General Discussion: 2018 Investor Roundtable

[dc_h]

Oh, I agree. (Thanks for clarifying that.) I think they'll easily be up to 7500/week before the end fo the year.

The need for three lines has simply caused me to be worried about margins. I'm hoping they can get back to the planned margins.

I think this is just final assembly, not the whole line, so minor impact on costs. As noted the line design is improved, so next shutdown will likely deploy new line process to one of the original lines and then repeat on the second original line to hopefully get to 10k per week.

 

[Waiting4M3]

I think this is just final assembly, not the whole line, so minor impact on costs. As noted the line design is improved, so next shutdown will likely deploy new line process to one of the original lines and then repeat on the second original line to hopefully get to 10k per week.

If it's just modifying existing line going forward, they may not need to do a full shutdown for future changes.

 

[mongo]

If it's just modifying existing line going forward, they may not need to do a full shutdown for future changes.

Perhaps, likely depends heavily on the part conveyance method...
If the new line is on the far side of the existing lines from the parts source, that lends itself to less disruption...

 

[tander]

...And then you want continual learning, something that Tesla obviously does not do. Ie. You want the car’s NN to learn as it drives. But we are so far away from that.

Don’t get me wrong, what Karpathy is doing is absolutely right. It is a little bit depressing that he and thus Tesla only started this 11 months ago...

I haven't watched the video yet, but isn't continual learning the basic strategy for autopilot? I thought the plan was basically you get x,000 cars on the road driving, basically collecting data in shadow mode, and once you get 5? billion miles of data it's demonstrably safer then a person. I thought the plan was to basically collect an arsenal of data that makes a) a public duty/service to prevent accidents b) indisputably proven to be safer for insurance legal purposes.

 

[EinSV]

Pulling this to the General thread

I'd love to see the market try to sort out who makes the better battery at what price. My perception is that so far Tesla is not valued well for the quality of battery that it brings to market. Much more discussion has revolved around the cost of the car relative to ICE vehicles. But more sophisticated investors will care about who makes the higher quality battery that truly supports the residual value of EVs.

But on purely on the price dimension, note that BNEF has been forecasting 2025 as the year when average battery pack cost falls to $100/kWh (when it becomes cheaper to build a BEV than a comparable ICEV). They have noted that average decline of pack cost is 20% per year, but have failed to note that at this rate the average price hits $107/kWh in 2020, five years earlier than their 2025 projection. As CATL blazes along its hype cycle, they too will be on a fast track to sub-$100 batteries, even if they have to go with lower quality to get there. So Tesla hits sub-$100 by 2019. Maybe CATL gets there by 2020 as well. This will put pressure on all other battery makers to stay in the race. Before long BNEF will have to bump up its 2025 target to 2022 or earlier. This will create the perception that EVs are really accelerating (despite the fact that they have been growing very fast all along). So what I see here is an explosion of investor interest in all names that look to win in the acceleration of the EV market. Remember that EVs have the potential to take over a multi-trillion auto industry plus a huge chunk out of the multi-trillion dollar oil industry. Tesla investors have known this all along, but most other investors believe that this is really along way off. They might even believe the 2025 BNEF target and think that it is too soon to invest deeply into EVs. But as that date is bumped up to 2022 or earlier, this will shake up expectations. Suddenly it is no longer a sensible idea to wait on the EV sidelines for some magic technology to arrive. My view is actually that sub-$100 batteries arrive in 2020. If correct, I think most investors will be blindsided and playing catch up. So I am looking forward to a pretty rapid upward reappraisal of the EV market. And Tesla stock at $330 is still a damn good vehicle for riding this wave.

These are great points. A related point that I think the market also fails to appreciate are the implications of battery and cost superiority for the next huge market after Model Y and Semi — electric pickups and truck-like SUVs.

Battery cost and performance superiority as exemplified by the Semi specs should also provide Tesla an enormous advantage in pickups and and SUVs based on the Tesla Pickup platform. We will probably start hearing more about this next year from Tesla after it gets past the Model Y launch.

This is a very lucrative market, a huge source of GHG emissions and ripe for disruption. If no one can compete with Tesla on the Semi they will have little chance in applications like pickups and truck-like SUVs that create significant battery challenges. This should be huge for Tesla.

 

[vgrinshpun]

Analysts from Berenberg hit all the nails on the heads - highlighting how competition is ***STRUCTURALLY NOT COMPETITIVE***

Vladimir Grinshpun on Twitter

 

[BioSehnsucht]

I haven't watched the video yet, but isn't continual learning the basic strategy for autopilot? I thought the plan was basically you get x,000 cars on the road driving, basically collecting data in shadow mode, and once you get 5? billion miles of data it's demonstrably safer then a person. I thought the plan was to basically collect an arsenal of data that makes a) a public duty/service to prevent accidents b) indisputably proven to be safer for insurance legal purposes.

Watch the video, it will explain in more depth. But in short, you don't learn from raw data, you must provide at least some direction on what is what (labeling for example, all the cars, all the road lines, etc) so that the neural net optimizer can process this labeled data to create the actual neural net you use in the vehicle. You give it labeled data and it learns to identify things that aren't labeled based on the labeled data. You can use the raw data to validate that the NN is identifying things as you intended, but the raw data can't be used for teaching.

If you think about it, humans don't learn from raw data either. We point out things to our children and tell them "car" or "tree", etc, until they learn to identify previously unseen ones based on their features. We don't just abandon them in the wilderness until they learn to understand the modern world with no help at all.

 

[pz1975]

Analysts from Berenberg hit all the nails on the heads - highlighting how competition is ***STRUCTURALLY NOT COMPETITIVE***

Vladimir Grinshpun on Twitter

Nice to see you back here!

 

[elliott82]

I wonder how far they can go using realtime computer graphics as a data source to assist in optimization. It would already be labeled and could provide an additonal vast amount of data, especially edge cases like he described in the video (weird traffic lights, car carrier trailers, etc) I'd imagine using something like Unreal Engine which has super realistic realtime rendering. Now that I think about it, it doesn't even need to be realtime... could be pre-rendered to make it even more realistic... It does take a significant amount of manpower to build the assets but I think it would be worth it. Any thoughts?

 

[hobbes]

Breaking news: Daimler recalls 774,000 cars in Europe

Daimler ruft 774.000 Autos in Europa zurück

Side note: As interesting as the recall itself is that the German minister of transportation (christian democrat/conservative) summoned Daimler boss Zetsche into his office today to give him that news. An expert on the evening news is saying he expects this to carry over to the US.

Meanwhile, the home of Audi boss Stadler has been searched as he is under investigation for fraud corresponding to the cheating devices.

 

[Zhelko Dimic]

I wonder how far they can go using realtime computer graphics as a data source to assist in optimization. It would already be labeled and could provide an additonal vast amount of data, especially edge cases like he described in the video (weird traffic lights, car carrier trailers, etc) I'd imagine using something like Unreal Engine which has super realistic realtime rendering. Now that I think about it, it doesn't even need to be realtime... could be pre-rendered to make it even more realistic... It does take a significant amount of manpower to build the assets but I think it would be worth it. Any thoughts?

Point of using vast amount of data is to discover new situations you wouldn't think of, which your simulation wouldn't provide. Otherwise, maybe, but I don't think Tesla lacks for data.
BTW, I think Google used approach you suggest, as they had 'real' miles and 'simulated' miles...

 

[elliott82]

Hey, thanks for the reply.

Why wouldn't the simulation provide new situations? You could simulate different lighting conditions, rain, snow, flying pigs, cars jumping over medians, suddenly breaking, etc etc...

Didn't Karpathy say they did in fact lack data in edge cases?

 

[Waiting4M3]

Point of using vast amount of data is to discover new situations you wouldn't think of, which your simulation wouldn't provide. Otherwise, maybe, but I don't think Tesla lacks for data.
BTW, I think Google used approach you suggest, as they had 'real' miles and 'simulated' miles...

I can see some value in simulating a new situation under different weather/lighting conditions. Once one of their cars encounter a corner case with unique road configuration/markings, simulation can help quickly "beef up" that case and teach the AI what that particular case would look in different time/season/weather, and wouldn't need to rely on more Tesla cars to go through that same location under actual time/season/conditions.

 

[Waiting4M3]

Hey, thanks for the reply.

Why wouldn't the simulation provide new situations? You could simulate different lighting conditions, rain, snow, flying pigs, cars jumping over medians, suddenly breaking, etc etc...

Didn't Karpathy say they did in fact lack data in edge cases?

For example in AK's video
Building the Software 2.0 Stack by Andrej Karpathy from Tesla

from around 21 to 24 minute mark he talked about the real world corner cases.

 

[aubreymcfato]

For example in AK's video
Building the Software 2.0 Stack by Andrej Karpathy from Tesla

from around 21 to 24 minute mark he talked about the real world corner cases.

They can probably use simulations to create new useful data for existing edge cases, ie. they can take an already difficult road and simulate different weather/lightning, etc. But to create a new edge case, you actually need to know that.

It's probably useless if not dangerous to feed into the machine pictures of a road with random lines, the NN could learn a reaction to a pattern that doesn't exist in real life, and so misinterpret a real situation that resemble that simulated one... I think they have a more "controlled" approach when they use simulations to create data they already know they need. So simulations are always "clean" data, so to speak. But you need to know what "clean" means beforehand.

 

[Mike Smith]

Tesla Accuses UAW of Orchestrating an Anti-Elon Musk Infomercial

 

[vgrinshpun]

Nice to see you back here!

 

[Zhelko Dimic]

For example in AK's video
Building the Software 2.0 Stack by Andrej Karpathy from Tesla

from around 21 to 24 minute mark he talked about the real world corner cases.

What he said
Agreed that known edge cases could be enhanced by simulations

 

[elliott82]

Thanks for the answers!

 

[Remus]

I haven't watched the video yet, but isn't continual learning the basic strategy for autopilot? I thought the plan was basically you get x,000 cars on the road driving, basically collecting data in shadow mode, and once you get 5? billion miles of data it's demonstrably safer then a person. I thought the plan was to basically collect an arsenal of data that makes a) a public duty/service to prevent accidents b) indisputably proven to be safer for insurance legal purposes.

What you described is basically reinforcement learning, a form of un-supervised learning. Despite of its rising popularity thanks to Alpha Go, it remains a toy in academia.