Compute for experiment about how steganography in large language models might arise as a result of benign optimization | grantmaking.ai