Comments by "XSportSeeker" (@XSpImmaLion) on "Tesla autopilot causes 8 car pileup" video.

  1. I'm absolutely not an expert... or even an intermediate in any of these areas, but from whatever little I know about computer vision vs lidar, from start I always found puzzling Tesla's decision to go with the former.... or rather, abandoning one in favor of the other "just because". Also the reason why I just can't believe FSD is coming anytime soon unless really major changes happen in the field that would likely require Tesla, Google and others NOT being at the helm of it because their philosophies on the matter don't fit what is required for it to really happen. Let me elaborate a bit. It seems to me, from the time I was experimenting with AR markers back almost 20 years ago during my CompSci course up to today, that there is an almost impossible problem to solve with computer vision, which is 100% accurate object identification. This is what leads to false positives which leads to phantom breaking. The tech evolved a very long way from back then, but from stuff I've seen and read on current tech, at least the ones that are open for public like the nVidia Jetson nano devboard stuff, you can have a high accuracy for identifying things, but it's still nowhere near perfect. And it'll likely never be by itself. Which sure, human drivers also won't be... but an autonomous driving system and a driver will always be attacking the problem at different angles, and so it becomes a problem. Lidar, though far from perfect and having it's own set of issues, eliminates some of the misidentification problems by basically seeing everything as volumetric shapes. But this ends up being also a source of problem... Lidar is able to see in 3D, but it has problems telling what things are and setting the boundaries exactly since it's basically sonar with light, so it might mistake, for instance, a flurry of snowflakes or a bunch of insects flying close together to a person or a solid object. This is kinda why the mix of both was needed, but also the reason why combining both ends up accentuating false positives of both. Say Lidar identifies something in front of the car, but computer vision tells you it's just a bunch of dust, so it should keep going without hitting the breaks... but for both to work in combination like that they'd need to individually work almost perfectly. If they can't, they'll always go on the safe side and break whenever one or the other detects a problem. The situation you get at is not as bad as it may sound, but it'll always look pretty bad for us. Why? It's a bit like AI - autonomous cars will be far better than us at some things, but also completely incompetent in comparison to us at others. Because it looks at the problem of driving in a different way, and it has different strengths and weaknesses in comparison to a human driver. Now, I was saying I could see FSD happening without Tesla, Google and whatnot, right? It's a moonshot and even harsher attack at the problem by looking at it in a different way, far more costly and drastic, but because of the tech involved, I currently see it as more feasible. What we'd need is something closer to what you may have seen in Minority Report ages ago... 20 years ago if it's the Tom Cruise movie (yeah, it's that old already xD ). By which I mean, dedicated car-only lanes, with a perfect communication system that is both isolated and secure, in which individual cars can see each other in a digital map. Now, of course, this poses all sorts of side questions and different problems in itself. For instance, if we're going to go that far, it just makes far more sense to perfect public transportation instead, not end consumer cars. Making roads and lanes dedicated to only cars where you can't have pedestrians crossing and stuff like that is close to impossible... we'd have to rethink everything related to transportation and urban planning. Then comes the matter of a means of communication that is both all encompassing but also secure, which we are basically walking further and further away from as time passes. The more likely scenario is that we'll continue going the way FSD is already going, people will end up conforming to what it is, and we'll make the tradeoffs along the way. Just because even if accidents will happen which we will have to pose the question whether it'd still happen if a human driver was behind the wheel at the time, on the collective autonomous cars will still cause less accidents. I dunno how the matter of responsibility will be solved, but it's already being dealt with as we speak. Values will change overtime. For instance, this ghost breaking case. Say all cars there were driving autonomously. If one of the cars breaks all of a sudden, shouldn't all other cars break too - if they were all autonomous? Or perhaps not even be driving too fast due to conditions. Autonomous cars will continue being a cause of headaches and issues while you have the mix of human drivers and autonomous cars on the roads, because the way we drive is always fundamentally different to, and often in conflict with the way autonomous cars drive... but if you only have autonomous cars on the roads, things become a bit more predictable. So we're back at dedicating lanes or roads only for one type of driving versus the other, which is pretty harsh. But this is basically the only way I can see problems like these solved... until then, we'll have to deal with accidents like this one, if people continue accepting this path towards autonomous driving that is. There are big chances that not long into the future we'll have an autonomous car crash or a series of them so notorious and so bad in the public eye that everything will get scratched. This has always been the biggest danger for the tech in general, and haphazard treatment of it like Uber did and Tesla does by overpromising, using misleading names, and sometimes trying to hide facts from public view only increases the chances of that happening.
    8