@stewbits22
more from that month
Some new features of Teslas FSD architecture and why they are dumping radar. Vision is continuous input Radar is pulsed input. In fast changing situations like rapid breaking of a car Infront of you. Radar loses track of the braking car momentarily multiple times.. This data mismatch between vision and radar slows down decision making of the neural net. Teslas neural networks are evolving into branched structures where each branch is a semi autonomous net specialising in one task without affecting the rest of the nets performance. Tesla shows clips of FSD overriding driver error in car parks. The driver applies full acceleration at pedestrians by mistake but FSD puts the brakes on instead Much more detail on scalable auto labeling of objects, the dojo super computer they have built that processes all the data, a breakdown of the data stack etc. Video starts at 7hrs 51 min https://youtu.be/eOL_rCK59ZI
1 comment
What Tesla is doing is immensely rad. This is also why despite the inflated price, that TSLA stock is still a buy. They are going to own that corner of the market.