@hazlin the neuromorphic side tends to be more concerned about it. they constantly note the brain does not do calculus, has no backprop, there is no teacher to force an outcome, there are no "attention heads," it uses watts of power and it learns in dreams.

anthropic style ANN work is just based on unrolling everything through time and trying to brute force stability through parameter explosions. the actual thing has little tiny moderator signals that autonomously re-normalize what comes in to them.

problem is brute force gets you a nice publisher cadence. the real thing requires puzzling out chaotic, dynamical systems that only work in the most precise of tunings. :neocat_solder_googly: