@tao@mathstodon.xyz It feels like a distinction between NP hard and NP complete problems. The best way to work in the AI is to map a problem that cannot be done by the current level of Ai and breaking it down and to make a simialrity basis to problems that are already being solved. Directly tackling the hard problems will very much affect the current small labs very much. So for small labs it is best to use the current compute available for decomposed tasks that will combine itself into helping solve the large problems. But they have to also make sure the problem they are addressing can have the expected usage, and will be worth addressing,