FSD rewrite will go out on Oct 20 to limited beta

[Daniel in SD]

Another vid from Brandon in Sacramento. He will for sure be my baseline from now on about the safety of the system. The new update doesn’t seem to have much for him and the car is still pretty damn dangerous. If/when the beta starts working well for him, I’ll start feeling good about the beta being released to the general public. As is, it can’t be trusted at all.

This is the wrong way of looking at it. Even he drove 1000 hours straight with no problem the car could still not be trusted at all. You need a much larger sample size than any one person is capable of obtaining within a reasonable amount of time.

 

[diplomat33]

No, Waymo is testing their prototypes and mapping the roads in 25 metro areas.
That is not autonomous driving.
There is only one place they are doing autonomous driving without driver supervision, and that is 50 sq miles in the desert.

The cars in the 25 metro areas are using the same autonomous driving software as the driverless cars in Chandler. It is the same autonomous driving software with or without a safety driver. So yes, the test cars in those 25 metro areas are autonomous cars. They don't stop being autonomous just because a person is sitting in the driver seat.

 

[mikes_fsd]

Brandon in Sacramento. He will for sure be my baseline from now on about the safety of the system.

Also, note his tweet/response about FSD.

Meant to say, if he can notice the improvements already, in the incremental releases, that gives me even more hope that this will move quickly.

 

[Silicon Desert]

Hmmm, while there is some interesting information in this thread, it is starting to sound worse than a presidential debate sometimes.

 

[DanCar]

As is, it can’t be trusted at all.

To me that makes it safer. If it could be trusted people would be quick to abuse it.

 

[mikes_fsd]

Hmmm, while there is some interesting information in this thread, it is starting to sound worse than a presidential debate sometimes.

Good thing I do not watch TV...

 

[mikes_fsd]

This is the wrong way of looking at it. Even he drove 1000 hours straight with no problem the car could still not be trusted at all. You need a much larger sample size than any one person is capable of obtaining within a reasonable amount of time.

No, you are thinking about this the wrong way.
One persons driving is never going to be good enough even with 10000 hrs.

The fleet of FSD enabled cars across different geographical and terrain situations will be the true sample size and will measure in millions (at first ) and very quickly billions of miles of data.

 

[john5520]

Why does everyone insist on using the worst possible camera angles for these videos?

Something like this is REALLY easy to set up. If you want, you could move it slightly further back, to allow capture of the mirrors & side views as well. Obviously multiple camera angles would be nice, but doing production to merge the TeslaCam data with the video from inside the car isn't really expected for quick videos. Someone should do it though (assuming TeslaCam is enabled)!

BTW, this was the video that was posted earlier in this thread (guess it could have been a different related thread), showing the man with a bunch of shopping carts, which the car mistook for a truck on the sidewalk, and the one with the last minute navigation reroute which caused the car to try to turn left from the rightmost lane.

It also detected rings painted on the road in the middle of an intersection with traffic lights and a protected left turn lane as a roundabout, a mistake even an experienced driver can make.


View attachment 602365

Yeah, he couldn't even get the camera angle right which made the video a chore to watch. So I checked out his channel for his latest video. Guess what? The angle is still bad with the rear view mirror still in the way. And this is after tons of people mentioned it in his last video. Sigh!

 

[verygreen]

What evidence do you have that Tesla is using HD maps - "centimeter level accuracy" like you said in your tweet??

Yeah, all those road features, the cross walks, stop lines, traffic signs and more are recorded with great accuracy. If you reread what I said in the tweet, you can argue all you want this is not centimeter level accuracy, but it's at least multi-centimeter level or otherwise very precise.

I gotta say, you really love your audience to be obedient and only ask shallow questions, but when someone calls you out on your bs all you have is diversion to a different subject.

how is this a different subject? This is the subject you accused me of pushing. So I am trying to understand how you came up with this conclusion first. You cannot just put words into my mouth and expect me to blindly defend them.

 

[Daniel in SD]

No, you are thinking about this the wrong way.
One persons driving is never going to be good enough even with 10000 hrs.

The fleet of FSD enabled cars across different geographical and terrain situations will be the true sample size and will measure in millions (at first ) and very quickly billions of miles of data.

That's exactly what I'm saying. You can't determine that the system is safe enough to "trust" by watching YouTube videos.
People seem to forget that human drivers go 500,000 miles between accidents.

 

[ElectricIAC]

That's exactly what I'm saying. You can't determine that the system is safe enough to "trust" by watching YouTube videos.
People seem to forget that human drivers go 500,000 miles between accidents.

Huh? I can’t name a single person I know of that has met this lofty metric.

 

[mikes_fsd]

Yeah, all those road features, the cross walks, stop lines, traffic signs and more are recorded with great accuracy. If you reread what I said in the tweet, you can argue all you want this is not centimeter level accuracy, but it's at least multi-centimeter level or otherwise very precise.

Geo-coded maps of intersections, crosswalks and traffic controls do not an HD map make!
You know that. When Karpathy says straight up, to an audience of peers in the race to reach autonomy that "we do not use HD maps", all of them seem to be on the same page.
But for you, you want to stretch the definition (redefine really) for whatever reason. Why?

Karpathy, in the same presentation, goes on to say that they use maps, but not HD maps. So, why would you straight up accuse Karpathy of lying?

how is this a different subject? This is the subject you accused me of pushing. So I am trying to understand how you came up with this conclusion first. You cannot just put words into my mouth and expect me to blindly defend them.

Uhm, you might not have read the actual original post.

1. Either Tesla is not using HD Maps - just like they claim.
2. or, Tesla has figured out how to map every state and road so fast, that they are able to release the HD Maps to all these random cars in 9+ states now?

So, my comment about them having mapped the entire state, was a jab at point #2!

 

[Daniel in SD]

Huh? I can’t name a single person I know of that has met this lofty metric.

In the 3rd quarter, we registered one accident for every 4.59 million miles driven in which drivers had Autopilot engaged. For those driving without Autopilot but with our active safety features, we registered one accident for every 2.42 million miles driven. For those driving without Autopilot and without our active safety features, we registered one accident for every 1.79 million miles driven. By comparison, NHTSA’s most recent data shows that in the United States there is an automobile crash every 479,000 miles.

Tesla Vehicle Safety Report
I believe the number is actually police reported accidents. I'm sure people curb their wheels way more often than that (at least Tesla drivers do!).
If you believe Tesla (which I don't) even if they achieve FSD twice as safe as a human it will still be 5 times worse than a human using Autopilot!
Sometimes I wonder if the people that are so optimistic about FSD are actually really bad drivers.

 

[DanCar]

People seem to forget that human drivers go 500,000 miles between accidents.

I've seen numbers as low as 40K miles between accidents.
Google's Driverless Car Is Now Safer Than the Average Driver

the average U.S. driver has one accident roughly every 165,000 miles. Here's how we got that figure: our average mileage per year is 16,550, according to the Federal Highway Administration; the average length of time we go between traffic accidents is 10 years, according to Allstate.

NHTSA is a small subset of accidents.

 

[verygreen]

You know that. When Karpathy says straight up, to an audience of peers in the race to reach autonomy that "we do not use HD maps", all of them seem to be on the same page.

Well his only definition of HD maps I saw was pretty much "We do some detailed mapping, but we don't consider that to be HD maps"

Well, I do consider those to be HD maps.

Karpathy, in the same presentation, goes on to say that they use maps, but not HD maps. So, why would you straight up accuse Karpathy of lying?

That's just your imagination. Of course by his definition they are not using HD maps because that's just impossible by the said definition.

But like I said in my tweet - you may argue as much as you want about centimeter accuracy and other straws, they are doing high definition maps. they are a lot more detailed that just pure navigation maps. Tesla just does not want to call them HD maps for whatever reason, That's fine. They can call them "not-HD maps" or whatever.

Uhm, you might not have read the actual original post.

1. Either Tesla is not using HD Maps - just like they claim.
2. or, Tesla has figured out how to map every state and road so fast, that they are able to release the HD Maps to all these random cars in 9+ states now?

I am not sure why you think 1 implies 2 and I asked you about it in the previous two posts and you evaded the answer. so I'll ask you again, where was it stated that Tesla's current "FSD beta" does not work in absence of HD maps? The whole claim is they use HD maps when available.

 

[Mardak]

This is my new FAVORITE video of FSD. Nothing in the navigation, the car is literally deciding where to take him. The owner seems just as amused as I am!

[

I'm not sure why I'm as surprised as I am about this realization, but Autopilot is driving primarily on predicted neural network outputs now. Clearly "old" Autopilot has been making predictions about lane lines and predicted path to stay in lane as visualized in verygreen's videos, and at Autonomy Day, Elon Musk highlighted how the neural network predicts curvy lines even when not directly visible - blocked by the hill.

So similarly, Autopilot driving itself without a destination is just another neural network prediction of "most of the time you go straight" or "when the lane splits, bias this way" or "clearly you cannot go straight, so right turn is most likely." That's somewhat the purpose of using a neural network as it learns common behaviors even for new situations it has never been trained on.

Here's another instance of the neural network making predictions of things it cannot see:

By design, many roundabouts are made to not allow the driver to see across, and the neural network knows that roundabouts look a certain way. Notably, the far end of the circle will have a median/curb even if it's physically impossible to see from the current position.

Overall, there are many things the neural network could predict about an intersection even without having clear vision. This gives me hope that even with blurry camera views, e.g., from rain or snow or just dirty, the neural network will be able to correctly infer many things to drive with high confidence.

 

[mikes_fsd]

I am not sure why you think 1 implies 2 and I asked you about it in the previous two posts and you evaded the answer. so I'll ask you again, where was it stated that Tesla's current "FSD beta" does not work in absence of HD maps? The whole claim is they use HD maps when available.

My entire claim, is quoting the person architecting and implementing the entire stack at Tesla -- Karpathy.
By comparison... Your claim of HD maps was quoting a TSLAQ cult member!

I will take Karpathy's definition of HD maps over yours, but I believe I have enough information to filter your posts at this point.
I hope the rest of the forum has been as enlightened as I in these past few "revelations" of yours!

 

[Huskyf]

Another vid from Brandon in Sacramento. He will for sure be my baseline from now on about the safety of the system. The new update doesn’t seem to have much for him and the car is still pretty damn dangerous. If/when the beta starts working well for him, I’ll start feeling good about the beta being released to the general public. As is, it can’t be trusted at all.

Is car have a problem not possible I saw today 30 minutes drive left/turn right turn not protected red turn on strong rain and no arbort !
And if you take a look Brand doesn't sent message when an error occur on FSD to Tesla not pressing the camera icon on the right of the T others videos I saw today test drivers press evertime they saw an error on the drive but Brandon never !

 

[P3dStealth]

Another vid from Brandon in Sacramento. He will for sure be my baseline from now on about the safety of the system. The new update doesn’t seem to have much for him and the car is still pretty damn dangerous. If/when the beta starts working well for him, I’ll start feeling good about the beta being released to the general public. As is, it can’t be trusted at all.

Wow the influencers are making it seem like FSD works and this guys video it does horrible. I don't see this as more useful than me just driving around turning on autopilot myself after turning. As is.... it is not useful at all and it is nowhere near ready for public release. None of those were difficult situations.

 

[mikes_fsd]

That's just your imagination.

I just realized that you might not have seen the presentation. I assumed you had. That is my bad.
Here are the 2 presentaions I am basing my argument of no HD Maps in Tesla's.