self driving has now devolved into a computing power problem soon overcome. this also means that all cars on the road will be able to comm each other as well. commercial availability is close at hand then with about a decade rollout to make it universal.
right now it is down to aking big data big enough.The same tech also applies to all aircraft as well.
Also the obvious transition needs to include a driver in position on standby as a safety measure. essentially our pilots have done this forever.
The takehome is that we are already there.
Expert Says Tesla FSD 12.X Will Achieve Better Than Human Driving
March 7, 2024 by Brian Wang
https://www.nextbigfuture.com/2024/03/expert-says-tesla-fsd-12-x-will-achieve-better-than-human-driving.html#more-193756
James Douma is an AI expert and he believes that Tesla FSD V12.X will achieve better than human driving with more data and compute to refine the system. It will need 1% or less interventions than FSD V11. If interventions have been at 1 every drive then 1% interventions would mean one intervention every 3 months or 1 intervention every year.
Tesla has installed over 15000 Nvidia H100 chips to reach over 50 Exaflops of training compute.
There are other reports of 5-10 times faster progress each month. This would suggest that the 1% intervention level could be reached by June 2024. This would be enough to get FSD out of beta and to have systems that would rapidly reach good performance across the world by adapting to all driving areas. The robotaxi level might take another 6-24 months, if James Douma and the reported faster progress is correct.
If this estimate of 2023 Nvidia H100 chip purchases is correct then Tesla has at least 15,000 Nvidia H100s. H100s are 12X Nvidia A100.
This would means that Tesla would have installed those chips by now and be at about 65-70 Exaflops of training compute. If Tesla was able to buy another 10,000 Nvidia H100 chips then they would reach the 100 Exaflop target from last year.
No comments:
Post a Comment