Will Tesla be able to deliver FSD with HW3.0 and current Model 3 sensor suite, ever?

[electrifyme]

Nope. It's a bunch of false promises and people paying to fund the development. I'd rather get a full working function than hoping it will come. Heck, you may even sell the car before it arrives... so there goes your money. I'll wait for it to be proven. I'd rather use that money for investments than a sunk cost for false promises.

 

[SergeyUser]

Interesting partial list of examples, which I agree with. What is becoming more clear to me is that the diversity of often nonstandard roadsigns out there, sometimes with conditional driving directions, means that recognizing a sign against a pattern may not be actually get us there. There may need to be text recognition and comprehension functions added to the AI in order to deal with these and other nefarious cases, given that sign creators seem to delight in inventing ever more edge cases for sign recognition.

Such examples also illustrate, I think, that relying on a remote database is unlikely to ever be a satisfactory solution. I can't think of any other technology but vision systems that will ever compete with the proliferation of states sign types, conditions, and especially with dynamic sign types (construction on this mile today, on the next mile next week, ...)

Duplicating the efficacy of human recognition and understanding (when the human is focused and not distracted, sleepy, or intoxicated) is no small task.

Actually, I don't see a problem maintaining correct database of every sign, speed limit, etc. The only problem is that while States know where they put signs, this information is lost and has to be re-collected by private companies. As soon as governments will realize that maintaining and publishing the info for everyone to use will increase public safety - they will probably do it.

 

[SammichLover]

Interesting partial list of examples, which I agree with. What is becoming more clear to me is that the diversity of often nonstandard roadsigns out there, sometimes with conditional driving directions, means that recognizing a sign against a pattern may not be actually get us there. There may need to be text recognition and comprehension functions added to the AI in order to deal with these and other nefarious cases, given that sign creators seem to delight in inventing ever more edge cases for sign recognition.

You need to keep in mind just how massive Tesla's data gathering is on this, every vehicle they ship that doesn't opt out is a source, and how well NN does integrating edge-case and variety in sets of the same type.

 

[SergeyUser]

BTW, I think the most basic FSD failure that is purely software-related is inability of the existing Autopilot/FSD implementation to stay in the lane without seeing lane markings before intersections and in other scenarios (someone may call it "split the difference" problem). Basically, Tesla cannot deduce where your current lane is, without swinging. Vision (cameras) are fine, but the software is not. It is just not "smart enough" to deduce such simple logic pretty much any driver can understand. I mean, come on. Yes, many people are worse than that, and I'm not indicating this as an actual issue per se (but for another reason - people are just lazy and they choose not to focus on the task, while Tesla software just can't). This is just an illustration how difficult it is to make the software "smart". In a words of someone who worked in AI space (The End of Starsky Robotics): The biggest problem, however, is that supervised machine learning doesn’t live up to the hype. It isn’t actual artificial intelligence akin to C-3PO, it’s a sophisticated pattern-matching tool.

I believe this is the issue at hand, as far as software is concerned. There is no AI right now, in a way people understand intellect. Very far from it. But it is very easy to believe in "exponential rate of improvement" and stuff like that.

 

[SammichLover]

BTW, I think the most basic FSD failure that is purely software-related is inability of the existing Autopilot/FSD implementation to stay in the lane without seeing lane markings before intersections and in other scenarios (someone may call it "split the difference" problem). Basically, Tesla cannot deduce where your current lane is, without swinging. Vision (cameras) are fine, but the software is not. It is just not "smart enough" to deduce such simple logic pretty much any driver can understand. I mean, come on. Yes, many people are worse than that, and I'm not indicating this as an actual issue per se (but for another reason - people are just lazy and they choose not to focus on the task, while Tesla software just can't). This is just an illustration how difficult it is to make the software "smart". In a words of someone who worked in AI space (

The End of Starsky Robotics): The biggest problem, however, is that supervised machine learning doesn’t live up to the hype. It isn’t actual artificial intelligence akin to C-3PO, it’s a sophisticated pattern-matching tool.

I believe this is the issue at hand, as far as software is concerned. There is no AI right now, in a way people understand intellect. Very far from it. But it is very easy to believe in "exponential rate of improvement" and stuff like that.

Perhaps said "people" don't really understand human intellect as it actually is?

Superior pattern processing is the essence of the evolved human brain

Setting to the side mysticism and suppositions of souls and such, at our core we're really just extremely good pattern matcher engines (lots and lots of them, in a web of engines functioning roughly in unison).

Does HW3 have enough pattern recognition power to function sufficiently in this narrow application? ¯\_(ツ)_/¯ I remain skeptical to a degree but I have to admit I'm less than I was 2 years ago (when I gave a pass on buying FSD at all) and even 1 year ago (when I did buy FSD, but made the purchase as what was in hand and on spec of the stoplight/stop sign functionality), and it is very likely the right overall conceptual path to the end, ultimately.

EDIT: BTW AP does indeed suppose lanes without markings. AP1 did it via following other vehicles (leading to some hilarious results at times) but the current AP does suspect, with a reasonable level of accuracy, where the actual lanes are. As far as I can tell it is using edge detection and "deducing" from the overall dimensions. Enough that it will continue to drive successfully in its lane on a 2 lane road when the lines disappear. You can also see them on roads where it can't get to with AP still engaged. AP won't let you engage though, as at this time the confidence judging gives it a fail, still. You can also see this stuff there on the AP data dumps that the hackers do. So it is there in the wings, waiting.

 

[SergeyUser]

Perhaps said "people" don't really understand human intellect as it actually is?

Superior pattern processing is the essence of the evolved human brain

Setting to the side mysticism and suppositions of souls and such, at our core we're really just extremely good pattern matcher engines (lots and lots of them, in a web of engines functioning roughly in unison).

Does HW3 have enough pattern recognition power to function sufficiently in this narrow application? ¯\_(ツ)_/¯ I remain skeptical to a degree but I have to admit I'm less than I was 2 years ago (when I gave a pass on buying FSD at all) and even 1 year ago (when I did buy FSD, but made the purchase as what was in hand and on spec of the stoplight/stop sign functionality), and it is very likely the right overall conceptual path to the end, ultimately.

Interesting thought and article. Maybe you are right and we certainly don't understand how brain works. I can think of other principal ways why car is limited in its ability to match humans, but that would be very far away from FSD topic and theoretical. Right now I just see how very simple tasks from a human prospective are slow and difficult to solve, with very little progress being made (yes, it is exciting, but it depends on your baseline, and if we use human as a baseline - this is certainly not that exciting).

 

[SergeyUser]

BTW, is there a way to place monetary bets on what will happen with FSD, the same way people bet on sports? I would love to place some bets against people who think otherwise. That would be fun.

 

[SammichLover]

BTW, is there a way to place monetary bets on what will happen with FSD, the same way people bet on sports? I would love to place some bets against people who think otherwise. That would be fun.

Yes. Buy FSD while having your personal valuation of the current below the current list price.

Other than that and the indirect bet via going very long/short on TSLA, I don't think so? I checked at a couple of the UK betting sites and I don't see anything like this. Probably too long a window until resolution for them to bother with.

 

[Tessaract]

Actually, I don't see a problem maintaining correct database of every sign, speed limit, etc. The only problem is that while States know where they put signs, this information is lost and has to be re-collected by private companies. As soon as governments will realize that maintaining and publishing the info for everyone to use will increase public safety - they will probably do it.

Really? It's a huge problem to maintain an up-to-the-minute, 100% accurate database. Firstly, relying on government to run this database would mean that it's not even close to being up-to-the-minute. We'd be lucky if it were up-to-the-month. Secondly, many signs are not erected by state governments, so now you gotta coordinate unknown many agencies. Given the problems even two government agencies have with coordination, the chances successful multi-agency coordination are zero. Thirdly, lives depend on the information to be accurate. What government agency is 100% accurate. Again zero chance.

I think it complete fantasy to think anyone would trust such a database with their lives. Except for those who play video games while their car runs beta SW to drive for them.

 

[Tessaract]

You need to keep in mind just how massive Tesla's data gathering is on this, every vehicle they ship that doesn't opt out is a source, and how well NN does integrating edge-case and variety in sets of the same type.

While the gathering data is great, gathering meaning is different from gathering data. And that's precisely the point. It's not an easy task to extract meaning from any arbitrary sign. Merely matching it against other signs will not always accomplish this, because the sign in question could be new, or in a different font, or the same sentence split over multiple lines differently... To accurately extract meaning from any arbitrary sign requires ultra-reliable text recognition, and advanced linguistic parsing. Not impossible, but certainly not a reality in any Tesla FSD SW yet.

 

[SergeyUser]

Really? It's a huge problem to maintain an up-to-the-minute, 100% accurate database. Firstly, relying on government to run this database would mean that it's not even close to being up-to-the-minute. We'd be lucky if it were up-to-the-month. Secondly, many signs are not erected by state governments, so now you gotta coordinate unknown many agencies. Given the problems even two government agencies have with coordination, the chances successful multi-agency coordination are zero. Thirdly, lives depend on the information to be accurate. What government agency is 100% accurate. Again zero chance.

I think it complete fantasy to think anyone would trust such a database with their lives. Except for those who play video games while their car runs beta SW to drive for them.

I assume state and local governments already have some inventory in place. It should be some internal process before a sign is erected (approvals, work orders, etc). If it can be published to the public - what is wrong with this?

 

[SammichLover]

While the gathering data is great, gathering meaning is different from gathering data. And that's precisely the point. It's not an easy task to extract meaning from any arbitrary sign. Merely matching it against other signs will not always accomplish this, because the sign in question could be new, or in a different font, or the same sentence split over multiple lines differently... To accurately extract meaning from any arbitrary sign requires ultra-reliable text recognition, and advanced linguistic parsing. Not impossible, but certainly not a reality in any Tesla FSD SW yet.

Actual sentence parsing isn't really what we're talking about, here. Defaulting to the low speed for complex schedule info (for example School zones) is an entirely viable solution, pending manual reporting to update the static background geographic database. Especially if you add tags for the sign in the database to note the very oddball sign for future scans.

You don't need truly 100% sign reading to represent an improvement over humans, because let's face it, humans certainly don't do that on the whole and very rarely in the individual.

 

[JayNJ]

Nope, not on any paperwork I have and I keep everything. Maybe written in blogs like this. Or in the press. I dunno. In any event I am not mad about being misled by Tesla Sales Consultants and Elon with ??promises?? ="declare that something will happen" (they quoted timelines)

I just would like ONE promise kept and that was about a timely re tool of my hw 2.0 up to 3.0 promised many time by Tesla over the past 2 years. (yes, before it even existed...but known that an upgrade was required). Just keep that promise so I can see all the stuff that the Model 3 owners are giddy about. After all, they took my $150,000 for the car, and any promised features lightning-fast.

I do agree if FSD gets litigated yet again.....Tesla will come across badly and customers as sympathetic. A mess I don't want to be a part of.

On the order page.

 

[Kognos]

No matter how the random elements of the FSD issue lands, I absolutely hate that it requires a multi-thousand dollar investment where we are constantly reminded that the price rises, and to "Buy now, before it's too late!!"

It's just snarky. It's the least sexy thing about Tesla.

Considering the hardware is already present in the system, this should both be a subscription service and also it should move with the owner, and not be tied fundamentally to the car.

 

[Nuclear Fusion]

Considering the hardware is already present in the system, this should both be a subscription service and also it should move with the owner, and not be tied fundamentally to the car.

Disagree. At the end of the day Tesla is a business, not a charity. That vision system hardware is there to make profits
They didn’t have to put the hardware in the car for anything beyond autopilot

 

[mtndrew1]

I assume state and local governments already have some inventory in place. It should be some internal process before a sign is erected (approvals, work orders, etc). If it can be published to the public - what is wrong with this?

Bless your heart.

-Government employee

 

[Kognos]

Disagree. At the end of the day Tesla is a business, not a charity. That vision system hardware is there to make profits
They didn’t have to put the hardware in the car for anything beyond autopilot

That is flat out not true. With the exception of modernizing and upgrading the CPU from (HW2.5 or whatever to 3.0), if you order a Tesla M3 with nothing else, simply Autopilot -- purchasing FSD requires absolutely no hardware changes in your car. It's entirely software. It's not a charity when it's already present.

I get it though, as they add features they update the price, but it's not worth 8000 for what features it has now, and it doesn't help with the value of the car. It should still be a subscription service tied to the user in some way, because all the cars are inherently equal.

 

[WilliamG]

That is flat out not true. With the exception of modernizing and upgrading the CPU from (HW2.5 or whatever to 3.0), if you order a Tesla M3 with nothing else, simply Autopilot -- purchasing FSD requires absolutely no hardware changes in your car. It's entirely software. It's not a charity when it's already present.

I get it though, as they add features they update the price, but it's not worth 8000 for what features it has now, and it doesn't help with the value of the car. It should still be a subscription service tied to the user in some way, because all the cars are inherently equal.

That logic is akin to saying my computer has a (hardware) graphics card therefore I should get all the (software) video games for free.

I bought FSD, for the record.

 

[Nuclear Fusion]

That is flat out not true. With the exception of modernizing and upgrading the CPU from (HW2.5 or whatever to 3.0), if you order a Tesla M3 with nothing else, simply Autopilot -- purchasing FSD requires absolutely no hardware changes in your car. It's entirely software. It's not a charity when it's already present.

I get it though, as they add features they update the price, but it's not worth 8000 for what features it has now, and it doesn't help with the value of the car. It should still be a subscription service tied to the user in some way, because all the cars are inherently equal.

The car doesn’t need 8 cameras, ultrasonic sensors & a 144TOPS computer for autopilot

 

[Kognos]

The car doesn’t need 8 cameras, ultrasonic sensors & a 144TOPS computer for autopilot

That's not the point.
The point is that whether you have FSD or Autopilot, the same hardware is there.

That logic is akin to saying my computer has a (hardware) graphics card therefore I should get all the (software) video games for free.

I bought FSD, for the record.

I didn't imply that or mean to imply that. If you buy a new Tesla M3 right now you have HW3.0 and cameras available (arguably - some cars still come with 2.5 but I defer that argument) for full FSD functionality with or without the software/licensing for FSD. I dislike and abhor the concept that when buying the product -- which I DO support a purchase arrangement of some kind - that the price keeps elevating without a lot of ROI, and that this purchase is fundamentally tied to a car.

A lot of software is licensed to a user and stays with the user. If all the Tesla cars are fundamentally the same AP hardware at this point, why do I need to repurchase 8000 worth of software every time I get a new car, if the next Tesla is also going to have all the same hardware in it whether I choose to buy the license or not?

This came to a huge point for me when I attempted to sell my M3 (I was just seeing value)... there was no way at all for me to add value to the sale by saying it came with FSD. The car value was the sticker price without FSD every time. If I could get the money back for the investment, then great. But otherwise I'm throwing money away on features, and that's even before the point that this is really all about NoA functionality and driveway/parking lot trickery, and not a lot of actual driving intelligence that's usable to end users.