Is Tesla closer than we think?

[diplomat33]

Tesla has had Level 4 since at least 2016

Tesla also has a consumer autonomy product in 50 states, which is 5000% better than Waymo.

Tesla has not deployed any autonomy yet.

 

[shrineofchance]

Tesla has not deployed any autonomy yet.

Looks pretty deployed to me! No hands on the steering wheel.

 

[diplomat33]

Many, many academic and industry sources which I have cited many timed here on TMC and elsewhere.

No! You cite academic papers about ML that everybody is doing. Companies like Waymo, Mobileye and Cruise also have 4D vision, Deep Neural Networks, Pseudo-Lidar, Convolutional Neural Networks, Transformers, etc... None of that is unique to Tesla. It is not evidence that Tesla is ahead.

 

[diplomat33]

Looks pretty deployed to me! No hands on the steering wheel.

That was a 2016 staged demo. It was not deployed to the fleet.

 

[helvio]

Sorry but this is a tired argument that Tesla fans have been pushing for years that is total BS. If Tesla has such a colossal advantage, where is it? How come Tesla is not far ahead of Waymo and Cruise in autonomy? Yes, Tesla's data has allowed Tesla to improve features and deploy features to a large fleet. But those features are L2. Tesla's "data advantage" has not allowed Tesla to deploy any autonomy yet.

If Tesla were really labeling data from 1M drivers every day, for a couple years now, there should not be a single feature that Tesla has not labeled. Tesla should have solved vision already. Clearly that is not the case.

Put simply, Tesla supposedly has had this colossal "data advantage" for years now and they still only have L2 while other companies have L4? That does not make any sense.

Data doesn't equal knowledge, but it's a requirement. It takes time. It's not like engineers know everything, they also learn day by day, you know? like you and me. Something new happens every day and they have to find ways to capture that from the fleet, label it, deploy it, check it, tweak it... many cycles until it's acceptable and our cars can finally brake for passerby hummingbirds. it takes time and achieves little. many littles equals a big. many bigs might eventually equal to FSD

 

[diplomat33]

Tesla has had Level 4 since at least 2016

Tesla also has a consumer autonomy product in 50 states, which is 5000% better than Waymo.

That was a staged demo.

If Tesla has had L4 since 2016, how come our cars are still L2? What is Tesla waiting for?

Look, I know you want to defend Tesla but making crazy claims that Tesla has already deployed L4 is so obviously false.

 

[diplomat33]

Data doesn't equal knowledge, but it's a requirement. It takes time. It's not like engineers know everything, they also learn day by day, you know? like you and me. Something new happens every day and they have to find ways to capture that from the fleet, label it, deploy it, check it, tweak it... many cycles until it's acceptable and our cars can finally brake for passerby hummingbirds. it takes time and achieves little. many littles equals a big. many bigs might eventually equal to FSD

Agreed. I am not against data. I understand the importance of collecting data. You can't do ML without collecting good data. And yes, it takes time to collect and label data. And we've seen Tesla's software get incrementally better over time because of the data that Tesla collects. So collecting data is definitely helping Tesla. I am pushing back against the hype that Tesla's data advantage is so colossal that Tesla will solve FSD before everybody else. That, I am not seeing much evidence for yet.

 

[shrineofchance]

No! You cite academic papers about ML that everybody is doing. Companies like Waymo, Mobileye and Cruise also have 4D vision, Deep Neural Networks, Pseudo-Lidar, Convolutional Neural Networks, Transformers, etc... None of that is unique to Tesla. It is not evidence that Tesla is ahead.

It is a truism in deep learning that data, compute, and neural network design are the three factors that determine performance. To argue, in the absence of direct quantitative evidence, that one companies' NNs have dramatically better performance than another companies', one must construct a plausible explanation of how that could be the case based on these three factors.

In the autonomous vehicle application in particular, we know that, for major, multi-billion-dollar companies, training compute is by far the least significant constraint of the three. Abundant cheap compute exists in the cloud and these companies can set up their own GPUs or ASICs for training.

Neural network design is the most unpredictable and mysterious of the three factors. However, there is good reason to believe it is a much less significant source of competitive advantage among major companies than data.

As Karpathy stated in the tweet I posted a page back, most cutting-edge research in AI is conducted by either a) academic labs like Mila or b) industry labs like DeepMind. Counterintuitively, the industry labs publish a huge amount of research that is replicable by other labs based on reading their papers and often even open source their research.

This is not because these corporations are simply generous, but largely because there is a very powerful ethos among AI researchers of publishing replicable papers. If Alphabet suddenly forbade DeepMind from publishing their research, there would no doubt be an exodus of researchers from DeepMind to FAIR or somewhere else that still allowed publishing.

In other words, AI researchers as a subculture are ideologically committed to open science, and this puts pressure on companies that want to do AI research to allow their researchers to do open science.

This is why the primary competitive advantage in the current landscape is data. Compute is abundant and cheap, AI research is largely open, but data is relatively scarce, expensive, and can be jealously hoarded.

 

[shrineofchance]

That was a staged demo.

If Tesla has had L4 since 2016, how come our cars are still L2? What is Tesla waiting for?

Look, I know you want to defend Tesla but making crazy claims that Tesla has already deployed L4 is so obviously false.

If Waymo really has L4, why isn't it being rolled out to every major city in the contiguous U.S.? Why do they have fewer L4 cars than engineers?

 

[El joe]

We keep thinking about Tesla as trying to solve all of these situations and it's obvious that they are trying, but just like an 80 year old grandma, they could avoid certain dangerous situations and still get to the destination fine

I really hope the 80 year-old grandma is not their FSD benchmark.

 

[diplomat33]

If Waymo really has L4, why isn't it being rolled out to every major city in the contiguous U.S.? Why do they have fewer L4 cars than engineers?

Yes Waymo really has L4. But deploying a large fleet of robotaxis to every major city in the US would cost a lot of money. Also, you can have L4 but you still need to make sure the L4 is safe enough to deploy. Waymo needs to make sure their L4 is safe enough before they deploy in every US city. Surely, you understand that having FSD and having FSD that is safe enough are two different things.

 

[shrineofchance]

Yes Waymo really has L4. But deploying a large fleet of robotaxis to every major city in the US would cost a lot of money. Also, you can have L4 but you still need to make sure the L4 is safe enough to deploy. Waymo needs to make sure their L4 is safe enough before they deploy in every US city. Surely, you understand that having FSD and having FSD that is safe enough are two different things.

So, in other words, Waymo's L4 is not technologically mature enough (or at least not demonstrably so) to be deployed to more than a tiny, token number of people — fewer than the engineers working on the L4.

Why is that situation so different or so much better than Tesla's similarly small-scale testing of L4?

 

[diplomat33]

So, in other words, Waymo's L4 is not technologically mature enough (or at least not demonstrably so) to be deployed to more than a tiny, token number of people — fewer than the engineers working on the L4.

Why is that situation so different or so much better than Tesla's similarly small-scale testing of L4?

It's different because Tesla does not have L4. Tesla is not doing any L4 testing.

 

[shrineofchance]

It's different because Tesla does not have L4.

What is this?

 

[Tam]

First thing first, Tesla Smart Summon still has the problem of hitting the curbs, running over the lane markers, involved in minor accidents, all in very slow speed. That kind of slow speed would be ideal for the system to avoid accidents.

Same with summoning the car out from a garage, it can still cause accidents in a very short distance and slow speed.

Let's increase the speed to highway speed, Tesla's system still collides with stationary obstacles: parking firetrucks, police cars, left-turning trucks in front...

For the newest available FSD beta 8.2: It still has problems with basic rules such as driving on/aiming at the wrong side of the road, aiming to collide with obstacles:

The below scenerio clearly shows that the sensors do notice the red road shoulders guiding the road toward the right as a sharp right 45 degree turn, but the system insists to drive straight into the iron fence in front (as predicted by the straight grey line emanating from the car's icon in the instrument cluster.

In the scenario below, the system displays correctly all the red road shoulders, parked cars but it fails to display the fast-moving Toyota car coming from the right and quickly crossing the Tesla from its front. The driver quickly intervened to prevent a T-bone into that Toyota.

To compare that with Waymo's history that it never had any history of hitting stationary obstacles since 2009 and its 2019 driverless program with no human drivers, Waymo has never hit anything or anybody in its 50 square miles in Chandler, AZ.

The most basic skill that I want Tesla to attain is the ability to reliably avoid collisions, not just at high speed, but at very low speed such as getting out of a garage.

I don't argue about Tesla has all the modern technology and data advantage. However, despite all the advantages that Tesla has, its system is still in its very early phase.

 

[diplomat33]

What is this?

Tesla told the CA DMV that that video was L3.

And I am sure you are aware that Tesla also told the CA DMV that the FSD Beta testing is L2.

So no L4.

 

[Tam]

So, in other words, Waymo's L4 is not technologically mature enough (or at least not demonstrably so) to be deployed to more than a tiny, token number of people — fewer than the engineers working on the L4.

Why is that situation so different or so much better than Tesla's similarly small-scale testing of L4?

Waymo has attained the basic skills of not hitting a stationary obstacle but the system's intelligence is still very infantile comparing to human as shown by its riders: It could get stuck in a parking lot or in a coned-off lane (although it does not hit anything or anybody, it doesn't know what to do next).

It's safe in the sense of not hitting anything but to make it intelligent enough outside of 50 square miles of Chandler, AZ, it's still a challenge.

 

[shrineofchance]

To compare that with Waymo's history that it never had any history of hitting stationary obstacles since 2009

We have no way of knowing whether this is true because Waymo hasn't disclosed such information.

 

[powertoold]

It's rather simple at this point. If V9 is a lot better than 8.2, then it's possible Tesla will achieve human level safety this year. If it's not, it's unlikely for another year or two.

I don't think Tesla would wide release fsd beta until its failures are predictable, like how AP and NoA work currently. It's very predictable when they're likely to fail.

I anticipate that V9 will have similar to Waymo performance and reliability for comparable routes and road conditions. This will be a huge accomplishment.

 

[Tam]

We have no way of knowing whether this is true because Waymo hasn't disclosed such information.

It has disclosed all Waymo's collisions in its annual California DMV disengagement. It shows why the car was disengaged including collisions. However, none of those collisions are caused by its system vs stationary obstacles but because of others hitting it.

In other states, each time there's a Waymo accident, the press would report it. None that Waymo hit others but others do hit Waymo.

Safety goes beyond not hitting things because if the system is so cautious, it can cause road rage and other drivers might hit Waymo as a result too.

So the intelligence of not being too cautious and avoid road rage from others is still a challenge and might be addressed decades from now.