“Tom McGrath cites research from FAR AI which found that for internal probe-based training to be effective, the probe must be high quality and backpropagation should not be run through the probe itself.”