The main problem with the current technology, to my eye, is you need these huge multi dimensional models with extremely lossy encoding in order to implement t

The main problem with the current technology, to my eye, is you need these huge multi dimensional models with extremely lossy encoding in order to implement the system on a modern CPU which is effectively a 2.5D piece of hardware that ultimately accesses a 1D array of memory. Your exponential problems have exponential problems. Scaling this system is factorially hard.