Is it just me or do the problems at dawn and dusk just come across as under developed technology. Why don’t they have better light sensors? And more data? I feel like people are rushing when rushing is the last thing they should be doing.
Better sensors won’t help. It’s a problem of creativity and adaptability. AI does well when things are stable and predictable, as deep down they’re about pattern recognition and complex prediction models. But they can’t come up with something unique and rationally useful that adapts to an unpredictable or chaotic circumstance. Our current modeling paradigm for AI can’t do that yet. You can train one AI to drive in snow, but you can’t train an AI that drives well on a regular condition which could spontaneously adapt to an unforeseen snowy road. The model just breaks and stop working properly. While a human who drives well, could quickly adapt to a sudden snow pour even if it is the first time encountering it. As humans can be instructed (current AI can’t) via receiving compressed information about what to expect and how to react to a snowy road. An AI has to train (go through the real situation) or it can’t adapt.
Yeah I would definitely think that better LIDAR would be the answer for the problems with dawn and dusk, but the problems with turning might be trickier. It seems there are so many unpredictable variables at play in every situation, that it’s hard to model for all of them. But even with that, I would assume eventually with enough modeling that problem will be fixed.
Is it just me or do the problems at dawn and dusk just come across as under developed technology. Why don’t they have better light sensors? And more data? I feel like people are rushing when rushing is the last thing they should be doing.
Better sensors won’t help. It’s a problem of creativity and adaptability. AI does well when things are stable and predictable, as deep down they’re about pattern recognition and complex prediction models. But they can’t come up with something unique and rationally useful that adapts to an unpredictable or chaotic circumstance. Our current modeling paradigm for AI can’t do that yet. You can train one AI to drive in snow, but you can’t train an AI that drives well on a regular condition which could spontaneously adapt to an unforeseen snowy road. The model just breaks and stop working properly. While a human who drives well, could quickly adapt to a sudden snow pour even if it is the first time encountering it. As humans can be instructed (current AI can’t) via receiving compressed information about what to expect and how to react to a snowy road. An AI has to train (go through the real situation) or it can’t adapt.
Yeah I would definitely think that better LIDAR would be the answer for the problems with dawn and dusk, but the problems with turning might be trickier. It seems there are so many unpredictable variables at play in every situation, that it’s hard to model for all of them. But even with that, I would assume eventually with enough modeling that problem will be fixed.