Here's a question for you tech gurus: can FSD performance get better between software updates? I have seen what appears to be behavior change (for both better and worse) even though there has been no new software update installed. I guess don't understand how much of driving behavior is controlled by real-time interaction with the network, as opposed to local processing. Sorry if this has been answered elsewhere--I looked and didn't see an answer.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Does FSD get better between updates?
- Thread starter AceG8tr
- Start date
jebinc
Well-Known Member
Many would say the beta FSD software has regressed more than improved the last six months.
Fernand
Active Member
There are changes I can't explain. Like yesterday I got no blue nags. None. I have a 4 oz counterweight on the right steering wheel spoke, just enough to make anti-nag torque lighter, but yesterday all throughout an 11 mile drive on freeways and on city streets with pedestrians, sharp turns and speed bumps, not a single blue nag. Nada. I ended up with my right hand on my lap, just to provoke it, and still, nada. Something had changed. I also thought Nicki was driving exceptionally smoothly, without a single override, but who knows on that.
Last edited:
OncomingStorm
Member
i'm not sure anyone outside of tesla could answer this with any certainty however, i have a personal theory that all the cars, but espically the beta testers do have some built in, limited, local error checking or correcting that they do. since they are so AI heavy, my guess is, the computer is constantly running simulations essentially and where it notices things that it did or would have done wrong it MAY record and use that data when encountering that again as a way of proofing the simulation and then sending that data back to the mothership for review.... me and some friends of mine have noticed this behavior too. This is all wild speculation that i have no real facts to support lol, but from the anecdotal information i've seen, it would hold.
OR it could all just be situational changes... the car reacting to different things differently.... who knows
I assume the OP means is the car learning on its own and getting better without receiving the next update.
Based on what Tesla has publicly stated, no. The cars are just collecting data, and that data is used to train the AI at HQ. Only when the next firmware update is released will the car operate under different AI/logic.
But there are lots of anecdotes of people seeing improvement on a given version with more driving. It could all be a psychological trick (we see what we want to see), or maybe there's something going on at the car level that we don't know about. I tend to think it's a psychological trick.
Based on what Tesla has publicly stated, no. The cars are just collecting data, and that data is used to train the AI at HQ. Only when the next firmware update is released will the car operate under different AI/logic.
But there are lots of anecdotes of people seeing improvement on a given version with more driving. It could all be a psychological trick (we see what we want to see), or maybe there's something going on at the car level that we don't know about. I tend to think it's a psychological trick.
I’ve found improvements across most all of the updates. Not perfect but definitely improving. From the first versions of block art to today I am rather impressed. Problem is many view the software as all or nothing or not improving quick enough. This is a never before done feat that shouldn’t be rushed.
TresLA
Member
I’ve seen improvement after some updates. In general, FSD beta keeps getting better, But if you’re asking whether there is improvement of a car on its own between updates, Tesla has said no. Whatever is learned is only sent back to Tesla to help improve the next update. One thing to keep in mind about neural nets is that they’re not direct programming, so it’s less predictable. An improvement or regression can rarely be so easily attributed to an update. By nature, they don’t behave exactly the same in what we perceive as the same circumstances.
I should have been clearer--yes, I was referring to "non-update" changes in behavior. One example: on one of the earlier versions of the FSD beta software, my MY would try to move into the wrong exit lane when approaching a left turn (i.e., it tried to turn into an exit that was about a hundred feet before the "correct" exit lane). However, after several times where I disengaged and corrected the car--it stopped doing it and started waiting for the correct lane. But based on other responses to my OP, it sounds like any behavior changes should be due only to firmware updates, not to any local learning by the car.
EVNow
Well-Known Member
Tam
Well-Known Member
It should also be clear that your software is beta which means sometimes it works fine and other times it might not even in a very same situation.
That's why sometimes it got into an accident and sometimes it magically avoids accidents.
For example, a Tesla employee was fired after showing Tesla FSD beta collided with a green bollard. Why would he show that? He could have ONLY shown other scenes that it NEVER hit a green bollard before!
So the randomness of never hitting the green bollard before could give a false sense that the car has learned so well between updates.
Thus, you might want to look for some consistency: Like it would always reliably blow through a particular stop sign at a specific location. Just repeat the scene over and over again and see how well it would learn between updates.
In my experience since I first got Autopilot in 2017 and now the FSD beta, the improvement in between updates is random that the car could have done either way.
With my first experience of FSD beta, it could randomly turn in the wrong direction of a one-way street or the correct way. I could have said that it "learned" between updates so that's why it's choosing the correct direction for the one-way street. But that randomness of going the wrong way disappeared with the next update.
The competency to deal with a particular curve of death never happened in between updates. Those curves of deaths are dealt with in new updates.
The other one that it has never learned and never randomly done right is a particular intersection that it is supposed to drive straight according to the navigation and the FSD beta path shown on the instrument cluster, but the system consistently moves to the right-turn-only-lane before reaching the intersection (then it illegally corrects itself by not turning right and moves straight while positioned itself in right turn lane only). No matter how many updates so far, it does the same. As in the past with curves of deaths, I believe this particular intersection will be categorized and fixed in subsequent updates but I would not hold my breath for a quick timeline.
Thus, it is important to differentiate randomness vs learning. If I was losing 20 times straight with a slot machine in a casino, then I won the 21st, 22nd, and 23rd time, is that machine learning or randomness?
Last edited:
Thank-you, that was very helpful. Verygreen suggests that there *is* some local memory that could cause the car to change behavior between updates. That is consistent with what I've seen.
Yes and no. For instance, FSD beta has absolutely no idea how to handle free underpasses in South Texas, especially Houston, nor their u-turn lanes. FSD beta tries to stop entering u-turn lanes, doesn't yield when exiting them, tries to split traffic when they are backed up (they are only one lane but often wide so 18-wheelers can use them), etc. Additionally, if you turn left under a underpass FSD always says in the closest lane which puts you into another dedicated left turn lane - which then forces you to do a u-turn or make an illegal additional lane change in the intersection. There are hundreds of intersections/freeway underpasses in Houston that are like this and FSD can't deal with them., I sent autopilot snapshots literally hundreds of times when dealing with these situations and it doesn't improve.
It does improve on things like smoothness of breaking, handling stop signs and unprotected turns, and stuff like that.
It does improve on things like smoothness of breaking, handling stop signs and unprotected turns, and stuff like that.
The Dallas area has the same situation with turning into left turn lanes than immediately needing to shift right out of the lane.
Yes, and no
Yes -> During the initial day or two with a new version of SW its widely reported that performance does improve. People have tried to backup this claim by saying it redownloads the mapping tiles.
No -> There is no onboard learning so it has no mechanism to get better. Any change is just randomness of testing something X number of times.
Similar threads
