FSD v12.x (end to end AI)

[PianoAl]

The chess analogy is encouraging. For years people wondered if a computer would ever be able to beat a human at chess. Many people said no.

Currently, no one can beat the best computers at chess.

 

[powertoold]

The decision wobble is the most concerning flaw in 12.2.1.

I can't intuit a fix for it. It might be the Achilles' heel for V12 end-to-end on HW3.

The funny thing is that I think Tesla's attempt to mitigate the wobble led to the freezing problem at some traffic controls. Tesla likely tried to incorporate some "memory" in the decision-making (perhaps taking some past frames into consideration), so at times, the car just remembers it needs to stop and just keeps stopping.

The other funny thing is that the decision wobble is also closely related to lane selection decisions. After all this time, lane selection is still kicking us in the ass!

In the many videos I've seen, the wobble seems much less obvious with HW4 (mostly Omar videos), so there's hope there, but who knows

Tesla was so prescient in displaying this sculpture years ago. It seems the decision fork is still the biggest issue plaguing FSD:

 

[Daniel in SD]

All,

Before v12 came out, it was clear that the first release of v12 wouldn't be perfect. Tesla still has a lot of number crunching to do. Issues with the first iteration have no bearing on the success of the approach longer term.

In the beginning, champion Go players could beat DeepMind's AI-powered Go initially too. Then with further training it became unbeatable. Then it was doing things no expert Go player had ever seen...

Where does this fit in to the analogy?

Human player defeats AI at the board game Go by finding and exploiting a weakness in the system | cApStAn

In a series of matches held in February 2023 Kellin Pelrine, an American research scientist intern at FAR AI, defeated an AI system at the board game Go. The win by Pelrine was a surprise reversal of the 2016 computer victory by “AlphaGo” against world champion Lee Sedol. AlphaGo was the first...

www.capstan.be

I feel that end to end will never be able to handle enough edge cases to achieve one severe collision every few million miles (current human driven Tesla performance).

 

[Daniel in SD]

I am so thirsty.

The focus here is incredible. It’s good that they are not bothering with simulations. Would not want to pollute the training data I suppose.
You’d think they could check the results with simulations though.

I guess they are probably just generating training data.

I admit that if it does this with high reliability with significant traffic, I am going to be amazed. I especially don’t think it will be able to deal with turning traffic.

Overall I am bullish on my beer.
View attachment 1025894

I think they’re going for two nines of performance.

 

[mongo]

I feel that end to end will never be able to handle enough edge cases to achieve one severe collision every few million miles (current human driven Tesla performance).

If it can detect objects and trajectories reliably then it can avoid causing severe collisions.
End cases shouldn't play into that.

 

[JHCCAZ]

strange even more, the latest 2024.2 is 100, a big jump
2024.2.8 to 2024.2.100

I feel it coming, V12

LOL I'd like to support your hope, but I'm afraid it's not a sign of the Twelfth Coming.

Three-digit suffixes, usually .100, sometimes .200, .300 or .400, are used by Tesla for temporary factory software branches for new cars.

Standard software versions need to handle the ever-growing list of hardware configurations built by Tesla. When they make a running change in manufacturing that requires slightly different programming, perhaps substituting a different door lock module or a new vendor chip in the motor control board - things like that - it takes some time to incorporate the new configuration into the lookup tables or case statements that will need to be part of standard software releases is going forward.

Of course they don't want to wait for the next software release, nor do they want to rush it or bypass regression testing of the fleet software. So they solve this problem by creating a temporary customized software build, just to support the latest changes and keep shipping the cars. (Some people call it "break-in software", but I don't think there's any evidence that a traditional machinery break-in period has anything to do with it.)

Some new Tesla cars don't have a three-digit build out of the factory, but I think it's more common than not. It usually takes anywhere from a couple of weeks to a couple of months for the initial three-digit software to get updated, after which the car is on one of the normal software tracks with everyone else. In the meantime, some number of cars show up with these three-digit builds and they get reported on TeslaFi and Teslascope.

 

[Daniel in SD]

If it can detect objects and trajectories reliably then it can avoid causing severe collisions.
End cases shouldn't play into that.

If that was all that was necessary we would have had robotaxis a decade ago.

 

[AlanSubie4Life]

I think they’re going for two nines of performance.

If you want, you can accept defeat now, and we can go double or nothing on one 9.

It does seem like this is a problem AI could solve. But it is possible it is more complex than Go. Everything is so arbitrary. Infinitely more configurations. Crazy difficult. Though quite straightforward for a human.

Guess we’ll see. I want it to succeed.

 

[Todd Burch]

@AlanSubie4Life ,

Can you man up and explain why you give a "laugh" emoji for every post I make? Pretty lame behavior.

 

[AlanSubie4Life]

@AlanSubie4Life ,

Can you man up and explain why you give a "laugh" emoji for every post I make? Pretty lame behavior.

The validity of the analogy of Go & Chess has been discussed quite a lot. We don’t know whether it is applicable here. And notably we’re not dealing with hardware where the complexity can be scaled. I am extremely weird. So I laugh when I see people continue to bring it up as an analogy for the progress that will be made here with HW3.

Don’t get me wrong: I am certain that over the long term (excluding HW3 and HW4) much progress will be made, and this problem might even be solved. There could be breakthroughs that address the fairly evident shortcomings of the current approach (for a robust solution). In five years it will probably be really good, though existence of personal Robotaxis at that point is debatable.

It’s just an emoji. It’s important for some people to be optimistic. Keeps the human race moving forward.

 

[Supcom]

strange even more, the latest 2024.2 is 100, a big jump
2024.2.8 to 2024.2.100

I feel it coming, V12

.100, .200, etc. are versions put on cars prior to delivery. There's nothing special about these and you should ignore them.

 

[Supcom]

It’s good that they are not bothering with simulations.

You don't know that. They very likely are doing a lot of simulation work as well.

 

[JulienW]

strange even more, the latest 2024.2 is 100, a big jump
2024.2.8 to 2024.2.100

I feel it coming, V12

100s are ONLY factory delivered software. All new Tesla's come with 100 (sometimes 200 or 300) software at delivery when new. This has nothing to do with updates. Been this way for years.

You have bought a Tesla so you would have got this when new.

 

[Supcom]

If that was all that was necessary we would have had robotaxis a decade ago.

Well, there is a difference between avoiding collisions and getting to a destination.

 

[AlanSubie4Life]

You don't know that. They very likely are doing a lot of simulation work as well.

I didn’t say they weren’t doing simulations!

Have to read between the lines.

To be clear: I think they do simulations.

 

[kpanda17]

100s are ONLY factory delivered software. All new Tesla's come with 100 (sometimes 200 or 300) software at delivery when new. This has nothing to do with updates. Been this way for years.

You have bought a Tesla so you would have got this when new.

thanks for reminding me, yes I remember something like 44 or 38.100 back in Dec

 

[kpanda17]

won't it be amazing when Chuck gets his V12 and able to do that U turn he always shows us after the unprotected left turn
I'm a Chuck FSD fan

 

[JulienW]

won't it be amazing when Chuck gets his V12 and able to do that U turn he always shows us after the unprotected left turn
I'm a Chuck FSD fan

Too bad Elon's not. He might let Chuck get it early at some point but Elon can be petty and unforgiving.

 

[powertoold]

won't it be amazing when Chuck gets his V12 and able to do that U turn he always shows us after the unprotected left turn
I'm a Chuck FSD fan

It will be able to do that U-turn, no doubt

 

[FSDtester#1]

100s are ONLY factory delivered software. All new Tesla's come with 100 (sometimes 200 or 300) software at delivery when new. This has nothing to do with updates. Been this way for years.

You have bought a Tesla so you would have got this when new.

As I was screwing in my license plate picking my new S up an update was pending.
So I got my first update before I drove my car 1 inch.