Look through any of the single cell atlas papers that are published all the time and you'l see slightly different methods, called machine learning in the past, AI now, used to achieve exactly the same thing. Every form of AI has been continuously produced new results like this throughout the history of genomics.
The reason you are reading about this is because 1) Gemma has a massive massive PR budget, whereas scientists have zero PR budget, 2) it's coming from Google so it's not the traditional scientists and you know Google and when they publish something new, it's makes it to HN.
I don't see any reason to be excited by the result here. It is a workaday result, using a new tool. I'm usually excited by new tools, but given the source, it's going to take a lot of digging to get past the PR spin, so that extra needless work seems exhausting.
Facebooks's proteins language modelling, followed by Google's AlphaFold, did have really new and interesting methods! But single cell RNA models have been underwhelming because there's no easy "here is the ground truth" out there like there is for proteins. We won't know if this is a significant advancement until years of study show which of the many scRNA foundation models make better predictions. And there was a paper about a year ago that poured a ton of cold water on the whole field: replacing models with random weights barely changed the results on the very limited evaluation sets that we have.
Every AI post is either filled with negative comments stating that AI can just regurgitate stuff. This states otherwise and the comment I replied to tries to downplay that.
> It's not like the model has devised new knowledge. Kind of a low hanging fruit.
Just keep moving goalposts.