I suspect that if models were able to be aware of and answer questions about their internal vectors, they would be much smarter and able to progress faster.