Quote:
Originally Posted by bluejays
Thanks. I still may be a little slow admittedly, but what are we fundamentally expecting to change with self driving as it stands today? Say the machine learning today learns that if an upcoming bicyclists' left leg twitches they're likely to fall into the cars' path. The reaction the car should make is to swerve the left to avoid (assuming no car beside). Wouldn't that generally be good enough? What will AGI resolve? I'm still struggling to understand where we need that 1% over the top leap?
|
Because, as I mentioned, Tesla vision AI is trained on what it knows, and what it has seen. In theory it knows a stop sign means stop because it was trained for that. The problem is that it requires may iterations of each unique case to make a decision on what to do. So what does it do when it sees something novel, a situation it isn't trained for? It fails. It may fail successfully, or it may fail catastrophically. The issue is that as an autonomous vehicle, deployed to millions of vehicles, is that these failures can not happen because someone needs to be liable. We aren't going to accept a collision rate anything less than near perfect. The minute a Tesla creams a kid in a driveway or parking lot is going to be the end of it.
To overcome these flaws, other companies have several sensors, and in my opinion, more robust programing. I think they can get safe enough with these methods, at least in areas with no winter, which is an entire other area of problems.
AGI would bring a human like brain to a vision only vehicle. That brain would be able to reason that a stop sign advertisement on the back of a bus doesn't mean stop. Or that it's Halloween, so you will see some weird things, kids everywhere etc. Or the low sun on the horizon isn't a yellow light. Or that flashing yellow at this particular spot should be treated as a stop sign. Or in this city, if you don't rush the yellow, you WILL get rear ended. And it's not that you can't program these things in, but that there are just too many, it's known as the long tail. And they are constantly changing, such as pedestrians with cell phones. Or headphones on. So AGI will provide the reasoning that allows it to solve for unique issues on the spot, vs pre-decided actions.