Filip Sondej comments on Testing which LLM architectures can do hidden serial reasoning