When you look at models that were built for a specific purpose, closely intertwined with experts who care about that purpose, they absolutely propel communities to new heights. Consider the impact of alphafold, it won a Nobel prize, proteomics is forever changed.

The issue is that that's not currently the business model that's aimed at most of us. We have to have a race to the bottom first. We can have nice things later, if we're lucky, once a certain sort of investor goes broke and a different sort takes the helm. It's stupid, but its a stupidity that predates AI by a long shot.

We know that the model training on the model training on the model leads to model collapse...

Value is determined by what we value, it's a choice. If a bunch of scientists value good approximations for how a protein will fold, and then a model generates more such things in a year than those scientists could make in a century, that's a lot of value. Not extracted from anyone, created.