I like the sound of that. Go Dojo!
TL;DR: FSD development that would have taken ten years could now take one year or even one month.
NVIDIA" project. Same deal when they excised them from the cars with their own ASIC. NVIDIA gouges so hard on prices that it drives people to run their own in-house chip development.
I agree with you in principle. Nvidia may well outpace Tesla in hardware innovation. One challenge Tesla faces is the time it takes them (years) to implement their design.
NVidias dominance is not due to a chip. It's the whole eco-system (hw/sw/interconnects and networking). I doubt Tesla will get Dojo to scale in a meaningful way any time soon. I've said this since the dojo announcement, and i still believe it to hold true.
They don't need the video - that's for humans to consume and decide if your scenario should be trained on. If yes, they take the telemetry, including the ON, map/gps data, speed, etc. and place that in the simulation. They watch how FSD Beta handles the maneuver, and train based on the results.
I don't see an upload video button for effed up FSD driving. Didn't they take that away some # releases ago?
15 million miles total? With 400,000 cars, FSDb should be adding over 15 million miles every week. The latest video I saw, Ashok claimed 250 million miles.
I've gotten that. How does an audio report get fed into an AI computer or does video go up too?
that works out to 37.5 miles per car assuming people enable it. I maybe use FSD about 5-10 miles per week.
A single car going through a single intersection is not much help to them. James Douma has explained how Tesla uses video from the cars. The engineers ask for video for a particular situation then all the cars from the fleet that encounter that situation upload the video. He alludes to this in a recent interview he did with Farzad Mesbahi:
