> Finally, as someone working in AI, I’m actually very optimistic about their next approach of just learning an end to end model to drive the car. I think this is the only scaleable way to reach human level driving. Any set of rules based + optimization based planner will have those edge cases that ruin the cars capability, in fact it turns out the long tail is really long.
Tesla doesn’t have an end-to-end model. They call it “end-to-end AI”, which just means they’re applying AI to all subsystems, exactly like everyone else. All the big players already use heavy ML-based planning systems that are not “rules or optimization based” and which are way more sophisticated than anything Tesla has put out or is planning to.
We don't know if that's actually true. Tesla claims "end-to-end AI" in their next iteration, not an "end-to-end model". Wayve [1] is the only one who has an end-to-end model that takes video input and outputs a trajectory, but they are still doing one-off demos.
Tesla doesn’t have an end-to-end model. They call it “end-to-end AI”, which just means they’re applying AI to all subsystems, exactly like everyone else. All the big players already use heavy ML-based planning systems that are not “rules or optimization based” and which are way more sophisticated than anything Tesla has put out or is planning to.