Comment by mk_chan

Just shows the offset openai feels like it has to add to ‘equalize’ the average discourse of its training material