FSD v12.x (end to end AI)

[kabin]

I can almost taste the beer. We might even see Chuck's neighborhood is the perfect location to start the robotaxi service.

https://twitter.com/x/status/1754879057756315837

Wow! How many times have they collected data for Chuck's unprotected left? About 6 months ago they spent a number of days driving it and the results were still awful. The search for the mythical 'perfect driver' data continues and that's just 1 intersection.

Focused testing my arse. More like overtraining in desperation.

 

[aronth5]

Wow! How many times have they collected data for Chuck's unprotected left? About 6 months ago they spent a number of days driving it and the results were still awful. The search for the mythical 'perfect driver' data continues and that's just 1 intersection.

Focused testing my arse. More like overtraining in desperation.

Since any results were for V11 and not V12 your comment is now meaningless. Let's be patient and see how well FSD V12 does at Chuck's UPL intersection.
If V12 does poorly then hammer away at Tesla.

 

[Matias]

Incredible post, incredible insight

https://twitter.com/x/status/1695562811190210890

He works for Tesla.

 

[Usain]

So back to that question: how does one train for this with end-to-end? Your answer was simulations - hopefully that will work! (I think we kind of talked past each other here.)

Training for what to do when a pedestrian is thrown under the car can be done with simulations.

But as I said before, dragging the pedestrian isn't really what got Cruise shut down. Even a human can be forgiven for not making the correct split-second decision after a pedestrian is thrown under the car.

The thing that got Cruise in trouble is that they allegedly tried to hide what actually happened.

 

[Mardak]

I believe that Tesla can accomplish the launch in only 1 to 3 years

Do you think the time from now to that potential launch of a Tesla robotaxi service is primarily due to getting 12.x safe enough to deploy in some areas? Where presumably once Tesla sees end-to-end is on the right track with continuous iterations and training, they can prepare this service in parallel to be ready when 12.x is ready to remove the driver? Even before the launch, what indicators do you think we'll be able to see with 12.x even still as driver assistance feature or what I was getting at earlier of maybe we'll see existing owner robotaxi-type functionality before a Tesla-run fleet?

 

[Usain]

Do you think the time from now to that potential launch of a Tesla robotaxi service is primarily due to getting 12.x safe enough to deploy in some areas? Where presumably once Tesla sees end-to-end is on the right track with continuous iterations and training, they can prepare this service in parallel to be ready when 12.x is ready to remove the driver? Even before the launch, what indicators do you think we'll be able to see with 12.x even still as driver assistance feature or what I was getting at earlier of maybe we'll see existing owner robotaxi-type functionality before a Tesla-run fleet?

Yes, I think the launch will be due to getting 12.x safe and reliable enough to deploy in at least one city within 1 to 3 years.

And yes to your second question as well.

From what Tesla has said, it sounds like they intend to run the service and owners will be allowed to join in. It will be like Uber or Lyft in that regard.

 

[AlanSubie4Life]

Training for what to do when a pedestrian is thrown under the car can be done with simulations.

But as I said before, dragging the pedestrian isn't really what got Cruise shut down. Even a human can be forgiven for not making the correct split-second decision after a pedestrian is thrown under the car.

The thing that got Cruise in trouble is that they allegedly tried to hide what actually happened.

Not sure about that. Little A, little B.

 

[Daniel in SD]

Wow! How many times have they collected data for Chuck's unprotected left? About 6 months ago they spent a number of days driving it and the results were still awful. The search for the mythical 'perfect driver' data continues and that's just 1 intersection.

Focused testing my arse. More like overtraining in desperation.

Yep. I just need 90% percent success to get my beer though.
Obviously in the future they’ll need a more scalable solution.

Yes, I think the launch will be due to getting 12.x safe and reliable enough to deploy in at least one city within 1 to 3 years.

Can you quantify what level of performance you expect to see by the end of the year? And what level of performance would be necessary for deployment?

I do agree that solving the Cruise edge case might not be necessary to achieve much greater than human performance. On the other hand Cruise has only driven 5 million miles and they nearly killed someone. The human driven fatality rate is 1 per 100 million miles.

 

[aronth5]

How do you know that V11 similarly configured would not be just as assertive and smooth?

I'm not saying it would be - just pointing out we're not comparing apples to apples (released version, vs. alpha version in limited release to "vetted" (and I use the term loosely) users).

Of course we don't know until individually each of uses V12. I'm basing this optimistic viewpoint based on AI Drivers feedback where he agreed V12 was a significant upgrade and now Omar's father who uses V11. He said the same thing. And for now I'm ignoring Omar's feedback because he always says every version is great.

As to your question on V11 being similarly configured that is a non starter. How can V12 be similarly configured when V12 doesn't use 300k+ lines of code and makes driving decisions completely differently?

 

[AlanSubie4Life]

How can V12 be similarly configured when V12 doesn't use 300k+ lines of code and makes driving decisions completely differently?

I am just referring to the slider settings. Want to have those adjusted in a comparable manner and it seems unlikely that they are.

That being said the 12-second gap being missed suggests maybe it's still pretty conservative.

 

[aronth5]

I am just referring to the slider settings. Want to have those adjusted in a comparable manner and it seems unlikely that they are.

Your assuming the slider settings are the same? Why? For all we know V12 either doesn't use comparable slider settings or uses completely different settings. For that matter do you or anyone on this thread even know the details on how slider settings work? We'd need a Tesla FSD developer to share that information before speculating.

 

[powertoold]

Tesla has no incentive to quickly roll out mass robotaxis

L2 FSD will be very lucrative once V12 is dialed in and smooth as butter in all situations

BTW, I challenge anyone to record a 30-minute city manual-drive following the 0mph rule at all stop lights and signs, and then let us critique your driving. I've tried to drive like V12 is required to do by NHTSA, and it's very difficult to do consistently.

 

[AlanSubie4Life]

Your assuming the slider settings are the same? Why?

No, that was my point.

For that matter do you or anyone on this thread even know the details on how slider settings work?

Nope, no one knows.

 

[AlanSubie4Life]

BTW, I challenge anyone to record a 30 minute city manual drive following the 0mph rule at all stop lights and signs, and then let us critique your driving.

No one cares about it coming to 0mph at stop signs. It's all good. Nothing wrong with it.

 

[powertoold]

No one cares about it coming to 0mph at stop signs. It's all good. Nothing wrong with it.

I'm just referring to the difficulty of curating "good" driver data.

Try following the 0mph rule in a parking lot

 

[AlanSubie4Life]

Try following the 0mph rule in a parking lot

One of the places it is more likely to be followed. Just kind of natural.

 

[powertoold]

One of the places it is more likely to be followed. Just kind of natural.

Really? In an empty parking lot, people will stop to 0mph at stop sign lines?

I urge people to try it out for 15 minutes in the city. If you have supreme confidence, record it

 

[JHCCAZ]

We'd need a Tesla FSD developer to share that information before speculating.

May I remind you this is TMC

 

[MP3Mike]

One of the places it is more likely to be followed. Just kind of natural.

Not at all, even I can't stand to fully stop at stop signs in a parking lot when not necessary. But I do at stop signs on streets.

 

[AlanSubie4Life]

Not at all, even I can't stand to fully stop at stop signs in a parking lot when not necessary. But I do at stop signs on streets.

Huh? Talking about difficulty. Already only going 5mph so it is pretty trivial.

Note the apparently retroactive addition of “empty” to this conversation.

Anyway I guess everyone is different, but it just seems easier and more expected to stop in typical parking lots since it is basically zero cost.