There's an interesting caveat in there: “It generates its own training data and learns from that data to pick precursors that would give the brightest quantum dot” This is vulnerable to a kind of founder effect, where random early results determine the direction of later tests. #chemsky 🧪
The general issue of appropriate bias control becomes a more crucial one in AI applications and obviously critical in such self-learning approaches. One more nice example for applying „artificial integrity“.
@aellalabrys.bsky.social
Not sure if this is applicable here, but the LHC also uses machine learning to filter out data early. To combat this spiral and probably for calibration and other tasks, I think they randomly include a percentage of raw data bypassing these filters. Maybe totally off, but it sprung into my mind.
Yes, it's a common issue, it occurs in gravitational wave detection as well. The difference here is that there is a loop with AI generating new real world data and feeding itself with it.