Well, reading the thread, it's pretty clear that she was correct. There is a second step of metaoptimization which incentives the researchers to encode algorithms to produce algorithms that more easily bias along a certain way. And she is correct that benchmarking models against biased datasets can lead researchers to bias for architectures that are in turn more likely to have a certain bias.

I'm sure LeCun knows this too. It's a fact that has been known for almost 40 years now, that for a given generalization system multiple bias modes are are possible, and of course the process of optimizing architectures against a biased benchmark will change the learning bias of the model. It's not just datasets that are biased, generalization systems are necessarily biased too.

As for the last line, I don't know that it should be expected for employers to fire employees pursuing recourse against sexual harassment.

I'm sure we can agree that the tone wasn't good, but she does have a point in saying that it's not possible to reduce the bias to datasets used to train the final model.

The paper : http://www.cs.cmu.edu/~tom/pubs/NeedForBias_1980.pdf