3 ms·

FSD already understands some human gestures, and it learns this stuff by seeing how human drivers do things. Understanding a headlight flash doesn't seem like a

FSD already understands some human gestures, and it learns this stuff by seeing how human drivers do things. Understanding a headlight flash doesn't seem like a big ask. And recently, the biggest complain from FSD users was that it drove a little too aggressively. Because it copies human drivers, FSD actually was doing slow rolling stops for a while. Then people freaked out and pretended that was super unsafe, and the government made them do full stops.

deng 2y ago
But that's exactly the point: there's no general rule for this kind of thing, it depends on the concrete situation. There are definitely stop signs where you need to do a full stop because you cannot see any oncoming traffic. Basically: it's OK to do rolling stops until it isn't. Same goes for flashing headlights, these can mean different things, depending on the situation. It can mean: I see you, feel free to drive. It can also mean: what the hell are you doing? Or: you left your laptop on the car roof.
- DennisP 2y ago
  Learning that sort of context is what neural nets are good at. The whole point is that none of this is hardcoded; the AI just learns what humans do in similar situations. At the point they're not even hand-labeling anything. The only input is video. Flashing headlights can be ambiguous for humans too. There have been plenty of times when someone flashed lights at me and I had no idea why.
jajko 2y ago
I am not going to trust super expensive car and life/health of me and my family on 'its generally working ok, sort of' level. This ain't new headphones or TV, people apart from emotional early adopters have very different expectations when their lives are at stake. Namely absolutely stellar performance compared to average driver, being marginally better while struggling with tons of non-standard situations ain't cutting it.
FireBeyond 2y ago
Neither of those things were "learned adaptations" from human driving, and this is one of the biggest fallacies around FSD. People letting their FSD do stupid things and only intervening at the last moment (if at all) to let it learn from its mistakes. That it might use the direction of the car in front of it as a guidance track doesn't mean it understood the human gestures of the cop telling it to do that. Or in a concert parking lot where an attendant might be doing things like alternating cars to the left and right lots. > Because it copies human drivers, FSD actually was doing slow rolling stops for a while. No, this was programmed behavior, with an interface/config setting, to do a "rolling stop". People "freaked out" because Tesla was literally allowing the car to perform illegal traffic infractions, and if they'd do it for stop signs, what else would they do it for? But none of this is some Tesla "swarm" learning to do rolling stops. There is no adaptive learning happening. This is all trained from static models according to parameters from Tesla.
- DennisP 2y ago
  Version 11 had lots of hand-coded behavior. Version 12 is entirely a neural network, trained on human driving. Partly it comes from people running FSD and intervening sometimes, and partly it comes from just passive observation of a people doing their own driving. When FSD runs, video feeds into the neural net and it outputs vehicle controls, that's it. Learning to respond to gestures is just more training on video and car control data. It shouldn't be hard to believe given all the other things we're doing these days with large neural networks.

deng 2y ago

But that's exactly the point: there's no general rule for this kind of thing, it depends on the concrete situation. There are definitely stop signs where you need to do a full stop because you cannot see any oncoming traffic. Basically: it's OK to do rolling stops until it isn't. Same goes for flashing headlights, these can mean different things, depending on the situation. It can mean: I see you, feel free to drive. It can also mean: what the hell are you doing? Or: you left your laptop on the car roof.

DennisP 2y ago
Learning that sort of context is what neural nets are good at. The whole point is that none of this is hardcoded; the AI just learns what humans do in similar situations. At the point they're not even hand-labeling anything. The only input is video. Flashing headlights can be ambiguous for humans too. There have been plenty of times when someone flashed lights at me and I had no idea why.

jajko 2y ago

I am not going to trust super expensive car and life/health of me and my family on 'its generally working ok, sort of' level. This ain't new headphones or TV, people apart from emotional early adopters have very different expectations when their lives are at stake. Namely absolutely stellar performance compared to average driver, being marginally better while struggling with tons of non-standard situations ain't cutting it.

FireBeyond 2y ago

Neither of those things were "learned adaptations" from human driving, and this is one of the biggest fallacies around FSD. People letting their FSD do stupid things and only intervening at the last moment (if at all) to let it learn from its mistakes. That it might use the direction of the car in front of it as a guidance track doesn't mean it understood the human gestures of the cop telling it to do that. Or in a concert parking lot where an attendant might be doing things like alternating cars to the left and right lots. > Because it copies human drivers, FSD actually was doing slow rolling stops for a while. No, this was programmed behavior, with an interface/config setting, to do a "rolling stop". People "freaked out" because Tesla was literally allowing the car to perform illegal traffic infractions, and if they'd do it for stop signs, what else would they do it for? But none of this is some Tesla "swarm" learning to do rolling stops. There is no adaptive learning happening. This is all trained from static models according to parameters from Tesla.

DennisP 2y ago
Version 11 had lots of hand-coded behavior. Version 12 is entirely a neural network, trained on human driving. Partly it comes from people running FSD and intervening sometimes, and partly it comes from just passive observation of a people doing their own driving. When FSD runs, video feeds into the neural net and it outputs vehicle controls, that's it. Learning to respond to gestures is just more training on video and car control data. It shouldn't be hard to believe given all the other things we're doing these days with large neural networks.