Source: Facebook

The only value in nn’s is in symbol detection, and from there on out, the cost reduction in the use of manifolds (geometry of relations) is so much cheaper, faster, more powerful, and controllable, and auditable that it will defeat training in every discipline that has any limitations on closure. So much like our brains, it is more logical to create hardware that produces a limited number of symbols, and then use software to synthesize them. Our brains do this at very low cost by making profound discounts to cost, relying on memory substitution rather than input retention. This is why we are able to act quickly in some circumstances, but why we have such high error rates in anything of any complexity. So AI will be beneficial above and below human scale, but not so much *at human scale* because nn’s can defeat us in perception, and GAI’s may defeat us at reason. But those two technologies NN/ML, and Algorithmic Artificial Intelligence will perform better than NN’s unless there is some tremendous innovation in hardware that allows it to compete with the 20w of power consumption of the human brain, or the 200 watts of an ordinary processor processing symbolic data (akin to language) fed to it by extraordinarily good measurements, provided by some variation of nn and Bayesian hardware.