4 ms·
Their world model is shit apparently. It should be a hierarchical Bayesian model, constantly updated, and checked for likehood distance from priors, and when it
by pas 6y ago
Their world model is shit apparently. It should be a hierarchical Bayesian model, constantly updated, and checked for likehood distance from priors, and when it gets too big, alert the driver and disengage.
Cars don't randomly appear and disappear in the middle of the open road. If their model cannot interpret these events as "a big fuckin' problem, I'm out of my operational envelope, I need to stop", then that model is indeed dangerous.
Sure, this is easier said than done.
- frongpik 6y agoThe ML model may be shit and that's ok so long as it shows the path it's going to follow and sticks with it.
- pas 6y agoSure, if it's simply a follow the car in front of me model. (Which is lightyears from FSD.) Also then it needs to be super-super finicky about what it interprets as a car/road/etc, and start disengaging the instant it detects something is not right. But obviously Tesla/Musk decided to just tone down the carefulness.