FSD Beta 10.13

[AlanSubie4Life]

and dealing with left turning traffic, which can be hard for a human and is a case you have to watch out for with this strategy!),

Very curious in Clip #2 (Chuck pointed out this special situation), how this would have behaved if traffic in the first set of lanes had been clear, when the left turning traffic was approaching. You could do this turn many times and not have exactly the right situation: 1) near-side traffic clear, but 2) left turn traffic approaching on far side. Car has to perceive and predict that car is turning left, otherwise it could proceed and then have to halt in first set of traffic lanes to avoid a collision. Humans would rely on seeing the left turn signal in this case (or maybe wait for lane #1 on the far side to be empty to minimize risk since people often don't signal) if employing this turning strategy.

And it needs to be pretty quick about getting into the the position in the middle so as to not interfere with anyone turning left. As long as you get in position quickly enough they'll easily see you, and just turn behind you, rather than across you. Right now it takes 7-9 seconds (!!!) for the Tesla to cross the three lanes of traffic and get in position. This is an eternity relative to human performance, FWIW. If you're doing that, you need to be able to see traffic about 400-500 feet away on the far side and see them signaling (or, easier, determine that there is no traffic in that fast lane). Speed that movement up, and it makes it easier to make a safer move and reduces the burden on perception.

Tricky. Probably won't see that very frequent, but at the same time quite rare case tested though, for 10.13.

I have this situation all the time on my UPL, because turning traffic is more common. But visibility is excellent, and I'm always using the all-or-nothing strategy, since I have just four lanes and it's not as busy as this, so it's not an issue to just wait for the all clear. Hoping they don't make the "frogger" strategy a thing for all left turns, as it would potentially increase risk.

 

[clydeiii]

I’m feeling better about my bet but more pessimistic about Tesla’s approach.
If their simulation tools can't handle this scenario then what good are they? This seems like one of the easiest things to simulate.

Seems likely they created 10.13 with simulators, and now they're comparing the simulation to reality and tweaking where necessary.

 

[AutoRocket]

They take your human driving data whether or not you’ve paid for FSD

you can opt out…. supposedly.

 

[AutoRocket]

With 100k in the program and 10s of thousands more "chomping at the bit" to get in we know that's NOT going to happen.

people are Elon shills what can I say? fan bois will degrade themselves and pay for a feature that will never work in the life of their vehicle.

 

[Daniel in SD]

Seems likely they created 10.13 with simulators, and now they're comparing the simulation to reality and tweaking where necessary.

If they still have bugs in their simulator that can be found within hours of real world testing that’s a big problem. I went back and looked at what they said at autonomy day (now three years ago!) and they said the problem with their simulator was edge cases.

 

[tmoz]

people are Elon shills what can I say? fan bois will degrade themselves and pay for a feature that will never work in the life of their vehicle.

Tesla pulled off the ultimate Tom Sawyer: make your chore sound like so much fun, that your friends will pay for the privilege of painting that fence.

 

[BnGrm]

Right now it takes 7-9 seconds (!!!) for the Tesla to cross the three lanes of traffic and get in position. This is an eternity relative to human performance, FWIW.

I’ll take your word on the exact time. I didn’t count it. However, this was the part that bothered me the most. Maybe I just overly speed to the median.

 

[2101Guy]

Tesla pulled off the ultimate Tom Sawyer: make your chore sound like so much fun, that your friends will pay for the privilege of painting that fence.

LOL. But PERFECT analogy.

 

[dckiwi]

people are Elon shills what can I say? fan bois will degrade themselves and pay for a feature that will never work in the life of their vehicle.

This is not an airport, no need to announce your departure.

 

[S4WRXTTCS]

I haven’t requested to “volunteer” for the AI team by paying $12,000. If they want my data to improve their product they can pay me or offer the beta access free….

This whole bit about Tesla leveraging its customer base to do product testing is way overplayed. For most of us the entire data collection is passive where it collects data when people choose to use it.

In the entire history of the FSD Beta only one single person has gotten Tesla's attention to actually attempt to fix how it handles something, and that's the Chucks unprotected left turn.

The majority of us including me simply weren't willing to put work into something unless there was some validation for that effort. You mentioned payment, or free use of a product. Those are fair, but there are more reasons. I would have put more effort into testing it if it led to improvements in places I used it. Or if Tesla ever acknowledged any of it as being worked on or asked for more info. There was zero engagement so I stopped bothering with it.

Most of Chucks work didn't seem to be motivated from a desire to help Tesla. Instead it was more because of the engagement of the community. A lot of the upgrades to his process were for the benefit of the viewer, and not so much for Tesla.

At some point he reached such a following that Elon/Tesla couldn't simply ignore him anymore. In a sense the entire FSD Beta program up to that point seemed to be dominated by influencer types gushing on about how good it was.

How did an involvement of 100,000 people only result in a single test case being a focus? Sure it represented a lot of UPL's, but there are hundreds of other cases as well that are representative of similar problems. Every issue I had with FSD Beta could have been easily reproduced on another road/situation that was similar.

I think it came down to three reasons that Chucks UPL was focused on.

1.) Popularity to the point where Musk himself was made aware of it
2.) It was likely deemed a solvable problem. Now maybe not to the degree that I'd like to see it solved, but I think Engineers at Tesla saw it as being feasible. Meaning that with reasonably good weather there was no hardware limitation. There doesn't seem to be any visibility issues with the cameras.
3.) Easy reproducibility

In fact the best move for Chuck is to simply not test that UPL with 10.13, and instead test another similar one that Tesla never had a chance to test it on. A test case should never include a sample used to train unless you're focus is on validation.

 

[S4WRXTTCS]

This is not an airport, no need to announce your departure.

I must be missing something because I don't see a departure acknowledgement let alone a departure time or gate number. Not even a destination.

 

[Dewg]

This whole bit about Tesla leveraging its customer base to do product testing is way overplayed. For most of us the entire data collection is passive where it collects data when people choose to use it.

In the entire history of the FSD Beta only one single person has gotten Tesla's attention to actually attempt to fix how it handles something, and that's the Chucks unprotected left turn.

The majority of us including me simply weren't willing to put work into something unless there was some validation for that effort. You mentioned payment, or free use of a product. Those are fair, but there are more reasons. I would have put more effort into testing it if it led to improvements in places I used it. Or if Tesla ever acknowledged any of it as being worked on or asked for more info. There was zero engagement so I stopped bothering with it.

Most of Chucks work didn't seem to be motivated from a desire to help Tesla. Instead it was more because of the engagement of the community. A lot of the upgrades to his process were for the benefit of the viewer, and not so much for Tesla.

At some point he reached such a following that Elon/Tesla couldn't simply ignore him anymore. In a sense the entire FSD Beta program up to that point seemed to be dominated by influencer types gushing on about how good it was.

How did an involvement of 100,000 people only result in a single test case being a focus? Sure it represented a lot of UPL's, but there are hundreds of other cases as well that are representative of similar problems. Every issue I had with FSD Beta could have been easily reproduced on another road/situation that was similar.

I think it came down to three reasons that Chucks UPL was focused on.

1.) Popularity to the point where Musk himself was made aware of it
2.) It was likely deemed a solvable problem. Now maybe not to the degree that I'd like to see it solved, but I think Engineers at Tesla saw it as being feasible. Meaning that with reasonably good weather there was no hardware limitation. There doesn't seem to be any visibility issues with the cameras.
3.) Easy reproducibility

In fact the best move for Chuck is to simply not test that UPL with 10.13, and instead test another similar one that Tesla never had a chance to test it on. A test case should never include a sample used to train unless you're focus is on validation.

I agree with many of your points, but I think you're being overly pessimistic on how helpful FSD Beta testers are. If you were in a group of 100 dedicated testers, I would expect significant engagement with you. I envision that's how it is with internal testers, engineers, and employees at Tesla. There just isn't enough manpower to take all the reports from Beta participants and review/reply to all of them. The only comfort and assurance that we're being helpful is in the 10.13 changelog, which has a point:

- Upgraded the occupancy network to detect dynamic objects and improved performance by adding a video module, tuning the loss function, and adding 37k new clips to the training set.

I'm guessing we, the Beta testers, provided a great many of those 37,000 clips with our report button. I might be overly optimistic, and the truth might lie somewhere in the middle.

 

[momo3605]

There just isn't enough manpower to take all the reports from Beta participants and review/reply to all of them.

There should be enough Man power to review/respond to more than 1 use case (Chuck’s ULT). Without engagement from the FSD team and watching my intersections fail update after update for the past 12 months and 10 updates doesn’t instill too much confidence that they’re doing anything about my uploaded clips. But who knows

 

[__xyzzy__]

I'm new to Teslas and new to this forum. Seems like there's a lot of trolling and misunderstanding of people's motivations for buying FSD. So here's my perspective, for what it's worth.

I have a MYP on order (EDD Sept 13 - Oct 18 2022) and I am delighted to pony up the $12k for FSD.

Yes, I know FSD is not ready for general public use.
Yes, I know I will have to pay as much or more attention to FSD driving than my own driving.
Yes, I know it will make mistakes.
Yes, I know I'm paying for the privilege of running incomplete beta software.
Yes, I know I'm paying to train Tesla's AI stack.
Yes, I know there's a good chance FSD won't be ready for a loooooong time, despite what Elon has been saying for years and years.
Yes, I know it's not going to drive me around everywhere and take a burden off me.

And I'm STILL excited! I'm a tech nerd. I work in a very high tech computing field where reverse-engineering complex technology stacks is part of my day job. I'm thrilled to get a chance to play with FSD beta, to learn its quirks, foibles, and its successes. I'm stoked to watch what happens with every upgrade of the software and to live in it as it gets better and better.

Do I expect it to work well the majority of the time? Of course not. That's not why I'm buying it.

All of which to say: putting aside the trolls for now, if a reasonable person has no use for a not-fully-functional FSD, then I understand why such a person would be baffled by people paying $12k for it. Why pay $12k for a thing that's more work/stress than driving yourself?!? My wife says of FSD: "I'd never use it, it sounds terrible. I'd be too stressed out the entire time thinking I'm going to crash."

But if I can understand that point of view, I'm sure the nay-sayers - with a modicum of effort to put oneself in the shoes of others and entertain a different perspective - can come to realize that for some of us FSD isn't a chore. It's not a burden. It's not hard work. It's not even that we expect it to be functional. We don't feel ripped off. We're getting exactly what we paid for.

This kind of technology is exciting and interesting simply for its own sake. And I'm looking forward to the thrill of playing with a bleeding edge self-driving car!

 

[Dewg]

I'm new to Teslas and new to this forum. Seems like there's a lot of trolling and misunderstanding of people's motivations for buying FSD. So here's my perspective, for what it's worth.

I have a MYP on order (EDD Sept 13 - Oct 18 2022) and I am delighted to pony up the $12k for FSD.

Yes, I know FSD is not ready for general public use.
Yes, I know I will have to pay as much or more attention to FSD driving than my own driving.
Yes, I know it will make mistakes.
Yes, I know I'm paying for the privilege of running incomplete beta software.
Yes, I know I'm paying to train Tesla's AI stack.
Yes, I know there's a good chance FSD won't be ready for a loooooong time, despite what Elon has been saying for years and years.
Yes, I know it's not going to drive me around everywhere and take a burden off me.

And I'm STILL excited! I'm a tech nerd. I work in a very high tech computing field where reverse-engineering complex technology stacks is part of my day job. I'm thrilled to get a chance to play with FSD beta, to learn its quirks, foibles, and its successes. I'm stoked to watch what happens with every upgrade of the software and to live in it as it gets better and better.

Do I expect it to work well the majority of the time? Of course not. That's not why I'm buying it.

All of which to say: putting aside the trolls for now, if a reasonable person has no use for a not-fully-functional FSD, then I understand why such a person would be baffled by people paying $12k for it. Why pay $12k for a thing that's more work/stress than driving yourself?!? My wife says of FSD: "I'd never use it, it sounds terrible. I'd be too stressed out the entire time thinking I'm going to crash."

But if I can understand that point of view, I'm sure the nay-sayers - with a modicum of effort to put oneself in the shoes of others and entertain a different perspective - can come to realize that for some of us FSD isn't a chore. It's not a burden. It's not hard work. It's not even that we expect it to be functional. We don't feel ripped off. We're getting exactly what we paid for.

This kind of technology is exciting and interesting simply for its own sake. And I'm looking forward to the thrill of playing with a bleeding edge self-driving car!

Can I get an Amen up in here? Welcome to the forums. Try to ignore the haters. There are those people here who are not keen on FSD, but present logical and well thought-out discussions. I respect them and enjoy our conversations.

 

[AlanSubie4Life]

I'm new to Teslas and new to this forum. Seems like there's a lot of trolling and misunderstanding of people's motivations for buying FSD. So here's my perspective, for what it's worth.

I have a MYP on order (EDD Sept 13 - Oct 18 2022) and I am delighted to pony up the $12k for FSD.

Yes, I know FSD is not ready for general public use.
Yes, I know I will have to pay as much or more attention to FSD driving than my own driving.
Yes, I know it will make mistakes.
Yes, I know I'm paying for the privilege of running incomplete beta software.
Yes, I know I'm paying to train Tesla's AI stack.
Yes, I know there's a good chance FSD won't be ready for a loooooong time, despite what Elon has been saying for years and years.
Yes, I know it's not going to drive me around everywhere and take a burden off me.

And I'm STILL excited! I'm a tech nerd. I work in a very high tech computing field where reverse-engineering complex technology stacks is part of my day job. I'm thrilled to get a chance to play with FSD beta, to learn its quirks, foibles, and its successes. I'm stoked to watch what happens with every upgrade of the software and to live in it as it gets better and better.

Do I expect it to work well the majority of the time? Of course not. That's not why I'm buying it.

All of which to say: putting aside the trolls for now, if a reasonable person has no use for a not-fully-functional FSD, then I understand why such a person would be baffled by people paying $12k for it. Why pay $12k for a thing that's more work/stress than driving yourself?!? My wife says of FSD: "I'd never use it, it sounds terrible. I'd be too stressed out the entire time thinking I'm going to crash."

But if I can understand that point of view, I'm sure the nay-sayers - with a modicum of effort to put oneself in the shoes of others and entertain a different perspective - can come to realize that for some of us FSD isn't a chore. It's not a burden. It's not hard work. It's not even that we expect it to be functional. We don't feel ripped off. We're getting exactly what we paid for.

This kind of technology is exciting and interesting simply for its own sake. And I'm looking forward to the thrill of playing with a bleeding edge self-driving car!

Yeah I am of similar mindset in most ways, though I would not have paid $12k.

I would also add:

Yes, I understand I will not be allowed to use it (except for highway) with anyone else in the car since it is so unpleasant for a passenger, and it may remain that way forever.

Yes, I understand it will be a less comfortable experience on city streets than just driving myself even when it is not “making mistakes.”

Yes, I understand I am quite possibly increasing my accident risk by using it.

 

[eli_]

FSD is worth $12k just to see the cool 3D visualizations on the dashboard, honestly I think 99% of the brilliance of the FSD project is in the 3D maps it infers from the camera output, that's like pure magic to see in action. The rest of the stack, the control and planning... yeah that stuff seems super thin and naïve to me. But even if the self-driving stuff never works out, they can at least sell their accelerator chip and the BEVNets, that's like where all the value is.

 

[Silicon Desert]

insiders like their head of AI who decided to leave right before it ships? It’s nowhere near complete.

Sounds like he was making a sarcastic joke, but you missed it

 

[nvx1977]

This whole bit about Tesla leveraging its customer base to do product testing is way overplayed. For most of us the entire data collection is passive where it collects data when people choose to use it.

In the entire history of the FSD Beta only one single person has gotten Tesla's attention to actually attempt to fix how it handles something, and that's the Chucks unprotected left turn.

The majority of us including me simply weren't willing to put work into something unless there was some validation for that effort. You mentioned payment, or free use of a product. Those are fair, but there are more reasons. I would have put more effort into testing it if it led to improvements in places I used it. Or if Tesla ever acknowledged any of it as being worked on or asked for more info. There was zero engagement so I stopped bothering with it.

Most of Chucks work didn't seem to be motivated from a desire to help Tesla. Instead it was more because of the engagement of the community. A lot of the upgrades to his process were for the benefit of the viewer, and not so much for Tesla.

At some point he reached such a following that Elon/Tesla couldn't simply ignore him anymore. In a sense the entire FSD Beta program up to that point seemed to be dominated by influencer types gushing on about how good it was.

How did an involvement of 100,000 people only result in a single test case being a focus? Sure it represented a lot of UPL's, but there are hundreds of other cases as well that are representative of similar problems. Every issue I had with FSD Beta could have been easily reproduced on another road/situation that was similar.

I think it came down to three reasons that Chucks UPL was focused on.

1.) Popularity to the point where Musk himself was made aware of it
2.) It was likely deemed a solvable problem. Now maybe not to the degree that I'd like to see it solved, but I think Engineers at Tesla saw it as being feasible. Meaning that with reasonably good weather there was no hardware limitation. There doesn't seem to be any visibility issues with the cameras.
3.) Easy reproducibility

In fact the best move for Chuck is to simply not test that UPL with 10.13, and instead test another similar one that Tesla never had a chance to test it on. A test case should never include a sample used to train unless you're focus is on validation.

Here's how I see it:

Tesla chooses what to focus on for each release, and they don't necessarily prioritize beta tester feedback. Hypothetically if beta testers never submitted any feedback, Tesla would still be releasing updates. What is unclear is how much influence beta tester feedback has on their priorities. In 10.13, one line item I saw called out was improved performance on wide residential roads. This feels heavily influenced by beta tester feedback. Did a youtuber make their channel's mission to showcase wide residential roads for them to work on this issue? Clearly not. This was likely a ton of beta testers reporting the issue AND Tesla was ready to address it. UPLs happen to be one of the most difficult maneuvers for manual or auto driving. That's why Chuck focused on it, and even if Chuck never existed, Tesla would eventually have to put effort to improve it.

Just because there is a lack of transparency with how Tesla uses beta tester data and feedback doesn't mean we're being ignored. Their instructions were clear: just use the system and intervene when necessary. That alone gives them the data they need to prioritize what they want to focus on.

That said, I do believe Tesla might have fast-tracked their focus on UPLs because of Chuck's channel. So maybe we need to get some aspiring YT'er to start a channel showcasing nothing but rotaries. Or school zones and school buses. Finding a niche with FSDb has never been easier!

 

[__xyzzy__]

FSD is worth $12k just to see the cool 3D visualizations on the dashboard, honestly I think 99% of the brilliance of the FSD project is in the 3D maps it infers from the camera output, that's like pure magic to see in action. The rest of the stack, the control and planning... yeah that stuff seems super thin and naïve to me. But even if the self-driving stuff never works out, they can at least sell their accelerator chip and the BEVNets, that's like where all the value is.

Agreed! I’m grateful to be able to drop $12k on a frivolity, but completely understand how $12k could be absolutely out of reach for people, or even if within reach I can see how it just wouldn’t be worth it.

Or to be more tongue-in-cheek, as I say to my kids: “it’s ok to be wrong as long as you learn something from it!”