This is an interesting video:
https://youtu.be/dDUC-LqVrPU
TL;DW We're starting to see early evidence of diminishing returns with our current AI architectures. If this is true, then eventually they start to improve in a logarithmic manner, making superintelligence (at least using our current architectures) impossible to achieve from a practical standpoint. The issue is that we need too much data on specific things for it to perform well on them all.