My guess is several rounds of RLHF or fine-tuning with some kind of curated 'woke' dataset. Alternatively, they trained a critic model that penalizes showing wh

My guess is several rounds of RLHF or fine-tuning with some kind of curated 'woke' dataset. Alternatively, they trained a critic model that penalizes showing white people or other 'bad' representations and then fine tuned using that.