> I think unless we really understand what is happening inside GPT-3, You realise we have debugging tools that let us inspect what is going on inside transform

> I think unless we really understand what is happening inside GPT-3, You realise we have debugging tools that let us inspect what is going on inside transformer-based models like GPT-3 right? > why do we train a system that is supposed to be intelligent on such data? Because it's available and does interesting things. > we cannot really tell whether it is really dumb or just playing dumb for the above reason. No one who works in the field thinks anything approaching this. I think GPT-3 does raise questions about what intelligence is though.