-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD v12.x (end to end AI)
- Thread starter Buckminster
- Start date
-
- Tags
- #fixphantombraking autopilot chill is not chill elon is an ass fix fsd extreme late deceleration fsd does not commit to go fsd does not maintain proper speed fsd hard and late braking fsd has decision wobble fsd jerky driving fsd late deceleration fsd sometimes may not detect train crossings fsd stops too early too quickly fsd stops too far behind line fsd tailgating in city streets
I have no beef with him, he's a STEM guy (meaning his calculator has many buttons and symbols that you may not recognize). So big words are strictly for extra credit.
You, the other hand, are paid to interpret complex and deliberately obfuscated regulations, so I hold you to a high standard. Also, you've repeatedly explained that you need help to elevate your wife's opinion of you. I thought maybe you could impress her if you could join the ranks of the cognoscenti.
Think he would impress her more with a bigger cannoli vs a cognoscenti.
I keep hoping that it's more about right-sizing.
(Relax, Mods, we're just talking about the FSD computer...)
Control loop: some kind of input, followed by some processing/filtering, followed by some kind of output, then feedback. A standard way of looking at the world and doing things in engineering.
You want your FM radio to lock onto the subcarrier so it can extract the stereo audio? Control loops. You want a roller to squeeze hot steel to a given thickness in a steel mill? Control loops. These things are all over the place; there’s hundreds of them in the hardware of a PC.
There’s lots of differential calculus to analyze and design these loops. Some loops are linear, others are nonlinear, which leads to Fun.
Thing is, it’s quite possible to insert a human into a loop. The obvious: drive a car down a road with marked lanes. The human is looking out, seeing the lanes, moving the wheel back and forth to stay in the lane.
As a component in the loop, the characteristics of the human have to be taken into account. Delay, strength, resolution of vision, you name it. This isn’t an, “It’s obvious!” situation. It’s quite possible to design a vehicle that literally can’t be controlled by a human. If a car does something faster than a human can keep up with, well, it’s not going to sell well if people crash them on test drives. It’s why there are test pilots in aircraft: and some of them have had to eject if the designers didn’t get the ergonomics right.
Heck, about once or thrice a year we all hear reports of some over ambitious idiot in a super car with the traction control off wrapping said super car around a tree. This isn’t a hypothetical.
So, some gonzo on this thread says, “It’s not hard to control a car! Just twitch the muscles back and forth a bit! It’s easy!”
Like. Suuurree it is. And getting a computer to do all that image recognition, path recognition, avoiding pedestrians, not turning so fast as to roll the car, and all that jazz, is simple, too?
There’s nothing simple about any of it. Whether it’s a computer driving a car, a human driving a car, or a human just walking across the landscape.
We all can do the latter, mostly, but we’re all the product of several hundred million years of evolution that has pretty much wired us up to do the job; and the ones who didn’t do the job well got eaten, and generally didn’t pass their genes down as a result.
And it’s pretty blinking amazing that, in the space of 80 years or so, we’re building machines that can navigate down a highway, duplicating something that took us and our ancestors hundreds of millions of years to do.
5x-10x miles per intervention improvement from 12.3 to 12.4 with a similar "major" improvement going to 12.5? I wonder what the improvement was from 11.x to 12.x so far although that's a bit mixed with 11.x having a lot more highway miles and 12.5 theoretically restoring that functionality (as well as adding Cybertruck) starting in 6 weeks / 3 more 2-week development cycles.
Who in this thread is arguing that it's simple or unimpressive for a computer to drive a car? Trying to argue that physically turning a steering is difficult for a human being in a million words makes you look like a troll or a lunatic.
Well hopefully 12.4 will show up on TeslaFi by next week (if those employees weren't fired) and it will finally be "next week" from last week. Almost a record for Musk IF it happens but we will have to wait and see to believe. Interesting that we may see highways implemented on 12.5. Of course he doesn't mention ASS/Banish/Parking Lots which would be required for it to be a TRUE Single Stack but that didn't stop 11.x from being called a Single Stack. So guessing ASS/Banish/Parking Lots will be pushed back YET again.
Tesla you have been working that ASS out for almost 2 years now so it better be a hot, tight and sweet ASS when we see it.
Again Tesla give us some 12.3.7
to play with in the meantime. You don't really have to change anything since our placebo induced brains can add perceived improvements. We just need to feel some headway while we impatiently watch and wait on 12.4.x to make the slow rollout to us.
Last edited:
Mike_TV
Active Member
Probably for Robotaxi.
LSDTester#1
Ingesting large quantities of LSD
No, not more cameras, just different positions. Looks like it is to simulate for a shorter narrower vehicle.
LSDTester#1
Ingesting large quantities of LSD
PianoAl
Active Member
Note that you can't clear that nag by just torquing the wheel. You must torque the wheel AND look up at the road.
Yes, this invaluable in a strange town. I learned that here, when the nav said, "Enter the roundabout and take the second exit."
Thanks. I may perform a camera recalibration this weekend before we leave for a long 600 mile each way two week road trip, during which I plan to use FSDS quite a bit to aid with driver fatigue. I did a camera recal when 12.x was first released earlier this year, but I'm certainly willing to do so again especially right before a long road trip.
When entering our development entrance off of the main road - which requires a right hand turn via a painted turn lane - FSDS has shown varying behavior. About 50% of the time it will move right into the right hand exit lane and execute the right hand turn properly - moving entirely into the right hand turn lane promptly. The other 50% of the time, it will very gradually and tentatively move into the right hand turn lane - blocking cars behind us from moving through the light/intersection with a green light - and about 25% of the time - it only moves about halfway into the right turn lane until it actually has to execute the right hand turn - and slows way down prior to taking the right hand turn. Ironically, it seems to perform well when there are no cars behind our vehicle - but performs tentatively when there are cars behind our vehicle - which again seems to be caused by some kind of proximity issue that results in FSDS being too tentative when making lane changes of any kind when other vehicles are behind us.
Bigger picture, I've seen some YT content that basically outlines how FSDS 12.3.x doesn't really use inference very well right now. So in the example of the double left turn intersection that I outlined, it literally does not plan ahead because it doesn't "see" the fact that the right lane will come to an end with a solid white line soon after the turn is executed. This is apparently coming in 12.5 or 12.6 from what I've gathered (not 12.4). I was hoping to see 12.4 before 5/19 when we leave for our road trip - but I'm doubtful it'll actually come to fruition based on Musk time.
Last edited:
In at least half of the cases - the right hand lane was free of traffic - that's why the aggressive driver behind me was passing both our car and the lead car ahead of me via the right lane.
PianoAl
Active Member
Interesting. I find any long drive tiring and much less so on FSD. Might be age related.
Serious question: What about driving on a four lane for twelve hours? We did some twelve hour drives on a road trip, and I wouldn't have done that without FSD.
I am super freaked out that the latest firmware contains the latest FSD. We are in unchartered territory!
zoomer0056
Active Member
Really need to relax a little ... less coffee maybe. Writing "latest" should be replaced with numbers.
I think they sort of were forced to do that as 3/4 of the N. American fleet was on 2024.3.25 because of the free trial and a large portion of that isn't subscribing/buying FSD past the trial. So to get those non-FSD people onto the Spring update they needed to keep the same FSD as was in 3.25.
It is still a very slow trickle of 3.25 cars moving to 14.6 based on TeslaFi data. The huge surge of 14.6 yesterday was mostly 8.9 cars worldwide.
I think there is still a strong possibility that cars that are subscribed or purchased FSD may stay on 3.X and get FSD 12.4 ahead of anyone on 14.6 that may decide to subscribe later on.
