falcon_tech

Thats a facinating step in AI usage into different aspects. Just curious How will this process guarding against the failure mode where agents say " this is unusual " , and the judgement is itself biased towards known family . Its easy to flag something as intresting when it partialyl resembles an already kown system (CRISPR- like repeats , in this case) then to flag something unknown or novel. So Going from 200k RTs to 3500 candidates to 20 reports to 1 real hit very narrowing path basically it seems like the real risk ; the things thrown out of this filter as an outlair might be an genuinely novel and getting thrown out in early stages just because it didnt look like anything claude had knowledge on , will be an "false negetive".

I also ran into a similar problem in my research of detecting and localizing AI-Manipulated medical images, where models struggled to detect highly novel or known pattern of manipulations , surprisingly even when the anomoly was visually obvious to human afterwards.

curious if something similar shows up in survey process , and how they'd even know if it did