willow_hiller
Well-Known Member
I'm not partial to using "AI" myself, but it's just synonymous with "machine learning" these days. Pretty much any ML model that produces impressive results is now labeled as "AI."
-
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 jerky driving fsd late deceleration fsd stops too early too quickly fsd stops too far behind line fsd tailgating in city streets
I'm not partial to using "AI" myself, but it's just synonymous with "machine learning" these days. Pretty much any ML model that produces impressive results is now labeled as "AI."
AI is the new "Cloud" that everyone was using 10 years ago, erroneously.
diplomat33
Average guy who loves autonomous vehicles
Elon is livestreaming V12 on HW3 Model S. He says V12 is driving purely end-to-end from video training, no hardcoding at all. They did not specifically train the NN what a traffic light is, what a stop sign is, or a scooter is, etc V12 "learned" from video training how to drive.
diplomat33
Average guy who loves autonomous vehicles
Watching livestream of V12, it confirms what I've been saying about the E2E approach. It is impressive that the car can learn to drive in a very generalized way based only on video training. It does seem like a good way to "solve" autonomous driving. But I think the key will be getting the right data in order to get the reliability high enough to remove driver supervision. Right now, on the video training so far, V12 can drive from A to B, handling a lot of common cases, including roundabouts, intersections, pedestrians, cyclists, lane changes, etc... But it made some mistakes like driving over the lane line in turns. And the intervention rate is very poor. Elon had his first intervention after only 20 minutes driving when the car thought the green for the turn only lane was its light and tried to go on a red. So I think it is clear that E2E can do autonomous driving, the big challenge will be getting reliability good enough to remove driver supervision and how much training will that require. Is it doable? IMO, yes, the question is when.
willow_hiller
Well-Known Member
Thanks for the recap. Just tuned in, but I cannot tell whether the stream has ended, or whether there are just technical difficulties (on my end via Android app or on Elon's end).
Edit: working on Chrome on PC. Not on Android app.
willow_hiller
Well-Known Member
I guess this answers my question about whether V12 could run on the HW3 NPUs:
And from the livestream, the visualizations still show the usual perception that we've been used to with 11.x, so those existing intermediate predictions are still being run for end-to-end. At 9:15, they talk about how it's already running at full 36fps framerate but theoretically reach 50fps if cameras supported that.
Great summary.
It wasn't Omar in the car btw, maybe Ashok?
Last edited:
For this problem, any novel architecture will take literally years of scut work to tweak and tune it. A big rewrite coming out now means they're pretty far away from a reliable system
OMG they're going to need way more time.
At a minimum there needs to be tons of examples of bad drivers and negative reinforcement for those---and that's hard to get outside simulation obviously.
And even humans need to learn general rules, it's called "driving school" and is not instinctive and natural in neurology unlike 3 year olds learning natural language.
AlanSubie4Life
Efficiency Obsessed Member
Only thing worse than the driving was the video quality! My poor neck.
Light at 18:00 was particularly poorly handled. The car has to squeeze into the other side of the intersection! And then it shouldn't just go on a green arrow as was mentioned.
However, we finally can understand from this video why Elon thinks FSD is good! It's because he can't drive!!! Seems like bottom 10% at best. Who does that? Really bad, for sure.
Light at 18:00 was particularly poorly handled. The car has to squeeze into the other side of the intersection! And then it shouldn't just go on a green arrow as was mentioned.
However, we finally can understand from this video why Elon thinks FSD is good! It's because he can't drive!!! Seems like bottom 10% at best. Who does that? Really bad, for sure.
Demo was a complete disaster. A huge safety disengagement with multiple traffic infractions in just acouple miles of simple driving.
But to the prominent Tesla Faithful it was a declaration of victory.
Just shows how delusional and completely ignorant they are.
Elon could poop out a pile of dung and these Tesla fans would go on a frenzy and lap it up.
It never ceases to amaze me.
https://twitter.com/R6Alex/status/1695310876918595843
But to the prominent Tesla Faithful it was a declaration of victory.
Just shows how delusional and completely ignorant they are.
Elon could poop out a pile of dung and these Tesla fans would go on a frenzy and lap it up.
It never ceases to amaze me.
https://twitter.com/R6Alex/status/1695310876918595843
Last edited:
If the system is told that driving on the road is good, why does it have to be told that driving off a cliff is bad? Because it may have enough weak positive cues (e.g. poorly marked roads, etc) that it would entertain driving off the cliff? If so, can't that be addressed by complaining that the system doesn't have enough confidence in good actions that it requires the driver to take over? Isn't that part of the whole SAE driving levels thing?
This is what is known as the Elon DISTRACTION field.
I’m happy to see his car tried to run a RED LIGHT, that has been my one unacceptable action for nearly two years now.
No, but it mis-interpreted the LEFT TURN lane light that does go green, for the GO STRAIGHT lane that the car was in which had TWO SOLID REDS and started to proceed into the intersection (to to 7 mph), where clearly turning cars from the OPPOSING direction were coming from left to right in front of the direction that the EM TESLA was going to go straight on. So, it was a full intervention, EM even says “ok, intervention” but then says something like ‘let’s not call it an intervention” since I think he possibly correctly interprets it as a total fail, like say running a red light - which it would be considered, or stop sign, etc. It’s about 19:50-21:00
I was bullish before and now even more bullish. I think this might go to customer beta users like yourself in October and rollout to all for holiday update.
willow_hiller
Well-Known Member
This was a live preview of an early version that's only being internally tested right now. It's not being pushed to customers, and it hasn't yet been trained at scale.
If you expected more, you're the one with unrealistic expectations.
Similar threads
