Blog Musk Touts ‘Quantum Leap” in Full Self-Driving Performance

A “quantum leap” improvement is coming to Tesla’s Autopilot software in six to 10 weeks, Chief Executive Elon Musk said a tweet.

Musk called the new software a “fundamental architectural rewrite, not an incremental tweak.”

https://twitter.com/x/status/1294374864657162240

Musk said his personal car is running a “bleeding edge alpha build” of the software, which he also mentioned during Tesla’s Q2 earnings. 

“So it’s almost getting to the point where I can go from my house to work with no interventions, despite going through construction and widely varying situations,” Musk said on the earnings call. “So this is why I am very confident about full self-driving functionality being complete by the end of this year, is because I’m literally driving it.”

Tesla’s Full Self-Driving software has been slow to roll out against the company’s promises. Musk previously said a Tesla would drive from Los Angeles to New York using the Full Self Driving feature by the end of 2019. The company didn’t meet that goal. So, it will be interesting to see the state of Autopilot at the end of 2020.

 

[Mardak]

TESLA

MOBILEYE EyeQ4 (Q4 2017, 2.5 TOPS)

Perhaps you misunderstood Karpathy, but these curb detection examples currently used for Smart Summon are *not* the BEV network, and notably he says stitching up these individual views is not good for FSD needing accurate intersection predictions.

 

[powertoold]

In this cas

Intersection prediction that Tesla is using for their BEV was in production well before Tesla. EyeQ4 had it in Q4 2017 and this is using a chip that has 2.5 TOP imagine their networks in EyeQ5 (24 TOPs).

And in regards to the BEV network, Andrej said "Its relatively standard and well understood." This is quite the opposite of the networks Waymo are developing. They are literally inventing new techniques (through DeepMind) and implementing it at Waymo.

TESLA

MOBILEYE EyeQ4 (Q4 2017, 2.5 TOPS)

Unfortunately currently deployed lane keeping in Mobileye cars is still far behind Tesla. Sad.

Come up with more tier 1 excuses.

We've made it clear that a consumer deployed system should be the emphasis in these discussions. It doesn't matter that some company had something something in their basement.

This is Mobileye level "2+" lane keeping.

 

[Bladerskb]

In this cas


Unfortunately currently deployed lane keeping in Mobileye cars is still far behind Tesla. Sad.

Come up with more tier 1 excuses.

We've made it clear that a consumer deployed system should be the emphasis in these discussions. It doesn't matter that some company had something something in their basement.

This is Mobileye level "2+" lane keeping.

and i can show you 1,000 videos of AP failing on that same/similar curvy road.

 

[Bladerskb]

Perhaps you misunderstood Karpathy, but these curb detection examples currently used for Smart Summon are *not* the BEV network, and notably he says stitching up these individual views is not good for FSD needing accurate intersection predictions.

My point is that the BEV network is being feed similar data that was used to generate the road edge detection just with more labels such as which lane is going to where, oncoming lane, lane type, curb, crosswalks, divider, road edge, etc. Their BEV prediction neural network architecture isn't exotic.

For example, mobileye already has networks that does this and way more. Its quite easy to take this data to create a BEV predicted network. Even Andrej said so.

BEV prediction network from a random student on Medium Blog.

 

[diplomat33]

Elon continues to hype up the FSD rewrite:

 

[Knightshade]

Not really.

He used that exact language months ago to describe why the re-write was needed:
https://twitter.com/elonmusk/status/1294724702967984128

Autopilot was trapped in a local maximum, labeling single camera images uncorrelated in time. Now, it's not.

And it's even a term he's warned LIDAR folks like Waymo about going back to 2018-

Elon Musk still doesn’t think LIDAR is necessary for fully driverless cars

In my view, it’s a crutch that will drive companies to a local maximum that they will find very hard to get out of

One tiny area in Arizona seems like a pretty local maximum to me.

 

[boonedocks]

Reality is complicated... couldn't have seen that one coming!

To be honest I have seen a lot of drivers that couldn’t handle rust on an expressway

 

[banned-66611]

I am curious. When Elon says the beta will go out to "expert and careful drivers", is he implying that it won't go out to the Early Access group but instead will go out to a new group that Tesla considers to be safe drivers (based on data maybe)?

What he means is that they are still stuck at Level 2.

If FSD worked they wouldn't need beta testers.

 

[Snuffysasa]

What he means is that they are still stuck at Level 2.

You are just stating the obvious thing that no one is challenging. FSD/ this rewrite / this upcoming update.... area all of course level 2. Tesla, nor Elon, has never said anything otherwise.

If FSD worked they wouldn't need beta testers.

This doesn't make any sense... you always need to test software, doesn't matter what level it is

 

[banned-66611]

And it's even a term he's warned LIDAR folks like Waymo about going back to 2018-

He's talking out of his arse again. Tesla's "local maximum" problem was that they had a crappy image classifier that was unreliable and only worked on single images. The re-write is trying to expand it to tracking objects over multiple frames but Waymo has been doing that since way back when they had those cute little Google cars all those years ago.

One of the great advantages of lidar is that it gives you the location of the object as well as helping to identify it. Given a few simple assumptions about the speed at which things move (lamp posts are fixed, pedestrians don't go over about 25 kph) it's much much easier and less computationally intensive to track objects.

It looks like Musk has realized, years after everyone else, that he needs to build a 3D model of the environment and track objects in it. That's going to be much harder with only cameras.

 

[banned-66611]

This doesn't make any sense... you always need to test software, doesn't matter what level it is

Isn't it at all worrying or suspicious to you that he is claiming they have developed "full self driving" but have just finished doing a full re-write and not demonstrated anything yet?

I mean these things take everyone else years and years to develop and Musk is claiming the re-wrote it in a few months and now it's safe to test on public roads.

He promised it by the end of the year and is throwing out some half baked crap just to claim he kept his word. 100% certain that robotaxi isn't launching this year as he claimed though if they are only just hitting beta at level 2.

 

[Knightshade]

Isn't it at all worrying or suspicious to you that he is claiming they have developed "full self driving" but have just finished doing a full re-write and not demonstrated anything yet?

That is a gross misrepresentation of what has been said.

He's claimed they're developing full self driving. Not that it's done.

I mean these things take everyone else years and years to develop and Musk is claiming the re-wrote it in a few months and now it's safe to test on public roads.

The re-write has been ongoing for at least a year... in interviews in January he was mentioning it'd been underway for months already at that point- and obviously it leverages/builds lessons learned over the previous years of development- so again you are misrepresenting the facts.

He promised it by the end of the year and is throwing out some half baked crap just to claim he kept his word

He's been pretty clear the re-write wouldn't be great this year.... again you seem to be putting words in his mouth.

A while back when asked if the re-write would handle roundabouts for example he said it should attempt them this year, but not be good at them for probably another year- he's been very clear this isn't finished/ready/L5 or anything like it.

. 100% certain that robotaxi isn't launching this year as he claimed though if they are only just hitting beta at level 2.

If you look at the timing it seems pretty obvious that they thought HW3 would fix the local maximum problem in early 2019... then, it didn't. So they had to go all-in on the 360 rewrite that they're about to do narrow release on.

Will that fit it? We don't know yet. Elon probably doesn't either.


But one thing in his favor- when he finds out he's on the wrong path- he changes it.

See instead Waymo that's still stuck in their tiny square of Arizona with "working" L4 because they refuse to admit their plan can't scale massively.

 

[J1mbo]

The fact that FSD on a HW2.x car is still fundamentally the same as FSD on HW3 (except "visualisations", which I believe @verygreen said were artificially limited to HW3) suggests that the "rewrite" is actually the code release that was always planned for HW3.

It was probably behind the cool stuff we saw on the Autonomy day - AK's 3D pointcloud and the landscape visualisation in the cars, for example.

 

[banned-66611]

That is a gross misrepresentation of what has been said.

He's claimed they're developing full self driving. Not that it's done.

He said that there would be public beta in a few weeks. That means they must be extremely close to having it finished and presumably doing internal testing on public roads by now. Surely even Musk isn't mad enough to unleash an alpha quality "full self driving" on an unsuspecting public?

He's been pretty clear the re-write wouldn't be great this year.... again you seem to be putting words in his mouth.

**April 22nd 2019:** "We expect to be feature complete in self driving this year, and we expect to be confident enough from our standpoint to say that we think people do not need to touch the wheel and can look out the window sometime probably around the second quarter of next year."

**April 22nd 2019:** "We expect to have the first operating robot taxi next year with no one in them! One million robot taxis!"

**May 9th 2019:** "We could have gamed an LA/NY Autopilot journey last year, but when we do it this year, everyone with Tesla Full Self-Driving will be able to do it too"

**April 12th 2020:** "Robotaxis release/deployment... Functionality still looking good for this year. Regulatory approval is the big unknown.

https://twitter.com/elonmusk/status/1249210220200550405

**April 29th 2020:** "So I think we could see robotaxis in operation with the network fleet next year, not in all markets but in some."

**8th July 2020:** World AI Conference: "I remain confident that we will have the basic functionality for level five autonomy complete this year"


He's been slowly rolling it back from "definitely in 2019" to "maybe the basic functionality will be in there by the end of 2020". No announcement that robotaxi has been delayed though so the deadline for that is still 31/12/2020.

 

[Microterf]

Reading through past comments by Musk/ interpretations, I'm still very optimistic that we are getting close to Level 4 and that Tesla's/Comma.ai approach is the only way that works in most areas.

Self-driving is like a Rubik's cube. It is not solved until all 6 sides match. You can 95% of the way there using one strategy to find out that it fails and have to re-do everything and take a different approach. This is what Tesla did in 2016, and why he thought they would be able to do coast to coast in 2017 but were unable to deliver on that.

1. Listening to Karpathy speeches, he brings up all of the "edge cases" that we think about, and talks about how they are solving them. So it's not like they aren't thinking about these things or working on them and avoiding those cases.

2. Musk solves a problem, then admits where they were wrong to begin with. Based on his personality, I wouldn't have expected him to come out their approach was a dead-end until they knew the rewrite would work. At that point, you started to see him talk about the autopilot re-write, and how the old way wasn't going to work.

 

[diplomat33]

Reading through past comments by Musk/ interpretations, I'm still very optimistic that we are getting close to Level 4 and that Tesla's/Comma.ai approach is the only way that works in most areas.

L4 does not mean that it is almost self-driving. L4 means fully capable of driverless but not everywhere. So when you say Tesla is close to L4, what area do you think our Teslas will be driverless? Highway? City?

Self-driving is like a Rubik's cube. It is not solved until all 6 sides match. You can 95% of the way there using one strategy to find out that it fails and have to re-do everything and take a different approach. This is what Tesla did in 2016, and why he thought they would be able to do coast to coast in 2017 but were unable to deliver on that.

Not sure what you mean by "all 6 sides match". What 6 sides? Full self-driving has 3 pillars: perception, planning and driving policy.

• Perception means your car can reliably detect and track all relevant objects around it in real-time (lane lines, traffic lights, signs, other vehicles, pedestrians, etc...)
• Planning means that your car can predict the paths of other objects and figure out in real-time a safe path that avoids collisions, follows the rules of the road and follows your nav route.
• Driving policy means that your car has rules that ensure it drives safely, respects other road users and obeys traffic laws.

You need to implement those 3 pillars in a reliable way if you want a good full self-driving car.

 

[Microterf]

I was saying 6 sides of the Rubik's cube.

I say level 4 meaning that they will be Valenti drive everywhere, but not in every condition. (Heavy rain, snow, etc)

 

[Knightshade]

He said that there would be public beta in a few weeks. That means they must be extremely close to having it finished

Utter and complete nonsense.

FSD has been publicly beta since 2016.

Is it extremely close to finished 4 years running now?


He even explicitly mentioned one common road situation (roundabouts) that it's at least another year before it's good at it.

and presumably doing internal testing on public roads by now. Surely even Musk isn't mad enough to unleash an alpha quality "full self driving" on an unsuspecting public?

....you realize the re-write coming to a limited group in a few weeks is still LEVEL 2, not L4/L5, right?

Because it really sounds like you don't realize that and have invented an imaginary argument based on things nobody has actually claimed are true.

**April 22nd 2019:** "We expect to be feature complete in self driving this year

And it turned out that was a dead end development path and they had to completely re-write the system.

Which has been mentioned a number of times since the ~1.5 year old quote you're citing.

Much of the rest of your quotes seem to also show a lack of understanding between "exists/works at all" and works well"


"Car has a non-zero chance to do a specific thing" (say, turn right at an intersection)- that means the feature exists in the software.

"Car perfectly does a specific thing every time in nearly all circumstances"- that means it works well



that first one is what they're hoping to deliver with the re-write, and Musks been pretty clear that the version that has a non-zero chance to do EVERYTHING will still be quite some time away from doing those new things WELL.

Which is part of why the system will remain L2 for the foreseeable future.




Finally, regarding Elons timelines- He's told you himself don't believe him. So why do you keep doing that?

People should not ascribe to malice that which can easily be explained by stupidity." (LAUGHTER) So-- so it's, like, just because I'm, like, dumb at-- at predicting dates does not mean I am untruthful. I don't know-- I-- we've-- I never made a mass-produced car. How am I supposed to know with precision when it's gonna get done?

Now replace "mass-produced car" with "fully operational L4/L5 driving system" in that statement and you'll see Elons own thoughts on how much weight you should give to his predicted dates about it.[/QUOTE]

 

[zecar]

Finally, regarding Elons timelines- He's told you himself don't believe him. So why do you keep doing that?

His recent comment on share price suggests Musk doesn't believe FSD is workable. What would Tesla be worth today if FSD had a 70% chance of working in the next 3 years.

Musk got busy now on FSD announcements now because of Waymo progress. But unlike when he announced Cybertruck against Rivians building momentum, Musk has essentially no control of FSD's release. So he left with redefining the meaning of FSD to mitigate the accumulation of potential financial liability.

As every observant person driving a current Tesla should know, the car does not reliably sense all objects in its immediate environment. There is no reason to consider complicated edge case strategies when the car can't always be certain that it's driving path is clear. Tesla has demonstrated that cheap cameras and forward looking radar aren't good enough for the current state of the art in machine vision.

Waymo can move assertively because that system has high confidence about all object location within its braking distance. Tesla has repeatedly demonstrated their inability to do the same.

 

[banned-66611]

FSD has been publicly beta since 2016.

It's not FSD though is it? It's L2 driver aids.

They just called it FSD so that idiots would think they were making progress on actual FSD.