"Self" driving on dirt roads.....

[glhs837]

So you will quite often see articles saying how Mercedes or Ford or GM is farther ahead than Tesla....... Watch this video and tell me any other maker is remotely close to where Tesla is at. No markings, no pavement, no nothing. Yes, you are still in the loop, my point is that from a technical perspective, this tech is crazy. No radar, no lidar, just cameras.

Keep in mind, this isn't trained to these roads. The system works no matter where you drop it although I suspect Mumbia of Dehli might cause it to melt down. But in any civilized traffic situation, it just works.

x.com

x.com

 

[Clem72]

I've said it before. I have no doubt they will get to the 90% solution quickly, 95% a decade or two later and 99% never. That's fine, as long as it does the 95% safer than the average human it will be good enough. We can take over when we need to drive through a fire, or through a washed out road and up an embankment, or (slowly) through a crowd of violent protesters.

 

[PeoplesElbow]

Pretty sure from the road display that there is more than just camera, the car already knows that the road is there via gps or something. Unless it can see the curved etc in the road before I can watching the video.

 

[glhs837]

I've said it before. I have no doubt they will get to the 90% solution quickly, 95% a decade or two later and 99% never. That's fine, as long as it does the 95% safer than the average human it will be good enough. We can take over when we need to drive through a fire, or through a washed out road and up an embankment, or (slowly) through a crowd of violent protesters.

See I think your timelines are skewed these people have sold 6 million cars. 6 million cars are out there. Racking up billions of miles in data and when it was people doing the learning. Yeah then your timelines makes sense but now that AI is parsing the learning things are going to get better exponentially. Have you seen the AI data center that they've added on to the end of giga, Texas? It's insane. Now some of that compute is going to be destined for the robot I'm sure, but they'll still have plenty of leftover capacity for the driving.

 

[glhs837]

Pretty sure from the road display that there is more than just camera, the car already knows that the road is there via gps or something. Unless it can see the curved etc in the road before I can watching the video.

Well of course it has GPS it also has I believe Google Maps baked into the operating system. I was specifically referring to sensors that other self-driving vehicles have.

Not sure why everything's bolding but I'm too lazy to fix it

 

[Clem72]

See I think your timelines are skewed these people have sold 6 million cars. 6 million cars are out there. Racking up billions of miles in data and when it was people doing the learning. Yeah then your timelines makes sense but now that AI is parsing the learning things are going to get better exponentially. Have you seen the AI data center that they've added on to the end of giga, Texas? It's insane. Now some of that compute is going to be destined for the robot I'm sure, but they'll still have plenty of leftover capacity for the driving.

Doesn't matter how many cars they have, they all have the same sensors/cameras, run the same software, etc. There will be blind spots that they all have in common. They can learn one set of problems with 6 million sets of the same hardware running a million different courses. They will learn a different set of problems if they had 6 million different sets of hardware on the same course.

 

[glhs837]

Doesn't matter how many cars they have, they all have the same sensors/cameras, run the same software, etc. There will be blind spots that they all have in common. They can learn one set of problems with 6 million sets of the same hardware running a million different courses. They will learn a different set of problems if they had 6 million different sets of hardware on the same course.

But every car is someplace different. Doing different things and even the same car and the same route with the different driver is going to experience different things and different solutions to those problems. How much have you learned over your years of driving? Now imagine you could learn everything from 5 million other drivers? That's Musk's point humans. Learn how to do this with pretty good reliability with a really limited experience set and only two cameras.

 

[Clem72]

But every car is someplace different. Doing different things and even the same car and the same route with the different driver is going to experience different things and different solutions to those problems. How much have you learned over your years of driving? Now imagine you could learn everything from 5 million other drivers? That's Musk's point humans. Learn how to do this with pretty good reliability with a really limited experience set and only two cameras.

We'll see. One of the things I learned over my career was to control variables. If I have 10 guys performing 10 different tasks I will have a lot of data, but much of it I won't have appropriate context to understand. If I have 10 guys performing the exact same task I can learn a lot more about how efficient they were, or how safe, how effective.

Think about how they train machine learning (LLMs or otherwise), they typically start with very highly defined data sets. Images with every object identified and described, or all in a very narrow category (I.E. this is 100k x-rays of healthy lungs, this is 100k x-rays of cancerous lungs, etc.). What they don't do is, for instance, give their x-ray AI doctor uncategorized unlabeled pictures of every object or body part ever x-ray'd and expect it to magically pull details from the chaos.

 

[thurley42]

I've said it before. I have no doubt they will get to the 90% solution quickly, 95% a decade or two later and 99% never. That's fine, as long as it does the 95% safer than the average human it will be good enough. We can take over when we need to drive through a fire, or through a washed out road and up an embankment, or (slowly) through a crowd of violent protesters.

We've been at 95-99% for a decade at least, I can use a $500 autopilot and program a car to drive on it's own for thousands of miles. We've been having autonomous challenges across the Australian outback for years.....the hurdle isn't the autonomy, the hurdle is humans driving cars.....

 

[Clem72]

We've been at 95-99% for a decade at least, I can use a $500 autopilot and program a car to drive on it's own for thousands of miles. We've been having autonomous challenges across the Australian outback for years.....the hurdle isn't the autonomy, the hurdle is humans driving cars.....

BS, even the best new software can BARELY maneuver around an active parking lot. Just because it functions for "99%" of a particular journey doesn't mean it's a 99% solution. I was referring to 99% of all possible driving interactions, which I guarandamtee your driving on dirt roads in the Australian outback doesn't require. How well does your Tesla/Waymo handle a 4 way stop where the person with the obvious right of way just sits there trying to wave you forward? Does the AI vehicle know if it's more appropriate to hit a small animal than slam on the breaks or swerve at high speed? What if it's a skunk? What if it's a child? Does the car know how to break traffic laws if the situation demands, and which are more of a risk? Does it know if your vehicle is a BMW that you aren't supposed to signal before merging? If a confused old lady in front of you starts Driving in reverse does the car honk the horn? Reverse itself? accelerate around? Driving is a bit like conversing in this regard. There are more than 150k unique words in the English language, most people use less than 1k different words in a given day. Just because it can complete one driving task, or a hundred, doesn't mean its anywhere close to a 99% solution.

 

[CPUSA]

I've said it before. I have no doubt they will get to the 90% solution quickly, 95% a decade or two later and 99% never. That's fine, as long as it does the 95% safer than the average human it will be good enough. We can take over when we need to drive through a fire, or through a washed out road and up an embankment, or (slowly) through a crowd of violent protesters.

That's not how you spell swiftly...

 

[glhs837]

We'll see. One of the things I learned over my career was to control variables. If I have 10 guys performing 10 different tasks I will have a lot of data, but much of it I won't have appropriate context to understand. If I have 10 guys performing the exact same task I can learn a lot more about how efficient they were, or how safe, how effective.

Think about how they train machine learning (LLMs or otherwise), they typically start with very highly defined data sets. Images with every object identified and described, or all in a very narrow category (I.E. this is 100k x-rays of healthy lungs, this is 100k x-rays of cancerous lungs, etc.). What they don't do is, for instance, give their x-ray AI doctor uncategorized unlabeled pictures of every object or body part ever x-ray'd and expect it to magically pull details from the chaos.

So, its not being done like that.

Tesla's Latest Full Self-Driving Software Update: FSD (Supervised) v12.3.5 Unveiled

Tesla, the pioneering electric vehicle manufacturer, continues to push the boundaries of autonomous driving with its latest Full Self-Driving (FSD) software update, v12.3.5.

evpeiwe-tech.com

I don't begin to understand why its so different, but according to all I can see, it is. Buddy of mine let me take his for an SFSD ride to lunch from 2185 to Plaza Azteca two weeks back. I was pretty impressed.

At some point we'll need to decide exactly how much better than a human we need the systems to be.

 

[Sneakers]

I, for one, like driving and have no use for full self driving. I am on board and fully in agreement with collision avoidance like auto-brake to avoid rear-ending someone, and alert monitors. But I want full control of the normal driving.

 

[Kyle]

I, for one, like driving and have no use for full self driving. I am on board and fully in agreement with collision avoidance like auto-brake to avoid rear-ending someone, and alert monitors. But I want full control of the normal driving.

Took me a long time just to trust the Cruise Control.

I'll be damned if I turn over full operation to a computer.

"We're gonna hit that tree, HAL!!!

"I'm afraid I have to do it, Dave!"

 

[glhs837]

I, for one, like driving and have no use for full self driving. I am on board and fully in agreement with collision avoidance like auto-brake to avoid rear-ending someone, and alert monitors. But I want full control of the normal driving.

Love driving myself. Went down to Jax a few years back. 95% of the drive was crap. Straight line I-95.