Reciprocity
Active Member
No one, not even Tesla really knows how many miles because not all miles are the same. EAP can run in all HW2 cars in ghost mode so it doesn't have to be engaged to get the types of data they need. It's Certainly better when EAP is engaged because you get a unique data point which is disengagements and specifically what happened around that event. The data from the car that is really useful is the image data and point cloud created by the radar + vision combo. Machine learning requires a quantum fu**toone of image data to be able to identify objects and know how they act/react in the environment. Street signs don't move but bicycles do. The more variations of all those objects the better the system will be at knowing what they are and how they are likely to move. Machine learning is not a static tech either, as it gets better the amount of data you needs becomes less and less. Also, you could have to much data and not have the power to process the data fast enough. More is not always better, better is better. If your algo can see 100,000 stop signs and identify them at 99% then your a billion times better then the comp. Today the best algo needs to see millions and millions to identify 85% of something. Efficiently being able to process billions of images gives you a huge advantage. I think Tesla biggest advantage is that they definitely have the data and with the right people working on the problem, I hope they have the processing efficiency as well. It is certainly something that will improve alot as they go because algos get better and machines get faster. If it was 6B miles last year it could be 3B by now and only 1B next year. No one knows when they really started in ernest and no one knows how many miles, it's really a silly metric to use to estimate when they will have a good enough system, as defined as 10x better then humans.
As important as teaching the machine is, it's almost as important to be able to validate what the Nadine knows and maybe they have a better idea there how many miles they need to show the system is safe and find edge cases where the system fails. You run you're solution against this massive data and determine how many times your car would crash and feed the failures back into the system. How many miles do you need to have a good sampling of every situation? Trick question because the situations are infinite, but what they can show is that in 6B miles, the human had 1000 crashes, then machine had only 2. That would be statically relevant to show regulators the advantages. No doubt enough data to have the system act as a guardian angel and gather more data to show regulators. The guardian angel would intervene in situations where an accidental is imminent, like running a stop sign or light. Show a regulator 1000 of these situations and you will have regulators stumbling over themselves to approve it.
