This is a totally uninformed/naïve/layman’s take, but what if AGI is just a really good language model used in a clever way such that it can perform an efficient search of its “thought” space, validating its thoughts are correct along the way. Programming, logic, math, etc are perhaps the easiest forms of “thoughts” for a computer to validate, but given enough quality data maybe it could be good at all kinds of other tasks as well.