Training machine learning algorithms to avoid traffic accidents can be challenging because the rare occurrence of such events leads to the insufficiency of training data. We introduce the idea of applying instantaneous crowdsourcing to augment autonomous vehicles with collective human cognitive capability within super-human reaction time. However, because the instantaneous crowdsourcing system must prefetch possible futures in order to generate tasks, in complex real-world problems we would need to hire implausibly many workers to support this approach. In this work, we propose that predicting dangerous futures from crowd-worker input can help resolve this problem. In a formative study to inform the design of crowd prediction workflows, we found there are two main challenges: 1) false positives, which can initiate instantaneous crowdsourcing more than necessary, and 2) handling a large number of futures with multiple candidate objects in the scene.