>>11402666That's fair, although most tasks we are interested in don't require human-level intellect. There are "only" ~140 million neurons in the visual apparatus of the brain, which is in line with how many parameters modest-sized modern neural networks have. Surely this should be enough, backed with a small "reasoning core", to solve vision problems (semantic segmentation is still an open problem, so is 3D reconstruction from 2D images and in general, the inverse graphics problem).
Additionally, ravens can solve metric fuckload of tasks much better than modern neural networks can, despite their entire brains being about ~1.2 billion neurons, comparable to the larger modern neural networks. If it was purely about capacity, we've already reached the desired approximate level. Clearly engineering is required to match the raw network capacity to par with a raven's learning ability.
Additionally, I don't think comparing neuronal size in living organisms' brains to that of neural networks makes that much sense, again in light of engineering constraints on ANNs: they are fixed in topology, connectivity, architecture, and signalling (real neurons send out noisy signals and are time-dependent, as well as relying on other stuff, while ANNs are deterministic barring "useless" tricks, so the computing capacity on the same amount of neurons is not the same, not to mention activation and neuron types - real neural networks have many types of neurons that don't compute the same thing whereas usual ANNs have all nodes operate the same way).
Also I want to emphasize that I'm not saying that increased computing power is going to help, merely that real breakthroughs in performance definitely come from elsewhere. Whether these breakthroughs are unusable because of computing power is another question.
