I imagine paying attention to the capabilities of search engines will be important. Classical computing is motivated by a desire to retrieve information quickly

I imagine paying attention to the capabilities of search engines will be important. Classical computing is motivated by a desire to retrieve information quickly. Search engines are motivated by a desire to retrieve information using fuzzy semantic concepts like language, features of an image, etc. Much of modern deep learning is motivated by modeling the task of information retrieval as a differentiable matrix multiplication (e.g. self-attention) in order to back-propagate error to the parameters of a large graph using stochastic gradient descent. In theory, this can give us a single checkpoint, which runs on a single GPU, that does more-or-less all of what "core" Google search does. I don't think that quite guarantees a singularity. There will need to be a lot of work afterwards. Humans can update their priors by collecting new information about their environment in real-time. They can also (sort of?) simulate situations and update their priors from that. Reinforcement learning could be crucial to solving these issues as it allows agents to learn through both real and simulated environments. Robotics may need to catch up, although recent advancements are pretty crazy. Assuming we don't all die first, of course.