Perform Tractable Research While Avoiding Capabilities Externalities [Pragmatic AI Safety #4] — AI Alignment Forum