FSD v12.x (end to end AI)

[drtimhill]

Eventually the neural network will generalise to unknown failure modes also.

I've no idea what this can mean. The NNs do not, at present, learn by themselves. I dont see that happening for decades, and even then it will be very risky and tricky, since it would learn BAD things as equally as GOOD things.

 

[JulienW]

Wonder how much FSD Beta experience The Real Joey Davila has? I bet not a lot. Seems Tesla made a mistake by not releasing it to only original Early Access testers to start. New users just don't have the experience of knowing when to intervene and trusting an all new v12 system. Tesla should have been better at knowing the unknowns that could and will happen.

 

[kabin]

Can they adjust the matrix weights?

Yep. They might need to present more sets of data or new data but in the end the weights are tweaked and fingers are crossed that regressions are minimized.

 

[swedge]

I think (hope) the same will be true for autonomous driving. Over time, AVs will get better and safer as companies continually improve the software. The key will be for regulators and society not to pull the plug on AVs where there is an accident. We need to deploy responsibly but be willing to accept some risk and learn from failures. I hope eventually AVs will be like airplanes, where you summon a driverless car and you can sleep in the back seat with confidence that a chance of serious injury or death is virtually nil.



I think passenger comfort is important but it is a secondary priority after safety. Safety comes first. Once your safety is "acceptable" then you focus on passenger comfort. I do think eventually, once multiple companies have AVs that are "safe enough", passenger comfort, convenience, availability and cost, will be the main factors that separate different AV products.

I agree.

My hunch is that since safety is the first priority, FSD needs lots of experience with situations which require quick reactions so it can learn how to respond successfully. So, like test pilots who must fly to and beyond the safe edge of the envelope, FSD beta needs to allow situations to develop more than a seriously defensive driver would. After it gets good at dealing with close calls would be the time to work on smooth, comfortable driving.

The other thing I think is going on is that Tesla accumulating tons of miles of FSD driving, as well as accident data. Accident rates are measured in small numbers per millions of miles, so it takes many, many millions of miles to establish statistically significant comparisons which can be used in regulatory negotiations and liability defenses. And in defensible sales pitches.

Perhaps the current high pricing of FSD facilitates these opposing goals of collecting lots of miles while the driving is not very comfortable: The pricing selects for people with cash and attention to spend on a experimental version of a product still under development. Like the old Safety Score, the pricing is throttling back the uptake of FSD while weeding out the lazy and under-insured.

We are only slightly past the Wilbur and Orville phase, perhaps in the unpressurized prop driven airsickness phase. Complaining that FSD is not like a 777 (with it's crew of 15 and hot meals) is a bit premature, I think. It is worth noting the opportunities for improvement, but demanding perfection is unproductive, as is demanding comfort before safety is tested and proven.

 

[kabin]

Wonder how much FSD Beta experience The Real Joey Davila has? I bet not a lot. Seems Tesla made a mistake by not releasing it to only original Early Access testers to start. New users just don't have the experience of knowing when to intervene and trusting an all new v12 system and knowing the unknowns that will happen, Tesla should have been more selective.

There's some rock star TMC beta testers with strikes and some have reportedly struck out. Hopefully the latter is more related to vehicle/system anomalies.

 

[AlanSubie4Life]

Wonder how much FSD Beta experience The Real Joey Davila has? I bet not a lot. Seems Tesla made a mistake by not releasing it to only original Early Access testers to start. New users just don't have the experience of knowing when to intervene and trusting an all new v12 system and knowing the unknowns that will happen, Tesla should have been more selective.

A fair amount (since at least August 2023), and with a few forced disengagements. Seems like at a minimum they should have not rolled it out to anyone who had ever had a strike! Seems like a pretty easy and very relevant screening for early releases! It's fine for people who have "inadvertently" due to "hardware issues" to be kept out of the first rollout since they got a strike.

 

[diplomat33]

I've found with my passengers & wife the comfort issues are actually often about safety. Sure there are the rough hesitations & very unnatural behaviors which are annoying and not safety events. But in recent experiments with my brothers riding along, they commented and had a lot of pucker factor in situations where FSD was not driving defensively - which is a clear safety issue (e.g. not bumping lane position on freeway around other vehicles, driving in blind spots, cutting it close to vehicles about to reverse). Another example of a safety issues is surging forward at unprotected left turn with oncoming traffic too close (even though it probably would have used the buffer lane), resulting in disengagement. All of these situations dramatically raised the stress level of the passengers because it was clear to them that FSD v11 was not capable of driving like a competent human. They compared it to a very bad learner driver (one of my nephews recently learned to drive).

Passenger comfort and safety can be related. But when I talk about passenger comfort, I am talking strictly about things like jerkiness, hard braking and too fast acceleration, that sort of thing, that are not safety issues per se. What you mention, the sudden surging forward with oncoming traffic, certainly creates bad passenger comfort but I would put it in the safety issue category first since it directly relates to unsafe driving. It is unsafe driving that is also uncomfortable rather than driving that is safe but uncomfortable.

 

[Scott Chapman]

TeslaFi is still showing two people who haven't installed v12. How is that eve possible?

 

[JulienW]

A fair amount (since at least August 2023), and with a few forced disengagements. Seems like at a minimum they should have not rolled it out to anyone who had ever had a strike! Seems like a pretty easy and very relevant screening for early releases! It's fine for people who have "inadvertently" due to "hardware issues" to be kept out of the first rollout since they got a strike.

Even more thought. I like the "no strikes" idea and puts me in the running. But maybe Tesla should have even brought back the Safety Score for a couple of weeks and required the same 100 it did to start. Like everyone I hated it (and worse I live in the middle of a city) but would have fought through it to get v12.

 

[JulienW]

TeslaFi is still showing two people who haven't installed v12. How is that eve possible?

Vacation or OoT.

 

[drtimhill]

Passenger comfort and safety can be related. But when I talk about passenger comfort, I am talking strictly about things like jerkiness, hard braking and too fast acceleration, that sort of thing, that are not safety issues per se. What you mention, the sudden surging forward with oncoming traffic, certainly creates bad passenger comfort but I would put it in the safety issue category first since it directly relates to unsafe driving.

Burt isnt it more complex than that? You really have three categories: (a) "pure" comfort, where the passengers dont feel unsafe, they just feel jerked around, (b) "imagined" safety issues, when the car is executing a technically safe maneuver that would make a human uneasy (e.g. squeezing through a narrow gap where a human would be anxious), (c) "real" safety issues, when the car attempts something that is or could be dangerous.

I think there is overlap between all these categories, but the overlap between (b) and (c) is problematic, since it undermines confidence it the ADAS system (examples like driving closer to the car in front). In effect, the ADAS needs to be "de-tuned" so that it simulates a human driver even in circumstances when it could do better. Turning left into oncoming traffic comes to mind. The car can do the math to figure out it has plenty of time (and be right), but scare the passengers (and oncoming traffic) half to death since humans need a wider margin of safety.

Ultimately of course you end up with something like this:

 

[Daniel in SD]

Burt isnt it more complex than that? You really have three categories: (a) "pure" comfort, where the passengers dont feel unsafe, they just feel jerked around, (b) "imagined" safety issues, when the car is executing a technically safe maneuver that would make a human uneasy (e.g. squeezing through a narrow gap where a human would be anxious), (c) "real" safety issues, when the car attempts something that is or could be dangerous.

I think there is overlap between all these categories, but the overlap between (b) and (c) is problematic, since it undermines confidence it the ADAS system (examples like driving closer to the car in front). In effect, the ADAS needs to be "de-tuned" so that it simulates a human driver even in circumstances when it could do better. Turning left into oncoming traffic comes to mind. The car can do the math to figure out it has plenty of time (and be right), but scare the passengers (and oncoming traffic) half to death since humans need a wider margin of safety.

Ultimately of course you end up with something like this:

This certainly isn't a problem yet since FSD beta can't even estimate distances or speeds as well as a human. See the recent example of it hitting a stationary car and my bet with @AlanSubie4Life that it will get a single "9" of performance on Chuck's ULT.
If this issue does become relevant then Tesla should still operate within the safety envelope of humans until they remove the need for humans to monitor the system.

 

[Todd Burch]

Wonder how much FSD Beta experience The Real Joey Davila has? I bet not a lot. Seems Tesla made a mistake by not releasing it to only original Early Access testers to start. New users just don't have the experience of knowing when to intervene and trusting an all new v12 system. Tesla should have been better at knowing the unknowns that could and will happen.

That guy was pretty reckless to let that happen. It’s not like the car was moving fast or made an abrupt change of direction…

 

[Daniel in SD]

That guy was pretty reckless to let that happen. It’s not like the car was moving fast or made an abrupt change of direction…

He probably thought it was an "imagined" safety issue.

(b) "imagined" safety issues, when the car is executing a technically safe maneuver that would make a human uneasy (e.g. squeezing through a narrow gap where a human would be anxious)

 

[Wnuk]

Being that is just dash cam video of the incident, how can it be verified FSD was actually in use?

 

[Pdubs]

Wonder how much FSD Beta experience The Real Joey Davila has? I bet not a lot. Seems Tesla made a mistake by not releasing it to only original Early Access testers to start. New users just don't have the experience of knowing when to intervene and trusting an all new v12 system. Tesla should have been better at knowing the unknowns that could and will happen.

I would like to shout this from the top of the launch pad. this friends n family sht is slowing down the development. get it in our hands

 

[stopcrazypp]

The "crash" video is up. Unfortunately just dash cam footage, so no actual FSD Beta visualization to confirm it was running. That, and the story is nonsensical. His friend is claiming he was expecting V12 to perform a "3-point turn."

If this really was on V12, the driver must have been asleep or otherwise distracted to allow such an obvious, slow-paced accident.

https://twitter.com/x/status/1761811308176196082

FSD Beta doesn't make 3 point turns, right? Nor U-turns until recently? Or have people experienced it making 3-point turns and expect it to do so?

 

[OxBrew]

I thought you said you were done with Tesla and FSD?

lol, more of an "I'm over it" sense. Still following FSD development, since there's so many on the road, need to watch out for them.
Still have to keep the Y for 18 months or more, until a real option is on the Supercharger network. New smaller Rivian maybe, IDK.

Sorry, this is getting really OT, wont happen again.

 

[JulienW]

FSD Beta doesn't make 3 point turns, right? Nor U-turns until recently? Or have people experienced it making 3-point turns and expect it to do so?

It doesn't reverse at all (a tiny rollback of an inch or so can happen) so one more reason this person likely shouldn't have v12.

 

[TK211X]