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.

 

[heltok]

I used to develop self driving cars. Been thinking about the 4D system a lot. What strikes me is how quickly the problem of self-labelling becomes very complex. You need to label an intitial set of data, then you use autolabelling to to expand this dataset and humans to fix where the self labelling is not sure what to do. Then you add new features to this, such as lane modifiers to speed limit signs forcing you to redesign the labelling again. Over time with more data you will develop better algorithms to generate the 4D point cloud and now you have different point clouds than when you started. Then suddenly you have California Haze and you need to add another air quality to your system and make sure you have enough partially covered stop signs under red haze. And who knows what more problems reality will throw at the developers...

Anyway, I guess my point is that this is not a 1000 IQ120 engineers problem that VW etc can just throw money at fixing, you really need a few IQ 140-160 experienced jedi engineers like Karpathy to come up with a good strategy. Otherwise you very quickly get lost in technical debt, missed deadlines, new goals, rewrites etc where you add more todos than you get rid of, while having to release something to the public and dealing with bugs in that.

I can see why this has been delayed. I assume their engineers have been working very hard under a lot of challenges trying to make this 4D system a reality. And once it is done in 2-4weeks Tesla will have grown their lead even more. And I am happy I don’t have to lose life expectancy trying to make this a reality.

 

[powertoold]

Anyway, I guess my point is that this is not a 1000 IQ120 engineers problem that VW etc can just throw money at fixing, you really need a few IQ 140-160 experienced jedi engineers like Karpathy to come up with a good strategy.

Not only do you need jedi engineers, but you also need a large fleet to gather hundreds/thousands of examples of edge cases. No other company has this. I can't even fathom the complexity of the task, but then again, I'm not that smart lol.

 

[Daniel in SD]

Not only do you need jedi engineers, but you also need a large fleet to gather hundreds/thousands of examples of edge cases. No other company has this.

Mobileye has this.
Intel’s Mobileye has a plan to dominate self-driving—and it might work

 

[mikes_fsd]

Mobileye has this.
Intel’s Mobileye has a plan to dominate self-driving—and it might work

Can you explain how they get that data?
Vast majority of those cars do not have OTA capability to send that data.

Mobileye has bargained for access to sensor data from customer vehicles. Shashua says that Mobileye is already collecting data from Volkswagen, BMW, and Nissan vehicles. He says three other unnamed carmakers have also agreed to participate.

The scale of this program is massive. Mobileye says it is already collecting 6 million kilometers (3.7 million miles) of sensor data every day from vehicles on public roads. Mobileye expects to have more than 1 million vehicles in its European fleet by the end of 2020, and 1 million American vehicles the following year.

Wonder what it costs them to get access to the data.

MobileEye definitely has the right idea.

 

[powertoold]

Again, many fail to appreciate what Tesla has been able to do with autopilot. Not only do they have access to the data. They have access to the whole damn stack, the car, the chip, the wireless connectivity. They control everything. Not only do they control everything, they can update or request data within days or even hours.

Karpathy has mentioned that Tesla is constantly churning their data engine, and they have the ability to request data without even updating the firmware. The feedback loop at Tesla is incredible, and it's required to develop novel software that few think is possible.

 

[Daniel in SD]

Vast majority of those cars do not have OTA capability to send that data.

I'm pretty sure the vast majority of fancy cars these days have cellular modems. I quick glance at VW's site seems to indicate that connectivity is standard.

Wonder what it costs them to get access to the data.

No idea but since Mobileye basically has a monopoly on ADAS (except for Tesla of course) they have a pretty strong negotiating position.

 

[AlanSubie4Life]

many fail to appreciate what Tesla has been able to do with autopilot.

I definitely don’t appreciate what they’ve done so far. Can’t use it with my wife in the car due to phantom braking. And it can’t drive I-5 in the Central Valley properly (a trivial case!).

That being said, it is perfectly adequate as a lane keep assist and adaptive cruise for long road trips (as long as my wife is not in the car of course!). And it is kind of fun to use, just to predict when it will fail, and to laugh at the shortcomings. I would not be surprised at all if we conclude later that it improved safety overall (for some users it certainly makes it less safe though!), even with current significant limitations.

Looking forward to my HW3! I am sure it will fix everything.

 

[Daniel in SD]

I definitely don’t appreciate what they’ve done so far. Can’t use it with my wife in the car due to phantom braking. And it can’t drive I-5 in the Central Valley properly (a trivial case!).

That being said, it is perfectly adequate as a lane keep assist and adaptive cruise for long road trips (as long as my wife is not in the car of course!). And it is kind of fun to use, just to predict when it will fail, and to laugh at the shortcomings. I would not be surprised at all if we conclude later that it improved safety overall (for some users it certainly makes it less safe though!), even with current significant limitations.

Looking forward to my HW3! I am sure it will fix everything.

What's wild is that if people are impressed now I shudder to think about the level of hype that is coming. Once FSD can do someone's commute even once they will become convinced that "Level 5 no geofence" is right around the corner. Peak hype is going to be insane.

 

[diplomat33]

Elon gave an estimated timeline for FSD rewrite release:

 

[Daniel in SD]

Elon gave an estimated timeline for FSD rewrite release:

I wonder if the rewrite also runs on HW2.5?

 

[diplomat33]

I wonder if the rewrite also runs on HW2.5?

I would guess no. The rewrite does not sound like something that HW2.5 could handle.

 

[nukequazar]

nukequazar said:
These cars with current sensor hardware and available computing power will never get past Level 2 Autonomy. They would need LIDAR and constantly updating 3D HD maps to get to "full self driving."

S0XY said:
Prove it

Time will tell... mark my words.

 

[Cheburashka]

Elon gave an estimated timeline for FSD rewrite release:

Didn't he give that same timeline a few weeks back?

Meaning the timeline is now n + 2 to 4 weeks, where n is today's date.

 

[diplomat33]

Didn't he give that same timeline a few weeks back?

Meaning the timeline is now n + 2 to 4 weeks, where n is today's date.

Yes. No surprise that it got delayed.

 

[Cheburashka]

Still better than 6 months maybe 12 months definitely. So we are definitely closer to release!

 

[Daniel in SD]

Yes. No surprise that it got delayed.

I really think the only reason he gives timelines is to put pressure on the engineering team. Smart Summon was 6 weeks away on October 31st 2018 and wasn't released until September 2019.

 

[Daniel in SD]

Still better than 6 months maybe 12 months definitely. So we are definitely closer to release!

It would be interesting for someone to study every date and calculate the true "Elon Time" conversion factor.
"3 months maybe, 6 months definitely" was Jan 23rd 2017.
Stop light and stop sign recognition was released April 26th 2020.

Scale factor is about 6? So maybe next fall for the rewrite? This forum should implement a gambling feature so we can all bet!
I feel like it will only be a couple months late this time.

 

[emmz0r]

I would guess no. The rewrite does not sound like something that HW2.5 could handle.

That's going to be very interesting regarding smart summon and the other features promised to EAP owners. I would think sign reading and other things would be integral to deliver EAP feature set.

 

[mikes_fsd]

That's going to be very interesting regarding smart summon and the other features promised to EAP owners. I would think sign reading and other things would be integral to deliver EAP feature set.

I believe it will run on 2.5 for EAP and to allow the better NN's to run on all cars for the safety features (AEB, etc).
I also believe that was a small portion of why they bought DeepScale Deepscale

Their purpose is to make NN's smaller and thus run on older/less powerful hardware.

Some info Tesla just bought an AI startup to improve Autopilot—here’s what it does

 

[AlanSubie4Life]

Smart Summon was 6 weeks away on October 31st 2018 and wasn't released until September 2019.

What is there to disagree with here? Mysterious.