Erik Hoel at The Intrinsic Perspective:
While GPT-4 does have a radically different architecture from our own biological brains (like being solely feedforward and synchronous, whereas our own organic brains have a lot of feedback and asynchronous processing), AIs are neural networks inspired from biological ones. AIs have a learning rule that changes the strengths of the connections between their neurons, just like us—except their learning rule is applied from the outside by their engineers during a training phase, unlike us, who are forever learning (and forgetting). When artificial neural networks are trained, they often develop the properties we associate with real neural networks like grid-cells, shape-tuning, and visual illusions, which is why researchers in 2019 proposed a “deep learning framework for neuroscience.” There are even some prominent arguments that our own brains follow learning rules not too dissimilar. This similarity is why I advanced the Overfitted Brain Hypothesis, which argues that dreaming is a form of data augmentation, like a noise injection, that makes our nighttime experiences sparse and hallucinatory, and evolved to prevent overfitting.
more here.