Mobileye CEO Criticizes Tesla's AI FSD 12 While One Wall Street AI Analyst Praises FSD 12

[Crissa]

I've followed that problem for a long time because it sucked for me being without for a couple years. Do you have a source about how Tesla got around it.

They use a combination of recognition of shapes (signs with different shapes mean different things) via context in appropriate samples in the machine learning, and mapped signs and traffic information (much of it is from Google).

They don't need to read S-T-O-P to recognize an octagonal red blur.

Alot of the stupid behavior of FSD Beta (and other self-driving systems) is based on this.

-Crissa

Sponsored

 

[scottf200]

They use a combination of recognition of shapes (signs with different shapes mean different things) via context in appropriate samples in the machine learning, and mapped signs and traffic information (much of it is from Google).

They don't need to read S-T-O-P to recognize an octagonal red blur.

Alot of the stupid behavior of FSD Beta (and other self-driving systems) is based on this.

-Crissa

I can see stop signs. I've experienced occluded ones that are still known about (displayed on visuals). This seemed to be mapped data.

What about rectangular speed limit sites with numbers on them? As well letters on them for trucks, or night time or school or ...

Do you have a source of this info?

 

[Crissa]

I can see stop signs. I've experienced occluded ones that are still known about (displayed on visuals). This seemed to be mapped data.

What about rectangular speed limit sites with numbers on them? As well letters on them for trucks, or night time or school or ...

Do you have a source of this info?

Speed limits are usually map data.

Which info do you need a source for?

-Crissa

 

[scottf200]

Speed limits are usually map data.

That seems unlikely for all of them because I see it change on my display precisely when I just meet any speed limit sign. I think speds are in map data for general segments of roads. I've seen that in the old Tesla tile data that we used to be able download and someone wrote a program to break it down (segments, etc).

Truck, construction, etc vary a lot. I've seen just truck or other temp signs that seem to be 'read' as I get right up to them.

Which info do you need a source for?

Re: Teslas still cannot read signs, no one but Mobileye can, they can only try to recognize signs.
In this post I quoted it and asked for the source. https://www.cybertruckownersclub.co...t-ai-analyst-praises-fsd-12.9317/#post-179874

It took Tesla multiple years to give me speed limit sign reading between my 2016 and 2017 X during the hardware change. It may have been 3 or even 4 years.

 

[scottf200]

Mobileye seems to be gaining traction in the past year.

Porsche, Zeekr, others.

Mobileye acted as a Tier 0.5 supplier amid the cooperation with Geely Lynk & Co. Previously, Mobileye only supplied semi-finished components to Tier 1 suppliers, but now it is responsible for the complete solution stack for the first time, including hardware, software, drive strategy and control. Mobileye will also provide a multi-domain controller and provide software OTA updates after the system is deployed.

http://www.researchinchina.com/Htmls/Report/2021/70664.html

 

[scottf200]

1 day ago ...
Title: We can move faster in China and from there go global, says Mobileye CEO Amnon Shashua

 

[Crissa]

That seems unlikely for all of them because I see it change on my display precisely when I just meet any speed limit sign.

My Googlemaps does, too. It has precisely zero cameras to tell where it is.

Mobileye has lost and gained partnerships repeatedly. They don't have any L3 or L4 cars on the roads except maybe in China, having had all of theirs in the west pulled from the roads.

-Crissa

 

[scottf200]

Teslas still cannot read signs, no one but Mobileye can, they can only try to recognize signs.

Can I see this source please?

 

[scottf200]

Mobileye has lost and gained partnerships repeatedly. They don't have any L3 or L4 cars on the roads except maybe in China, having had all of theirs in the west pulled from the roads.

I was referring the the past year as their architectural approach and technical abilities are starting to be recognized (thus the acceleration of partnerships). Also see the China related video above.

Aside:
Out of curiosity do you have much Tesla FSDbeta experience as a driver or passenger in different environments?

Just wanted to understand your first hand experience. That gives a lot of insight when seeing what all these systems do and how to judge the complexity of the situations and handle it.

 

[Crissa]

Can I see this source please?

https://www.teslarati.com/tesla-autopilot-update-speed-limit-sign-detection/

https://patents.google.com/patent/US20080137908A1/en

There are others. They interpret this to mean they own the reading of signs, although they've lost a couple cases so far but this one has held.

-Crissa

 

[Crissa]

I was referring the the past year as their architectural approach and technical abilities are starting to be recognized (thus the acceleration of partnerships). Also see the China related video above.

Swapping out partnerships is not 'accelerating', despite how often they put out shiny press releases.

They have Intel backing them now, but... They've collapsed financially several times.

-Crissa

 

[scottf200]

https://www.teslarati.com/tesla-autopilot-update-speed-limit-sign-detection/
https://patents.google.com/patent/US20080137908A1/en
There are others. They interpret this to mean they own the reading of signs, although they've lost a couple cases so far but this one has held. -Crissa

Ok, thanks for pointing to those. They don't seem to be saying what you were tho. Example:

For example, Karpathy stated Tesla holds the most extensive set of “Except Right Turn” signs in its database. These signs often say the same thing but are different sizes, fonts, or shapes, all of which are factors that can spell trouble for a software system.

PLEASE NOTE:
The old highway AP/Autopilot speed limit handling was pretty solid and usable.
The new highway FSDbeta speed limit handling is AWFUL because it changes speeds abruptly based on nearby streets or under/overpasses or something. It is inconsistent.
I've seen this in the past couple weeks on a short road trip across a couple states. I'm on current FSDbeta version.
The speed will drop all over a suddenly form your set 75 or whatever to 60 or 45 and your car will suddenly slow down (regen) even with cars right behind you. Many are reporting it.

 

[scottf200]

Swapping out partnerships is not 'accelerating', despite how often they put out shiny press releases.

They have Intel backing them now, but... They've collapsed financially several times. -Crissa

Interesting. It is hard to know if you are being objective or what the basis is for your dislike of Mobileye. No big deal. Just interesting as I've read similar dozens of times in other forums.

Obviously, intel backing them has huge ramifications for hardware. As well as for financial backing.

Have you watched many of their presentations? For example, how they gather their data for the HD maps as only one part of their input (with other being vision/cameras ... and depending on supervision or chaffer level ... then radar on front/all 4 corners ... then lidar at the unsupervised level). They send back tiny bits of data packets vs massive video. They have their hardware cameras and data bits that get sent back in millions of cars (big manufacturers) for many years.

They are at a lot a ADAS conferences with big relative audiences (vs just a Tesla investor event). They are really transparent and methodical with several types policies and models. This instills way more confidence to me than Teslas method of just try things and throw into a beta test group ... then reverse course (HD maps, radar, USS, etc backtracks). So many of us with FSDbeta have drastically scaled back our belief of FSDbeta in the past couple years because of hyped promises that in the real world are not delivering.

I've followed Mobileye for a few years and have several personal years of experience with AP and FSDb. I use it easily 95% of the time for Eyes ON, Hands ON ... Level 2 driving.

FSD may end up being "good enough" for the way I use it and want to use it. ie. if I end up deciding on a CT and get FSD. I think Mobileye has a much better methodical and professional plan for getting there.

Interesting discussion. No big deal. I just follow several companies in this space and see how they have changed over time.

Aside: Semi-truck highway ADAS is pretty great to see evolve (big implications since the USA ships so so much via highways, and I do road trips where I wish computers were assisting semi-truck drivers keep lane centering, distance following, etc)

 

[PilotPete]

Interesting. It is hard to know if you are being objective or what the basis is for your dislike of Mobileye. No big deal. Just interesting as I've read similar dozens of times in other forums.

Obviously, intel backing them has huge ramifications for hardware. As well as for financial backing.

Have you watched many of their presentations? For example, how they gather their data for the HD maps as only one part of their input (with other being vision/cameras ... and depending on supervision or chaffer level ... then radar on front/all 4 corners ... then lidar at the unsupervised level). They send back tiny bits of data packets vs massive video. They have their hardware cameras and data bits that get sent back in millions of cars (big manufacturers) for many years.

They are at a lot a ADAS conferences with big relative audiences (vs just a Tesla investor event). They are really transparent and methodical with several types policies and models. This instills way more confidence to me than Teslas method of just try things and throw into a beta test group ... then reverse course (HD maps, radar, USS, etc backtracks). So many of us with FSDbeta have drastically scaled back our belief of FSDbeta in the past couple years because of hyped promises that in the real world are not delivering.

I've followed Mobileye for a few years and have several personal years of experience with AP and FSDb. I used easily 95% of the time for Eyes ON, Hands ON ... Level 2 driving.

FSD may end up being "good enough" for the way I use it and want to use it. ie. if I end up deciding on a CT and get FSD. I think Mobileye has a much better methodical and professional plan for getting there.

Interesting discussion. No big deal. I just follow several companies in this space and see how they have changed over time.

Aside: Semi-truck highway ADAS is pretty great to see evolve (big implications since the USA ships so so much via highways, and I do road trips where I wish computers were assisting semi-truck drivers keep lane centering, distance following, etc)

If that’s your complaint with V11.Xbeta, then you should be loving v12 when it is released. Imagine AI learning from 1M video clips of drivers adjusting to a speed limit change. That is how the AI is learning to deal with speed limit changes. So take the average of how almost everyone does it, and that’s how the AI is going to handle it. As has been posted here already, they had to handpick the Stop Sign clips because the AI was learning to “roll/creep” through the stop instead of coming to a full and complete stop (like the law says). I would anticipate a no-regen speed reduction at worst and a “reduced throttle” speed reduction at the other end.

 

[scottf200]

If that’s your complaint with V11.Xbeta, then you should be loving v12 when it is released. Imagine AI learning from 1M video clips of drivers adjusting to a speed limit change. That is how the AI is learning to deal with speed limit changes. So take the average of how almost everyone does it, and that’s how the AI is going to handle it. As has been posted here already, they had to handpick the Stop Sign clips because the AI was learning to “roll/creep” through the stop instead of coming to a full and complete stop (like the law says). I would anticipate a no-regen speed reduction at worst and a “reduced throttle” speed reduction at the other end.

I agree it is interesting. I did watch the 'v12' drive where that aspect was talked about. However, most people just use TACC/ACC to set a speed at or a little higher than they want to go and then follow cars which keeps your distance and speed to the other traffic level.

FSDb speed limit control is absolutely a mess compared to how it use to reliably work. I think it will get worked out eventually but I can't imagine how the Tesla engineers could drive in places like problems that many of us are seeing and think it is any good at all.

The full stop or California stop stop sign issue is one of the main points of the opening post and blog post tho. If you just try to do AI/ML without other controls over it then you have a hard time dealing with restrictions / regulations / laws. I have a feeling it will come back and bit them.

My current stance: FSD may end up being "good enough" for the way I use it and want to use it. ie. if I end up deciding on a CT and get FSD. I think Mobileye has a much better methodical and professional plan for getting there.

Good summary of points and reference links in the blog IMO.
https://www.mobileye.com/opinion/are-we-on-the-edge-of-a-chat-gpt-moment-for-autonomous-driving/

Sponsored