When AI meets the real world. FSD tester and progress tracker, Builder of @CamCutApp dashcam manager and telemetry visualizer. Tesla fan
Joined November 2013
- Tweets 11,896
- Following 423
- Followers 15,262
- Likes 28,267
964 Photos and videos
Pinned Tweet
May 30
I love FSD so much. Mesmerizing night drive through the LA freeways, with the gloomy skyline in the background and Heart Drives on the track.
7
3
42
4,211
Jun 12
Impressive restraint by FSD 14.3.3 on this unprotected left turn as the light turned yellow.
The approaching vehicle looked like it might stop, but it ended up running the red light. A less patient turn here could have ended very differently.
Telemetry visualized with @CamCutApp
2
2
82
10,734
Edge Case retweeted
Jun 11
Some lane picking indecision is back in v14.3.3 affecting smoothness of the ride
Jun 11
Gotta report warts when I see ‘em.
FSD 14.3.3 had a freakout at the tollbooth. As you can see, it couldn’t decide where to go. I’ve gone through many times without issue.
Any thoughts as to what might cause this kind of behavior? @Tesla_AI
Also thanks @CamCutApp!
1
1
12
2,162
Jun 11
The 14.3 branch of FSD is noticeably more assertive at yellow lights. In this case it committed and cleared the intersection, though it had plenty of room to stop. The adjacent vehicle continuing through the yellow made the decision much more comfortable shielding us from the van making the unprotected left turn.
5
1
43
4,069
Jun 11
FSD testers are the real QA engineers for Tesla. 🫡
FSD reaction time testing @ ~45 mph
@DavidMoss and I were determined to kill this dummy
2
1
27
2,602
Jun 10
Another user @realCarnasis submitted this crash footage of his Tesla being T-boned while FSD was engaged. Raw video analyzed using @CamCutApp telemetry visualization.
The other driver was ultimately found at fault after entering the roadway and colliding with the Tesla while FSD v14.2.2.5 was engaged.
What makes this crash interesting is that there appear to be multiple contributing factors:
• The driver was pressing the accelerator before the collision, causing the vehicle to accelerate to roughly 40 mph. Had the throttle not been overridden, FSD may have approached the area at a lower speed.
• A portion of the striking vehicle was visible between the stopped cars before it fully emerged into the roadway. FSD did not react until the vehicle entered its path, leaving only a few hundred milliseconds to respond.
By the time emergency braking began, there simply wasn’t enough distance left to avoid the collision.
The other driver caused the crash, but it’s an interesting example of how throttle overrides, speed, and anticipation of partially occluded vehicles can all influence the outcome.
@Tesla_AI
64
9
406
61,622
Jun 10
Driver and passenger were unharmed.
$20,000 of damage $3500 in PPF and tint. Took about 3 months to get it repaired.
6
37
3,578
Jun 10
Joe shared the raw footage with me and fed it through @CamCutApp
Based on the telemetry it does appear that FSD turned the wheel back in the last moment before the driver took over. This happened really fast and probably pretty scary to experience.
FSD seems to have reacted to a phantom orb and veered towards the oncoming vehicle for a moment.
Worst FSD critical takeover I've ever had – potentially fatal – and I've been an OG beta tester across multiple vehicles.
On 14.3.3, the car appears to misread round red glare spots as a small animal and jerks hard into oncoming traffic. Had I not intervened, a head-on was likely. Reported as critical.
This is the regression you get when you raise the gain on small-animal detection: a phantom obstacle triggering an evasive swerve into oncoming cars.
Exactly why we still need supervision and why unsupervised FSD for owners will need FSD 15, at least.
@Tesla_AI @aelluswamy @wholemars @SawyerMerritt @AIDRIVR
37
14
320
112,644
Jun 10
Ok there seems to be about a 27 degree lag from when the wheel is yanked until FSD telemetry logs disengaged state. See wheel yank test below.
This means that @DocBorelli did indeed turn the wheel when FSD was still heading toward the car.
Thanks @DevinOlsenn for pointing this out!
1
1
38
5,177
Jun 10
FSD pulled onto the grass to let another car pass, but failed to recognize the drop-off hidden by the grass and bottomed out.
Jun 10
This is a very serious safety issue for FSD in China and needs immediate attention.
When tall grass or dense vegetation grows very close to the edge of a narrow road, FSD can sometimes misinterpret it as part of the drivable surface. In reality, the road often ends abruptly, with a drop-off or unsupported edge beyond it.
During oncoming traffic situations, FSD may move too far to the side and attempt to drive beyond the actual road boundary. In my experience with V14, I have already had to perform three critical interventions because the vehicle was about to drop a wheel off the road edge.
This is not just a comfort issue—it can potentially cause wheel, suspension, or even battery damage, making it a significant safety concern that should be addressed as soon as possible.@edgecase411 @wholemars @aelluswamy @elonmusk
11
1
44
5,450
Jun 8
Below is official crash footage of my dad being rear-ended while using FSD v12.6.4 on his HW3 Model 3 earlier this year. The telemetry overlay was created using the new crash visualization tools in @CamCutApp v1.1.
Unfortunately, this experience completely destroyed his trust in FSD, and it’s going to be difficult to earn that trust back.
In this case, traffic ahead was slowing down normally. There was no sudden stop or hard braking. For reasons that aren’t entirely clear, FSD delayed braking until Automatic Emergency Braking (AEB) intervened and forced a much more aggressive stop. (Watch the autopilot status switch to “Aborting”) The driver behind failed to react in time and rear-ended my dad’s car.
One of my biggest concerns with FSD v12 on HW3 is that its reaction timing could occasionally be inconsistent, making these situations difficult to supervise. Everything can appear normal until the system reacts a second or two too late.
FSD v14 on AI4 has been a significant improvement, which is one of the main reasons I upgraded seven months ago. But i think it’s important not to forget that there are many people out there that are still using this outdated software and should be warned about its shortcomings.
To my knowledge, CamCut is the first commercially available tool that can combine Tesla crash footage, telemetry logs, and official crash report data into a synchronized visualization of the entire event.
The crash analysis tools are currently in beta. If you have Tesla crash footage you’d be willing to share to help improve the feature, I’d greatly appreciate it. Feel free to DM me.
189
11
381
100,083
Jun 8
@DevinOlsenn this is what happens when AEB takes over controls from FSD. It’s red hands event on the screen
17
6,126
Jun 8
Watch out @nursedanakay I’m coming for your views! 😂
My Tesla with 14.3.3 will brake stab for a squirrel on the shoulder 100 yards ahead, yet yeets me nearly 30 mph over speed bumps. 🤷♀️
5
1
20
4,037
Jun 7
There is simply no scenario in the United States that can match this level of chaos.
The training needs to come from local drivers in China.
FSD is very timid getting through the crazy cross traffic and it really annoys the drivers behind.
It’s safe but it also needs to keep moving x.com/CyberTshr/status/20629…
4
5
32
2,642
Jun 7
Insanely quick FSD reaction time when a pedestrian suddenly jumps out in front of it.
From the moment the person’s head first becomes visible from behind a parked car to the moment FSD starts braking: 0.17 seconds.
This dashcam video clip, including the reaction time measurement, is available in the new @CamCutApp v1.1. Anyone can download it and explore the full project themselves.
@bilalsattar and I ran a series of reaction tests a few weeks ago. You really have to be inside the car to get the full experience
15
59
469
46,470
Jun 4
FSD reacted quickly and swerved once the vehicle started pulling out of the driveway, but a human driver anticipates it earlier from the headlights reflecting across the pavement before the vehicle was visible.
Training FSD to recognize indirect cues like headlight shadows could help it slow down sooner and be better prepared for emerging hazards. @Tesla_AI
10
10
152
13,851
Jun 3
FSD appears to have trouble detecting barrier arm from a distance. It reacted only at the last second at high speed, attempted to move over, and ended up striking it.
This looks like a supervision failure as much as an FSD failure.
227
34
765
224,944
Jun 3
Be honest how many of you are gonna let this slide to keep your FSD streak alive
81
3
155
34,089
Jun 3
I’ve been waiting to see videos like this out of China.
Dense urban environments with constant cross traffic, scooters, three wheeled bikes, and pedestrians expose one of FSD’s biggest weaknesses: when everything is moving at once, it can become overly timid. Sometimes the only way forward is to confidently and assertively claim space.
Jun 3
A portion of the livestream that many people have been asking for: FSD V14 handling complex intersections and challenging road scenarios in China. The audio has been translated into English.
Follow me for more videos like this
#FSD @Tesla_AI @Tesla @edgecase411 @wholemars
3
2
38
4,772
Jun 2
Great compilation of edge cases from China. FSD can literary drive anywhere
4
5
33
3,384