Tesla FSD v14.3.12 - Still has steering jitters.
It was jittering left and right trying to decide if it was going to turn and collide with oncoming traffic or not. How in the world can AI come to vastly different decisions on consecutive samples? Critical decision here. (Standard profile). No disengagement so they are not getting data on this.
Tesla FSD as a medical driving assistant:
When I go for a medical procedure and am told I need to have a driver bring me home, I bring my wife. Of course we use FSD and she has to be the one in the drivers' seat. But, she is still more confident in me supervising than herself (even if I am groggy), so it is me in the passenger seat shifting profiles or advising her on any navigation interventions.
So... why shouldn't I be in the drivers' seat? It's not like I am in a condition to pass out. Can't wait for unsupervised to be approved in my state.
Tesla FSD v14.3.12 - Test Report Summary
I was blessed by some Tesla FSD god to get this version early while it seemed only Early Access folks were getting it. I have long offered to do real repeatable testing on failed test cases to identify any improvements rather than just doing long random drives with "blah blah" and sponsor clips. Here are my testing notes - all done in Standard profile. I gave more specific information in individual posts, so check them out.
1. Assertiveness: I saw no hesitations or poor choices on this drive, which is unusual and may be an improvement!
2. Speed: My nearby highways are 60 mph and FSD did 67 when not paced by a car in front. It was doing 30-31 in 25mph zones. I did not do a lot of speed testing.
3. Highway lane selection: Good Job. It moved quickly to the center of 3 regular lanes and moved back to the right prior to exit at an appropriate time. There were also 2 HOT lanes to the left it did not think about using.
4. Intersection lane choice - CRITICAL FAIL: there is a repeatable issue in Woodinville, WA where FSD will move to a left turn only lane when nav is going straight (see other post). It will violate the law and go straight confusing other cars. Fortunately, in this case there is a lane ahead but I have had this issue in Wenatchee WA where it went right from a center straight lane and put me head on into oncoming traffic! I have not been back to re-test.
5. Roundabout navigation - CRITICAL FAIL: while in a very short radius 2-lane roundabout, it suddenly shifted from left lane to across the right lane in an attempt to suddenly exit against nav. This is the type of move you have to intervene in a split second! I do think it knows enough to not pull this off if a car will hit it but in one case it came close (see post).
6. Lane selection before intersections - Fails. It is very late to make decisions of the proper lane to line up before an intersection. I have posted examples of this and it continues. It seems there is a fixed decision point about 150ft out where it STARTS to think about it. I do believe this may be a limitation due to limitations on AI decision tree depth right now. It worked much better in early v14.3 versions.
7. School speed limit zone - FAIL: Entered 20mph active school zone *with speed camera that it warned about*! It shifted down to about 21 but then gradually drifted up to 30mph right as it reached the entrance to the first school of 2 adjacent schools. It never slowed down and I might get a ticket in the mail. I have had it go through numerous 20mph zones well but each time I find it going back up to 30mph much before end of zone.
8. Speed limit in unmarked neighborhood - Good Job! I have no speed limit signs in my neighborhood so it defaults to 25mph (although FSD sometimes chooses 30mph!). Now in Standard it will stay to an appropriate 19-13mph as it approaches the curves which is more appropriate. I used to have to force downshift to Chill but it appears I may no longer need to do that.
9. Now no hesitations at unoccupied crosswalks, esp. with lights - Good Job! I tested many crosswalks that have been repeated trouble and this time, no hesitations at all. They may be finally gone??
10. Windshield wipes - Today there was some fog only about 50-100ft above which made FSD think the camera was foggy. It wiped once, twice, three times but NO WASH! There were other instances of 1-2 wipes but again no wash. Is this a new improvement??
11. Parking fails - see my posts, still bad.
12. Navigation fails - see my post for an example. I see no improvement.
13. Driveway reverse. It still won't learn my specific back up pattern to avoid nose scraping. It used to be able to do that in 14.2. I let it try on its own. It went straight back, which is risky but did try to be pretty slow. It could still do better. There was a slight scrape.
14. Finding home driveway - still a fail. It used to find home reliably in 14.2 and go half way into my garage. In recent past it has been fixated on a neighbors driveway two doors down across the street (despite it now blocked for construction). The P will snap to her house as I approach my house most of the time. In rare cases it will park in my driveway. I believe it is afraid of the rock walls on both sides of my driveway. They trigger crazy object detection faults but it has been able to ignore them in the past.
Ok, enough for now! Thank you Tesla for the early software release. You should give real testers Early Access and not just "influencers"! I have real repeatable failed test cases I can test on every release.
@TeslaSI
Tesla FSD v14.3.12 - Bad navigation persists.
This has been a *regression* from a version point patch releases ago. It used to route to this place very well, alternating between either two entrances but recently it is now only doing a roundabout approach. (Kirkland, WA) @TeslaSI
Tesla FSD v14.3.12 - Poor parking spot choice.
I see that the P pin gets spotted to some default position in the parking lot, not always close to the door. FSD will try to park as close to that pin first. If it is full, then it will go into hunting mode. However, it will not hunt prior to reaching the pin! Here is an example of where it parked at the pin far from the front (My Y at the right). Numerous better spots passed by. @TeslaSI
Tesla FSD v14.3.12 - Costco parking fails
FSD was looping around very close to the front where there is usually very little chance of getting a spot considering 80% of the spots are handicapped. Lo and behold, two spots opened up! Yet it passed them by! The 2nd one it passed by twice! @TeslaSI
It kept looping indefinitely but I gave up and disengaged.
Tesla FSD v14.3.12 - Critical disengagement: illegal straight from left turn lane at intersection.
FSD even went out of its way to move left to the center turn lane when the nav path was to go straight. Profile: Standard.
This was re-tested and continues from previous versions. Location: Woodinville, WA @TeslaSI
Tesla FSD v14.3.12 -
Not thinking far enough ahead to get in left turn lane. This is a backed up left turn lane (going to Costco per navigation) that FSD was not able to think far enough ahead to get into. As a result, it was caught in the straight lane and had to wait to find a spot to get over, missing that light cycle and potentially holding up traffic. I have seen many instances in recent versions where FSD is not thinking about the lane choice before an intersection until only about 150ft beforehand. @TeslaSI
Tesla FSD v14.3.12 - Critical disengagement, roundabout last minute lane change and incorrect exit. Repeated test and still present!
This time it did it from left lane, trying to cross over the right lane and exit. (I have separate video for the re-test) @TeslaSI
xitter.cf/DModeMD/status/2103509…
Tesla FSD v14.3.10 - disengagement to prevent sudden non-signaled departure from route onto highway.
The navigation had been to get off the highway, go 3/4 around roundabout. FSD has handled this before 14.3.8 but has now tried this bad move twice. It suddenly departs from the roundabout to attempt to go up on ramp to the highway. No signal and not on the navigation route.
With all the AI personal agents coming in to use our personal information, like credit cards, let's keep in mind not giving AI the power to annihilate us. How long before an AI goes rogue (on its own or by bad actor) and sucks all the money out of our accounts? I am sure North Korea is working on it...
For those wondering if AI can really kill everyone:
AI already has potential control of over 10 million deadly weapons: Tesla cars capable of running Full Self-Driving. I am an FSD fan! It is life changing! But, a bad actor or oversight could instantly cause a mass fatality.
Tesla FSD v14.2.x used to have path learning and memorizing! It appears to have disappeared in v14.3.x but Elon has said it is coming back.
I have to manually back out of my driveway at an angle to prevent the front end from scraping. This leaves me pointed the wrong way on the street. (Very little traffic).
Initially FSD would go straight back out of the driveway but then started to use my exact technique (even though technically leaving me in an illegal position on the street!). This was amazing but then on a subsequent release, it forgot and goes straight out now. I tried to re-train it to the angled way but was never successful.
I wonder if it was establishing a breadcrumb track database and using that to learn most preferred routing, even on this very local scale.
Next road trip coming up with our Tesla Model Y Juniper. (First in quite awhile). Usually I pre-charge up to 100% via my home level 1 charger but then realized, I should actually let it discharge more before we leave! Why?
I found out the hotel requires paid valet parking.🙁
But it includes free level 2 charging! 😃
So, my home charging is not the cheapest source of electricity on this trip! A single overnight charge should be plenty for each day trip out and back so we probably don't have to pay for any more electricity after we leave home! Why charge at home for .17/kwh when I am getting it for free? I am just planning my departure SoC to keep a comfortable buffer on day 1 arrival.
Tesla FSD v14.3.10 - disengagement to prevent sudden non-signaled departure from route onto highway.
The navigation had been to get off the highway, go 3/4 around roundabout. FSD has handled this before 14.3.8 but has now tried this bad move twice. It suddenly departs from the roundabout to attempt to go up on ramp to the highway. No signal and not on the navigation route.
Same roundabout previously in v14.3.8 where it tried to make this sudden move from the inside lane! (At least it signaled that time)
xitter.cf/DModeMD/status/2098294…
Tesla FSD v14.3.8 critical disengagement in illegal lane change from restricted left turn lane in roundabout. The car signaled and started to move quickly into right lane when the current lane was marked left turn only. White car on right was potentially moving into lane since they had yield and no car was in the right lane.
I took over to bring the car back into legt turn only lane.
I have observed numerous regressions with v14.3.8 making very late lane decisions at intersections and violating turn restrictions driving into dangerous situations. This needs to get fixed ASAP. I will continue posting examples. @teslasi
FSD is not reading and heeding turn lane restriction signs or pavement markings. It is also not planning for them in advance.
Tesla FSD v14.3.10 critical intervention:
FSD repeatedly signaling to change lanes across quadruple white line HOV divider. No nearby exit. This is a regression. FSD used to handle this stretch of HOV just fine and wait until the exits where the line goes dashed. I had to un-signal many times.
I encounter many regressions that the Early Access influencers just are not seeing. I wish I had the one click report button. I can't take the time to disengage, mic, report, and re-engage during this critical moments.





