[Symmetry]

A working memory that can preserve information indefinitely outside a particular context window and which can engage in multi-step reasoning that doesn't show up in its outputs.

GPT4o's context window is 128k tokens which is somewhere on the order of 128kB. Your brain's context window, all the subliminal activations from the nerves in your gut and the parts of your visual field you aren't necessarily paying attention to is on the order of 2MB. So a similar order of magnitude though GPT has a sliding window and your brain has more of an exponential decay in activations. That LLMs can accomplish everything they do just with what seems analogous to human reflex rather than human reasoning is astounding and more than a bit scary.

[datameta]

I'm curious what resources led you to calculate a 2MB context window, I'd like to learn more.

[Symmetry]

Looking up an estimate of the brain's input bandwidth at 10 million bits per second and multiplying by the second or two a subliminal stimuli can continue to affect a person's behavior. This is a very crude estimate and probably an order of magnitude off, but I don't think many orders of magnitude off.