The original concept was a chip that re-programs its instruction set based on the tasks it's being asked to do most. 

 FPGAs are fascinating.  I hear they aren't that great for LLM stuff though.  CUDA is apparently the king.