The Tesla Thread

[zamler]

Originally Posted by Fuzz

I'm not sure what you mean? It seems that since that date, that's how you sign up. If you have other facts, maybe provide them?

You used that article to say anyone can download and use the beta, beta download button never happened. There are about 2000 people (excluding Tesla employees) currently testing beta 9.

Elon Musk Confirms Tesla FSD Wide Release ‘Download Button’ Will Not Come Until FSD Beta V10/11

[Fuzz]

Thanks, that's more useful. Everything I had seen made it seem like it was available to all. Another interesting bit from that article:

However, according to Elon, currently, FSD beta v9 is only using the pure vision production code for highway driving. Musk says “Beta 10 hopefully (Beta 11 definitely) will use one stack to rule them all – city streets, highway & complex parking lots.”

Nothing else I had read or seen from these beta tester videos indicate that. They all say "it's only using vision cameras" even when driving city streets. Which makes that "almost hitting the rail pylons" incident even more confusing. Wouldn't radar have detected that? I assume this also means the beta is not available yet no new Model 3's and Y's, since they don't have radar sensors, and version 9 software would still need them on city streets?


Just wondering if the testing videos we are seeing are what they claim they are. From the video I linked, "This is totally FSD 8 behaviour"...probably because it is. So beta 10 will actually be where they go full vision only. I guess this is what happens when you don't have a PR department, and do it all through tweets and "influencers".

[PeteMoss]

Originally Posted by zamler

Co-pilot is the perfect name Tesla dropped the ball here.

I predict it will come out this car didn't have full self driving and was not in autopilot during the crash. And there was a person in the drivers seat or possibly there was a sack of something on the seat and the belt was buckled.

Replying to months old message, but if they named it something different it wouldn't get near the attention it does now which is bad in the sense of arguments among online people, but also makes casuals think they are further along then they really are.

Sent from my Pixel 4 XL using Tapatalk

[zamler]

Originally Posted by Fuzz

Nothing else I had read or seen from these beta tester videos indicate that. They all say "it's only using vision cameras" even when driving city streets. Which makes that "almost hitting the rail pylons" incident even more confusing.

Beta 9 is vision only it works on the latest models that have no radar. This is not an issue of "seeing", the 8 cameras see everything it's a coding/AI challenge. The FSD computer saw the pillars but didn't understand exactly what they were. Radar doesn't solve this at all as I said the more sensors you have the harder it is to decide which is correct. Radar is not some magic signal where you bounce it off something and it automatically tells you, hey that's something you can crash into. The signal has to be interpreted and acted on, which is something traditionally a human does people see the radar map and decide.

Wouldn't radar have detected that? I assume this also means the beta is not available yet no new Model 3's and Y's, since they don't have radar sensors, and version 9 software would still need them on city streets?

Beta 9 only uses cameras! And again it's not a detection issue it's something more training will solve.

Just wondering if the testing videos we are seeing are what they claim they are. From the video I linked, "This is totally FSD 8 behaviour"...probably because it is. So beta 10 will actually be where they go full vision only. I guess this is what happens when you don't have a PR department, and do it all through tweets and "influencers".

Again, beta 9 works on cars that have no radar sensors which is every Model 3 and Y currently being made. FSD non beta package also works with said models as does the standard collision avoidance system. Vision only, no radar. It was recently tested and did at least as well as the previous version that used radar.

What Tesla is attempting is way beyond any real world AI I've seen so far, it does great 98% of the time it's that last 2% that is hard. That's why I said recently I think we are a year+ out before this works better than a human driver on average.

[Wormius]

How exactly does radar not solve an issue with a solid object being a potential obstacle? It seems like this is exactly what radar would solve in this exact situation where pillars are the same colour as the roadways.

[zamler]

Tesla's head of AI Karpathy goes through why pure vision is the best solution, time will tell what the correct approach is. Can't compare Tesla's FSD to anything else though since only Tesla has a system that will drive any road regardless if it has been on that road before and will drive on unmapped roads. Every other system is geofenced.

[Fuzz]

Originally Posted by zamler

Tesla's head of AI Karpathy goes through why pure vision is the best solution, time will tell what the correct approach is. Can't compare Tesla's FSD to anything else though since only Tesla has a system that will drive any road regardless if it has been on that road before and will drive on unmapped roads. Every other system is geofenced.

Obviously the guy who's career is on the line is going to say the system he designed is the best way to do it. Right?

[zamler]

Originally Posted by Fuzz

Obviously the guy who's career is on the line is going to say the system he designed is the best way to do it. Right?

Which is why I said "time will tell what the correct approach is". I'd be curious as to why you think Elon is making the wrong call here and should instead go all in on a multi sensor and camera system.

[Fuzz]

Originally Posted by zamler

Which is why I said "time will tell what the correct approach is". I'd be curious as to why you think Elon is making the wrong call here and should instead go all in on a multi sensor and camera system.

Articles like this:

https://www.inverse.com/innovation/b...ving-companies
Tesla is the loudest, but plenty of companies are going about the work quietly, and I beleive they will have more success.


There are plenty of opinions that run counter to the Tesla hype, I've read plenty. I happen to think that these things are true:
1)true AI is needed to replicate the human driving experience with our senses(Tesla's goal)

2)true AI is so far away that you can't plan an FSD around it
3)without true AI, you must make it up elsewhere
4)additional sensors are one way to get around the limitations of not having true AI.


This race is far from over, I just don't think Tesla is on the right path to reach the finish line, and don't think a handful of $30 cameras can get them there. Many others agree.

The logic behind collecting ever more data is that you will capture more of the rare scenarios that could flummox your AI, but there’s a fundamental limit to this approach. “Eventually you have unique cases. And unique cases you can’t train for,” says Weinberger. “The benefits of adding more and more data are diminishing at some point.”

https://spectrum.ieee.org/cars-that-...y-self-driving

[zamler]

Originally Posted by Fuzz

Tesla is the loudest, but plenty of companies are going about the work quietly, and I beleive they will have more success.

You believe based on? Have you seen any of these systems navigate city streets?

There are plenty of opinions that run counter to the Tesla hype, I've read plenty.

Hype has zero to do with the actual technology don't know why you brought this up except I guess Tesla "hype" bothers you.

I happen to think that these things are true:
1)true AI is needed to replicate the human driving experience with our senses(Tesla's goal)

2)true AI is so far away that you can't plan an FSD around it
3)without true AI, you must make it up elsewhere
4)additional sensors are one way to get around the limitations of not having true AI.

Get around the limitations how? How can more sensors enable handling of complex driving situations?

This race is far from over, I just don't think Tesla is on the right path to reach the finish line, and don't think a handful of $30 cameras can get them there. Many others agree.

As I said time will tell.



I do find it comical that Waymo and Nvidia are leading FSD. How? Waymo is an on the rails system it may was well be a roller coaster.

[Wormius]

Originally Posted by zamler

Tesla's head of AI Karpathy goes through why pure vision is the best solution, time will tell what the correct approach is. Can't compare Tesla's FSD to anything else though since only Tesla has a system that will drive any road regardless if it has been on that road before and will drive on unmapped roads. Every other system is geofenced.

This doesn’t really explain the claim that radar wouldn’t help in that scenario. It seems like it would most certainly have identified those as solid, vertical structures.

[zamler]

I've already explained clearly why lack of radar is not the problem.

[Wormius]

Originally Posted by zamler

I've already explained clearly why lack of radar is not the problem.

It looks like other car manufacturers, and even Tesla, had it working before. Don’t look now which sensor to believe? Maybe go with the one that actually detects solid objects?

This honestly just screams that they’ve got supply constraints and dealt with it in the poorest way possible.

I have a person like this in my work life, “can’t get that part for a few weeks? we’ll just build everything from now on without it and figure out how badly that screws us later”.

[Fuzz]

Originally Posted by zamler

You believe based on? Have you seen any of these systems navigate city streets?

Based on what I've seen and read. I for opinions the same way most people do. You seem flabbergasted that I've come to different conclusions than you,despite others in the industry comnig to similar conclusions.

Originally Posted by zamler

Hype has zero to do with the actual technology don't know why you brought this up except I guess Tesla "hype" bothers you.

Get around the limitations how? How can more sensors enable handling of complex driving situations?

The same way we do. Having 2 eyes is better than one, because you get depth perception. Adding ears is better than not, because you take in audio cues. Even your nose kicks in occasionally, to get out from behind a dirty diesel. It's baffling to think that more sensors are worse.



I understand the issues with sensor confusion, but if you get that with 3 sensors when one is wrong, how to you get around it when one sensor is wrong, and it's all you have?

Originally Posted by zamler

As I said time will tell.

I do find it comical that Waymo and Nvidia are leading FSD. How? Waymo is an on the rails system it may was well be a roller coaster.

I'm not sure what you mean by"on rail" system. From everything I have seen, it navigates similarly to Teslas. I know they need more work done mapping the back end, so it can't currently be placed anywhere(I don'[t think) but it's not just following one road.

[zamler]

Originally Posted by Wormius

It looks like other car manufacturers, and even Tesla, had it working before. Don’t look now which sensor to believe? Maybe go with the one that actually detects solid objects?

What do you consider a solid object and a not solid object.

[Wormius]

Originally Posted by zamler

What do you consider a solid object and a not solid object.

Personally, I consider a monorail pillar to be a solid object.

The machine vision AI that mistook a monorajl pillar for just part of the road, obviously considered it a “not solid object”.

[Fuzz]

https://twitter.com/user/status/1418413307862585344


A good example of why this stuff is really really hard. It's so obvious to us, but without context, the "AI" is baffled.

[zamler]

That's cool, and concerning. Don't think I've ever seen the moon that yellow.

best comment in that twitter thread

No that is for the radar to detect. Oh wait a minute....

[Fuzz]

FSD Beta 9.1 patch notes:added celestial body positioning to improve traffic light detection.

[Wormius]

On smokey days the moon will be pretty dark orange in the sky. Even the sun will be pretty tones down and just look like a red circle in the sky. I guess at least with gps they can at least consult a celestial objects map and not get confounded by the moon and sun impersonating traffic lights.