I've spent a LOT of time with 3.6 and 5.4.1. Overall 5.4.1 is much much better, and that is the overall consensus online. Yes there are some quirks with green lights, but overall it's an incredibly capable build. Also you're wrong, E2E didn't get added until 12.5.6, and it looks like 12.5.6.1 will be the one to go wide (hopefully this week). critiquing any of its highway driving is kind of a moot point now, since it will essentially be replaced entirely with E2E. Anyways your title is garbage, but I guess that's what gets clicks and engagement nowadays - so it works.
I’m letting the AI generate my title, sorry clicks is what the AI is looking for I do disagree with one point you made regarding 12.5.4.1 being moot. A few months ago, I would have agreed with you, but recently we've observed that we need to consider the majority of users and the version number they are using, as this will affect their experience with FSD. My wife was very accustomed to driving on version 12.3.6 when it first came out because we paid for a month of Full Self-Driving (FSD). Now, she is on her free trial of 12.5.4.1, and her car behaves as if it was hit over the head with a blunt object. It doesn’t perform anywhere near as well as mine, even though we are using the same version of FSD. As a result, we have two cars that are acting completely differently despite having the same software. I’m currently making a video about this situation because she refuses to use FSD after our drive this past weekend.
The software engineering behind FSD is flawed, and it will never be fully autonomous and non-supervised. The latest software still has many obvious issues .
Something feels off. I don't understand how you can fix some issues but end up breaking other things that were working. If this is a neural network and the car is learning, then once it resolves an issue, it shouldn't revert back to the previous problem. It's as if every time there's a major release of FSD, they hit the restart button and delete all the data out.
@@FSDdriverI am not a ML specialist, but I have some knowledge about that. My feeling is that they train the ML model to satisfy a range of test conditions (where a test means whatever inputs the model is to act upon, at least video). I believe this training dataset is being continuously modified to add new borderline scenarios. However, the issues you experience is because they won't ever be capable of creating test cases for each and every situation. Hence, with each new release they will solve issues they did observe, but they will give birth to other issues in real life, issues they were never aware before, simply due to the way an ML model reacts in general (it is not a deterministic system overall) . I strongly believe that this game Tesla does cannot result in a perfect self driving system, eg perfect so that ie never creates accidents. However, I'm also at a loss in suggesting an alternative... :(
Even with the model you described, it's not encountering new scenarios, which makes it difficult to drive. The problem lies in the fact that things it used to handle well are now completely absent. For example, my car would navigate through intersections when the light was green, but after the recent update, it slams its brakes at green lights. This suggests there may be an error in the latest models regarding how they capture data, leading to this abrupt behavior. In my opinion, they should maintain the current working model permanently while continuing to process green light intersection data as a secondary source in their training. It feels like they've discarded this crucial data altogether, and as a result, we end up with less effective training models.
@@FSDdriver I understand now and thanks for detailing this. This speaks of a big failure in their training data. I don't imagine how this could be possible, but one lesson I learned across the years is that one should not trust until it sees (I'm referring to their software dev process in this case)
i agree
It's a hard thing to understand. Every FSD update is like teaching a newborn child how to do everyday things.
I've spent a LOT of time with 3.6 and 5.4.1.
Overall 5.4.1 is much much better, and that is the overall consensus online. Yes there are some quirks with green lights, but overall it's an incredibly capable build.
Also you're wrong, E2E didn't get added until 12.5.6, and it looks like 12.5.6.1 will be the one to go wide (hopefully this week).
critiquing any of its highway driving is kind of a moot point now, since it will essentially be replaced entirely with E2E.
Anyways your title is garbage, but I guess that's what gets clicks and engagement nowadays - so it works.
I’m letting the AI generate my title, sorry clicks is what the AI is looking for
I do disagree with one point you made regarding 12.5.4.1 being moot. A few months ago, I would have agreed with you, but recently we've observed that we need to consider the majority of users and the version number they are using, as this will affect their experience with FSD.
My wife was very accustomed to driving on version 12.3.6 when it first came out because we paid for a month of Full Self-Driving (FSD). Now, she is on her free trial of 12.5.4.1, and her car behaves as if it was hit over the head with a blunt object. It doesn’t perform anywhere near as well as mine, even though we are using the same version of FSD. As a result, we have two cars that are acting completely differently despite having the same software. I’m currently making a video about this situation because she refuses to use FSD after our drive this past weekend.
The software engineering behind FSD is flawed, and it will never be fully autonomous and non-supervised. The latest software still has many obvious issues .
Something feels off. I don't understand how you can fix some issues but end up breaking other things that were working. If this is a neural network and the car is learning, then once it resolves an issue, it shouldn't revert back to the previous problem. It's as if every time there's a major release of FSD, they hit the restart button and delete all the data out.
@@FSDdriverI am not a ML specialist, but I have some knowledge about that. My feeling is that they train the ML model to satisfy a range of test conditions (where a test means whatever inputs the model is to act upon, at least video). I believe this training dataset is being continuously modified to add new borderline scenarios. However, the issues you experience is because they won't ever be capable of creating test cases for each and every situation. Hence, with each new release they will solve issues they did observe, but they will give birth to other issues in real life, issues they were never aware before, simply due to the way an ML model reacts in general (it is not a deterministic system overall) . I strongly believe that this game Tesla does cannot result in a perfect self driving system, eg perfect so that ie never creates accidents. However, I'm also at a loss in suggesting an alternative... :(
Even with the model you described, it's not encountering new scenarios, which makes it difficult to drive. The problem lies in the fact that things it used to handle well are now completely absent. For example, my car would navigate through intersections when the light was green, but after the recent update, it slams its brakes at green lights. This suggests there may be an error in the latest models regarding how they capture data, leading to this abrupt behavior.
In my opinion, they should maintain the current working model permanently while continuing to process green light intersection data as a secondary source in their training. It feels like they've discarded this crucial data altogether, and as a result, we end up with less effective training models.
@@FSDdriver I understand now and thanks for detailing this. This speaks of a big failure in their training data. I don't imagine how this could be possible, but one lesson I learned across the years is that one should not trust until it sees (I'm referring to their software dev process in this case)