FSD rewrite will go out on Oct 20 to limited beta

[82bert]

I'm willing to take a bet on your myopicness.

Sure. Although I’m not sure you understand myopic. But sure.

 

[karbomusic]

I’m not going to claim that true level 5, fully autonomous driving is right around the corner, but whenever I hear someone say that it may not happen for 50 years, I can’t help but think how myopic they are. I’d be shocked if it’s not here within 10 years. Really and truly.

I'm curious as to what is stopping Elon from getting closer and closer to L4/5 but keeping the designation as L2 for no other reason than keeping the accountability on the driver? Seems like based on the SAE Levels there's what the car can actually do the vast majority of the time, and who gets the blame, where the functional level could be higher than the rated Level +/- some small margin of error hence keeping it L2.

Also, pardon my lack of SAE research but what is the acceptable failure rate for L4 or L5? It can't be 0% in this universe just like planes can't be 0% - Is there some criteria such as demonstrably as good or better than a human based on some test? What about gray areas where it's ambiguous as to whether a human could have done better in crash scenario X. I'm asking all of this because it seems plausible FSD could exceed L2 in some form yet stay at L2 for responsibility reasons.

 

[Daniel in SD]

I’d be shocked if it’s not here within 10 years. Really and truly.

That's what people said ten years ago.
You should look at what AI researchers said 50 year ago (History of artificial intelligence - Wikipedia)

 

[EinSV]

I'm curious as to what is stopping Elon from getting closer and closer to L4/5 but keeping the designation as L2 for no other reason than keeping the accountability on the driver? Seems like based on the SAE Levels there's what the car can actually do the vast majority of the time, and who gets the blame, where the functional level could be higher than the rated Level +/- some small margin of error hence keeping it L2.

Also, pardon my lack of SAE research but what is the acceptable failure rate for L4 or L5? It can't be 0% in this universe just like planes can't be 0% - Is there some criteria such as demonstrably as good or better than a human based on some test? What about gray areas where it's ambiguous as to whether a human could have done better in crash scenario X. I'm asking all of this because it seems plausible FSD could exceed L2 in some form yet stay at L2 for responsibility reasons.

Once FSD can handle a region safely as Level 4 autonomous Tesla could open a Robotaxi network in that area. Also, they could release Level 4 FSD to regular Tesla customers in that region and it would be a huge sales draw.

Tesla has very strong financial incentives to get Level 4 FSD up and running as soon as they can safely do it and secure any necessary regulatory approvals.

 

[Daniel in SD]

I'm curious as to what is stopping Elon from getting closer and closer to L4/5 but keeping the designation as L2 for no other reason than keeping the accountability on the driver? Seems like based on the SAE Levels there's what the car can actually do the vast majority of the time, and who gets the blame, where the functional level could be higher than the rated Level +/- some small margin of error hence keeping it L2.

Also, pardon my lack of SAE research but what is the acceptable failure rate for L4 or L5? It can't be 0% in this universe just like planes can't be 0% - Is there some criteria such as demonstrably as good or better than a human based on some test? What about gray areas where it's ambiguous as to whether a human could have done better in crash scenario X. I'm asking all of this because it seems plausible FSD could exceed L2 in some form yet stay at L2 for responsibility reasons.

The level of a driving automation system feature corresponds to the feature’s production design intent. This applies regardless of whether the vehicle on which it is equipped is a production vehicle already deployed in commerce, or a test vehicle that has yet to be deployed. As such, it is incorrect to classify a level 4 design-intended ADS feature equipped on a test vehicle as level 2 simply because on-road testing requires a test driver to supervise the feature while engaged, and to intervene if necessary to maintain safe operation.

From a practical standpoint Tesla needs their cars to be legal on public roads. It seems likely that if such a L2 system was abused enough to result in a decrease in road safety it would be banned.
There is no performance requirement for the SAE levels. Tesla could operate their vehicles without drivers in many states with no regulatory approval at all. Obviously it would present some major liability issues though!

 

[mikes_fsd]

Karpathy joined Tesla (from OpenAI) in June 2017, shortly after he wrote this Medium post:
Software 2.0

Fascinating. Considering when watching the videos from this year re how the rewrite is going, it all has his signature.

I think the AGI (Artificial General Intelligence) stuff Elon got from Karpathy (not the other way around)!

 

[powertoold]

Karpathy joined Tesla (from OpenAI) in June 2017, shortly after he wrote this Medium post:
Software 2.0

Fascinating. Considering when watching the videos from this year re how the rewrite is going, it all has his signature.

I think the AGI (Artificial General Intelligence) stuff Elon got from Karpathy (not the other way around)!

Elon has many talents. One of them is getting the best people to fulfill his vision. Chris Lattner was in and out in no time. Lattner is very well regarded in the CS community, but he wasn’t the right guy for the job.

 

[DanCar]

... I think the AGI (Artificial General Intelligence) stuff Elon got from Karpathy (not the other way around)!

AGI doesn't exist, except for science fiction books, peoples dreams, and what if discussions. It is the target of significant research. And there has been some discussion of a U.S. government Manhattan style project to develop it.
United States should make a massive investment in AI, top Senate Democrat says

The top Democrat in the U.S. Senate wants the government to create a new agency that would invest an additional $100 billion over 5 years on basic research in artificial intelligence (AI).

Trump version: Artificial Intelligence for the American People | The White House

 

[mikes_fsd]

"it [FSD beta release] does not really change much wrt [with regard to] what we have seen before"
We have not see any of the functionality the FSD Beta has shown a Tesla can do.

I tried to find some list of what we have seen FSD Beta do in the short week (and stuff that we have not seen before, contrary to what green said)!

1. Right turns
2. Right turns on RED
3. Left turns
   1. Unprotected
   2. Protected
4. 4-way stop
   1. with other cars
   2. without other cars
5. slowing down for speedbumps
6. passing bikes
   1. going over double yellow to do so
7. passing pedestrians
8. stopping at crosswalks with ppl walking
9. navigating roundabouts without a lead car
10. respecting traffic lights

This is just what I could dump off the top of my head (guaranteed there is more than that)!

 

[mikes_fsd]

AGI doesn't exist, except for science fiction books, peoples dreams, and what if discussions. It is the target of significant research.

Did you read the article?
I am not here to discuss AGI, just made an observation from the article and what Elon had said in the past.

 

[powertoold]

I tried to find some list of what we have seen FSD Beta do in the short week (and stuff that we have not seen before, contrary to what green said)!

I’ve noticed that the FSD beta seems to improve without a software version update. I’ve heard and read from multiple beta users that the software is not able to do a maneuver on one day and is able to do it another day? Some kind of secret sauce is going on.

Plus, Elon tweeted that his beta software is a few days ahead of the current users. Why would he even say that if it’s only a few days and what does that even mean?

 

[qdeathstar]

Personally i think level 5 autonomy in passenger vehicles isn’t as big a value add as people are making it out to be. There is some demand, but it is limited in scope. I doubt most cars will be self driving 100 years from now.

 

[Daniel in SD]

I’ve heard and read from multiple beta users that the software is not able to do a maneuver on one day and is able to do it another day? Some kind of secret sauce is going on.

That's the beauty of neural nets. You never know what you're going to get!

 

[GoSharks]

I tried to find some list of what we have seen FSD Beta do in the short week (and stuff that we have not seen before, contrary to what green said)!

1. Right turns
2. Right turns on RED
3. Left turns
   1. Unprotected
   2. Protected
4. 4-way stop
   1. with other cars
   2. without other cars
5. slowing down for speedbumps
6. passing bikes
   1. going over double yellow to do so
7. passing pedestrians
8. stopping at crosswalks with ppl walking
9. navigating roundabouts without a lead car
10. respecting traffic lights

This is just what I could dump off the top of my head (guaranteed there is more than that)!

Tesla showed a lot of this during autonomy day. What they haven't demonstrated is a hands-off/unattended drive, which is what Green appears to be focusing on.

 

[EinSV]

I tried to find some list of what we have seen FSD Beta do in the short week (and stuff that we have not seen before, contrary to what green said)!

1. Right turns
2. Right turns on RED
3. Left turns
   1. Unprotected
   2. Protected
4. 4-way stop
   1. with other cars
   2. without other cars
5. slowing down for speedbumps
6. passing bikes
   1. going over double yellow to do so
7. passing pedestrians
8. stopping at crosswalks with ppl walking
9. navigating roundabouts without a lead car
10. respecting traffic lights

This is just what I could dump off the top of my head (guaranteed there is more than that)!

A few more relatively sophisticated maneuvers (also off the top of my head):

11. Going around illegally parked car blocking lane (safely crossing double yellow to do so)
12. Managing one-lane, two-way traffic by waiting for oncoming traffic to pass before proceeding
13. Creeping up to intersection after initial stop to see around other cars before proceeding/turning

 

[DanCar]

That's the beauty of neural nets. You never know what you're going to get!

Would be cool if Tesla implemented some kind of micro updates, where only some of the machine learning weights and biases were updated nightly. Similarly there is incremental / continual learning:

1. Incremental learning - Wikipedia
   
2. Why Continual Learning is the key towards Machine Intelligence

 

[powertoold]

That's the beauty of neural nets. You never know what you're going to get!

I wonder though, if they deploy robotaxis, they may need a way to update the navigation logic quickly in case there’s a problem... they may have built in a way to do that, although I don’t have the slightest clue how. Perhaps some form of intersection maneuvering templates

 

[mikes_fsd]

We saw a lot of this during autonomy day. What we haven't seen is hands-off, which is what Green appears to be focusing on.

Oh, you are going to there.
Let me find the threads on this forum of "experts" dressing down Tesla, how gamed that was and how big of a farce that video is.
Oh, and how many attempts it took them to get that single shot.

Is that where he is getting his "previous" experience with FSD on Tesla from? This seems very doubtful to me.

And then seeing it in the hands of Tesla owners using on their own cars, at their discretion (as cringe worthy as some of the first videos were) to say "it [FSD beta release] does not really change much wrt [with regard to] what we have seen before"

Really!?

 

[mikes_fsd]

Would be cool if Tesla implemented some kind of micro updates, where only some of the machine learning weights and biases were updated nightly. Similarly there is incremental / continual learning:

1. Incremental learning - Wikipedia
   
2. Why Continual Learning is the key towards Machine Intelligence

This is what I think is happening. (but I have no inside knowledge)
The NN's stay the same but they toss out a new set of weights/parameters that the car will prioritize outputs from these 8 NN's at priority 9 and these 20 at priority 5 and the rest at priority 3.

 

[willow_hiller]

Would be cool if Tesla implemented some kind of micro updates, where only some of the machine learning weights and biases were updated nightly. Similarly there is incremental / continual learning:

Correct me if I'm wrong, but isn't this already incorporated into the "DeepRain" neural network? The last update on that feature does some active reweighting when it's active but a driver taps to wipe the windshield anyway.