How many of the AlphaFold protein folding predictions have been actually verified by a lab? And are any of those actually new proteins or just tiny variations on things in the training set?
> Shoichet agrees that AlphaFold predictions are not universally useful. “There were a lot of models that we didn’t even try because we thought they were so bad,” he says. But he estimates that in about one-third of cases, an AlphaFold structure could jump-start a project. “Compared to actually going out and getting a new structure, you could advance the project by a couple of years and that’s huge,” he says.
AI is definitely capable of solving novel programming problems and I'm constantly surprised how well Copilot understands the esoteric context of my code.
GP could've read the papers, or read commentary of the papers, or read about the 13th Critical Assessment of Structure Prediction (CASP) contest.
Instead, "llms are bad, so isn't this bad too"
I don't believe that. I think you are being fooled by the enormity of the training set, followed by the non-novelty of things you personally haven't seen before.
I'm not saying AI tools aren't useful though. I'm just saying they are what they are: a fuzzy search in an enormous problem space. This is impressive and cool, but not the same as being able to solve truly novel problems.