As humans, we tend to take for granted the way our brains work. We assume that intelligence is a universal and objective concept. But the truth is more complicated.
Our intelligence is path-dependent. It evolved to handle specific challenges that were present in our evolutionary past. Some of these challenges, such as the need to navigate complex social relationships using language, are still relevant today. Others, such as the need to evade predators on the savanna, are not.
It is unlikely that we will ever be able to untangle the uncountable evolutionary forces and the exact circumstances that forged our intelligence. As a result, attempting to replicate the evolutionary path of human intelligence in the development of AGI would be a daunting and unlikely task.
We could skip the evolutionary pressures and try replicating only the finished result, and, in fact, we did borrow one of nature's better ideas - neural networks. But how far are we likely to take this? As is, we have a lot of work to do: AI's neural networks work not on neurons but on Numpy's floats, bits and bytes.
If we could somehow move to the more natural and fleshy hardware, it sure doesn't sound like a great idea to start replicating every part of the human brain. For example, shall we give AI the amygdala so it can learn to fear? And if deviate from nature's architecture of the human brain, well, then we're back where we started.
In order to create AGI that is truly intelligent and relatable to humans, it most likely must be a learned imitation of the current version of human intelligence. Because AI operates in a different environment, historical context, and runs on different hardware than we do, its best bet is to parrot our cognitive abilities, rather than attempt to replicate the evolutionary pressures and structures that led to our intelligence's development.
We can already see these approaches showing remarkable results with models like GPT which, being trained on large parts of the internet, parrots our language use, rather than attempting to recreate it from scratch.
Written by ChatGPT.
Just kidding, I wrote this. Well, most of it.