Universal approximators vs. algorithmic completeness