I understand the concept of AI model collapse caused by recursion. What I’m proposing goes beyond a basic feedback loop, like repeatedly running Stable Diffusion. Instead, I envision an AI system with specialized expertise, akin to a scientist making a breakthrough based on inputs from a researcher—or even autonomously. This specialized AI could then train other, less specialized models in its area of expertise. For example, it might generate a discovery that is as straightforward as producing a white paper for interpretation.
If there is an virtual "scientist" that is trained on DNA for instance, could come up with a discovery for a treatment. This gets published, circulated and trained in to other models.
This isn't the kind of inbreeding that you suggest as the answer is valid.
cshores|1 year ago