AI companies keep talking about how dangerous their models are, and they just keep building them anyway.