I agree that the mathematics motivating neural networks are well understood. We understand the heuristics behind them; obviously, we created them and they work.

I agree that the mathematics motivating neural networks are well understood. We understand the heuristics behind them; obviously, we created them and they work. I more just mean at the most fundamental level, do we understand intelligence; could we reconstruct it without just ‘optimizing’ towards it. If we understand intelligence then we should be able to deterministically explain all behaviour of neural networks/LLM’s, instead of just poking them from the outside and seeing how they react.