do we need 10 or 100M more vehicles streaming data, or do we need to use 1000x more transistors daily for the AI annealing? 10,000x the number of parameters?
Maybe things have to be a billion times better before we get real FSD?
One problem with this kind of estimation is that it's not actually clear that we can get there from "here" using current approaches. If it turn we can't, estimation is pretty hopeless if history tells us anything about putting a timeline on breakthroughs...
Autonomous vehicles by and large aren’t a data problem anyways - it’s a robotics problem conditioned on the output of ML models for perception and prediction. The majority of the work is just implementing all of the bizarre edge cases inside of the robotics stack.
The idea of tackling this with an e2e ML model is a pipe dream touted by people that aren’t familiar with the space or trying to hype up their approach. Whenever this approach is attempted teams will very quickly realize how untenable it is and return to implementing individual robotics modules.
tl;dr there’s nothing special about Tesla’s approach to AVs or machine learning and you shouldn’t expect them to leapfrog the competition because of some nebulous ML-based advantage.
Coming from someone who has seen the interiors of both the FSD beta and the stack at an actual AV company.
Except Andrej Karpathy. Why dumb down if they know it isn't working, now that they have all the financial resources to pivot and overtake the other manufacturers (who have not achieve anything anyway)? It makes no sense, unless Karpathy does not know what he's doing.
Edit: the argument that others don't agree is not convincing because Tesla should not have succeeded by others' standards either.
If you listen to any of AK’s talks he only really talks about perception and he’s basically said that right now none of their planning is based on ML models. This is backed up by what reverse engineers have found in their Tesla’s.
Could you speed up labelling by using a combination of eye tracker cameras, road cameras, and brain signal sensors on real drivers? That way you could quickly understand what road footage corresponds with certain brain signals. Assuming brain signals are easier to label.
The current deep learning approach is a dead end.