100 more years to understand general intelligence and build it from first principles? It's not a prerequisite - AGI should be buildable soon without any deep understanding of its internals.
Re "no free lunch": the world doesn't produce all datasets; it produces world-like datasets, with stable objects, generic kinds and causal mechanisms, invariances, symmetries, etc.. These define all-purpose inductive biases...somehow 😁
100 more years to understand general intelligence and build it from first principles? It's not a prerequisite - AGI should be buildable soon without any deep understanding of its internals.
To build a "Monolith" instead of a "HAL9000" will need to know how it works.
Re "no free lunch": the world doesn't produce all datasets; it produces world-like datasets, with stable objects, generic kinds and causal mechanisms, invariances, symmetries, etc.. These define all-purpose inductive biases...somehow 😁