Thoughts on Tesla AI day and Full Self Driving:
The upgrade of large language model architecture into the lane segmentation challenge is a breakthrough
The upgrade of large language model architecture into the lane segmentation challenge is a breakthrough
In effect Tesla is training the system to generalize a βgrammarβ of intersections;
this feeds into the vehicle's understanding of potential trajectory choices
and (will?) inform predicted trajectories for other vehicles
this feeds into the vehicle's understanding of potential trajectory choices
and (will?) inform predicted trajectories for other vehicles
Itβs possible that some of the traditional ML used for path planning could get pulled into this architecture since the grammar of an intersection is contextually dependent upon the vehicles within it
(Think about intersection traversability when somebody βblocks the boxβ)
(Think about intersection traversability when somebody βblocks the boxβ)
HW4 will almost certainly devote some silicon to transformer-optimized compute, which should meaningfully improve throughput for this portion of the model.
At some point I think Tesla will subject the entire pre-HW3 fleet to an upgrade cycle
At some point I think Tesla will subject the entire pre-HW3 fleet to an upgrade cycle
(And they almost certainly will as cars fall off of lease.)
Loading suggestions...