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

[PilotPete]

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/

Whether people like v12 FSD or not may come down to this; “How well do you tolerate someone who drives very well, just not ’like you do’?”. It may drive better/safer/more predictably than you do, but it still may well come down to your tolerance for someone else’s driving.

As for the AI/ML question, they are hand picking many of the scenarios that are played and learned from. As I recall, of the original 10,000 A100 chips, 4,000 are dedicated to labeling the clips for learning. And if/when this method presents as a solution, it is only goign to get exponentially better with every load of new videos.

Sponsored

 

[scottf200]

Whether people like v12 FSD or not may come down to this; “How well do you tolerate someone who drives very well, just not ’like you do’?”. It may drive better/safer/more predictably than you do, but it still may well come down to your tolerance for someone else’s driving.

That is a good interpretation.

The problem is people drive differently ... and sometimes radically different.

• in the same geographical area
• 'young' vs 'old' ... calm vs aggressive
• with and without passengers in the front seat --- rear seat WAY worse for many (corners, braking, etc)
• in USA cities vs USA rural vs USA suburbia
• in different countries ... China vs India vs USA
• along the lines of above is different types of vehicles (ratio of scooters, tuktuks, etc to cars)

 

[Crissa]

I grow tired of presentations which fluff one and are negative of others.

I expect them to think their method is the best. Of course.

But when rubber meets the road, Tesla (and GM and others) have more units on the streets learning than Mobileye, and Mobileye depends upon process patent trolling to keep their edge rather than a better product.

That I will be biased against.

-Crissa

PS, 'having an extensive database of signs' is not the same as 'reading'.

 

[scottf200]

But when rubber meets the road, Tesla (and GM and others) have more units on the streets learning than Mobileye

FYI, REM (Road Experience Management) data gathering (10k data packets) has been going on for several years and includes uploads from Nissan, BMW, Volkswagon, and other OEMs.

Mobileye presentations show visually USA and EU (etc) maps on how much they gather in a day, week, etc. They get ton of data rapidly. It is pretty stunning on how quickly they get it from so many sources.

The next few years will be interesting for sure with ADAS. So many companies in China have caught up (and some passed) Tesla FSD as far as I can tell. I think FSD will be OK in L2, L2+ but not the best. I never see FSD doing driverless and suspect it is years before Hands OFF.

 

[PilotPete]

I think FSD will be OK in L2, L2+ but not the best. I never see FSD doing driverless and suspect it is years before Hands OFF.

Sorry, but my computer education and current experiences are going to step in here. Whether you see FSD getting to L5 or not is completely your personal bias and/or your misunderstanding of what is occurring.

I don’t care what company is involved, if you have dedicated the kind of resources that Tesla has in the past MONTH to AI learning, you are likely to leave everyone in the dust. The only way not to, is to be stupid in how you manage the resources. And of all the things I think EM is, stupid ain’t one of them. Think about it, 10,000 NVidia A100 processors, 10,000 H100 GPUs now online. And NVidia just updating the background software to DOUBLE the throughput of the H100. AND Dojo clusters are coming online. Your judgement of FSD is based on lines of code trying to cover every base, and even the YouTubers have been talking about how good the latest version is. But v12 is going to be AI driven (no pun intended), and that is creating a set of driving rules for FSD that is going to show exponential growth and capabilities. When you feed a “stop sign” video through, AI looks at how the stop sign was handled., AND how everyone else acted and reacted, AND the lane holding, and did they turn right or left, and how was the turn signal used, and…. and… and… it looks at and sees and learns from EVERYTHING. While you are training it on lane keeping on the freeway, it is learning what you and the other drivers are doing. It will self learn when a turn signal comes on, what to expect. Or brake lights, or debris on the road, or cones, or lane closures, or someone just being stupid and drifting over into your lane. And unlike you, it will be looking for all of this when it drives, 100% of the time, without looking down at the speedometer, or the radio, or… full attention all the time.

I’ve seen the very beginnings of AI driven flight assist stuff. And I’m here to tell you, the future is here, alive and well. I wouldn’t want to get into ACM against a AI opponent. It can see smaller hints and react faster than you or I could ever dream about. This is exactly what was needed to take them over the top on FSD. How long it takes them to get over the top is going to be measured in months and weeks, not years. L5 is going to ready in just a matter of 1-2 years, whether the regulators are or not! And, it may be less than a year, if they can force feed enough video through.

 

[Sirfun]

Sorry, but my computer education and current experiences are going to step in here. Whether you see FSD getting to L5 or not is completely your personal bias and/or your misunderstanding of what is occurring.

I don’t care what company is involved, if you have dedicated the kind of resources that Tesla has in the past MONTH to AI learning, you are likely to leave everyone in the dust. The only way not to, is to be stupid in how you manage the resources. And of all the things I think EM is, stupid ain’t one of them. Think about it, 10,000 NVidia A100 processors, 10,000 H100 GPUs now online. And NVidia just updating the background software to DOUBLE the throughput of the H100. AND Dojo clusters are coming online. Your judgement of FSD is based on lines of code trying to cover every base, and even the YouTubers have been talking about how good the latest version is. But v12 is going to be AI driven (no pun intended), and that is creating a set of driving rules for FSD that is going to show exponential growth and capabilities. When you feed a “stop sign” video through, AI looks at how the stop sign was handled., AND how everyone else acted and reacted, AND the lane holding, and did they turn right or left, and how was the turn signal used, and…. and… and… it looks at and sees and learns from EVERYTHING. While you are training it on lane keeping on the freeway, it is learning what you and the other drivers are doing. It will self learn when a turn signal comes on, what to expect. Or brake lights, or debris on the road, or cones, or lane closures, or someone just being stupid and drifting over into your lane. And unlike you, it will be looking for all of this when it drives, 100% of the time, without looking down at the speedometer, or the radio, or… full attention all the time.

I’ve seen the very beginnings of AI driven flight assist stuff. And I’m here to tell you, the future is here, alive and well. I wouldn’t want to get into ACM against a AI opponent. It can see smaller hints and react faster than you or I could ever dram about. This is exactly what was needed to take them over the top on FSD. How long it takes them to get over the top is going to be measured in months and weeks, not years. L5 is going to ready in just a matter of 1-2 years, whether the regulators are or not! And, it may be less than a year, if they can force feed enough video through.

That was my motivation for helping my mother-in-law study and pass her driver's license renewal test this summer. She now has a license to legally drive until she's 92. Even though regulators won't accept this technology for a while. At least she will be able to use it, and to go where she wants to go and feel independent. She still is a pretty good driver on her own, but with FSD I can be much more confident she will be fine.

 

[scottf200]

Sorry, but my computer education and current experiences are going to step in here. Whether you see FSD getting to L5 or not is completely your personal bias and/or your misunderstanding of what is occurring.

I don’t care what company is involved, if you have dedicated the kind of resources that Tesla has in the past MONTH to AI learning, you are likely to leave everyone in the dust. The only way not to, is to be stupid in how you manage the resources. And of all the things I think EM is, stupid ain’t one of them. Think about it, 10,000 NVidia A100 processors, 10,000 H100 GPUs now online. And NVidia just updating the background software to DOUBLE the throughput of the H100. AND Dojo clusters are coming online. Your judgement of FSD is based on lines of code trying to cover every base, and even the YouTubers have been talking about how good the latest version is. But v12 is going to be AI driven (no pun intended), and that is creating a set of driving rules for FSD that is going to show exponential growth and capabilities. When you feed a “stop sign” video through, AI looks at how the stop sign was handled., AND how everyone else acted and reacted, AND the lane holding, and did they turn right or left, and how was the turn signal used, and…. and… and… it looks at and sees and learns from EVERYTHING. While you are training it on lane keeping on the freeway, it is learning what you and the other drivers are doing. It will self learn when a turn signal comes on, what to expect. Or brake lights, or debris on the road, or cones, or lane closures, or someone just being stupid and drifting over into your lane. And unlike you, it will be looking for all of this when it drives, 100% of the time, without looking down at the speedometer, or the radio, or… full attention all the time.

I’ve seen the very beginnings of AI driven flight assist stuff. And I’m here to tell you, the future is here, alive and well. I wouldn’t want to get into ACM against a AI opponent. It can see smaller hints and react faster than you or I could ever dram about. This is exactly what was needed to take them over the top on FSD. How long it takes them to get over the top is going to be measured in months and weeks, not years. L5 is going to ready in just a matter of 1-2 years, whether the regulators are or not! And, it may be less than a year, if they can force feed enough video through.

I appreciate you taking the time to write up those thoughts. I follow this field (ADAS) fairly closely.

My profession is/was software development tho not in AI/ML. I've been learning about it a few years now tho.

My personal bias is based on reading how others perceive and use FSD over the past few years and my many miles of experience. I use it easily 95% of my driving. I have 90K miles on my current 2017 TMX with upgraded cameras, upgraded FSD computer, and upgraded MCU2 (info and main interface for drivers display and main display). I've probably driven half on AP and half on FSD (current 11.4.7). My previous 2016 TMX with 36K miles probably had AP driven on it mostly as well. Many suburbia and roadtrips.

I understand your comments about Elon and have similar ones. However, he clearly is a harsh boss and overrides anything he wants because he is the 'boss'. Meaning even if there are training professionals with more knowledge and skill he certainly demands they do whatever he thinks is the target.

Not sure if you read the blog. People including me realize how powerful AI/ML is by feeding it examples. That is understood. It is really a matter of if you can architect it in such a way that you can have 'guardrails'. Stop signs are an obvious example. There are laws, regulations, and restrictions that Tesla *must* follow or they can't allow it to be used. This seems clear to me. Then there are many user preferences. The current FSD barely gives anyone behind you to react (change lanes, slow down) before it turns on the blink *right* before the corner you are about to turn as another example. Exiting timing of on-ramps is another or lane changes for various reasons (upcoming turn, exit, slow vehicles, etc.) are other examples.

I just don't think FSD v12 one-size-fits-all is going to be easy to manage or control. It sounds like you think this is easy and this knob will be just giving it different videos and rebuilding it as fast as possible. I get some of that. I think this will only work in some scenarios.

Thanks for the conversation! Interesting times!!

P.S. Re: even the YouTubers have been talking about how good the latest version is.
I used to watch a lot of these and now do some. Many "influencers" really seem to show the best examples but not the worse. Many deviate from the real world experience of myself and many others on other forums. My lastest FSD 11.4.7 still screws up regularly in pretty simple ways.

 

[scottf200]

That was my motivation for helping my mother-in-law study and pass her driver's license renewal test this summer. She now has a license to legally drive until she's 92. Even though regulators won't accept this technology for a while. At least she will be able to use it, and to go where she wants to go and feel independent. She still is a pretty good driver on her own, but with FSD I can be much more confident she will be fine.

That is pretty fantastic. Certainly a great case IF we can get older folks to actually use it. Much safer for them AND the others on the road!

 

[PilotPete]

That was my motivation for helping my mother-in-law study and pass her driver's license renewal test this summer. She now has a license to legally drive until she's 92. Even though regulators won't accept this technology for a while. At least she will be able to use it, and to go where she wants to go and feel independent. She still is a pretty good driver on her own, but with FSD I can be much more confident she will be fine.

THIS right here is the best damn reason to make it happen!!!

 

[Sirfun]

That is pretty fantastic. Certainly a great case IF we can get older folks to actually use it. Much safer for them AND the others on the road!

I think my strategy will be that when we go places together I have her drive.

And I'll sit in the passenger seat and watch nervously at first.

 

[PilotPete]

I appreciate you taking the time to write up those thoughts. I follow this field (ADAS) fairly closely.

My profession is/was software development tho not in AI/ML. I've been learning about it a few years now tho.

My personal bias is based on reading how others perceive and use FSD over the past few years and my many miles of experience. I use it easily 95% of my driving. I have 90K miles on my current 2017 TMX with upgraded cameras, upgraded FSD computer, and upgraded MCU2 (info and main interface for drivers display and main display). I've probably driven half on AP and half on FSD (current 11.4.7). My previous 2016 TMX with 36K miles probably had AP driven on it mostly as well. Many suburbia and roadtrips.

I understand your comments about Elon and have similar ones. However, he clearly is a harsh boss and overrides anything he wants because he is the 'boss'. Meaning even if there are training professionals with more knowledge and skill he certainly demands they do whatever he thinks is the target.

Not sure if you read the blog. People including me realize how powerful AI/ML is by feeding it examples. That is understood. It is really a matter of if you can architect it in such a way that you can have 'guardrails'. Stop signs are an obvious example. There are laws, regulations, and restrictions that Tesla *must* follow or they can't allow it to be used. This seems clear to me. Then there are many user preferences. The current FSD barely gives anyone behind you to react (change lanes, slow down) before it turns on the blink *right* before the corner you are about to turn as another example. Exiting timing of on-ramps is another or lane changes for various reasons (upcoming turn, exit, slow vehicles, etc.) are other examples.

I just don't think FSD v12 one-size-fits-all is going to be easy to manage or control. It sounds like you think this is easy and this knob will be just giving it different videos and rebuilding it as fast as possible. I get some of that. I think this will only work in some scenarios.

Thanks for the conversation! Interesting times!!

P.S. Re: even the YouTubers have been talking about how good the latest version is.
I used to watch a lot of these and now do some. Many "influencers" really seem to show the best examples but not the worse. Many deviate from the real world experience of myself and many others on other forums. My lastest FSD 11.4.7 still screws up regularly in pretty simple ways.

Scott,

I don’t think it is “easy” by any stretch. But I do think it is “doable” With enough resources. And I I think Tesla has those resources in place. I recently saw an autopilot NOT certified for landing follow the AI directed flight commands and make a perfect landing. And the AI only had about 20,000 landing videos shoved through it. If they had the processing horsepower and the amount of data Tesla has, then I’m starting to fear for my job! The more of this I see, the more I know that even my optimism is well short of where this can be and when.

I appreciate your conversation as well. Remember the Chinese curse (no, not THAT one!) “May you live in interesting times.”

 

[FutureBoy]

Mobileye seems to be gaining traction in the past year.

Porsche, Zeekr, others.

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

OMG

People complain about Tesla calling FSD Beta and that the promises of delivery have not been hit. Here is the thing though. Tesla has never claimed that their system is anything above Level 2. And since it has been available to the public, it has always been labeled Beta.

Now let's look at Mobile Eye:

No one wants to say anything about them claiming to be "NEAR-COMPLETE" since 2016?

And I hear now that Mercedes is selling it's system as being Level 3. See what Zack and Jesse have to say about it:

 

[FutureBoy]

I'm not going to claim any knowledge as to who is going to win the robo-taxi race. I do believe that whoever does it first and well will win the market in near totality. Right now my guess is that it will be Tesla but everything I'm basing that on is 2nd or 3rd hand.

Neil deGrasse Tyson gives a little historic context though and what he thinks is going to happen to people driving in a very short time span.

 

[PilotPete]

Neil deGrasse Tyson gives a little historic context though and what he thinks is going to happen to people driving in a very short time span.

I’m sorry, but NdGT wants to be a celebrity so bad. He doesn’t appear interested in all about the actual science business, just getting guest spots and cable shows.

The don’t ride horses on the streets, they go to the stables? These are the words of a man who NEVER gets out of big cities. I live in Southern California, a rather busy suburban area. Where I live, 9,000 sqft home lots are standard. There are condos in my area that are 7 digits in value. Guess what, you can go into the downtown area and see horses on the streets from time to time. During certain parts of the year, they are there on a regular basis. Are they the standard mode of transportation? Nope, but there are a few of them tied up next to the bar and barbershop on certain days. And his 20 year timeline might hold up in the urban areas, but in rural America, not so much,

 

[Crissa]

I’m sorry, but NdGT wants to be a celebrity so bad. He doesn’t appear interested in all about the actual science business, just getting guest spots and cable shows.

The don’t ride horses on the streets, they go to the stables? These are the words of a man who NEVER gets out of big cities. I live in Southern California, a rather busy suburban area. Where I live, 9,000 sqft home lots are standard. There are condos in my area that are 7 digits in value. Guess what, you can go into the downtown area and see horses on the streets from time to time. During certain parts of the year, they are there on a regular basis. Are they the standard mode of transportation? Nope, but there are a few of them tied up next to the bar and barbershop on certain days. And his 20 year timeline might hold up in the urban areas, but in rural America, not so much,

I live in the mountains, we have horses here. They only ride them in the street insofar as it gets them to the trails, to cross the highway, and that's it. A few people have rubber soles to ride in urban areas, but they're the minority. Horses are basically pedestrians.

The vast majority of riding isn't in streets, but trails and specific stables and often private land just set aside for that.

Of course, he's using 'street' to mean 'road', in the Strongtown version of the word.

-Crissa

Sponsored