Navigate on Autopilot is Useless (2018.42.3)

[DrDabbles]

This is absolutely incorrect 3 times over. Comparison to general purpose compute cores is completely unfounded. You're just not paying attention. Custom ASIC's used for AI training is because it's doing an entirely different task. It's got nothing to do with VC's. It's all about how many transistors you can buy per dollar and how many you can put to doing work. The eventual conversion is from kWh->trained models.

Dude. It's my job to know these things, you're supremely far off base. You realized that the TRIP chips are custom ASICs with a bunch of prepackaged ARM IP, right? Nothing to do with training, everything to do with running the various networks necessary.

It's all about how many transistors you can buy per dollar and how many you can put to doing work.

The per dollar metric is going in the wrong direction as node processes continue to shrink, and the concurrent use is going down as heat becomes an extremely limiting factor. More of the circuitry needs to be in low power or power off states than ever before.

The eventual conversion is from kWh->trained models.

If we're going to talk training, then I have absolutely nothing good to say about the amount of energy used to train pointless models to promote mostly sham companies.

 

[dfwatt]

Statements like these make me root for Tesla to finish quickly even though I'm spending this thread calling out their obvious flaws. It's a talent limitation not a hardware limitation.

So are you rooting against humility in the face of nature? The achievement of human-like capacities in some of AI's more impressive achievements does not mean we are not a very long way from truly duplicating what a conscious agent does. Ultimately whether it's a talent limitation or a hardware limitation, neither of those contradict what I said earlier about the deceptive complexity of duplicating what a biological brain does while driving a car.

 

[drtimhill]

You have to wonder if people notice that there are basically two distinct and incompatible positions in this thread:

1. Tesla need to sort themselves out and finish FSD and fix the bugs. It's ridiculous how long its taking them!
2. Tesla (and Elon) are way over their heads and will never deliver FSD because they underestimated the fundamental difficulties in ADAS.

I don't think anyone here (myself included) is in a position to accurately judge the feasibility of either position. Those in group 1 are probably seriously underestimating the complexities involved in what they think are "simple things to fix" (Dunning Kruger, anyone?), while those in group 2 might well be overestimating (perhaps because they would enjoy seeing Tesla flunk?).

As I have said before, no-one has ever attempted anything like this (forget Waymo with geofenced cars stuffed with every sensor known to man). Tesla are quite literally trail-blazing, and while its fun to be a critic, it's vastly harder being the artist, and actually having to invent new techniques and approaches as you machete into unknown territory.

Right now, I can get in my car and have it drive me to the store, or a restaurant, and 9 times out of 10 it will get me there comfortably all by itself. Sure, it should be 10 times out of 10, and we all are excited for that day, but only 10-15 years ago, even doing it ONCE would have been jaw-dropping. Think about that. You are almost literally driving science-fiction, guys.

Does this mean we shouldn't point out the obvious flaws, or the delays from (over-optimistic) dates? Of course not! I myself get mad when the car bizarrely moves out of the lane it needs to be to make an upcoming turn, or leaves it late to try to merge, or puts on a turn signal for no discernible reason. But I'm also aware of the massive complexities and problem space Tesla are attacking, and with each release they get closer and closer. Do you think they dont know the car does these things? Or care? Of course they do, but it's the usual old story: they have finite resources and the list of stuff they work on is prioritized, and, like it or not, some of the "embarrassment" issues are pushed down the stack.

And one other thing to consider. They have persisted. And are getting closer. Crazy though it is, that does take a bombastic, over-optimistic, driven, slightly crazy person such as Elon (though I personally detest him in many ways). Ford, or GM, or Toyota would have quietly shuttered a program such as FSD years ago. Finally, there are all those FSD up-front $$ buy-ins, that make so many people here angry. Without debating the dubious ethics of this, the very fact that Tesla have taken so much in up-front buy-in has kind of forced their hand here. They are committed, both by Elons (over) optimism and the financial fallout and backlash that would occur if they tried to pull out. Sure, they have weaseled a bit from L4/5 to L2/3, but technically these levels are far closer together than many (maybe even including Tesla) realized.

So yes, debating this stuff is fun, and even useful. But let's keep it in context .. we are watching, and taking part in, an incredible experiment. So why not (literally) enjoy the ride!

 

[dfwatt]

Statements like these make me root for Tesla to finish quickly even though I'm spending this thread calling out their obvious flaws. It's a talent limitation not a hardware limitation.

I think I may not have been totally clear on what I was trying to say. AI will definitely solve autonomous driving. But it's taken orders of magnitude time more than Elon or other people for that matter predicted because people underestimated the difficulty of the problem. Do you have an issue with that statement? if so what is the factual basis for that difference?

 

[eli_]

Agree with it being a talent limitation, that's why they keep doing Autonomy Days and Elon was hosting hackathons at his house, and they've got a bipedal robot they trot out whenever there is a recruit event. I think they're having trouble convincing more people to come work there for average pay, long hours, no remote work days, and you can make much more money at Google or Apple. "Yeah you're going to be fixing driving bugs for the next 5-10 years but check it out, we got a sweet robot!"

 

[DrDabbles]

Tesla (and Elon) are way over their heads and will never deliver FSD because they underestimated the fundamental difficulties in ADAS.

To be clear, Elon promised Level 5 with existing hardware. That challenge is well beyond driver assistance and was never, ever, going to happen. Ever.

perhaps because they would enjoy seeing Tesla flunk?

Nobody on this forum wants Tesla to fail at this, and certainly none of us that bought the features want any such thing. The constant need to be victims is eye roll inducing.

As I have said before, no-one has ever attempted anything like this

I'm going to go ahead and disagree. CMU had NavLab driving on the open highway in the early 1990s. And to say "forget Waymo" shows a complete lack of understanding of the players in this space or their accomplishments. If someone is seriously comparing Tesla to Waymo, they might as well be in a different universe. They are nothing alike- Waymo is actually doing driverless rides in multiple markets. They are running a business based on their service, and Tesla is struggling to not hit curbs or parked cars.

Either way, if you consider FSD comfortable, you either drive a route that Tesla has lidar mapped, or your idea of a comfortable ride is more like an amusement park ride. The sawing at the wheel during turns, the stopping in the middle of a turn, delayed lane changes, braking behavior, creeping into an active intersection with vehicles approaching quickly, etc. The lack of pre-planning and the constant reactive behavior makes the ride unbearable for every passenger that was in my Tesla.

I think there are two separate groups represented in this thread-

1. Realists that see 63 pages of comments since 2018 and accept the fact the software doesn't do what we were told it already did

2. People that need to justify the fact they paid money for a set of features that don't work as described.

 

[verygreen]

So, your argument that has there has been no progress is to unilaterally declare that the progress that has been made is not "progress" at all, but some kind of dead end that is in no way related to the technology needed to self-driving. Do I hear the distant crash of goal posts being shifted???

it's progress, it's just towards some other goals, not towards some actual autonomy. the distant crash is the target hitting against sidelines as it departs the path it was promised to be on more and more towards the actual goals it was always moving to.

In other news, ALL of medical science is a complete failure because humans still die.

bad argument. Not all humans have died yet (demonstrated by at least me), compare that to all prior humans on the old medical science that are all dead now.

 

[verygreen]

As I have said before, no-one has ever attempted anything like this (forget Waymo with geofenced cars stuffed with every sensor known to man). Tesla

This argument is basically "if we ignore all the serious players in the field then Tesla is the only contender and is vastly ahead".

 

[AWDtsla]

Dude. It's my job to know these things, you're supremely far off base. You realized that the TRIP chips are custom ASICs with a bunch of prepackaged ARM IP, right? Nothing to do with training, everything to do with running the various networks necessary.



The per dollar metric is going in the wrong direction as node processes continue to shrink, and the concurrent use is going down as heat becomes an extremely limiting factor. More of the circuitry needs to be in low power or power off states than ever before.



If we're going to talk training, then I have absolutely nothing good to say about the amount of energy used to train pointless models to promote mostly sham companies.

I'm talking about training because that's been the core problem. Nothing so far has shown HW3 inference does not have enough power, plausible but a different topic.

Energy to train a model of a certain size is the right metric. THAT has been scaling very very well. But you were probably thinking about inference. Again, not shown to be the issue, and clearly has not been for as long as HW3 has been around. When FSD is legally deployed we'll know for sure.

 

[DrDabbles]

I'm talking about training because that's been the core problem.

Nope. You think FSD's problem is just a matter of training??

Nothing so far has shown HW3 inference does not have enough power, plausible but a different topic.

The fact they need both TRIPs to handle the current workload and it's still not fast enough, still not running enough networks, still can't drive unaided suggests otherwise.

THAT has been scaling very very well.

 

[verygreen]

The fact they need both TRIPs

all four actually (there are two TRIPs per SoC on hw3). They need both SoCs

 

[AWDtsla]

The fact they need both TRIPs to handle the current workload and it's still not fast enough, still not running enough networks, still can't drive unaided suggests otherwise.

You're making some unfounded assumption if HW3 had 100x as much compute as it does today that the car would suddenly drive itself. Extremely flawed view. BTW almost all the failures I experience are in static and low speed environments, not that this will change your view.

 

[drtimhill]

So are you rooting against humility in the face of nature? The achievement of human-like capacities in some of AI's more impressive achievements does not mean we are not a very long way from truly duplicating what a conscious agent does. Ultimately whether it's a talent limitation or a hardware limitation, neither of those contradict what I said earlier about the deceptive complexity of duplicating what a biological brain does while driving a car.

While I agree with your overall assessment of AI (which, anyway, should be called Artificial Mimicry, not AI), I'm not so sure about driving. Pretty much everyone who drives does so with the mental equivalent of autopilot anyway .. how often have you got into the car and driven to work while (a) musing on what tasks you have to do and (b) listening to the news on the radio? And we can do that because driving is essentially a mechanical process, and THAT level of AI is probably quite practical.

Sure, when something odd happens we "come into focus" on the driving task to handle this, but in the L2/L3 case of ADAS this is handled by the human also.

 

[drtimhill]

I'm going to go ahead and disagree. CMU had NavLab driving on the open highway in the early 1990s. And to say "forget Waymo" shows a complete lack of understanding of the players in this space or their accomplishments. If someone is seriously comparing Tesla to Waymo, they might as well be in a different universe. They are nothing alike- Waymo is actually doing driverless rides in multiple markets. They are running a business based on their service, and Tesla is struggling to not hit curbs or parked cars.

Either way, if you consider FSD comfortable, you either drive a route that Tesla has lidar mapped, or your idea of a comfortable ride is more like an amusement park ride. The sawing at the wheel during turns, the stopping in the middle of a turn, delayed lane changes, braking behavior, creeping into an active intersection with vehicles approaching quickly, etc. The lack of pre-planning and the constant reactive behavior makes the ride unbearable for every passenger that was in my Tesla.

If you had actually read and digested my post, you would see why neither of these remarks has any relevance whatsoever, regardless of their (in)accuracy.

Enjoy your Rivian.

 

[dfwatt]

While I agree with your overall assessment of AI (which, anyway, should be called Artificial Mimicry, not AI), I'm not so sure about driving. Pretty much everyone who drives does so with the mental equivalent of autopilot anyway .. how often have you got into the car and driven to work while (a) musing on what tasks you have to do and (b) listening to the news on the radio? And we can do that because driving is essentially a mechanical process, and THAT level of AI is probably quite practical.

Sure, when something odd happens we "come into focus" on the driving task to handle this, but in the L2/L3 case of ADAS this is handled by the human also.

I think what you mean to say is that driving is mostly the stringing together of a whole lot of habitual and thus overlearned operations that have low to moderate attentional load at least some/most of the time. Attention is the bandwidth Limited process that hogs conscious resources and it's virtually impossible to have multiple attentionally demanding processes running concurrently- or at least it's impossible to do them well because of their competition with each other.. There are probably multiple streams of processing going on while we're driving and we can tolerate a certain amount of competition as long as those other streams are not too attentionally demanding. Hence the old aphorism that we can walk and chew gum at the same time. But we can't walk and chew gum, do mental arithmetic, and compose a symphony all at the same time. How many times have we experienced it either as a driver or a passenger that we're having a conversation and then the driver comes to a difficult intersection and tells everybody to shush. Same basic principle

 

[DrDabbles]

You're making some unfounded assumption if HW3 had 100x as much compute as it does today that the car would suddenly drive itself.

I made absolutely NO such assumption. You could give HW3 all the compute in the world and it wouldn't drive itself.

 

[scottysize]

You have to wonder if people notice that there are basically two distinct and incompatible positions in this thread:

1. Tesla need to sort themselves out and finish FSD and fix the bugs. It's ridiculous how long its taking them!
2. Tesla (and Elon) are way over their heads and will never deliver FSD because they underestimated the fundamental difficulties in ADAS.

I don't think anyone here (myself included) is in a position to accurately judge the feasibility of either position. Those in group 1 are probably seriously underestimating the complexities involved in what they think are "simple things to fix" (Dunning Kruger, anyone?), while those in group 2 might well be overestimating (perhaps because they would enjoy seeing Tesla flunk?).

As I have said before, no-one has ever attempted anything like this (forget Waymo with geofenced cars stuffed with every sensor known to man). Tesla are quite literally trail-blazing, and while its fun to be a critic, it's vastly harder being the artist, and actually having to invent new techniques and approaches as you machete into unknown territory.

Right now, I can get in my car and have it drive me to the store, or a restaurant, and 9 times out of 10 it will get me there comfortably all by itself. Sure, it should be 10 times out of 10, and we all are excited for that day, but only 10-15 years ago, even doing it ONCE would have been jaw-dropping. Think about that. You are almost literally driving science-fiction, guys.

Does this mean we shouldn't point out the obvious flaws, or the delays from (over-optimistic) dates? Of course not! I myself get mad when the car bizarrely moves out of the lane it needs to be to make an upcoming turn, or leaves it late to try to merge, or puts on a turn signal for no discernible reason. But I'm also aware of the massive complexities and problem space Tesla are attacking, and with each release they get closer and closer. Do you think they dont know the car does these things? Or care? Of course they do, but it's the usual old story: they have finite resources and the list of stuff they work on is prioritized, and, like it or not, some of the "embarrassment" issues are pushed down the stack.

And one other thing to consider. They have persisted. And are getting closer. Crazy though it is, that does take a bombastic, over-optimistic, driven, slightly crazy person such as Elon (though I personally detest him in many ways). Ford, or GM, or Toyota would have quietly shuttered a program such as FSD years ago. Finally, there are all those FSD up-front $$ buy-ins, that make so many people here angry. Without debating the dubious ethics of this, the very fact that Tesla have taken so much in up-front buy-in has kind of forced their hand here. They are committed, both by Elons (over) optimism and the financial fallout and backlash that would occur if they tried to pull out. Sure, they have weaseled a bit from L4/5 to L2/3, but technically these levels are far closer together than many (maybe even including Tesla) realized.

So yes, debating this stuff is fun, and even useful. But let's keep it in context .. we are watching, and taking part in, an incredible experiment. So why not (literally) enjoy the ride!

You forgot group 3. Tesla is making headway and I for one am happy with the progress.