Elon: "Feature complete for full self driving this year"

[mongo]

Tl;dr; want some peanut butter?

You literally just confirmed every single thing i said, he taught a CNN 101 class that got really popular. That is his only achievement. His claim to fame and why he was hired by Elon. Infact at Autonomy Day, Elon basically said that. Yet Tesla fans hail him as some savior. He isn't done ANYTHING in the ML community. He has not invented anything, HE hasn't done any breakthroughs, any advancements. Do you know how many others have created classes better than that? Thousands..

Like i said Waymo partnering with Andrej would be equivalent to partnering with a random college student who interned here and there which all college student do.

You have a stange definition of confirm.
You:

There's nothing that andrej has done other than teach a basic CNN 101 course that got really popular.

He co-created the course, that is more than just teaching.
It was not a 101 class, it was a 231. And how many of those thousands of classes were created after his???

You:

All NN advancement and breakthroughs (RL, GANs, Transformers, AlphaGo, AlphaGoZero, AlphaStar, NLP, WaveNet, Duplex, etc) has come from Google Brain and DeepMind engineers and I mean ALL of it.

Andrej worked at Google and DeepMind, so he is in the set of engineers you are extolling. But it was 'only' an internship, so I guess you don't care.

He also got a PhD, which is no small feat, and much more than a 'random college student' such as youself. Although, you consider a PhD a waste of time/ money (at least in machine learning) so I wouldn't expect you to care about that.

You probably also don't care that the professor that Andrej co-created the class with (and who directed his doctiral program), Fei-Fei Li, was the driving force behind InageNet and, on her sabbatical from Stanford University from January 2017 to fall of 2018, Li joined Google Cloud as its Chief Scientist of AI/ML and Vice President.

He is also listed in the authors of 11 publications that have been cited over 25 thousand times.
Andrej Karpathy - Google Scholar Citations

But he never worked full time at Google, so his past and present accomplishments are meaningless.

 

[Bladerskb]

Tl;dr; want some peanut butter?


You have a stange definition of confirm.
You:

He co-created the course, that is more than just teaching.
It was not a 101 class, it was a 231. And how many of those thousands of classes were created after his???

You:

Andrej worked at Google and DeepMind, so he is in the set of engineers you are extolling. But it was 'only' an internship, so I guess you don't care.

He also got a PhD, which is no small feat, and much more than a 'random college student' such as youself. Although, you consider a PhD a waste of time/ money (at least in machine learning) so I wouldn't expect you to care about that.

You probably also don't care that the professor that Andrej co-created the class with (and who directed his doctiral program), Fei-Fei Li, was the driving force behind InageNet and, on her sabbatical from Stanford University from January 2017 to fall of 2018, Li joined Google Cloud as its Chief Scientist of AI/ML and Vice President.

He is also listed in the authors of 11 publications that have been cited over 25 thousand times.
Andrej Karpathy - Google Scholar Citations

But he never worked full time at Google, so his past and present accomplishments are meaningless.

Google has thousands of interns a year. My own company i used to work with had hundreds of interns every summer.
Tesla hires hundreds of Autopilot interns a year.

Infact even Elon himself says PHD is absolutely worthless and he despises it. When Andrej tried to talk about his PHD he literally interrupted him. So even your God doesn't like PHDs.

Internship means nothing. Everything you listed means nothing because we are not talking about some ML engineer which all ML engineer go through the same exact process. We are talking about Top researcher or as Willow put it. The best researcher. Yet this best researcher has done absolutely nothing other than what a random college student would do. Yet you again dismiss this fact that he has made ZERO contribution to NN advancements or any breakthroughs. Which is the whole point of the topic. But keep regurgitating how many internships he has done. I hate to break it to you thousands have done internship at Google, Deepmind and Google Brain aswell.

 

[mongo]

Google has thousands of interns a year. My own company i used to work with had hundreds of interns every summer.
Tesla hires hundreds of Autopilot interns a year.

Infact even Elon himself says PHD is absolutely worthless and he despises it. When Andrej tried to talk about his PHD he literally interrupted him. So even your God doesn't like PHDs.

Internship means nothing. Everything you listed means nothing because we are not talking about some ML engineer which all ML engineer go through the same exact process. We are talking about Top researcher or as Willow put it. The best researcher. Yet this best researcher has done absolutely nothing other than what a random college student would do. Yet you again dismiss this fact that he has made ZERO contribution to NN advancements or any breakthroughs. Which is the whole point of the topic. But keep regurgitating how many internships he has done. I hate to break it to you thousands have done internship at Google, Deepmind and Google Brain aswell.

'Foremost' may be a little strong. Though it may depend on the specific application one is talking about. I don't see any other vehicle company with a system as large at Tesla's. Regardless, that evaluation can exist with needing to attempt to discredit AK's accomplishments.

What is a PhD candidate, if not a researcher?
You are misstating Elon's view on PhDs (and Master's) (hint: he does not despise them or think them totally worthless)
What SpaceX CEO Elon Musk Says About Getting a PhD
A few years in a degree program could be more more productive than a few years at a failed AI company.

 

[Dutchie]

Elon Musk on Twitter

"Still on for end of year. Just software limited. Any Tesla car with HW2 (all cars built since Oct last year) will be able to do this."
3:34 AM · May 22, 2017

(he was talking about the coast to coast fully autonomous drive demo)

My goodness, that was three years ago! And now Tesla is still struggling with basics like traffic lights and Stop-signs. Yet, Elon maintains that Robotaxi functionality is still on for end of this year...

And they can't even make the windshield wipers go properly with NN architecture!

Sorry, Elon, you have lost all credibility. Better say nothing for a while.

 

[banned-66611]

There are probably 2 motivations for this tweet:

1) Elon is making an educated guess based on the current state of AP development, especially the progress with the new rewrite. So based on that progress, Elon believes that robotaxis can still happen by the end of this year.

2) As the CEO, Elon is known for imposing tough and ambitious deadlines as a way of getting stuff done. Elon is mostly likely giving the AP team the internal deadline of making robotaxis happen by the end of this year, whether or not it can actually be done, in order to push them hard to accomplish the goal.

3) Tesla needs continued inward investment and for people not to start suing them for not delivering Full Self Driving for 4 years.

 

[Daniel in SD]

Interesting also that Elon is not backing down on the robotaxi claims. I thought the AP rewrite meant the timelines even for pure functionality would be pushed back, but he’s still maintaining this year: Elon Musk on Twitter

I want to know why he thinks regulatory approval will be a problem? Other companies have gotten regulatory approval. Is there some change to the current regulations that he's lobbying for? Is there a country or state that has regulations that he thinks should be a model?

 

[diplomat33]

I want to know why he thinks regulatory approval will be a problem? Other companies have gotten regulatory approval. Is there some change to the current regulations that he's lobbying for? Is there a country or state that has regulations that he thinks should be a model?
View attachment 531845

I could be wrong but I interpret the tweet that Elon thinks that the 2 FSD features on the website, "traffic light & stop sign response" and "automatic driving on city streets", are on track to be ready this year and Elon thinks they will be enough to do robotaxis. So Elon expects Tesla to finish those features and make them reliable enough, and then that regulators will give Tesla permission to deploy these features as a robotaxi service without a driver.

But it does seem odd since as far I as I know, anybody can get regulatory approval if they have a safety driver. But maybe Tesla is hoping to get permission to go straight to a driverless robotaxi without a safety driver?

But I have to say that if Elon really is planning to finish "traffic light & stop sign response" and "automatic driving on city streets" and thinks that will be enough for robotaxis, and assuming the features are reliable, hopes regulators will give Tesla permission to deploy driverless robotaxis, that seems incredibly naive.

 

[mongo]

I want to know why he thinks regulatory approval will be a problem?

Tesla sells cars to people in all 50 states and multiple other countries. Do they all have FSD regulations in place, and are their requirements achievable? If not, big unknown.

 

[Mo City]

I want to know why he thinks regulatory approval will be a problem? Other companies have gotten regulatory approval. Is there some change to the current regulations that he's lobbying for? Is there a country or state that has regulations that he thinks should be a model?
View attachment 531845

I have no doubt regulatory approval for driverless cars will be very difficult once requested. But when Elon says it will be the main issue, he's blowing smoke.

 

[Daniel in SD]

But I have to say that if Elon really is planning to finish "traffic light & stop sign response" and "automatic driving on city streets" and thinks that will be enough for robotaxis, and assuming the features are reliable, hopes regulators will give Tesla permission to deploy driverless robotaxis, that seems incredibly naive.

I have to say that I think Elon is smarter than that. I believe there are states such as Florida where Tesla could legally deploy robotaxis today with no approval needed from regulators.

 

[Daniel in SD]

Tesla sells cars to people in all 50 states and multiple other countries. Do they all have FSD regulations in place, and are their requirements achievable? If not, big unknown.

That's a good question! That's exactly what I'm curious about. To me it would make sense to deploy in a place with little to no regulation first to prove that the system works. Tesla already geofences features so that shouldn't be a problem.

 

[diplomat33]

That's a good question! That's exactly what I'm curious about. To me it would make sense to deploy in a place with little to no regulation first to prove that the system works. Tesla already geofences features so that shouldn't be a problem.

I seem to remember Elon mentioning that they could deploy geofenced robotaxis first. So what you suggest could very well happen. And it would make Elon's tweet a bit more sensible if he is merely suggesting that the first supervised geofenced robotaxis could happen this year.

I have to say that I think Elon is smarter than that. I believe there are states such as Florida where Tesla could legally deploy robotaxis today with no approval needed from regulators.

Well, yeah, I figure that Elon is smarter than that. But his tweet still makes me wonder because he is either planning to release the traffic light and stop sign feature and "City NOA" and thinks it will be good enough for robotaxis, or the AP rewrite is hiding a massive jump in FSD capability that we don't know about.

 

[DanCar]

I want to know why he thinks regulatory approval will be a problem? Other companies have gotten regulatory approval. Is there some change to the current regulations that he's lobbying for? Is there a country or state that has regulations that he thinks should be a model?
View attachment 531845

Robotaxis are legal in Arizona according to a post a saw from Waymo, long ago.
I hope nobody is taking Elon seriously on this. He has been wrong every time on this subject since 4+ years back. What do think he is all of sudden correct? zero

 

[diplomat33]

I have no doubt regulatory approval for driverless cars will be very difficult once requested. But when Elon says it will be the main issue, he's blowing smoke.

Yes. The main issue will be actually getting safe and reliable autonomous driving. If and when that is achieved, getting regulatory approval will be super easy.

 

[Daniel in SD]

Robotaxis are legal in Arizona according to a post a saw from Waymo, long ago.
I hope nobody is taking Elon seriously on this. He has been wrong every time on this subject since 4+ years back. What do think he is all of sudden correct? zero

I think you need regulatory approval though. In some states, like Florida, I think no regulatory approval is required. I think states have been competing to have the most lax regulations to attract companies away from California. Which is fine by me, I wouldn't want them testing Smart Summon at 45mph here...

 

[banned-66611]

I want to know why he thinks regulatory approval will be a problem? Other companies have gotten regulatory approval.

Other companies have got regulatory approval for proven safe technology that has only ever had one very minor accident (hit a bus at 2 MPH). They are being extremely careful about rolling it out.

Musk wants to send out some beta quality software and a little disclaimer saying if you die in a fireball then you should have been paying attention hands on the wheel fully alert. Tesla is already under multiple investigations for autopilot fatalities and various regulators are looking at how they can force them to do better.

Maybe this is his plan. Make some rubbish software that barely works and then blame the regulators for not allowing it. Nobody can sue because it was always sold on the basis of having to wait for the regulators.

 

[diplomat33]

Other companies have got regulatory approval for proven safe technology that has only ever had one very minor accident (hit a bus at 2 MPH). They are being extremely careful about rolling it out.

Musk wants to send out some beta quality software and a little disclaimer saying if you die in a fireball then you should have been paying attention hands on the wheel fully alert. Tesla is already under multiple investigations for autopilot fatalities and various regulators are looking at how they can force them to do better.

Maybe this is his plan. Make some rubbish software that barely works and then blame the regulators for not allowing it. Nobody can sue because it was always sold on the basis of having to wait for the regulators.

I would not go as far as to say that Elon is planning to release bad features on purpose just to blame regulators. But I do interpret his "regulator approval is the big unknown" as basically saying "I have no idea if the features will even pass muster with regulators."

It is worth noting that this is not the first time that Elon has said that FSD was almost ready but that a lack of regulatory approval might prevent the release. It's a clever way of deflecting blame for any delay or problem.

Tesla does have a long history of releasing AP features as "beta" that require driver supervision. So it is entirely possible that Elon is planning to release robotaxis also as "beta" requiring driver supervision and he's admitting that he does not know if regulators will go for it.

In a previous earnings call, Elon mentioned the 3 stages of FSD development. First stage is "feature complete", second stage is "Tesla data says ok to remove driver supervision" and third stage is "regulatory approval to remove driver supervision". He has also said that they might roll out robotaxis in some geofenced areas first.
.
So I think the following is a pretty reasonable assumption for how robotaxis might roll out:
1) Tesla will release "traffic light/stop sign response" and "automatic city driving" and claim robotaxis are "feature complete". But FSD will be deemed "beta" and require driver supervision.
2) Presumably, based on disengagement feedback from the fleet, when the disengagement rate meets some internal criteria, Tesla will claim that the robotaxis are good enough to remove driver supervision at least in some geofenced areas to start with.
3) Tesla will ask regulators for permission to officially remove the nags and the requirement for driver supervision at least in those geofenced areas where their data shows that they think the robotaxis are safe enough.

 

[mongo]

So I think the following is a pretty reasonable assumption for how robotaxis might roll out:
1) Tesla will release "traffic light/stop sign response" and "automatic city driving" and claim robotaxis are "feature complete". But FSD will be deemed "beta" and require driver supervision.
2) Presumably, based on disengagement feedback from the fleet, when the disengagement rate meets some internal criteria, Tesla will claim that the robotaxis are good enough to remove driver supervision at least in some geofenced areas to start with.
3) Tesla will ask regulators for permission to officially remove the nags and the requirement for driver supervision at least in those geofenced areas where their data shows that they think the robotaxis are safe enough.

Given the podcast talks Elon has given (supervised, sleep in backseat capable, regulator approved), I don't think he would call #1 a robotaxi, since it still requires a driver and is not reliable. I think he would reserve that for #2.

 

[diplomat33]

Given the podcast talks Elon has given (supervised, sleep in backseat capable, regulator approved), I don't think he would call #1 a robotaxi, since it still requires a driver and is not reliable. I think he would reserve that for #2.

True. I just meant that Elon will probably call it "feature complete" for a robotaxi.

 

[mars_or_bust]

Here is a part of the Third Row Podcast that I think gives some good insight into the process of doing Smart Summon:

Q: "What's so hard about curbs?"

A: "When we first started on Smart Summon, and you all know since you have the cars, it was laggy. Not perfect. We were just trying as much as we could to make it smooth. With the older Autopilot system that we had before the rewrite, using the cameras, for curbs and stuff, you can only use the side repeater cameras and maybe the back-up camera and the angle at which you are seeing the curb, 80% seeing the actual curb and the rest you are guessing because of the distance, how high the curb is, and all that. That is why it such a hard thing to label and teach the neural net. When doing all the labeling, it is very very complicated. The accuracy has to pretty much be 100% or else there is going to be some flaw in the car driving around the curb or seeing the curb. With the new rewrite, it will be a lot better because we can look around, zoom in, zoom out, and label that curb in 3D and make it much easier."

I guess now we know why Smart Summon sucked when it was released. He basically admits that the cameras could only see 80% of the curb!! WOW!! Frankly, it is a bit shocking that Tesla released Smart Summon with this obvious flaw!

I think his answer is also a perfect microcosm for the entire AP development. Tesla releases stuff before it works reliably, the sensors are not really good enough but Tesla makes it work as best they can, and then later, Tesla eventually figures out how to do the feature the right way. I am not being anti-Tesla. He admits it!!!

It is encouraging though that the rewrite will apparently allow the system to zoom in and out and thus make it possible to see curbs better which should lead to better Smart Summon. Hopefully, the rewrite will indeed lead to much better Smart Summon and much better AP.

@nepenthe This quote illustrates why I make "absolutist" statements about Tesla's hardware not being good enough for L5. This is a former Tesla employee who worked on AP, admitting that the car only has the side repeaters and maybe the back up camera to see curbs when it is backing out in Smart Summon and that the cameras could only see about 80% of the curbs!! That is not good hardware for doing FSD! Now, Tesla is apparently going to fix this in the upcoming rewrite but it is telling that the current sensors have these issues. With additional parking cameras, this problem could have been fixed immediately and Tesla could have avoided all the problems with Smart Summon. And of course, lidar would have solved this problem as well since lidar would have painted a 3D map and shown exactly and accurately where the curbs are. My point being that with more sensors and better sensors, Tesla could have much higher confidence in solving perception and solving FSD.

In the same video the guest, who just landed a new job at Waymo, stated that Tesla is quite a bit ahead of Waymo in self driving

And basically confirmed that Waymo gave up on having its own car fleet - they will sell their solutions to other car companies.And that they have some lidar problems, which they attempt to fix with even more lidars. And their current self driving cars are all supervised.

So the conclusion seems that Tesla has problems with self driving, but is way ahead of everybody else, and is working faster towards that goal than everybody else.

Waymo had to intervene after the video got released, and voice over each time the "Waymo" phrase got mentioned - pathetic, and a very indicative insecure, move by Waymo.