FSD discussion

[Stretch2727]

Not an AI or NN expert but 2 questions with regards to v12 FSD.

1. How does the neural net end up being a relatively small dataset/program compared to all the video data input? Does the NN scale linearly with more data training or is it a fixed "system" just learning more information. In other words is there a constraint when doing more and more training that the NN could be come too large to run in the car.

2. How will they train for accident avoidance? Will there need to be some programming for this or will they need to find many situations where a good driver avoided an accident or at least minimized the consequences.

 

[Buckminster]

https://twitter.com/x/status/1696914493241213084

 

[heltok]

https://twitter.com/x/status/1696833182421672165

The V12 he was driving was already pretty good. Maybe just training the same model for longer with the same balance of data would have fixed the one intervention, but rather they will not retrain and just keep training on the same model but add a bit extra data. I assume the end2end will be very data/training hungry. Then at some point they will probably increase the network size from 50fps to 36fps size and retrain from scratch requiring even more data and longer training. As they haven't decided to do that yet is probably because they don't have enough compute/data to train an even larger model... Here is good graph showing how scaling the dataset and model size improved performance, the probably are at the blue line right now not seeing the end the improved performance...

 

[willow_hiller]

Not an AI or NN expert but 2 questions with regards to v12 FSD.

1. How does the neural net end up being a relatively small dataset/program compared to all the video data input? Does the NN scale linearly with more data training or is it a fixed "system" just learning more information. In other words is there a constraint when doing more and more training that the NN could be come too large to run in the car.

2. How will they train for accident avoidance? Will there need to be some programming for this or will they need to find many situations where a good driver avoided an accident or at least minimized the consequences.

For very large neural networks (in the billions of parameters), the weights do end up approximating something like a database. That's how LLMs end up remembering specific facts about topics.

In terms of file size for those weights, we can look at Llama v2. Their 70 billion parameter model has 16 bit floating point precision, so it works out to about 130 GB.

I believe HW3 has 64 GB of flash storage, and even if it could hold 64 GB of weights, it wouldn't be able to process them due to 8 GB of memory per chip.

But this NN size constraint is a part of what induces generalization. The network isn't capable of solving the task it's being trained to do by memorizing specific answers, so it minimizes the loss by learning generic answers. And some of these generic learnings may help v12 avoid collisions in the first place.

 

[Buckminster]

FSD reduced to $12k

https://twitter.com/x/status/1697461111489225012

 

[Buckminster]

To solve the lack of stalks "crisis":

Could Tesla introduce a feature where the vehicle signals automatically for manual drivers when nav is in use?

 

[DarkandStormy]

https://twitter.com/x/status/1479595757388845059

Per Elon, this means FSD is regressing.

 

[jeewee3000]

https://twitter.com/x/status/1479595757388845059

Per Elon, this means FSD is regressing.

Not necessarily.

Most likely it is an incentive to get more HW4 FSDb intervention data. The FSD take rate has dropped since the optimism from Autonomy Day/AI Day has faded and people think FSD is years away as opposed to around the corner.

So HW4 being so new Tesla currently has very little HW4 testers using FSDb. (shadow mode they can still pull clips, but Tesla gets no intervention data. V12 automatically sends interventions to the mothership in order to train on improvements. This principle was done in prior versions as well, but now it is a lot more automated.)

So yeah, I don't think Elon states FSD is regressing. But they might not have as many takers as they like for the full FSD 15k package, and therefore want to create some demand.

 

[DarkandStormy]

Not necessarily.

Most likely it is an incentive to get more HW4 FSDb intervention data. The FSD take rate has dropped since the optimism from Autonomy Day/AI Day has faded and people think FSD is years away as opposed to around the corner.

So HW4 being so new Tesla currently has very little HW4 testers using FSDb. (shadow mode they can still pull clips, but Tesla gets no intervention data. V12 automatically sends interventions to the mothership in order to train on improvements. This principle was done in prior versions as well, but now it is a lot more automated.)

So yeah, I don't think Elon states FSD is regressing. But they might not have as many takers as they like for the full FSD 15k package, and therefore want to create some demand.

https://twitter.com/x/status/1262530589506273280

Why are they dropping the price on something that's going to be nearly $100k in value so soon?

 

[Buckminster]

https://twitter.com/x/status/1698212328091640205

https://twitter.com/x/status/1698216951074730347

 

[BitJam]

Why are they dropping the price on something that's going to be nearly $100k in value so soon?

Good question. I imagine it's because they want to get more data to feed their ever growing computers for training. They may also want to get data from more diverse regions. I'm reminded of Tom Sawyer whitewashing his Aunt Polly's fence.

 

[uscbucsfan]

So Gali no longer thinks major FSD progress is happening any time soon. Dubious position for a Tesla bull.

We don't need robotaxi for Tesla stock to start soring. We just need to see major FSD progress for the market to put 2 and 2 together.

(In this case, 2 + 2 = 1000)

Define progress? Even Omar's videos don't budge the needle for the stock and they are what many hope V12 is for actual drives (hands free, no disengagements, etc).


It seems any stock jump tied to FSD would be in the driverless form, not any sort of level 2, regardless of how good it is. At the minimum, level 3, which if Elon sticks to his guns we will never see on a Tesla.

 

[Usain]

Define progress? Even Omar's videos don't budge the needle for the stock and they are what many hope V12 is for actual drives (hands free, no disengagements, etc).


It seems any stock jump tied to FSD would be in the driverless form, not any sort of level 2, regardless of how good it is. At the minimum, level 3, which if Elon sticks to his guns we will never see on a Tesla.

I think major progress would be wide release of V12 where Tesla allows everyone to be hands free with the promised smooth, human-like behavior and very rare need for intervention. Wall Street would not ignore that kind of performance for very long.

I believe that will be the point where FSD has its "ChatGPT" moment.

 

[uscbucsfan]

I think major progress would be wide release of V12 where Tesla allows everyone to be hands free with the promised smooth, human-like behavior and very rare need for intervention. Wall Street would not ignore that kind of performance for very long.

I believe that will be the point where FSD has its "ChatGPT" moment.

I disagree. We see people like Omar touting this for years...those who aren't Tesla followers don't know his rides are curated. Also, there's no promise that v12 is fully hands free.

I don't think FSD leads to a large stock jump until it's licensed by someone else (confirmed) or there is actual driverless functionality. Regardless of what Elon, Omar, or others say or show about FSD now, it hasn't moved the needle for the stock much.

 

[Usain]

I disagree. We see people like Omar touting this for years...those who aren't Tesla followers don't know his rides are curated. Also, there's no promise that v12 is fully hands free.

I don't think FSD leads to a large stock jump until it's licensed by someone else (confirmed) or there is actual driverless functionality. Regardless of what Elon, Omar, or others say or show about FSD now, it hasn't moved the needle for the stock much.

Elon did promise hands-free coming soon. Of course, that was awhile ago, but I believe that the promise will be fulfilled with V12.

Omar's videos are not influential to the greater investment community.

But think about what happens when FSD gets really, really good. Word spreads. A-list reporters will do stories on it. That's all it will take for the market to wake up and say, "Oh, that FSD thing is for real now. I can't ignore it any longer."

The same thing happened with the technology that became ChatGPT. It got ignored for a long time before suddenly everyone realized its potential.

 

[uscbucsfan]

Elon did promise hands-free coming soon. Of course, that was awhile ago, but I believe that the promise will be fulfilled with V12.

Omar's videos are not influential to the greater investment community.

But think about what happens when FSD gets really, really good. Word spreads. A-list reporters will do stories on it. That's all it will take for the market to wake up and say, "Oh, that FSD thing is for real now. I can't ignore it any longer."

The same thing happened with the technology that became ChatGPT. It got ignored for a long time before suddenly everyone realized its potential.

He promised it in January, but that didn't happen.

He never said if it would be situational, which it likely will be as the current nags are more frequent with difficult situations.

How much do you think a solid level 2 ADAS boost the stock for a company like Tesla, logically, regardless of reviews? Again, it's my opinion that licensing or driverless L4+ will be when we see a huge leap and without diving into that too much, I think we are years away. Deliveries, profits, additional products, and adoption are the biggest drivers for TSLA in the very near future. In the long run, I do see actual FSD being very good to us financially.

 

[AndrewZ]

I disagree. We see people like Omar touting this for years...those who aren't Tesla followers don't know his rides are curated. Also, there's no promise that v12 is fully hands free.

I don't think FSD leads to a large stock jump until it's licensed by someone else (confirmed) or there is actual driverless functionality. Regardless of what Elon, Omar, or others say or show about FSD now, it hasn't moved the needle for the stock much.

I'm not sure what people are expecting out of a hands-free urban Level 2 ADAS, I personally would not pay one red cent extra for that.

A Level 3 system that allows me to stop paying attention to the road for long haul highway driving, that's something I'd pay for, but I loathe to think about how long it'll be before a company takes responsibility for what vehicles are doing at highway speeds.

 

[uscbucsfan]

A Level 3 system that allows me to stop paying attention to the road for long haul highway driving, that's something I'd pay for, but I loathe to think about how long it'll be before a company takes responsibility for what vehicles are doing at highway speeds.

Things change, but Elon has been pretty vocally against the idea of level 3 for Tesla.

 

[AndrewZ]

Things change, but Elon has been pretty vocally against the idea of level 3 for Tesla.

Level 3 might happen in the city before it happens at highway speeds where reaction times need to be shorter and the consequences of mistakes tend to be much higher because of the speeds involved.

Level 4-5 on highways, who knows when that will happen

 

[jerry33]

Level 3 might happen in the city before it happens at highway speeds where reaction times need to be shorter and the consequences of mistakes tend to be much higher because of the speeds involved.

Level 4-5 on highways, who knows when that will happen

As it is now, there are virtually no takeovers on the highway.