I remember joking that this paper puts us marginally closer to the Men in Black memory wiping device, but in the context of this thread that joke seems a bit dark.
I think it's even worse actually, because this is not something that needs to be done explicitly or on purpose.
Simply training AIs with a target goal of maximizing engagement could lead to models discovering and exploiting superstimuli or superstimuli-like bias in humans.
With the reservation that I only skimmed the article, it seems like what they produced was visual stimuli resulting in patterns of neural activation/non-activation at a rather limited number of sample points in a higher order visual area (macaque V4).
Not to take away from their results, such as they are, but it is very expected that visual inputs should have specific and fairly predictable effects on V4, and one could probably have designed such patterns manually from known perceptual psychology in a few iterations, with feedback from the recording electrodes for the details.
It is not at all obvious to me, and indeed not even plausible, that they'd be able to control arbitrarily chosen neurons this way.
I think it's even worse actually, because this is not something that needs to be done explicitly or on purpose.
Simply training AIs with a target goal of maximizing engagement could lead to models discovering and exploiting superstimuli or superstimuli-like bias in humans.