The case for more Alignment Target Analysis (ATA) — LessWrong