[lifeplusplus]

i dont a human can create a system this complicated... i dont think we can even map it out. if it's done, it'd be done by AI and it'd be a black box. we couldn't create an autonomous car, we need ai to come up with million rules and parameters. same story here just order of magnitudes more complex.

[visarga]

Can't do it intentionally, but evolution can do it unintentionally.

Nature solved this problem of radical open-ended exploration through evolution. It takes a lot of time but in a single run evolution created all species and life on earth. Evolutionary methods also created AlphaGo, AlphaTensor and AlphaFold - three AIs that surpass all humans put together in specific fields.

Basically learning from massive search and problem solving outcomes - that's how you solve hard problems. And if you think about it, that's how science works too - so many papers, many contradicting each other or trying completely different approaches adds up to evolution in the space of scientific ideas. Any one human alone can't do it, but all of us over a long enough time can do so much more, diversity and open-endedness are key.

[lifeplusplus]

Ai kinda follows iterative exploration approach too

[chaxor]

Don't forget that having a GNN (or perhaps Sheaf NN) model the system can allow decomposition such that an analytical expression can be extracted to explain the dynamics of the model quite well. This was shown via Cranmer et al, and continues to be a good way to find explanatory models that both interpolate as well as the NN and extrapolate past the provided data.