1:36:14 tesla has a mobile service team that already services their entire fleet. Also charging, cleaning, etc. Tesla by far has the infrastructure advantage
The thing with LiDAR vs only camera is not just cost (which is absolutely massive when scaled to the millions annually) but it’s also processing speeds. If you solely focus on everything hinging on vision than everything you do goi g forward will revolve around perfecting vision. If you rely on lidar for safety than you’re handicapping yourself in terms of vision while simultaneously bogging down the entire processing network as the two systems compare every frame with one another to concur that they’re right. LiDAR is not bad by any means, but in terms of future proofing autonomy it is like a training wheel that demands more energy, cost and processing power which could otherwise be used more efficiently on purely vision. We’ll see in the long run but I imagine that LiDAR will not be used in autonomous vehicles in 5+ years as all switch to the vision based approach.
@@apple1231230 I think 5 years is a bit optimistic. LiDAR is beginning to be introduced on consumer cars and may well be a big part of making Level 3 cars work. I think LIDAR won’t be needed in 20 years but the only autonomous cars around at the moment rely heavily on it.
@@williamlowry2487 look up "FSD V 12.5.3 drives" you'll see entire drives across the USA regularly being done by teslas with zero input from the driver, across the USA. my favorite are the Manhattan videos, its tense. these are vision only and represent the overwhelming majority of autonomous miles being driven in the USA and growing. Waymo is good, but i would say all the extra tech is bogging the "magic" down, that being an AI capable of driving vs code.
23:08 This guest seems to have some fundamental misconceptions around how Tesla accrues training data. He seems to think that FSD needs to be engaged in order for it to learn. The opposite is true. Their system is trained while humans are driving. When a human takes over from FSD, this is treated as a strong signal to pay attention to how a human driver behaves. You tried to correct his misconception but he incorrectly doubled down on it.
This is a topic I'm really interested in. I have a fair bit of experience with Waymo, Tesla, and even Cruise. My 2cents: I'm glad Cruise is off the roads. It was a really uncomfortable experience. It couldn't even stop at a stop sign without slamming on the brakes. Waymo is much better. Very smooth like a professional driver, but very limited service area. You both mentioned that you haven't tried the latest version of FSD, and I would suggest experiencing it. The latest version is very good for a generalized solution. Of course it's not as good as Waymo, but it can also go many places that Waymo can't.
Interesting how in the equation 1. Car - 2. Driving - 3. Traffic (incl. regulations)... AV developers focus on 2. replacing the human driver by some sort of technology, and on Traffic (3.). The Car (1.) is completely left out of the equation, like there is no fundamental difference between maneuvering a Cybertruck and a VW Beetle through dense city traffic. Would be interested to hear your comments on this... Cheers.
Level 3 is weirdly described imho. The main difference between L3 and L4 is that the ODD is much smaller and the car can leave the ODD rolling using a controlled handover, where the computer is still driving and is liable until complete. Also, Waymo has loads of safety rails around the machine learning and redundancy (such as several sensing modalities and more and better cameras) that Tesla's L2 solution doesn't need (as they can fail, beep and blame the human at all times). I doubt we'll see a pure end-to-end system driving this decade. There are no safety critical pure ML apps at all afaik, not even radiology.
@@StefanNorberg It’s quite possible that Level 3 may never take off! We could easily end up with cars that have very good Level 2 systems but the driver is always responsible and Level 4 where they aren’t! Level 3 manufacturers might struggle to justify the higher prices that they would need to sell their cars for.
Good conversation, but I think you didn’t talk enough about scale. Waymo need to premap a city. And that can’t be automated, cause the system can’t drive there. And waymo miles and tesla fsd miles aren’t even comparable. In a chart you almost can’t see the waymo miles
I would suggest trying out the latest in self driving before doing an AI self driving podcast where you openly admit you haven’t tried FSD in a year (it was literally heuristics last year vs AI this year)
@TheDoggydogworld That’s an interesting concept considering he said he tested in in March and FSD 12.5 didn’t release until July 29th. No he was on 12.3 and we can’t even verify if it was 12.3.6 which was significantly better than 12.3 and significantly worse than 12.5.3 currently is.
@@apple1231230 - you're right. The host said 12.5 just as TBL answers the "what version" question. Listening carefully I think I can hear TBL say 12.3.
1:36:14 tesla has a mobile service team that already services their entire fleet. Also charging, cleaning, etc.
Tesla by far has the infrastructure advantage
The thing with LiDAR vs only camera is not just cost (which is absolutely massive when scaled to the millions annually) but it’s also processing speeds. If you solely focus on everything hinging on vision than everything you do goi g forward will revolve around perfecting vision.
If you rely on lidar for safety than you’re handicapping yourself in terms of vision while simultaneously bogging down the entire processing network as the two systems compare every frame with one another to concur that they’re right.
LiDAR is not bad by any means, but in terms of future proofing autonomy it is like a training wheel that demands more energy, cost and processing power which could otherwise be used more efficiently on purely vision.
We’ll see in the long run but I imagine that LiDAR will not be used in autonomous vehicles in 5+ years as all switch to the vision based approach.
@@apple1231230 I think 5 years is a bit optimistic. LiDAR is beginning to be introduced on consumer cars and may well be a big part of making Level 3 cars work. I think LIDAR won’t be needed in 20 years but the only autonomous cars around at the moment rely heavily on it.
@@williamlowry2487 look up "FSD V 12.5.3 drives"
you'll see entire drives across the USA regularly being done by teslas with zero input from the driver, across the USA. my favorite are the Manhattan videos, its tense.
these are vision only and represent the overwhelming majority of autonomous miles being driven in the USA and growing.
Waymo is good, but i would say all the extra tech is bogging the "magic" down, that being an AI capable of driving vs code.
23:08 This guest seems to have some fundamental misconceptions around how Tesla accrues training data. He seems to think that FSD needs to be engaged in order for it to learn. The opposite is true. Their system is trained while humans are driving. When a human takes over from FSD, this is treated as a strong signal to pay attention to how a human driver behaves. You tried to correct his misconception but he incorrectly doubled down on it.
Agree
This is a topic I'm really interested in. I have a fair bit of experience with Waymo, Tesla, and even Cruise. My 2cents: I'm glad Cruise is off the roads. It was a really uncomfortable experience. It couldn't even stop at a stop sign without slamming on the brakes. Waymo is much better. Very smooth like a professional driver, but very limited service area. You both mentioned that you haven't tried the latest version of FSD, and I would suggest experiencing it. The latest version is very good for a generalized solution. Of course it's not as good as Waymo, but it can also go many places that Waymo can't.
Thanks for sharing.
Interesting how in the equation 1. Car - 2. Driving - 3. Traffic (incl. regulations)... AV developers focus on 2. replacing the human driver by some sort of technology, and on Traffic (3.). The Car (1.) is completely left out of the equation, like there is no fundamental difference between maneuvering a Cybertruck and a VW Beetle through dense city traffic. Would be interested to hear your comments on this... Cheers.
Level 3 is weirdly described imho. The main difference between L3 and L4 is that the ODD is much smaller and the car can leave the ODD rolling using a controlled handover, where the computer is still driving and is liable until complete. Also, Waymo has loads of safety rails around the machine learning and redundancy (such as several sensing modalities and more and better cameras) that Tesla's L2 solution doesn't need (as they can fail, beep and blame the human at all times). I doubt we'll see a pure end-to-end system driving this decade. There are no safety critical pure ML apps at all afaik, not even radiology.
@@StefanNorberg It’s quite possible that Level 3 may never take off! We could easily end up with cars that have very good Level 2 systems but the driver is always responsible and Level 4 where they aren’t! Level 3 manufacturers might struggle to justify the higher prices that they would need to sell their cars for.
Good conversation, but I think you didn’t talk enough about scale. Waymo need to premap a city. And that can’t be automated, cause the system can’t drive there. And waymo miles and tesla fsd miles aren’t even comparable. In a chart you almost can’t see the waymo miles
Right off the bat with the SAE levels already with the bad takes lol
I would suggest trying out the latest in self driving before doing an AI self driving podcast where you openly admit you haven’t tried FSD in a year (it was literally heuristics last year vs AI this year)
But FSD isn’t a self driving system!
@@williamlowry2487 okay
His 45 minute SF drive with intervention was FSD 12.5.
@TheDoggydogworld
That’s an interesting concept considering he said he tested in in March and FSD 12.5 didn’t release until July 29th.
No he was on 12.3 and we can’t even verify if it was 12.3.6 which was significantly better than 12.3 and significantly worse than 12.5.3 currently is.
@@apple1231230 - you're right. The host said 12.5 just as TBL answers the "what version" question. Listening carefully I think I can hear TBL say 12.3.
Timothy B. Lee has a nails on the chalkboard vocal fry voice. Tough to listen to.
This is a lotta talk about a technology that is worse than trains.
Trains only work in countries with top 10 state capacity.
It’s not autonomous vehicles or trains. Also Google is a private company free to pursue what they want