"More data" doesn't necessarily solve anything at all. It's tempting to think that, and that if there's just enough data we can solve all problems. But it's very, very far from the truth. Machine learning (AKA glorified curve fitting) is interesting, but the reality is that it's an approach for a pretty limited set of problems. "How do I pilot a complex vehicle in a complex world from A to B without killing people," just likely isn't one of them, even if it might be applicable to ask, "Is this a raccoon or a baby human?" from small snapshots along the way to solving that problem. Not only that, but datasets are usually pretty specific to the algorithm they are applied to, and they require a huge amount of manual testing and maintenance to apply to them to say "Yes, this really does look like a solution," or, "No, you're coming to insane conclusions," at every step along the way. In other words, the bigger the dataset, the more the algorithm's hand must be held too (not necessarily a linear relationship, but a positive correlation if you want that data to actually be useful for anything...).
(There's more involved in autopilot software than just machine learning, but that's what the "big datasets" thing applies to.)
I’d imagine they are using a bunch of narrow AI functions for various parts of the machine, but still hand crafting the core algorithm. I remember the first self-driving car was really just a robot attached to a steering wheel and a dashcam. So what it was learning to recognize was just which way the road was curved. It drove at a fixed 5mph and was still really dangerous to ride in even after training. That was like 50 years ago but still
i mean, yeah, thats what captcha is for. and idk, more data does in fact make it safer. look at research papers of it. sure, itll never get perfect but at least its safer than humans most times. but what would be even safer is trains :train-shining:
"More data" doesn't necessarily solve anything at all. It's tempting to think that, and that if there's just enough data we can solve all problems. But it's very, very far from the truth. Machine learning (AKA glorified curve fitting) is interesting, but the reality is that it's an approach for a pretty limited set of problems. "How do I pilot a complex vehicle in a complex world from A to B without killing people," just likely isn't one of them, even if it might be applicable to ask, "Is this a raccoon or a baby human?" from small snapshots along the way to solving that problem. Not only that, but datasets are usually pretty specific to the algorithm they are applied to, and they require a huge amount of manual testing and maintenance to apply to them to say "Yes, this really does look like a solution," or, "No, you're coming to insane conclusions," at every step along the way. In other words, the bigger the dataset, the more the algorithm's hand must be held too (not necessarily a linear relationship, but a positive correlation if you want that data to actually be useful for anything...).
(There's more involved in autopilot software than just machine learning, but that's what the "big datasets" thing applies to.)
I’d imagine they are using a bunch of narrow AI functions for various parts of the machine, but still hand crafting the core algorithm. I remember the first self-driving car was really just a robot attached to a steering wheel and a dashcam. So what it was learning to recognize was just which way the road was curved. It drove at a fixed 5mph and was still really dangerous to ride in even after training. That was like 50 years ago but still
i mean, yeah, thats what captcha is for. and idk, more data does in fact make it safer. look at research papers of it. sure, itll never get perfect but at least its safer than humans most times. but what would be even safer is trains :train-shining: