But I do … 
I agree that we both understand the technical issues, but I think that since the different strategies imply quite different issues, e.g., when expanding to a new area, then the use of the same term for both obscures this potentially significant difference.
Why? It’s the right term. It’s fine to use the word “surgery” to describe both an open-heart procedure and an arthroscopic knee procedure, even though they’re both radically different types of surgeries. It doesn’t obscure the differences between them - it’s just the correct term for the type of thing that they are.
Both Tesla and Waymo engage in geofencing. Or will, one presumes - Tesla hasn’t actually started yet. They will do it in different ways, based on some different factors (and some in common), and for some different reasons - but they’ll both end up with geofenced operations.
1 Like
One modifies the term surgery with modifiers like open heart in order to make it clear what is involved. Were there such modifiers here, one could use the common term. But, without such modifiers, we have Tesla which seems only has to change where the line is drawn vs Waymo which has to go through a non-trivial prep process to change where the line goes.
2 Likes
Again, I think that’s mistaken - it’s perfectly appropriate to say you’re having surgery next week without having to provide a modifier, even though there’s lots of different types of surgeries. This is just a pedantic quibble, though - you’re assuming that when someone uses a general term that covers a wide variety of different things that either they are presuming (or others will assume) that the things are identical. I think that’s incorrect, any more true than merely telling someone you’re having surgery next week means you are blurring the distinction between different types of surgeries.
But more importantly, leaving aside the semantic quibble, your assessment may be wrong.
Based on their public statements, Tesla has drawn their line around the areas where their cars are currently capable of self-driving. There is, one presumes, a reason why they are being limited to that area and are not being operated elsewhere in the city. A possible (and I think likely) explanation for that reason is that the cars’ “brains” are not capable of handling every type of street or intersection in the city. Same as Waymo.
Waymo expands their service area with mapping, which is non-trivial. But Tesla may have to seriously upgrade their cars’ AV driver to move from one service area to a broader one, which is also non-trivial.
The Waymo vs. Tesla approach reflects skepticism about how well and fast the cars’ brains will develop. Waymo says that cars’ brains are going to develop slowly, so in order for them to self-drive we need to give them more sensory inputs and lots of mapping data to compensate. Tesla thinks cars’ brains are going to develop really fast, so we don’t need to mess around with the extra inputs or maps. So Waymo needs to add maps to add service area…but Tesla needs to add car brainpower to add service area. I’m not sure that Tesla’s approach doesn’t involve a much harder lift at the margins than Waymo’s, and that Waymo is the one that will be able to scale much faster.
1 Like
Or, the obvious reason would be that they picked a smallish area in which handling exceptions would be simple and where it is likely that a significant number of endpoints would be within the area. I.e., if they manage a shortish period well, they can quickly and easily expand.
1 Like
But that’s not the reason they gave. They specifically said that it was picked to limit the environment to scenarios that the driver could handle:
“When we deploy the cars in Austin, we are actually going to deploy it not to the entire Austin region but only to the parts of Austin we consider to be the safest,” Musk said on CNBC. “So we will geo-fence it.”
He added: “It’s not going to take intersections unless we are highly confident it’s going to do well with that intersection. Or it will just take a route around that intersection.”
Now, again, that might not entirely be the truth - Tesla may just be playing their cards close to their chest, and could deploy to any physical environment in the city if they wanted to. But their public statements indicate that the geofence is there to limit the driving area to where their cars can handle it, not just to optimize trip ends.
I just saw albaby’s post. Kind of nullifies what I just wrote, but I’ll leave it here.
I read somewhere (can’t remember where, sorry) that Tesla is specifically is avoiding problematic areas, like tricky intersections where FSD has trouble. They haven’t yet released their service area maps, which indicates to me they are still working on them.
I think you are correct, in that Tesla is hoping for a generalized solution where this type of mapping isn’t required and then they can expand where ever they want.
I believe Waymo is also hoping for a generalized solution, but their tech isn’t ready for that.
If I had to guess, I’d say Tesla doesn’t think their tech is ready either, simply because they don’t appear to have initiated the AV ride hailing permitting process outside of Texas (where special permits are not rquired).
So yeah, I also meant to say that the area would be less complicated.
Right - which is the point I’m making. The main reason you geofence is that you want to limit the cars to the areas they’re capable of navigating. Waymo increases the area their cars are capable of navigating by mapping, Tesla increases it by upping the AI driver’s capabilities to be able to handle more complicated driving situations. We have no real idea yet which is harder to do.
Well, yes, that is one criteria. But, there are other factors like ease of service and likelihood of a substantial traffic from endpoints within the area. One of the issues which I suspect will contrast the two is that Waymo’s map will have a bunch of detail exceptions, but Tesla’s will be mostly just a line around a likely area.
Again, we’ll have to see. I think it’s equally likely that Tesla’s map will have some holes in it as well. Not at first, of course - I’m sure they’ll start with a reasonably compact area within the city that doesn’t have anything that FSD can’t handle - but once they get beyond the pilot, then certainly.
I don’t think FSD is level 5 autonomy yet. That means there will be some non-trivial number of driving scenarios that it cannot handle at an acceptably safe rate to be allowed to drive itself there. So there will be areas “within the line” that FSD will have more trouble handling than most - and it wouldn’t be shocking if its geofence simply didn’t let the car go to those areas.
If Waymo does not have a detailed map of the area, it cannot drive. This also means that if it does have a map and the road changes, it gets confused. It needs to not only be geo fence but needs the latest hd maps and a sensor suite to operate.
Tesla does not need maps. It is like a human. It sees and navigates. If it has not seen scenario before, it may get confused. This is why they spent $10B on massive data and training infra. They took out most of the code. It is neural nets.
Right. And so the real question is which approach will work better.
To anthropomorphize a bit, the Waymo driver is designed to be able to function with a dumber brain. It’s got extra sensors and a map to guide it. Because it’s got that extra information, it can navigate to an acceptable degree of intervention with a “lower” level of brain power.
Tesla’s approach is to make the brain better, so that it doesn’t need the extra sensors and map. They were expecting the brain to get good enough, fast enough, that it would be able to drive itself anywhere without the extra help. But that hasn’t happened yet. We have no idea how good FSD is at driving itself, rather than driving in conjunction with a human behind the wheel to take over immediately. But we do know it’s consistently been nowhere near good enough to do what Musk has forecast in his public statements.
So this is the question - is Tesla’s AV “brain” good enough to drive on its own consistently enough and with few enough interventions, compared to Waymo’s AV “brain” that wouldn’t be able to do that with just vision but can do that with the aid of extra sensors and maps? Tesla has an advantage in that its cars are cheaper and it doesn’t need maps - but is that advantage undone by a higher intervention rate, and thus a higher operating cost to keep the cars on their way?
We don’t know - Tesla doesn’t release data on FSD’s intervention rate. We won’t know from the Austin pilot, unless some of the information the city gets provided is required to be disclosed to the public - there’s no way someone from the outside would be able to determine the rate that teleoperators intervene. But if they do expand to California, then we’ll have more of an answer.
I think you hit it on the head. Go back and read all of his posts and it actually is hilarious. He get’s frustrated and then starts throwing out threats and insults. This has been a very enjoyable thread just to read. You can really see when someone doesn’t have a clear grasp of the subject.
1 Like
Geofencing is a technology that uses GPS, RFID, Wi-Fi, or cellular data to create virtual geographic boundaries around a physical location. When a mobile device or tag enters or exits this boundary, pre-defined actions can be triggered, such as sending alerts, notifications, or targeted advertisements.
This has the ring of “makin’ stuff up.”
What is your definition of “dumber brain”, of lower level of brain power? Some measure of size of neural network?
What is your definition of “extra information”? Giga-something per second?
More generally, what is your basis for the claim that “Waymo AI is dumber” but has “extra information?”
1 Like
I didn’t make that claim. I said that it’s designed to be able to function with a dumber brain. It gets more sensory inputs, and has a high definition map - so it has extra information that a car relying on visual information doesn’t have. So for any given level of functionality, it will be able to perform that level of functionality with a “dumber” brain than a car that lacked the extra sensory data and map information.
I’m not making any claims about what, specifically, makes a brain “dumber” or not. I’m deliberately speaking colloquially. If you provide the car with a “cheat” like a digital map, it can get away with having less analytic power - there’s stuff it can “know” without having to infer it from visual data. The extra information isn’t “giga-something per second” - it’s getting LIDAR and radar info and has access to a HD map.
Well, all of my responses below are correct (even in a non pedantic sense), unless someone can provide a real basis of support otherwise (which no one has).
We have no idea what Waymo AI was “designed for” with respect to a “dumber brain” or otherwise, unless Waymo disclosed it somewhere.
You can’t/won’t even define “dumber.” How are we even supposed to know what you mean by AI being “dumber” if we can’t define it?
Why are we talking about “dumber” (artificial) intelligence and we can’t define “dumber”?
That’s somewhat ironic.
You have no idea about this (how sophisticated/dumb a given AI needs to be to deliver a given functionality with different kinds of sensor data), even if you define “dumber.”
Same answer, we have no idea here.
Same answer again, by what measure better? We have no idea.
I get your gist, but I think we really have no idea on these statements you are making. Without definitions of key terms (such as “dumber” in a discussion of intelligence), we really have no idea.
You could very well be correct for some definition(s) of “dumber” and “extra information,” but you don’t provide any basis for your claims, let alone definitions.
Two basic counterarguments.
First:
System 1 could be very high resolution camera data, high spatial and temporal resolution. System 2 could be lower resolution camera with LIDAR data. Does S1 or S2 have more “information” with respect to AVs?
I would say you and I don’t really know this, and can’t even make a statement about this (unless maybe you want to define terms more precisely).
Second:
Processing sensory inputs (including synchronizing the data in space and time, and also calculating correlated, interacting features) from multiple sensor types (camera + lidar) might require more analytical processing (somehow appropriately defined) relative to processing from a single sensor type (camera only).
One can imagine additional neurons and connections and specialized features (in artificial and/or biological intelligence) to process vision plus sound, versus either alone.
I would say you and I just don’t know.
First step in discussion: define terms so we can discuss the “same things,” especially on a technical topic.
1 Like
I understand perfectly what he’s saying, and he’s saying it in “message board” language, not for publication in a scientific journal.
There are multiple ways to solve the problem, and I expect there will be some flavor of both AI and mapping involved in most of them. Obviously computer power is the most important, and while the costs are still declining and the sizes getting smaller, it still takes prodigious energy and computational ability to run AI, even smaller (defined) sets of it.
That said, and based on reports from users here, it appears Tesla is getting close, but then we don’t really know “how close” it’s going to have to be to pass muster with regulators and with public opinion. If they really can put it all on one board, with only vision as the input, then that’s huge.
But “fencing” is a possibility, even “infinite fencing”, like when you watch your GPS map crawl along with you as you drive. It would be near impossible to put the entire country (world!) in memory, especially given changes in construction etc, but it would be possible to have discrete blocks of maps available for download as you roll close to a geofenced border. Yes, that would also require bandwidth, but there’s lots of that around. (The GPS mapping for my boat GPS does this, creating “blocks” of water maps which are updated as you travel.)
As for those temporary blocks like construction, well Waze relies on social media, there’s no reason a central AI server couldn’t note that when any one of the cars “finds it” and update the map in real time for everyone else. Yes, the first discoverer would be slightly inconvenienced, but for 99.9% of users they wouldn’t even notice.
The analogy I would use is that Tesla is going for the home run, while the Waymo technology is more about singles and doubles. It’s possible to win the game either way, and there are likely other approaches not yet implemented (and which may be too expensive to do.) This market seems to be a natural duopoly at best, perhaps even a monopoly if one gets it right and the other can’t cross the finish line timely. After that point it will be vastly expensive for any pretender to come up with an alternate, and the network effect of cars available would create a statis in the growth of competitors.
But anyway, yeah “dumber.” I get what he’s saying. Totally. So do you, I think.
1 Like
The problem I have (more generally) with using the term dumber, is that in computer science one often implements a “dumb” algorithm and it requires more compute power.
A very simple example is searching through a sorted list to find a specific item.
The dumb way is to just read each record and compare it to what you are looking for. A list of 1000 records will take an average of 500 comparisons to find the result.
However a “smarter” method is to do a binary search which will return any result in 10 tries.
Thus the smart method can use what one may call a less powerful or dumber processor and still appear to be 50x as good.
Now scale up your list to a million records and the dumb, brute force algorithm takes an average of 500K compares and the smart algorithm only takes 20 attempts.
Of course there is some extra overhead required…the smart algorithm with the dumber processor has to maintain the list in sorted order.
Mike
2 Likes