely nothing, just a simple term he figured everyone could understand and HW4 may have been designed from the ground up to be in "HW3 emulation mode" efficiently. I agree that this would make the most sense.
My guess is he means they are downsampling the camera video to the same resolution as HW3 before feeding it to the NNs. This allows them to train/test the NNs ONCE for both H3 and HW4, reducing training time and cost (which is very high).
 
I'm curious how they're stitching the HW4 feeds to the NN in an emulated fashion. My understanding is that HW4 has 1 less front camera - HW3 had 3 forward cameras, 1 wide angle, 1 standard, and 1 narrow focus camera. HW4 has two front facing cameras (not sure if they are both standard, or is one wide angle).
My guess is he means they are downsampling the camera video to the same resolution as HW3 before feeding it to the NNs. This allows them to train/test the NNs ONCE for both H3 and HW4, reducing training time and cost (which is very high).
If we assume that HW4 employs two cameras of different focal lengths (i.e. not the unlikely scenario of a left/right merged pair), then it's clear that one of these must be wide angle with essentially the same FOV as the HW3 wide camera. That camera can easily be binned or down-sampled to emulate the HW3 wide camera.

The remaining question is whether the second front HW4 camera has a medium or a narrow (tele) FOV (or perhaps some compromise intermediate FOV). Any of these possibilities can achieve HW3 emulation/compatibility, with the simplest solution being
  • 2x2 -> 1x1 binning to emulate whichever HW3 camera whose FOV closely matches one of the HW4 cameras.
  • A simple 2X crop (possibly shifted up or down for aim-point matching) to emulate either
    • the medium HW3 cam using the wide HW4 cam, or
    • the narrow HW3 cam using the medium HW4 cam.
Again it may not be quite this simple based on the actual HW4 setup - but something tells me that this simplification opportunity would have been carefully considered by the team during the HW4 planning, so I don't think it's an unreasonable speculation.

The dummy third camera location may be evidence that the final design came late in the HW4 development. Or, it may be evidence that something else is yet to come, like a super night-vision camera, or a supplemental sunblind-resistant camera, or a redundant independent forward-vision capability to satisfy critical system-failure scenarios.
 
If we assume that HW4 employs two cameras of different focal lengths (i.e. not the unlikely scenario of a left/right merged pair), then it's clear that one of these must be wide angle with essentially the same FOV as the HW3 wide camera. That camera can easily be binned or down-sampled to emulate the HW3 wide camera.

The remaining question is whether the second front HW4 camera has a medium or a narrow (tele) FOV (or perhaps some compromise intermediate FOV). Any of these possibilities can achieve HW3 emulation/compatibility, with the simplest solution being
  • 2x2 -> 1x1 binning to emulate whichever HW3 camera whose FOV closely matches one of the HW4 cameras.
  • A simple 2X crop (possibly shifted up or down for aim-point matching) to emulate either
    • the medium HW3 cam using the wide HW4 cam, or
    • the narrow HW3 cam using the medium HW4 cam.
Again it may not be quite this simple based on the actual HW4 setup - but something tells me that this simplification opportunity would have been carefully considered by the team during the HW4 planning, so I don't think it's an unreasonable speculation.

The dummy third camera location may be evidence that the final design came late in the HW4 development. Or, it may be evidence that something else is yet to come, like a super night-vision camera, or a supplemental sunblind-resistant camera, or a redundant independent forward-vision capability to satisfy critical system-failure scenarios.
As noted, one of the advantages of having higher-resolution cameras is that they can serve redundantly for the narrower-FOV cameras, and that's probably how Tesla was able to reduce from three to two physical front cameras from HW3 to HW4. What's more, it could allow the simulated narrow-FOV cameras to be "steered" within the wider-angle window. I've occasionally wondered whether a literally steerable telephoto camera might be a good idea, e.g. for reading street signs at a greater distance, though it's far from obvious how to train a network to effectively use such a thing? (But it is exactly how our eyes work; our brains control in real-time where to direct our "telephoto" foveas, which are a lot higher resolution than any FSD cameras.)
 
Scary situation on 12..3.6 today. It was at what I call an Optional Protected Left Turn: the left turn signal has a green arrow and a yellow arrow, but no red arrow. The yellow arrow was lit when I entered the turn lane, and had gone dark by the time I reached the intersection. I expected Tesla to stop, but it just slowed down, and then slowly turned in front of opposing traffic just as they started to advance. I had to gun the accelerator, because if I had hit the brake I would have been blocking opposing traffic. I was shaken.
1716267435490.png
 
  • Informative
Reactions: RabidYak and mgs333
V12.3.6 kindly gave me my 3rd strike today due to a comedy of errors and a navigation change. Does anyone know the duration of time the strikes last? My understanding they are indefinite or until you get 5 at which time you are suspended for 5 days. I have looked for this documentation, but have been unsuccessful in finding it.
Lastly, when V12.4 rolls out soon and I download it, that should erase any previous strikes, correct?
 
Today with V12.3.6, I encountered a red light at an intersection with a "No Turn on Red" sign. My car stopped, and then attempted to make the turn. I'm assuming the sign was not properly processed, so still a few corner cases.
Doesn't read no turn signs, it uses map data. Incorrect or not current in your case. In sections of ATL on Jan 1st ALL Right Turn on Red will become illegal. This will require a big update.
 
I finally got my free trial on 2024.14.6 and for the first time I could see paying for this someday. My previous drives on FSD v10 and v11 in Tampa have been extremely dangerous.
  • No longer begins stopping at 55mph from 50 feet away - reduced hard breaking and lots more use of regenerative breaking. Not perfect but much improved.
  • For the first time it can enter a double left turn lane on my main route - it used to stay in the lane that goes straight. It still hesitates - enters the left turn pretty late and does a half lane change and straddles bot lanes for 10feet or so before committing.
  • No longer tries to signal and almost change lanes every single time a left turn lane appears 2 miles before a left turn.
  • No psychotic lane choices - like taking the extreme right lane of a 4 lane road 3 miles before a left turn.
  • First time successful UPLs.
Not actually going to pay until
  • It goes the max speed I choose and not 15mph below school buses and cops.
  • It stops changing lanes when traffic is backed up right before a turn and you either have to douchebag merge back into the target lane or use an alternate route.
  • Fix selecting a lane that is ending in 50ft after a turn
  • Manual lane changes work - when I use the stalk to initiate a lane change it does half a lane change and straddles 2 lanes and then 50% of the time completes the lane change and 50% aborts the lane change. Happens even where there is 0 traffic. When FSD initiates a lane change it almost always works even in heavy traffic.
  • It stop tailgating - doesn't happen too often but super annoying.