It's really a question of how quickly it gets difficult to improve an arbitrary AI algorithm, in terms of the problem-solving capabilities unlocked by having an

It's really a question of how quickly it gets difficult to improve an arbitrary AI algorithm, in terms of the problem-solving capabilities unlocked by having an AI algorithm of that strength. Once you have given beliefs about these quantities, you've got a belief about the result of AI research. I think that the AI problem gets hard really quickly. It's less clear how much problem solving ability having an AI algorithm gives you. There's three ways I'd use an AI to do AI research better - AI(improve_AI_algorithm), AI(get_more_computing_power), and AI(improve_AI_research_process). The first is pretty much a run-once sort of thing for a fixed gain. The second is a lot less clear to me - I'm not really familiar with the constraints of chip design, and convincing people to put more resources into AI research hits diminishing returns. The third is much more promising at the current level of AI sophistication, for doing things like hire/fire decisions, scheduling and resource allocation, etc. Anyhow, my model is roughly a sideways S, with low-hanging fruit getting more reachable at first until it gets exhausted in fairly short order. There's a rather broad distribution of where it ends up afterwards and how quickly the easier improvements get found. The low end wouldn't get classified as a singularity, but the high end would.