Musk has said similar things in the past; that cars should be able to do everything human drivers can do with only vision, because that's all human drivers have (discounting the far more minor sensory inputs like sound and touch).
I find that take fascinating. Its, in effect, saying "well, computer hardware and the human brain are similar enough; the human brain just got an evolutionary headstart on its software, so once we solve that we're golden". Not only is there no evidence to support this conclusion; its trivial to recognize that the human brain and computers operate differently, at a fundamental, logical level. The brain has no transistors. It doesn't care about XOR or solid state memory. There is literally no part of the human brain which even remotely resembles any internal part of a computer; not remotely.
Why is it so accepted that this problem can be solved in software? Or, maybe: it can be solved, but it will take so much computing power that any computer which could solve it in real time couldn't fit in or be powered by a car? Virtualizing linux on my Windows PC incurs a double-digit percentage performance penalty, and that's on the same ISA; what is the human brain's ISA? AI is an emulation of human behavior; what is it's performance penalty? Are we just surfing Moore's Law hoping to outrun it?
I just find the hubris of his statement baffling; that we, meaning Our Scientific Community, literally does not completely understand why or how acetaminophen works to alleviate pain in the human brain; yet they and Musk are wholly confident that given the same inputs their software can emulate the human brain's outputs.
I'll commit this to the public record, and feel free to revisit in ten years; if a synthetic brain system ever hits performance parity with the human brain in a task like driving or consciousness, it will be running on hardware that looks nothing like what I'm typing on. I suspect that even calling it "hardware", with emphasis on the word "hard", will seem antiquated. Maybe self driving is 100% a software problem; but the brain doesn't really differentiate.
It would be interesting if the thing wrong with the "eyes are just cameras" logic was the subtle differences. Like how retinas are not flat, but concave or the way our eyes continuously reposition to fill in gaps in our vision. Such differences seem incidental or compensatory, but it's not impossible that our brain is able to make better distance/position judgements because of them.
The other missing piece is just how good the brain is at object recognition and persistence. A huge advantage when you have a 2d image and you need to gauge distance (at least partially) by size or when you need to avoid a moving obstacle by predicting its trajectory.
Your eyes 100% do several other things to give depth cues to your brain. That's why you can cover one eye and still judge depth pretty well. In fact, maybe two eyeballs really are all you need to do safe driving, but 2 CCD image sensors are NOT an adequate replacement for all the other things eyeballs give you.
Your eyes know the distance of something because they know how they had to move to bring the object into focus. Your eyes also have to be at slightly different angles to look at a distant object, which also is a depth cue, as well as all the temporal depth info you get simply as a byproduct of having an insanely powerful and robust object classification and world simulation engine in your skull.
I remember seeing some studies suggesting that vision training of artificial neural networks forms "similar" structures to the human brain. I don't think the brain is that magical compared to modern DL.
> it will be running on hardware that looks nothing like what I'm typing on.
not to mention the fact that humans with just sight and sound for sensory input are pretty crappy drivers - just look at accident statistics. Don't we want our AI drivers to be better than humans.
I find that take fascinating. Its, in effect, saying "well, computer hardware and the human brain are similar enough; the human brain just got an evolutionary headstart on its software, so once we solve that we're golden". Not only is there no evidence to support this conclusion; its trivial to recognize that the human brain and computers operate differently, at a fundamental, logical level. The brain has no transistors. It doesn't care about XOR or solid state memory. There is literally no part of the human brain which even remotely resembles any internal part of a computer; not remotely.
Why is it so accepted that this problem can be solved in software? Or, maybe: it can be solved, but it will take so much computing power that any computer which could solve it in real time couldn't fit in or be powered by a car? Virtualizing linux on my Windows PC incurs a double-digit percentage performance penalty, and that's on the same ISA; what is the human brain's ISA? AI is an emulation of human behavior; what is it's performance penalty? Are we just surfing Moore's Law hoping to outrun it?
I just find the hubris of his statement baffling; that we, meaning Our Scientific Community, literally does not completely understand why or how acetaminophen works to alleviate pain in the human brain; yet they and Musk are wholly confident that given the same inputs their software can emulate the human brain's outputs.
I'll commit this to the public record, and feel free to revisit in ten years; if a synthetic brain system ever hits performance parity with the human brain in a task like driving or consciousness, it will be running on hardware that looks nothing like what I'm typing on. I suspect that even calling it "hardware", with emphasis on the word "hard", will seem antiquated. Maybe self driving is 100% a software problem; but the brain doesn't really differentiate.