Re:Ask HN: What is the weirdest or most surreal (recent) tec...
by phkahler on Tuesday January 24, 2023 @02:48PM
(#34509167)
>> I feel this way about self driving cars as well. Its like even 10 years ago, the capabilities they have today would be considered mind blowing, yet they receive so much hate because they aren't yet adapted to all situations. Its baffling to me, every time I watch a video of one I get all giddy and I am like we are living in the future! Yet even from tech circles, they get so much criticism because they aren't 100% there yet.
This is one I take the side of the skeptic, and I can explain why. In college ~1994 a friend entered an autonomous ground vehicle contest. It was a big deal and our university hosted it each year - this was the 2nd year. He had cobbled together an unusual electric wheelchair base, a 486 PC and a tripod with an old camcorder looking down and forward with a frame-grabber card. The course was about one lane wide, bounded by 2 white lines painted in a grass field. There were some other obstacles and stuff (later years they had sections of dashed white lines and other complexity). We had talked about line following algorithms and such, and he had tried some things with limited success.
One day I said I had an idea, and he said OK but I need the hardware back in an hour. I don't recall if it was C or Pascal, but it was familiar enough that I coded something in an hour - with a visual display showing the 15 horizontal lines of pixels my code analyzed (at about 10 fps). We added some stuff later to incorporate sonar data, but that hour of code got us second place. I'm not here to brag about my fabulous code, I'm here to tell you that "faking it" with lane following and simple driving skills is really easy. There was zero "understanding" in what it was doing, zero ability to recover when it went wrong, zero of just anything but the simplest of code to enable moving forward without crossing the lines. I could write pseudo-code for it here in a comment in 10 minutes. When you think about it, driving safely on real roads requires semantic understanding of a lot of things and it's impossible to observe a vehicle driving and have any idea what level of comprehension it has. This combined with personal experience knowing how easy it is to do rudimentary lane following causes me to set the bar very high indeed. Parlor tricks are not safe, and whatever they have on the road today surely falls somewhere between that and fully sentient AI. But where is it exactly? What is good enough for your spouse and kids? And what is good enough for someone else to be letting drive on the road with them? Or while they cross the street on foot?
When a car can't "see" a semi crossing the road, or one doesn't identify something as human because it's not in a crosswalk, or another runs into and pushes an airplane, no I'm not impressed in the slightest at what these things CAN do reasonably well. The instances of what they get wrong tell me so much more about how far they have to go.
This is one I take the side of the skeptic, and I can explain why. In college ~1994 a friend entered an autonomous ground vehicle contest. It was a big deal and our university hosted it each year - this was the 2nd year. He had cobbled together an unusual electric wheelchair base, a 486 PC and a tripod with an old camcorder looking down and forward with a frame-grabber card. The course was about one lane wide, bounded by 2 white lines painted in a grass field. There were some other obstacles and stuff (later years they had sections of dashed white lines and other complexity). We had talked about line following algorithms and such, and he had tried some things with limited success.
One day I said I had an idea, and he said OK but I need the hardware back in an hour. I don't recall if it was C or Pascal, but it was familiar enough that I coded something in an hour - with a visual display showing the 15 horizontal lines of pixels my code analyzed (at about 10 fps). We added some stuff later to incorporate sonar data, but that hour of code got us second place. I'm not here to brag about my fabulous code, I'm here to tell you that "faking it" with lane following and simple driving skills is really easy. There was zero "understanding" in what it was doing, zero ability to recover when it went wrong, zero of just anything but the simplest of code to enable moving forward without crossing the lines. I could write pseudo-code for it here in a comment in 10 minutes. When you think about it, driving safely on real roads requires semantic understanding of a lot of things and it's impossible to observe a vehicle driving and have any idea what level of comprehension it has. This combined with personal experience knowing how easy it is to do rudimentary lane following causes me to set the bar very high indeed. Parlor tricks are not safe, and whatever they have on the road today surely falls somewhere between that and fully sentient AI. But where is it exactly? What is good enough for your spouse and kids? And what is good enough for someone else to be letting drive on the road with them? Or while they cross the street on foot?
When a car can't "see" a semi crossing the road, or one doesn't identify something as human because it's not in a crosswalk, or another runs into and pushes an airplane, no I'm not impressed in the slightest at what these things CAN do reasonably well. The instances of what they get wrong tell me so much more about how far they have to go.