>> I'll respond clearly: We're multiple computer vision leaps forward away from what Elon needs for success. They're easily half a decade behind Lidar based systems. And people die as a consequence of putting those systems on the road.
>Custom ASICs for ML that Tesla is building is fairly well-understood tech at this point. High-end smartphones have used similar tech for years now
joshuamorton is right and you didn't really address the point made. It isn't about understanding the hardware, it's about understanding the software. You point me to any individual who can tell you concretely how a specific convolutional neural network (CNN) arrives at a given solution and I'll take this all back.
We know so little about how CNNs work, and plane old neural networks (NNs) for that matter, it's embarrassing. They're prone to adversarial attacks, not only that, you can successfully mount the same attack against _any_ NN that was trained on a common dataset. We don't know how to effectively defend against these yet in a white box setting.
We have no idea what the solution space looks like. We barely understand why one of the most simple optimization algorithms outperforms almost all others on at these tasks. We barely understand why randomizing data visitation results in solutions that perform completely differently.
The counter argument goes something like "we tested it on a big eval set and it aced it" Well I'm here to tell you that things change. Assumptions made in creating that eval set might not generalize to all real-world cases. And in this case bad assumptions result in death.
As someone who has spent a good portion of my life working on this stuff I've learned that sometimes (most of the time) the best choice is the obvious, known, simple one. The fact that a good portion of cars may be controlled via something we know so little about should worry you.
On a different note, I think you're also missing why Tesla have opted to invest in a percept suite that doesn't use LIDAR. That reason is cost. Tesla needs to sell cars now. They and their customers cannot afford to put LIDAR on their current platform. At the same time they need to move the metal and they know their customers want AEVs. I think their strategy is sound from a purely business cost/benefit analysis. It's a risk but from a financial perspective a good one, because if it works they hit pay dirt.
> The fact that a good portion of cars may be controlled via something we know so little about should worry you.
I bet that we know much, much more about CNNs than about our brains.
And today pretty much all the cars are controlled by something that we don’t know basically nothing about. Why this doesn’t worry you?
The fields of neuroscience and psychology are far more advanced compared with the infantile field of deep learning.
Humans also learn independently from one another which means outlier events aren’t as much of a problem. You have the ability to observe the world around you, say you see an accident due to some rare weather event, and learn some abstract lesson extremely quickly. A CNN has to learn this over millions of samples out of band.
Based on the number of actual reproducible controlled experiments done on CNNs, I'm not convinced the field is less advanced than psychology at this point.
Well, we know quite a bit about how humans perform at the relevant levels of abstraction, and that's the level at which driving automation systems should be evaluated. They are not there yet, but there's no reason to believe the problems are insurmountable.
> I think you're also missing why Tesla have opted to invest in a percept suite that doesn't use LIDAR. That reason is cost.
Even beyond pure cost, one problem Tesla has is that they already SOLD tens of thousands of FSD option packages, predicated on NOT needing to retrofit the cars with LIDAR.
So arguably, Tesla might be worse off if they deliver a LIDAR based FSD (and need to retrofit tens of thousands of cars, or pay off the owners), than if they just plod on with a camera based FSD that never quite works safely.
>one problem Tesla has is that they already SOLD tens of thousands of FSD option packages
It's always seemed sketchy to me that they're selling FSD option packages when they don't know when FSD will actually happen or whether the current in-car hardware will be adequate to support it once it does happen.
He's likely banking on a bimodal distribution where FSD either happens very soon (within a year or so) in which case the hardware likely will be fine OR not for 5+ years in which case these cars likely will not be on the road any more and it becomes a moot point.
>Custom ASICs for ML that Tesla is building is fairly well-understood tech at this point. High-end smartphones have used similar tech for years now
joshuamorton is right and you didn't really address the point made. It isn't about understanding the hardware, it's about understanding the software. You point me to any individual who can tell you concretely how a specific convolutional neural network (CNN) arrives at a given solution and I'll take this all back.
We know so little about how CNNs work, and plane old neural networks (NNs) for that matter, it's embarrassing. They're prone to adversarial attacks, not only that, you can successfully mount the same attack against _any_ NN that was trained on a common dataset. We don't know how to effectively defend against these yet in a white box setting.
We have no idea what the solution space looks like. We barely understand why one of the most simple optimization algorithms outperforms almost all others on at these tasks. We barely understand why randomizing data visitation results in solutions that perform completely differently.
The counter argument goes something like "we tested it on a big eval set and it aced it" Well I'm here to tell you that things change. Assumptions made in creating that eval set might not generalize to all real-world cases. And in this case bad assumptions result in death.
As someone who has spent a good portion of my life working on this stuff I've learned that sometimes (most of the time) the best choice is the obvious, known, simple one. The fact that a good portion of cars may be controlled via something we know so little about should worry you.
On a different note, I think you're also missing why Tesla have opted to invest in a percept suite that doesn't use LIDAR. That reason is cost. Tesla needs to sell cars now. They and their customers cannot afford to put LIDAR on their current platform. At the same time they need to move the metal and they know their customers want AEVs. I think their strategy is sound from a purely business cost/benefit analysis. It's a risk but from a financial perspective a good one, because if it works they hit pay dirt.