“The threshold for AI recursive self-improvement to accelerate dramatically is relatively low, requiring a model to simply be better than the smartest humans at identifying and solving research bottlenecks.”