I believe for AI we have to step back and look at smaller but important problems that AI can't solve right now. Like basic arithmetic, unlike humans who by grasping concept of arithmetic can extrapolate to any arbitrarily large numbers of operands, neural nets can't do that especially if you're input data is just a bunch of arithmetic pairs and output. But that's changing promising research from Waterloo changes that: https://arxiv.org/abs/2502.16763 . This can be further worked on for other arbitrary algorithms and even languages as they conform to grammatical rules and structure.