Because Musk has a weird obsession that everything needs to work with just cameras, and no other sensors can help. While it might work someday in the far future with proper AGI (e.g Delamain from Cyberpunk 2077), until then it’s a pretty hopeless endeavour.
The problem is computer vision has a LONG way to go before it’s truly on par with human eyesight. Musk loves to crow how cameras are sufficient since we use our eyes to drive.
The thing is, eyes have special neural circuits that detect motion. They essentially filter out unnecessary information and send just the motion details to the brain. This prevents the brain from being overloaded with every detail the eye constantly sees.
And being overloaded with everything is exactly what computer vision currently does. It’s just a stream of images that the computer must analyze completely. So it’s working exactly opposite to how the eye & brain works.
Long way, but it’s not an impossible task, as at the core the eye is nothing but a bunch of light sensors that spit out a result, we just need to figure out how to calculate that result ourselves. Motion amplification could be one solution given enough computing power to do it in real time, for example.
But we agree, safe and accurate camera based self driving isn’t going to happen in a long, long time.
I’m not sure the approach Elon Musk has to developing self driving will ever work.
From what I’ve heard about how they “teach” the AI, it probably won’t, because loading massive amounts of new data is rewarded, but there is no proper qualitative control.
Because Musk has a weird obsession that everything needs to work with just cameras, and no other sensors can help. While it might work someday in the far future with proper AGI (e.g Delamain from Cyberpunk 2077), until then it’s a pretty hopeless endeavour.
The problem is computer vision has a LONG way to go before it’s truly on par with human eyesight. Musk loves to crow how cameras are sufficient since we use our eyes to drive.
The thing is, eyes have special neural circuits that detect motion. They essentially filter out unnecessary information and send just the motion details to the brain. This prevents the brain from being overloaded with every detail the eye constantly sees.
And being overloaded with everything is exactly what computer vision currently does. It’s just a stream of images that the computer must analyze completely. So it’s working exactly opposite to how the eye & brain works.
Long way, but it’s not an impossible task, as at the core the eye is nothing but a bunch of light sensors that spit out a result, we just need to figure out how to calculate that result ourselves. Motion amplification could be one solution given enough computing power to do it in real time, for example.
But we agree, safe and accurate camera based self driving isn’t going to happen in a long, long time.
I’m not sure the approach Elon Musk has to developing self driving will ever work.
From what I’ve heard about how they “teach” the AI, it probably won’t, because loading massive amounts of new data is rewarded, but there is no proper qualitative control.