>>13496 (OP)The biggest genuine successes in AI are the projects that try to solve problems with narrower scope.
There was a project for simulating protein folding on a computer, it proved too complex to be solved by the standard programming approach, but with machine-learning ai approach it got solved with sufficient accuracy to be genuinely useful. Compared to the aspirations of general AIs that are supposed to do "everything" this was a very specific task with clearly defined constraints. And it was done at relatively low cost with a small amount of GPU cards.
I think we should go that route. Instead of trying to make AIs do everything, we should break it down into lots of relatively narrow scope tasks and build up a collection of task-specific AIs that can be combined for more advanced stuff.
I think this would work well for coding AIs. You could have lots of different Code-Ais similarly like you load a bunch of software-libraries in software development.