Profound progress towards FSD

[Grifon]

During today’s Q2 earnings call, Elon mentioned several times how the new 4D version of FSD is a profound improvement over the current 2.5D stack. He he a damn good salesman and has me convinced.

It will be interesting to see what Tesla can achieve in the next 6 - 12 months.

 

[bd7349]

Also worth noting that he’s still confident in FSD by end of the year. He said he’s driving the current alpha builds ever day and it takes him through complex intersections, narrow streets, construction and can almost make it with zero interventions. Sounds like they’re trying to get the number of interventions down before wide release in a few months. Exciting times!

 

[diplomat33]

During today’s Q2 earnings call, Elon mentioned several times how the new 4D version of FSD is a profound improvement over the current 2.5D stack. He he a damn good salesman and has me convinced.

It will be interesting to see what Tesla can achieve in the next 6 - 12 months.

Yes, he is a great salesman for sure. I want to get excited but I've been burned before. Let's wait and see.

It will be interesting to see the rewrite in our cars. I have no doubt that it will be a big improvement to what we have now. Video instead of images is a must-have to do FSD.

Getting the reliability so good that there are zero interventions will be the super hard part, especially for complex urban driving. In fact, it will be the real test of Tesla's camera-only approach to see whether they can get the necessary reliability for true driverless with the current sensors or whether they hit a reliability wall.

Based on Elon saying that he cannot do his commute with no interventions, there disengagement rate is still pretty bad. So Tesla has a long way to go in reliability. Nobody has achieved 100% zero interventions. The question is how long can you go before an intervention. The leaders in FSD (Waymo and Cruise) can go about 10k miles in busy city driving before an intervention. I hope Tesla starts releasing some data on disengagements soon. That will really show their progress after "feature complete" is done.

 

[diplomat33]

Also worth noting that he’s still confident in FSD by end of the year. He said he’s driving the current alpha builds ever day and it takes him through complex intersections, narrow streets, construction and can almost make it with zero interventions. Sounds like they’re trying to get the number of interventions down before wide release in a few months. Exciting times!

He said that he could do his commute from his home to work with almost no disengagements. I don't know how far away Elon lives but that would actually not be a very good disengagement rate.

 

[Mardak]

Well… at least before, it sounded like Elon Musk stays at a hotel close to the Palo Alto HQ, which would be a 2 mile drive from El Camino, so I guess that means on average about 1 mile between disengagements. Although on the bright side, just a few months ago before stopping at traffic lights and stop signs, there would have been a disengagement for ~10 of those intersections, so roughly 0.2 miles between disengagement. So extrapolating this 5x improvement in disengagement, Tesla will be multiple orders of magnitude better later this year!

 

[HenryFarkas]

Also worth noting that he’s still confident in FSD by end of the year. He said he’s driving the current alpha builds ever day and it takes him through complex intersections, narrow streets, construction and can almost make it with zero interventions. Sounds like they’re trying to get the number of interventions down before wide release in a few months. Exciting times!

si
Had two interesting STOP sign experiences in my M3 today. There was a flagman on the left side of the road holding up a STOP sign.. My M3 didn't see him or the STOP sign even though it looked like a regulation STOP sign with octagonal shape, white letters on red background in the regulation all caps sans seriph font.. Later in the same drive, I came across a STOP sign that had been knocked down into the weeds presumably by a vehicle that had crashed through the metal barrier and took out the STOP sign. The sign was invisible, but my M3 knew it was supposed to be there because it warned me that it was going to stop in 500 feet, and it did stop right where the STOP sign used to be before it was knocked over. This makes me think that Tesla programs its cars to stop at mapped STOP signs, but not to recognize STOP signs.

 

[diplomat33]

Well… at least before, it sounded like Elon Musk stays at a hotel close to the Palo Alto HQ, which would be a 2 mile drive from El Camino, so I guess that means on average about 1 mile between disengagements. Although on the bright side, just a few months ago before stopping at traffic lights and stop signs, there would have been a disengagement for ~10 of those intersections, so roughly 0.2 miles between disengagement. So extrapolating this 5x improvement in disengagement, Tesla will be multiple orders of magnitude better later this year!

It's all speculation since we don't know the exact route. A 2 mile commute would not be a worthy test. Most likely, Elon drives around town for a bit, hitting as many intersections as he can, to give the software a decent test. But even he drives around for say 20-30 miles before getting a disengagement, that would still be an atrocious disengagement rate. And I checked google maps for the route from his home in Bel-Air to the Tesla HQ in Palo Alto. That's 352 miles of mostly highway. Even if he did that route with just 1 disengagement, that would still be a pretty bad disengagement rate.

 

[lascavarian]

I don't know how far away Elon lives but that would actually not be a very good disengagement rate.

I guess you are saying that you don't really know. Since we know it is not feature complete then he would have to fill in for those instances where the feature is not fully functional. It is sort of pointless to get ahead of things at this point.

What you call the disengagement rate is what EM seems to refer to as the "March of 9's" and that is yet to come some time after completion of the code rewrite into 3D or 4D from 2.5D and the achievement of feature completeness as I understand it. Still seems to be plenty to do but the plan seems to be moving.

Not everyone is happy with the abilities of autopilot but I am stoked by the progress I have seen lately.

 

[Big Dog]

During today’s Q2 earnings call, Elon mentioned several times how the new 4D version of FSD is a profound improvement over the current 2.5D stack. He he a damn good salesman and has me convinced.

It will be interesting to see what Tesla can achieve in the next 6 - 12 months.

Elon mentioned 'end of year' again. (so, 5 months and counting....)

 

[diplomat33]

I guess you are saying that you don't really know. Since we know it is not feature complete then he would have to fill in for those instances where the feature is not fully functional. It is sort of pointless to get ahead of things at this point.

No. Elon said that the rewrite can do "pretty much everything" and that once the rewrite is done, all is left is improving reliability ("march of 9's"). So I take that to mean that the full rewrite will basically be "feature complete".

What you call the disengagement rate is what EM seems to refer to as the "March of 9's" and that is yet to come some time after completion of the code rewrite into 3D or 4D from 2.5D and the achievement of feature completeness as I understand it. Still seems to be plenty to do but the plan seems to be moving.

The disengagement rate is simply the number of miles that the car can drive autonomously without a disengagement. The "march of 9s" is the process of improving reliability by solving cases that cause disengagements. So they are related but not the same. The disengagement rate can be used to measure roughly the progress of the "march of 9's".

 

[Puddles]

*sigh* the hardest part of these major (sometimes even the minor) revisions is the retraining and reacclimatizing of the driver

 

[Rogue one]

If he says end of year ever year he will eventually be right!

 

[powertoold]

If he says end of year ever year he will eventually be right!

Even if he isn't, hopefully they can at least reduce phantom braking and enable more sharp turns within lane.

 

[animorph]

This is why I wanted to get HW3 in both or cars. I think 2.0/2.5 will see some further development but won't be able to keep up with HW3. even for basic Autopilot.

I assume the new software implements the end-to-end NN processing. After that it's "just" a matter of training it for all the behaviors they want. So yeah, everything will be ready to go full FSD, but the training may take a long time, and take several iterations, or even require better hardware to get those final 9's. But we'll be along for the ride.

 

[powertoold]

Anyone willing to bet that hw3 won't be enough to process all the 4D nets at 36hz? It doesn't seem like they designed hw3 with this new rewrite in mind. Elon would have mentioned it a long time ago.

 

[bd7349]

Anyone willing to bet that hw3 won't be enough to process all the 4D nets at 36hz? It doesn't seem like they designed hw3 with this new rewrite in mind. Elon would have mentioned it a long time ago.

In the WAIC video Elon said they only just recently started utilizing the second computer on the FSD chip. That, to me, makes it sound like they have plenty more headroom to work with.

 

[powertoold]

In the WAIC video Elon said they only just recently started utilizing the second computer on the FSD chip. That, to me, makes it sound like they have plenty more headroom to work with.

The second chip is supposed to provide redundancy though. On Autonomy Day, I believe they said the 2nd chip processes the same data as the 1st, and they check if they agree.

 

[bd7349]

The second chip is supposed to provide redundancy though. On Autonomy Day, I believe they said the 2nd chip processes the same data as the 1st, and they check if they agree.

The way he said it made it sound like they moved on from that and now intend to use it to provide more compute power for FSD neural nets. It’s anyone’s guess at this point about what they’ll do though, we’ll have to wait till this is rolled out and they talk about it more.

 

[thewishmaster]

Anyone willing to bet that hw3 won't be enough to process all the 4D nets at 36hz? It doesn't seem like they designed hw3 with this new rewrite in mind. Elon would have mentioned it a long time ago.

Unlikely they would invest so much into hardware and then not even target it during development

 

[Buckminster]

So the 0.5 in 2.5D is time.

Does everyone agree that Alpha is still mostly 2.5D?