Update: had a friend with access send me a PDF of the study and looked through it. It seems that the big breakthrough is only half AI — the other half being loo

Update: had a friend with access send me a PDF of the study and looked through it. It seems that the big breakthrough is only half AI — the other half being looking directly at time series of fMRIs instead of static images with features in them manually selected for relevance, because how the various circuits in the brain operate and circulate over time is important information. Also they got this to replicate well with the same people at different times, and also generalize to two other cohorts, consistently, and also used XAI to check what the AI was keying off, to make sure it wasn't going off something nonsensical, and directly used those features with success as well. It seems like an extremely carefully controlled and designed study tbh.