Weird AI safety projects need funding

The article challenges conventional AI safety funding by arguing that it often assumes 'business as usual,' which may be inconsistent with the timelines and probabilities of doom many people hold. It suggests that even if one does not expect imminent catastrophe, there are conceivable scenarios where radical approaches would be justified. The core argument is that weird projects, despite being likely useless by default, offer a longer and fatter tail of upsides due to their higher variability. This means that reallocating some funding to them could be rational, not despite but because of the uncertainty. The piece calls for explicit, substantial effort to explore such strategies, pushing the Overton window in AI safety. It does not prescribe specific projects but rather argues for the need to consider them seriously.