The topic discussed in the Twitter thread, on the other hand, is much more interesting. I am a layman in ML (although I have implemented a NN myself, but I am not a data scientist) so I offer my view hoping to get some feedback and be corrected. My view is that the original post was correct and LeCun was not. I just can't see what sort of objective function you can use to solve every potential problem arising in the future about similar issues. You want a "totally unbiased" model? Then, you use total diversity in your training set? Can't it happen that the model will then enforce diversity in a non-diverse picture? Is that your ideal result?
That might be a silly example and maybe not 100% correct for the example discussed, but, in general, my layman feeling is just that ML cannot help in cases where moral judgement is required. It can provide a good tool in any other case where, for example, a quadratic error is a reliable error measure, but that seems a hard limitation.