“Progress in AI model capabilities for long-horizon tasks may not follow a clean scaling law and could instead exhibit phase transitions, where a certain level of capability unlocks the ability to handle much longer tasks.”