My comment is a hypothesis about how exactly the genome solved information inaccessibility in the form of locating learned WM concepts (ie through proxy matching), so it does more than merely suggest, but I guess you may be leaning heavily on reliably? As it’s only reasonably reliable when the early training environment is sufficiently normal: it has the expected out of distribution failures (and I list numerous examples).
My comment is a hypothesis about how exactly the genome solved information inaccessibility in the form of locating learned WM concepts (ie through proxy matching), so it does more than merely suggest, but I guess you may be leaning heavily on reliably? As it’s only reasonably reliable when the early training environment is sufficiently normal: it has the expected out of distribution failures (and I list numerous examples).