> Have you kept up with recent ML papers like MindEye, which have managed to reconstruct seen images using image generator models conditioned on fMRI signals?

Not really. I left the field mostly because I felt bitter. I find that most papers in the field are more engineering than research. I skimmed through the MindEye paper and don’t find it very interesting. It’s more of mapping of “people looking at images in a fMRI” to identifying the shown image. They make the leap of saying that this is usable to detect actual mind’s eye (they cite a paper where they requires 40 hours of per-subject training, on the specific dataset) which I quite doubt. Also we’re nowhere near having a portable fMRI.

As for portable MEG, assuming they can do it: it would be indeed interesting. Since it still relies on synchronized regions I don’t think high level thinking detection is possible but it could be better for detecting motor activity and some mental states.